11

GossipTrust for Fast ReputationGossipTrust for Fast ReputationAggregation in Peer-to-Peer NetAggregation in Peer-to-Peer Net

worksworksRunfang Zhou,Runfang Zhou, Kai Hwang, and Min CaiKai Hwang, and Min Cai

University of Southern CaliforniaUniversity of Southern California

IEEE Transaction on Knowledge and Data EngineeringIEEE Transaction on Knowledge and Data Engineering, Feb. , Feb. 20082008

22

OutlinesOutlines

IntroductionIntroduction

GossipTrust system architectureGossipTrust system architecture

Gossip-based reputation aggregationGossip-based reputation aggregation

Simulated performance resultsSimulated performance results

ConclusionsConclusions

33

P2P Reputation SystemsP2P Reputation SystemsReputation managementReputation management in P2P networks encourages resource sharing among well-behaved peers and fighting against selfish (free riding) or malicious peers.

Existing Approaches:Existing Approaches:

– Collecting, aggregating and disseminating feedbacks among peersCollecting, aggregating and disseminating feedbacks among peers

– EigenTrust[16], PeerTrust[35], PowerTrust[40], GossipTrust, etc.EigenTrust[16], PeerTrust[35], PowerTrust[40], GossipTrust, etc.

Common limitations:Common limitations:

– Ignore the feedback properties of P2P systems Ignore the feedback properties of P2P systems

– Assume an arbitrary feedback distributionAssume an arbitrary feedback distribution

– Not in agreement with the reality!Not in agreement with the reality!

Goal: Goal: Scalable, Robust, and SecureScalable, Robust, and Secure reputation management system reputation management system

44

Transitive Reputation AggregationTransitive Reputation AggregationAsk friend’s friends about the reputation of a peer in the system

Ask your friends j

What they think

of node k

And weight each friend’s opinion by

how much you trust him

j

jkij rrrik

'

0

0

0

0

0

0

0

0

0

0

0

0

Node 1

Node 2

Node 4

000000

0

Node 6

j

jkij rrrik

'

While |V(i) – V(i-1)| > δ,

V(i+1) = RTV(i)

55

The eBay Reputation SystemThe eBay Reputation SystemCrawling the feedback information from eBay UsersCrawling the feedback information from eBay Users

66

Power-Law Distribution of Peer FeedbacksPower-Law Distribution of Peer Feedbacks (10,000 eBay peers)

Feedback amount di — feedback amount of node i

Feedback frequency fd — the number of nodes with feedback amount d

77

Relative Performance of P2P Reputation SystemsRelative Performance of P2P Reputation SystemsComparison of PowerTrust with EigenTrust and PeerTrust Systems

PowerTrust[40]PowerTrust[40]

88

GossipTrust and PowerTrust GossipTrust and PowerTrust Reputation SystemsReputation Systems

P2P File Sharing P2P GridsOther P2P

Applications

Applications

Local Feedback

Generation

Global Reputation

Aggregation

Reputation Storage

Malicious NodeDetection andPunishment

Distributed Hash Table (Chord) / Unstructured Topology

PowerTrust / GossipTrust

((IPDPS-2006, IPDPS-2007IPDPS-2006, IPDPS-2007, , IEEE-TPDSIEEE-TPDS[40][40] April 2007) April 2007)

99

Reputation Systems for Reputation Systems for Unstructured Unstructured P2P NetworksP2P Networks

MotivationMotivation

– The Peer-to-Peer (P2P) architectures that are most prevalenThe Peer-to-Peer (P2P) architectures that are most prevalen

t in today’s Internet are decentralized and t in today’s Internet are decentralized and unstructuredunstructured

Challenges :Challenges :

– Short of secure hashing and fast lookup mechanismsShort of secure hashing and fast lookup mechanisms

– Most Scalable Reputation System were designed for Most Scalable Reputation System were designed for structstruct

ured ured (DHT-based) P2P networks (DHT-based) P2P networks

EigenTrust, PeerTrust, PowerTrust , …….EigenTrust, PeerTrust, PowerTrust , …….

1010

Reputation Systems for Reputation Systems for Unstructured Unstructured P2P NetworksP2P Networks

Meet 3 basic requirements– 1) computation and communication overheads for global reputati

on aggregation have to be considerably low. – 2) the amount of space for each node to store collected reputatio

n data has to be reduced as small as possible, preferably using a constant memory space.

– 3) the global reputation scores must be sufficiently accurate.

This paper proposes a gossip-based reputation system (GossipTrust) for fast aggregation of global reputation scores. It leverages a Bloom filter based scheme for efficient score ranking, thus GossipTrust does not require any secure hashing or fast lookup mechanism.

