Catch me, if you can: The long path from reputation to trust

Transcription

1 Catch me, if you can: The long path from reputation to trust Karl Aberer EPFL joint work with Zoran Despotovic, Le Hung Vu, Thanasis Papaiaonnou 1

2 MOTIVATION 2

3 Trust in Online Systems No trusted authority Selfish behavior Examples Peer-to-peer system Social networks Ecommerce sites Internet gaming 3

4 Selfish Behavior Example P2P: to conserve Own bandwidth Own CPU cycles Freeriding To exploit Greedily consume other s resources... before others can use that opportunity Problem 4

5 Exampleof Selfishness Freeriding in Gnutella [Adar, Hubermann 2000] Cooperation needs to be built: but how? 5

6 Cooperation Based on Trust Trust: the extent to which a peer believes the other peer will cooperate Gain is to be shared with the other peer But the peer is exposed to a risk of loss 6

7 Cooperation Based on Trust Trust: the extent to which a peer believes the other peer will cooperate Gain is to be shared with the other peer But the peer is exposed to a risk of loss 7

8 Cooperation Based on Trust Trust: the extent to which a peer believes the other peer will cooperate Gain is to be shared with the other peer But the peer is exposed to a risk of loss 8

9 Cooperation Based on Trust Trust model: the way in which the above belief is inferred Inference could be: if Bob cooperates with Alice then Bob will also cooperate with me with high probability 9

10 Cooperation Based on Trust Trust model: the way in which the above belief is inferred Inference could be: if B cooperates with C then B will also cooperate with A with high probability 10

11 Cooperation Based on Trust Trust model: the way in which the above belief is inferred Inference could be: if B cooperates with C then B will also cooperate with A with high probability 11

12 Cooperation Based on Trust Trust model: the way in which the above belief is inferred Inference could be: if B cooperates with A then B will also cooperate with C with high probability Reputation-based trust: Inference based on the collective earlier actions of a peer: it s reputation 12

13 Doesitreallymatter? Example: ebay ebay reputation profiles predictive of future performance [Resnick et al., 2002] Prices positively correlated with the feedback [Melnik and Alm, 2002] 13

14 Overview MOTIVATION REPUTATION-BASED TRUST IN ONLINE COMMUNITIES TRUST WITH RATIONAL PEERS TRUST AND IDENTITY CONCLUSIONS 14

15 REPUTATION-BASEDTRUST IN ONLINE COMMUNITIES 15

16 The Path Towards Trust Problem Countermeasure Selfish Behavior Reputation 16

17 On-line Reputation Common characteristics open community: interactions with different parties rarely repeated interactions with the same party Therefore need for recommendations of others possibility to misbehave and misreport 17

18 The Path Towards Trust Problem Countermeasure Selfish Behavior Manipulation of Reports Reputation 18

19 Model Peers provide services to other peers Know other peers they interacted with Store feedback w i on interactions withpeer i Feedback on service and reporting A social network (trust graph) 19

20 Social Networking Approach Computeglobal trust values t j for peer j By aggregating feedbacks Weighting by the trust in the recommender Similar to PageRank computation w 1 w 2 i in ( j ) w 3 peer j t j = w t i = wi i i in ( j) k in( i) t t k [Xiong, Liu TKDE 2004] 20

21 Social Networking Assessment Problems solved 1. Automatic aggregation of reputation data 2. Does not require central authority 3. Possibility of misreporting considered Shortcomings Costly(global) computation Trust values have no further meaning but ranking No distinction between propensity to provide poor service and to misreport 21

22 ProbabilisticEstimation Approach Assume probabilistic peer behavior, eg. P[peer jperforms service honestly] = θ j P[peer klies when reporting] = l k Probabilityof report y k on peer jfrompeer k P [ Y k = y k lk (1 θ j ) + (1 lk ) θ j ] = lkθ j + (1 lk )(1 θ j ) if if y y k k = 1 = 0 peer k l k = 0 report y k = 0 peer j θ j = 0 [Despotovic, Aberer Journal of Computer Networks 2004] 22

