Experience with an Object Reputation System for Peer-to-Peer Filesharing

1. Experience with an Object Reputation System for Peer-to-Peer Filesharing Kevin Walsh and Emin Gun Sirer Presented by Steve Ko

2. Problem • Object reputation or authenticity (“pollution” in p2p) • Q: downloaded file == what it claims to be • superman_returns.mpg could be a virus or a malware • The cause • Filesharing apps use meta-data for searching • Meta-data like name, encoding, content-hash, etc • Users blindly believe the meta-data. • Need a mechanism to evaluate the authenticity of downloaded objects

3. Possible Approaches • Use sharing (popularity) as an indicator • Not reliable • A large # of “attackers” can share the same malicious file. • Angry users can also share  • Use voting • A user casts a vote on the downloaded object’s authenticity. • Others see the votes and decide the authenticity. • Problem: trust!

4. Credence • Use voting and deal with the trust problem • How? (A bird’s eye view) • Compare voting history of two peers • Trust peers with identical votes more • If they don’t share enough history, build a trust relationship graph and trust multi-hop peers (transitive trust)

5. Credence • Voting history (1 == correct, 0 == incorrect) Local Trust Local Trust A B C Transitive Trust

6. Architecture • Vote database • Stores votes • (file content hash, timestamp, own vote, list of other votes) • Trust relationship graph • Who trusts whom by how much • Correlation table • Stores history similarities (i.e. degree of trust) • (peer id, correlation value) • Correlation values derive from an equation • Extended by the trust relationship graph

7. Correlation Computation • Local trust (between peer A and B) • θ = (p-ab) / √a (1-a) b(1-b) • a = (agreeing votes) / (all votes of peer A) • b = (agreeing votes) / (all votes of peer B) • p = (agreeing votes) / (all shared votes between A and B) • This gives values in [-1, 1] • Positive correlation (> 0): two peers agree • Negative correlation (< 0): two peers disagree • Transitive trust (among peer A, B, & C) • θac = θab * θbc

8. Algorithm • Assume a filesharing app with search capability (e.g. Gnutella) • Voting • After downloading, cast a vote – either thumbs-up or thumbs-down • Store the vote locally to the vote database

9. Algorithm • Evaluating • Send out a vote-gather message on a specific file (using the propagation method of search) • Receive votes and store them in the vote database • Calculate the weighted average of votes using correlation values

10. Evaluation • Modified Gnutella client • 10,000 public downloads since Mar, 2005 • Results from 9-month data • Using a crawler to gather data • 1200 clients, 39000 votes, and 84000 shared files

11. Trust Relationship Graph • Positive & negative correlation

12. User Classification • 35% of altruistic users, 50% of non-participants, and 15% of attackers

13. Local & Transitive Correlation • Not many high-quality correlations % of peers with valid correlation values

14. Vote Classification • Coordinated attackers cast a lot of votes # of votes cast

15. File Popularity • By number of times shared • By number of hosts

16. File Popularity by Voting • More negative votes than positive votes

17. Sharing and Voting • They are independent # of votes cast # of files shared

18. Decoy Attacks • Detect most decoy attacks Decoy attacks encountered Size of decoy attacks

19. Conclusion • A small fraction of users share malicious files • Malicious files are more popular than normal files • A voting system can be used to mitigate the attack