Bloom Filter http://en.wikipedia.org/wiki/Bloom_filter

1111

The GossipTrust SystemThe GossipTrust SystemScalable, Robust and Secure reputation system for P2P networksScalable, Robust and Secure reputation system for P2P networks

1212

Gossip-based Reputation AggregationGossip-based Reputation Aggregation

d ≤ logb with b = λ2

/ λ1, where λ1 and λ2 a

re the largest and se

cond largest eigenva

lues of the trust matr

ix S

g = O(log2n)

1313

1414

Gossip Protocol for Reputation AggregationGossip Protocol for Reputation Aggregation

– Tolerate the network link and node failuresTolerate the network link and node failures

– Support the computation of aggregate functions like Support the computation of aggregate functions like

weighted sum, average value and maximum over large weighted sum, average value and maximum over large

collection of distributed numeric values collection of distributed numeric values

– One thread sends the halved One thread sends the halved gossip pairgossip pair {½ {½ xxii ( (kk), ½ ), ½ wwii ( (kk)} )}

to itself (node to itself (node ii) and to a randomly selected node in the ) and to a randomly selected node in the

network.network.

Another thread receives all halved pairs from other Another thread receives all halved pairs from other

nodes and computes the updated nodes and computes the updated xxii((kk+1) and +1) and wwii((kk+1) +1)

– xxii is the is the gossiped global scoregossiped global score and and wwii is the is the gossiped gossiped

weightweight

1515

Gossip AggregationGossip Aggregation

vv11((tt)=0.5)=0.5 v v22((tt)=0.33 )=0.33 vv33((tt)=0.17)=0.17

ss1212=0.2 =0.2 ss2222=0 =0 ss3232=0.6=0.6

The updated global score of node The updated global score of node N2N2

is calculated as is calculated as

vv22((t+1t+1) = ) = vv11(t)×(t)×0.20.2 + + vv22(t)×(t)×0 + 0 + vv33(t)×(t)×0.6 0.6

= 0.5= 0.5××0.2 + 0.330.2 + 0.33××00+ + 0.170.17××0.6 = 0.2 0.6 = 0.2

Figure 2. Gossip aggregation to generate the global score at node N2, where weighted {local score, weight} are distributed at (a) Step 1 and (b) Step 2 subsequently to reach the consensus of a global score 0.2 at node N2.

1616

GossipTrustGossipTrust

GossipTrust is an extension of gossip protocolGossipTrust is an extension of gossip protocol– It uses the gossip protocol to aggregate reputation scIt uses the gossip protocol to aggregate reputation sc

ores.ores.– Each peer Each peer ii has normalized local score has normalized local score ssijij for all other for all other

peer peer jj– Global scores Global scores vvjj((tt) = ) = vvjj((tt-1)*-1)*ssijij

– Exchange with other peers to update Exchange with other peers to update vvjj((tt))– Stop if |Stop if |vvjj((tt) – ) – vvjj((t-t-1)| < threshold1)| < threshold

It treats all opinions in gossip procedure with the It treats all opinions in gossip procedure with the same weight regardless of the sources of the opisame weight regardless of the sources of the opinions.nions.

1717

Distributed Aggregation AlgorithmsDistributed Aggregation Algorithms

1818

Distributed Aggregation AlgorithmsDistributed Aggregation Algorithms

1919

Bloom-filter based Reputation StorageBloom-filter based Reputation Storage Bloom filters for reputation rankingBloom filters for reputation ranking – Example:Example:

a P2P network with 6 nodes, labeled as {0, 1, ..., 5}. a P2P network with 6 nodes, labeled as {0, 1, ..., 5}.

v0v0 = 0.05, = 0.05, v1v1 =0.2, =0.2, v2=v2=0.3, 0.3, v3v3 = 0.1, = 0.1, v4v4 =0.3, =0.3, v5=0.05v5=0.05..

Categories 1: {0, 1, 3, 5} and Category 2: {2, 4}Categories 1: {0, 1, 3, 5} and Category 2: {2, 4}

mm = 8 bits per filter. = 8 bits per filter.

h1h1((xx) = ) = x x ModMod (8) and (8) and h2h2((xx) = ) = xx+2+2

2020

Bloom filter-based Reputation Storage Bloom filter-based Reputation Storage

Figure 4 Bloom filter storage scheme for efficient retrieval of ranked peer reputations.

