REPUTATION SYSTEMS

1
REPUTATION SYSTEMS

• Gayatri Swamynathan
• CURRENT Oct 5 05

2
Outline

• Overview of reputation systems
• AEPP Paper

3
The Internet Today

• Growing popularity of online communities

• Centralized systems, e.g. EBay
• Decentralized and self-organizing networks

4
Decentralized networks qualities or setbacks?!

• Anonymous nodes
• Zero-cost identity switches

• Autonomous nodes
• Lack of Accountability
• Lack of a common driving goal

• Network is vulnerable to attacks from selfish and
  malicious peers
• Distrust runs high!

5
Research Focus

• Need for trust-mechanisms
• To assess trustworthiness of peers
• Use of reputation to build trust
• Higher the reputation, higher the reliability
• With a Reputation System
• Promote honest participation (avoid selfishness)
• Prevent malicious behavior (avoid maliciousness)
• Reputation Systems help secure decentralized
  networks

6
A Reputation System

• Helps establish mutual trust (distrust) by
  assigning a reputation to each peer
• How? Using post-transaction feedback
• Assumption past behavior is indicative of future
  behavior

7
Reputation Management
8
Reputation Management

• Production The generation of reputation
• Who generates reputation? How is it generated?

9
Reputation Management

• Generation The production of reputation
• Aggregation The reputation-based trust model
• What is the trust model? How are individual
  ratings converted to a reputation profile?

10
Reputation Management

• Production The generation of reputation
• Aggregation The reputation-based trust model
• Storage The storage of reputation data
• How is reputation data stored? What is stored?
  Where is it stored?

11
Reputation Management

• Production The generation of reputation
• Aggregation The reputation-based trust model
• Storage The storage of reputation data
• Communication The reputation exchange protocol
• How is reputation information exchanged? What
  parties are involved? What is transferred?

12
Reputation Management

• Production The generation of reputation
• Aggregation The reputation-based trust model
• Storage The storage of reputation data
• Communication The reputation exchange protocol
• Safeguarding reputation from attacks
• What are the security threats? How to protect
  the system from unfair ratings, and colluding
  parties? Is data integrity assured?

13
Production

• A reputation is generated by participants
  undertaking a transaction.
• Context of reputation resource-based OR
  peer-based
• Blacklisting peers
• Peer turnover

14
Aggregation

• Feedback/Ratings
• Number of transactions
• Credibility of feedback source
• Transaction context factor
• Size
• Time

15
Storage

• DHT-based approach
• Each peer holds trust data for multiple peers,
  and routing table
• Unstructured storage approach
• Certificates
• Experience Repository
• Data integrity redundancy, anonymity, digital
  signatures

16
Communication

• DHT-based approaches
• Search and update using query or insert messages
• Exploit communication protocol of base network
• Poll messages implemented on top of Query
  messages
• Specific exchange protocols

17
Reputation Management
Reputation Safeguards

• Dynamic peer personalities
• Unfair ratings
• Collusion
• Storage integrity
• Exchange integrity

Production
Aggregation
Communication
Storage
18
Reputation Management

• Production The generation of reputation
• Aggregation The reputation-based trust model
• Storage The storage of reputation data
• Communication The reputation exchange protocol
• Safeguarding reputation from attacks

19
Outline

• Overview of reputation systems
• AEPP Paper

20
AEPP 2005

• Decoupling Service and Feedback Trust in a
  Peer-To-Peer Reputation System
• (Gayatri S, Ben Y. Zhao, Kevin C. Almeroth)
• Robust to unfair ratings
• Dynamic peer personalities

21
Brief Background

• Reputation systems build trust by computing a
  peer's reputation as the average of its lifetime
  ratings.

22
Accounting credibility

• Correlated trust approach
• Credibility as service provider ? Credibility as
  feedback provider
• Weighted model

23
Decoupling Service and Feedback Trust
Service Reputation
Reputation profile
X
Feedback Reputation
24
Accounting credibility

• Decoupled trust approach

A
r1
Xs reputation profile
reputation (X) r1 feedback-reputation (A)
r2 f-reputation (B) r3 f-reputation (C)
X
r2
B
r3
C
25
Related Work

• PeerTrust
• a personalized similarity measure to more heavily
  weigh opinions of peers who have provided similar
  ratings for a common set of past partners.
• CONFIDANT
• a node's referral is interpreted subjectively per
  node
• not generic
• mobile ad hoc implementation

26
Reputation System

• Each peer OWNS two sets of reputation ratings
• A service rating (s-rating) and feedback rating
  (f-rating).

A
fAD -1
fAC 1
File (bad)
-1
B
C
D
1
-1
27
Trust Aggregation

• Decouple service and feedback trust
• Unfair ratings
• Weigh recent feedback more heavily
• Dynamic peer personalities

28
Performance Evaluation

• Decrease in the number of malicious transactions.
• Significant reduction in the number of false
  positives and negatives reported.
• In effect, increase in accuracy of reputation in
  the system.

29
Simulation Environment

• Simulations in C using tools built on the
  Stanford Graph Base (SGB)
• Graphs generated from the GT-ITM topology
  generator
• Model peer community
• Honest
• Dishonest
• Strategic

30
Simulation Environment
31
Experiment 1

• Measuring malicious transactions in a network
  with and without our reputation model
• 40 malicious nodes

32
Experiment 2

• Measuring malicious transactions in a network
  with and without our reputation model
• 50,000 transactions

33
Experiment 3

• Malicious transactions in networks with a
  conventional trust model and our decoupled model
• 50,000 transactions

34
Experiment 4

• Malicious transactions in networks with a
  conventional trust model and our decoupled .
• 40 malicious nodes, the percentage of strategic
  nodes varies.

35
Experiment 5

• False positives and negatives in a network with
  a conventional trust model and our decoupled
  model
• 40 malicious nodes, the percentage of strategic
  nodes varies

36
Open Problems

• Collusion
• Identity management (sybil-proofing)

37
At this point

• I hope to have familiarized you with
• Reputation systems
• A bit on my research
• Thank you!