23 ProbabilisticEstimation Approach Maximum Likelihood Estimation Determine l k by checking reports on own performance Collect all reports y 1, y 2,, y n on peer j Select θ j that maximizes the probability to obtain the observed reports L θ ) = P [ Y = y ] P [ Y = y ] LP [ Y = ( j n yn ] 23

24 Comparative Evaluation Precision in correctly assessing misbehavior Maximum Likelihood Estimation Social Networking 24

25 ComparingeBay and P2P Differences 1. no centralized authority to manage and verify reputation information 2. possibility to manipulate reports of others Question Which methods do exist for efficiently and automatically managing reputation data in the absence of any centralized infrastructure? 25

26 ImplementationIssues Variant 1: use trust graph as unstructured overlay network Simple maintenance, high cost of retrieval of reports Variant 2: use a separate structured overlay network Complex maintenance, efficient retrieval Additionalproblem: manipulation of stored reports and messages Use replication of reputation data 26

27 ProbabilisticEstimation Assessment Problems solved Efficient trust computation Trust values predict probability of cooperative behavior Distinction between propensity to provide poor service and to misreport Shortcomings Reputation-based trust detects malicious and selfish behaviors (signaling) but does not consider rationality of agents (sanctioning fostering honest behavior) 27

28 The Path Towards Trust Problem Selfish Behavior Manipulation of Reports Strategic Behavior Countermeasure Reputation Reputation of recommenders Computational trust models for signalling 28

29 REPUTATION-BASEDTRUST WITH RATIONAL PEERS 29

30 A Rational Agent Reputation Detection threshold 30

31 Rationality Actions of peers have associated utility Example: prisoners dilemma u 1,u 2 Main insights One shot game: no cooperation (Nash) Repeated game: cooperation tit-for-tat (Axelrod) Common explanation of the concept of trust 31

32 ExtendedModel Model Consumers provide feedback w i on peer i sservice Peer ihas incentiveto cheat: legal gain u i < illegitimate gain u i + v i Peer behavior Honest peers never cheat Malicious peers cheat probabilistically Rational peers optimize their utility [Hung, Aberer WI 2008, TAAS 2010] 32

33 Sanctioning peer i If t=1 and w=0 or t=0 and w=1: Peer i is cheating!!! 3 Most recent feedback w 1 peer i: k cheating detections Evaluate reliability t 2 of feedback w using computational trust model If peer I has less than k cheating detections: ok! 4 Severe sanctioning mechanism! 33

34 ComputationalTrust Model may use several information sources past rating behavior of the rater, past performance of sellers, trusted sources, own belief on environment s vulnerability, relations between peers may use variety of statistical models/heuristics probabilistic approaches, collaborative filtering, similarity of rater and own experience, clustering of ratings to isolate dishonest rater Accuracy measure (known to peers) P[est+ real-] < ε, P[est - real+] < ε 34

35 Trust Accuracyvs. Cooperation Theorem: if computational trust model sufficiently accurate and gains are bounded rational peers cooperate in all but the last Δ transactions. Bounded gains (u * <u<u*, v * <v<v*) 35

36 Incentive to Leave (Δ) Example: For ε< 0.2 and k= 1, 2 peerwill leave for the last 2 transactions accuracy of computational trust model 36

37 Emergence of Cooperation #transactions of an honest seller till mistakenly blacklisted Cooperation is enforced if peers stay infinitely or long enough given εsufficiently small resilient against rating manipulation malicious peers are eliminated Example: For ε= 0.2 and k= 10 peerwill be accidentally blacklisted after 2^22 transactions #cheats of a malicious seller till being eliminated whereas a cheater will be eliminated after 2^3 cheats 37

38 Accuracy-cost Trade-off Presence of high quality dishonesty detector may prevent rating manipulation by rational sellers High accuracy implies higher implementation cost Combine two computational trust models expensive/accurate dishonesty detector with probability c trust the rating with probability 1-c Result With very low crational peers still cooperate 38

39 Sanctioning: Assessment Considering rational behavior Enforces cooperation of rational peers Eliminates malicious peers Permits to optimize computational cost for dishonesty detection Shortcomings Requires shared, secure storage for reputation data Requires stable and global identities (like all other reputation-based trust mechanisms) 39

