Improving Peer-to-Peer Networks “Limited Reputation Sharing in P2P Systems” “Robust Incentive Techniques for P2P Networks”

Improving Peer-to-Peer Networks

“Limited Reputation Sharing in P2P Systems”

“Robust Incentive Techniques for P2P Networks”

In this talk …

• Peer-to-Peer Networks– Properties / Challenges– Threats– Solutions

Peer-to-Peer Networks

• Large Populations

• High Turnover

• Asymmetry of interest

• Zero-cost identity

Peer-to-Peer Networks (2)

• Threats– Malicious users: share bad resources– Free-Riders: selfish users

• Solutions– “Limited Reputation Sharing in P2P Systems”

by Sergio Marti and Hector Garcia-Molina

– “Robust Incentive Techniques for Peer-to-Peer Networks”by M. Feldman, K. Lai, I. Stoica and J. Chuang


by Sergio Marti and Hector Garcia-Molina

• Studies the performance of P2P resource-sharing network in the presence of malicious nodes

• Uses limited information sharing between peers• Studies the effect of reputation systems

System Model

• When a peer desires a file:– Queries all or a subset of the peers for the file– Collects responses (response set)– Selects a provider and accesses the resource– Verifies the authenticity of the file: V(R)– If V(R) determines that the file is invalid, the

peer fetches another file from the response set

• Peer wants to minimize the use of V(R)


Threat Model: Malicious peers

• Propagate fake files• Can pass false information• Three types of malicious nodes:

– N: Do not misinform– L: Lie about everyone– C: Collude and give bad ratings to good nodes and good

ratings to bad nodes

• Subversion Set: the set of unique fake files• πB: Percentage of malicious peers• pB: Probability of a file to be in the Subversion set


Reputation Systems

• Peers collect statistics about other peers and form opinions about them (reputation rating)

• 0 <= Reputation Rating <= 1• Peers can share reputation ratings• Selection procedure: selects from which peer to

download based on reputation ratings


Reputation Systems (2)

• Random Selection: no reputation used• Local Reputation System: uses only local

reputation ratings– File selection:

• Select-Best – high load on few nodes

• Weighted – tries to distribute load

– Friend-Cache (FC) with max size |FC|


Reputation Systems (3)

• Voting Reputation System: uses its own Reputation Rating R(q,r) and RRs of a quorum (Q) of its peers R(v,r). The combined rating is:

– Quorum selection:• Neighbor-voting• Friend-voting

• Additional parameters: 0: default rating T: selection threshold

• Whitewashing


Metrics• Efficiency

– Total number of V(R) evaluations:

– Verification ratio:

• Load– Load on node i:

– Average load:

• Message Traffic:


, rv 1

Metrics (2)Summary of Terms:


Simulation

• Fully connected power-law network– N = 1000 nodes– Max node degree dmax = 40– davg = 3.1– TTL = 5

• Local reputation: 100 queries from a single node; 1 turnover per step; all malicious nodes change identity every 10 steps

• Voter Reputation: 100,000 queries; 1 turnover per 1000 steps; all malicious nodes change identity every 10,000 steps


Simulation (2)


Voting System Parameters


Random:

rv = 28.2

Local Reputation

Voting System Parameters (2)


Malicious node type: N

Friend voting

Voting System Parameters (3)


Malicious node type: C

Friend voting

Efficiency Comparisons (1)



Efficiency Comparisons (2)



Load on Good Nodes (1)


Load on Good Nodes (2)


Message Traffic


Conclusion• Reputation rating decreases the times V(R) is performed

• Reputation sharing is less susceptible to malicious nodes

• White-Washing does have an impact

• Friend-Cache has to be big (around 100) to minimize Message Traffic. However, only the top few friends should be asked to vote

• Weighted and Best selection criteria have comparable efficiency, but Weighted selection distributes Load among nodes better