2121

Performance of Bloom Filter StoragePerformance of Bloom Filter Storage

2222

Peer CollusionsPower nodes : the most reputable nodes in a P2P reputation system– In GossipTrust, the power nodes are reelected periodically base

d on the updated global scores after each round of reputation aggregations.

We apply identity‐based encryption (IBC) to secure gossip message passing between nodes– The sender encrypts the message and signs the encrypt messa

ge. – The receiver verifies the message and decrypt message. – Some small delays may be introduced from signing, encryption,

verification, and decryption.

2323

Simulated Performance ResultsSimulated Performance Results

Environments:– We designed an event-driven simulator of GossipTrus

t. – We construct a Gnutella-like flat unstructured P2P net

work initially consisting of 1,000 nodes. – The simulation experiments run on a dual processor

Dell server with Linux kernel 2.6.9.– Each data reported is averaged over 10 simulation ru

ns.– We use 5 Bloom filters in each score-ranking class.

2424

Convergence Rate vs. Gossip ErrorConvergence Rate vs. Gossip Error

Figure 5. Gossip step counts of three GossipTrust configurations for various gossip error thresholds

•The convergence overhead increases with the decrease of gossiping error threshold and growth of network size.

2525

Figure 6 Gossip step counts of GossipTrust under various message loss rates. a) Network n= 1,000 nodes. (b) Fixed error threshold ε = 10-4

•As the percentage of message loss increases, the convergence overhead increases faster.•GossipTrust can tolerate moderate percentage of message losses up to 15% caused by the dynamism or unreliability of peers and link failure in the network.

2626

Figure 7: Effects on global aggregation errors by fake trust scores from malicious peers in a 1,000-node P2P network. a) Independent peers. (b) Collusive peers

•In general, the possibility of a peer behaving maliciously increases inversely with respect to its global reputation score.

2727

Aggregation Error vs. Malicious PeersAggregation Error vs. Malicious Peers

Figure 8. Effects on RMS aggregation errors due tofake scores by malicious peers in a 1000-node P2P network

•By Fig. 7 and Fig.8, we see that proper choices of α and θ can control the error in the global reputation aggregation process.

2828

Figure 9 Query success rate of a 1000-node GossipTrust system, compared with a no-trust system

•NoTrust system randomly selects a node to download the file without considering node reputation.•With the help of GossipTrust, even when the system has 20% malicious peers, it can still maintain around 80% query success rate.

2929

Figure 10. Average job turnaround time in PSA benchmark tests on a simulated P2P Grid. (a) Effect of job number. (b) Effect of malicious peer density.

•GossipTrust has 40% performance gains over the NoTrust system under 10% malicious peers. •Even when the malicious peers increase to 25%, the GossipTrust has a 75% performance gain.

3030

ConclusionsConclusions

This paper proposes a gossip-based reputation system (GossipTrust) for fast aggregation of global reputation scores. It leverages a Bloom filter based scheme for efficient score ranking.The simulation results demonstrate that GossipTrust has small aggregation time, low memory demand, and high ranking accuracy.

3131

ReferencesReferences[2] S. Boyd, A. Ghosh, B. Prabhakar, D. Shah, “Randomized Gossip Algorithms”, IEEE Trans. on Information Theory, June 2006, 52(6):2508-2530.[4] A. Broder and M. Mitzenmacher, “Network Applications of Bloom Filters: A survey”, Conf. on Comm., Control, and Computing, 2002.[14] M. Jelasity, A. Montresor and O.Babaoglu, “Gossip-Based Aggregation in Large Dynamic Networks”, ACM Trans. on Computer Systems, Vol.23, No.3, August 2005.[16] S. Kamvar, M. Schlosser, and H. Garcia-Molina, “The Eigentrust Algorithm for Reputation Management in P2P Networks”, ACM WWW’03, Budapest, Hungary, May 2003.[27] A. Singh and L. Liu, “TrustMe: Anonymous Management of Trust Relationships in Decentralized P2P Systems”, IEEE Intl. Conf. on Peer-to-Peer Computing, Sep. 2003.[35] L. Xiong and L. Liu, “PeerTrust: Supporting Reputation-based Trust for Peer-to-Peer Electronic Communities”, IEEE Trans. Knowledge and Data Engineering, Vol.16, No.7, 2004, pp. 843-857.[40] R. Zhou and K. Hwang, “PowerTrust: A Robust and Scalable Reputation System for Trusted P2P Computing”, IEEE Trans. on Parallel and Distributed Systems, April 2007, pp.460-473.[41] R. Zhou and K. Hwang, “Gossip-based Reputation Aggregation in Unstructured P2P Networks”, IEEE Int’l Parallel and Distributed Processing Symposium, March 2007.