40 The Path Towards Trust Problem Countermeasure Selfish Behavior Manipulation of Reports Strategic Behavior Whitewashing Reputation Reputation of recommenders Sanctioning 40

41 TRUST AND IDENTITY 41

42 Key Problem The problemof trust isinherentlylinkedto stable identification 42

43 How to prevent whitewashing? Remember: clients on ebay are willing to pay higher prices for reliable seller 8.1% according to a study of [Resnick2006] Can this help? Idea: sellers that stay longer in the system ask for higher prices Makes it unattractive to leave the system for whitewashing Does it work? 43

44 Identity Premium P(L) = u (1 -Ф) + f(l) L lifetime of seller f(0) = 0, f monotonically increasing 0 < Ф < 1, initial price below original price u Price P original price u price P(L) with identity premium Lifetime L 44

45 Cooperation Enforcement Theorem: If the identity costis sufficiently small there exists an identity premium function such that a rational provider will cooperate in every but the last interaction. Bound on identitycostand premium dependson Accuracy of dishonesty detector Potential cheating gain Initial price 45

46 46

47 Example Limit price above original price: Acceptable to buyers? Inefficiency? Limit price below original price: Acceptable to sellers? original price u=1 cheating gain 50% of original price Initalprice0.5 (Ф= 0.5) 47

48 Eliminating Inefficiency If the cheating gain is sufficiently small and the dishonesty detector is sufficiently accurate the price remains bounded and can approach any limit price by properly choosing Ф Thus for a finite number of services this inefficiency can be eliminated For infinite number it can be kept extremely small (order of provisioning a few services) Rationale to accept the scheme Providers: without premium no trade at all Consumers: no risk if providers are rational (provably) 48

49 CONCLUSIONS 49

50 The Path Towards Trust Problem Countermeasure Selfish Behavior Manipulation of Reports Strategic Behavior Whitewashing Reputation Reputation of recommenders Sanctioning Identity Premium 50

51 FromClosedto Open Trust So far trust managedin a closedsystem, but Multiple trust systems Transferrable identity Aggregate reputation from multiple sources Open Web Semantic Web: content-based entities All Web content on Tim-BernersLee Social Web: friend-to-friend authentication Social networks as computing substrate SensorWeb: linkthe web to physicalreality Soft biometrics, GPS, 51

52 New Opportunitiesfor Trust Systems Semantic, Social and SensorWebnot onlyhelp to buildtrust but are alsoin needof trust Semantic Web: is the content trustworthy? Social Web: are my social connections trustworthy? SensorWeb: is the measured data reliable? Trust S 3 W 52

53 Thank you for your attention! and we will continue to keep you busy for a while 53

54 Acknowledgements Work sponsored by Swiss National Science Foundation EU FP 6, projectsdip, Nepomuk NTT Docomo EuroLabs, Munich 54

55 Main Publications [Despotovic, Aberer 04] Z. Despotovic, K. Aberer: "P2P Reputation Management: ProbabilisticEstimation vs. Social Networks", Journal of Computer Networks, SpecialIssue on Management in Peer-to-Peer Systems: Trust, Reputation and Security, Elsevier, 50(4): , [Hung, Aberer 08] Le-Hung Vu, K. Aberer: «Effective Usage of Computational Trust Models in Rational Environments», Web Intelligence Conference (WI-IAT'08), December 9-12, 2008, Sydney, Australia. [Cudre-Mauroux et al 09] P. Cudré-Mauroux, P. Haghani, M. Jost, K. Aberer, H. de Meer: idmesh: graph-based disambiguation of linked data, WWW [Vu, et al 09] L.-H. Vu, T.G. Papaioannou and K. Aberer, Synergies of different reputation systems: challenges and opportunities, Proceedings of the 7th IEEE conference on Privacy, Security and Trust, [Vu, et al 10]L.H. Vu, T.G. Papaioannou and K. Aberer, Impact of Trust Management and Information Sharing to Adversarial Cost in Ranking Systems, Proc. of the 4th IFIP International Conference on Trust Management (IFIPTM 2010),