“Robust Incentive Techniques for Peer-to-Peer Networks”

by M. Feldman, K. Lai, I. Stoica and J. Chuang

• Game Theoretic approach to address free-riding in P2P networks

• Design better incentives based on Generalized Prisoner’s Dilemma

Model• Social Dilemma: Universal cooperation should maximize

social welfare, but peers should benefit exploiting the cooperation of others

• Asymmetric Transactions: Should handle asymmetric payoff

• Untraceable defections: defections should be transparent

• Dynamic Population: Peers should enter and leave the system as they please and be able to change their strategy

“Robust incentive Techniques for Peer-to-Peer Networks”

Generalized Prisoner’s Dilemma

• Client Always to cooperate (request service)• Server chooses whether to Cooperate or Defect based on

previous experience with the client• Client cannot trace defections• Agents alternate between client/server every game round• Agents can exploit blind cooperation from others


Population Dynamics• In each round of the game, each player plays one

game as client and one game as server• At the end of each round, players can:

– Mutate

– Learn

– Turnover

– Stay the same


Strategy• A strategy consists of:

– Decision function– Private or shared history– Server selection mechanism– Stranger policy

• Examples of Strategies:– 100% Cooperate– 100% Defect

• We want to strike a balance between the two:– Cooperate with good peers– Defect bad peers


Decision Function• For peer i:

– pi is a measure of services it has provides– ci is a measure of services it has consumed

• i’s generosity:– Discriminates against peers that (for the time being) have

consumed more than they have provided. Server will deny service even if it is of the same type

• Normalized generosity: – Servers treat peers with respect to their own generosity

• Reciprocative decision function = Cooperate with i with Pr. min(gj(i), 1)


Simulation

• Server Selection, History and Stranger Policy are shown through simulation results


Baseline Results


Baseline Results (2)


Server Selection

• Transactions are asymmetric– Most likely to receive service from peer that I gave

service

– Also likely to receive service from a peer that provided service earlier

• Store past servers and clients• Randomly choose a server or a client for new

transactions


Shared History

• Every player knows about transactions that other players perform

• Increases the size of “known” servers to select from

• Advantage: Scales better with:– Large populations

– High turnover

– Asymmetry

• Drawback: – Susceptible to collusion

– Harder to implement


Simulation Results (Shared History & Selection)


Simulation Results (Shared History & Selection) 2


Shared History Attack: Collusion

• Group of peers lie about transactions• Positive (bad peer claims another bad peer

cooperated)• Negative (bad peer claims that good peer

defected)• Bad for objective (trust everyone) reputation• Need subjective reputation mechanism


Shared History Attack: Collusion (2)

• Can use Maxflow graph algorithm as a distributed, subjective reputation system

• Reputationi(j) = Services Received/Services Provided =


Shared History Attack: Collusion (3)


Shared History Attack: False Reports

• One node may want to lower the reputation of another

• Easily identified: one node says transaction occurred and the other says it didn’t

• Solution: Modify MaxFlow so that a node always believes nodes closer to it.


Zero-cost identities

• Motivates whitewashing• Problem:

– a node generous to newcomers motivates whitewashing– A node stingy with newcomer discourages newcomers

to join the network

• Solution: Stranger Adaptive Strategy– Cooperate with stranger with Pr. min(ps/cs, 1)

• Susceptible to traitor attack• Solution: Bind ps + cs to a constant k and keep only ratio

r=ps/cs


Zero-cost identities (2)


Zero-cost identities (3)


Traitors


Conclusion• A strategy for P2P networks:

– Model over Asymmetrical Generalized Prisoner’s Dilema

– Decision function: • Reciprocative function = Cooperate with i with Pr. min(gj(i), 1)

– Shared (short term) history• Use maxflow to rate peer reputation

• Adaptive Stranger policy

– Discriminative selection mechanism


Summing it all up …

• For a good P2P system we need:– Reputation rating (based on outcomes of past

transactions)

– Reputation sharing between peers

– Subjective reputation (weighted function vs. maxflow)

– Eliminate White-Washing or decrease its effectiveness

– Short History

Improving Peer-to-Peer Networks “Limited Reputation Sharing in P2P Systems” “Robust Incentive Techniques for P2P Networks”

Documents