1. Ivan Osipkov Fighting Freeloaders in Decentralized P2P File Sharing Systems

2. Introduction Gnutella/KazAa: 70% freeloaders, 1% serve 50% requests, files satisfy Zipf distribution Need to spread files according to popularity and curb freeloading, in order to provide QoS and eliminate central nodes Properties sought in a solution: 1. Minimize polling 2. Distributed solution 3. Collusion resistant 4. Sybil attack resistant

3. User Behavior in Gnutella Average user is online for 1 hour User identity may change or several identities present at the same time Evolutionary Prisoner’s Dilemma Sims: 1.New users should be treated with distrust but they should be able to start quickly 2.A user needs to have some info about other peer’s interactions. Otherwise, the game will not scale as population rises and non-optimal strategies will be followed.

4. Proposed Solutions Polling: Our Goals: Witnesses 1.Overhead 2.Incomplete Picture 3.DoS 4.Who’ll participate? 1.Overhead 2.Who’ll participate? 1.No polling 2.Avoid public service

5. Contributions Off-line participation evaluation. No 3rd party brokers or polling: only PKI assumed Upload activity proportional to download activity. Popular files spread more, thus improved load-balance Progressive taxation on accumulated credit. Easy to start participating Distributed data for undeniable collusion detection Works on top of Gnutella-like protocol and can be integrated with it

6. Offline Participation Evaluation If peer A sends file F to peer B, A obtains a signed download receipt from B, and B obtains upload receipt. The receipts state the time-interval of the file transfer and IDs of peers involved. Each peer presents its receipts when requesting a download Credit of peer is calculated based on the receipts and the peer is put into a priority queue AB F AB {A,B,dld,T1,T2} sign(A) {A,B,upld,T1,T2} sign(B)

7. Credit Calculation Given a time frame (s,t) and current time T, credit contribution of this time-frame is f(s,t,T) where f "ages" as T increases. One adds all contributions of upload receipts, subtracts contributions of download receipts and of "unaccounted time-intervals“ Old receipts contribute little

8. Properties of Credits Old receipts can be discarded Dumping of download receipts leads to "unaccounted time-frames" for which peer may be charged even more. The more credit peer has, the more it loses to aging. Thus downloads cost more and uploads generate less credit: PROGRESSIVE TAXATION As a consequence of the formulas, credit has upper and lower limits.

9. General Bandwidth Let b Kb/sec be "unit" bandwidth. A user with n*b bandiwdth creates n virtual (but related) peers. Each virtual peer can be involved in a single transaction at a time which should be finished without interruptions. If peer A has n*b bandwidth and B has m*b where n<b, then n*b can be dedicated on each side for download, but accounting is done on each virtual peer separately.

10. New Users New users get more credit from uploads than the old users. Thus if it first provides uploads to others it will be able quickly to obtain a viable amount of credit. Need to have some (popular) files initially Dynamic vs potential credit Should initial credit be given?

11. Collusion Attacks If undetectable from receipts then they raise total credit insignificantly If detectable then proof is undeniable: use intersection of receipts Data needed for proof is distributed but within similar time- frame Peers holding the compromising receipts are interested in prosecution

12. Collusion Attacks (Cont’d) (colluder) A B (colluder) TIME T1T2T3 A C B finally downloads T4 A B A and B can’t dld/upld to others during this time T5 C B B: lousy service and no receipt from A

13. In the beginning, it takes B three rounds to download, I.e. bandwidth usage is 33% (at most) Continuous collusion will force bandwidth usage to 25% What if A says “listen B, I’ll download your files for you, just keep giving me receipts”? Simulate longer downloads? Collusion Attacks: Conclusion

14. Sybil Attacks Not alleviated by the above mechanisms Need to do clustering (e.g. using IP addresses which should be pinged) when calculating total credit. Or use central CA

15. Simulations Setup 200 peers Files have Zipf popularity distribution Topology is random with average degree 3.4 Every user has initial files chosen at random but with popularity in mind 4 Levels of priority queue based on credit Gnutella routing/discovery simulated 1.65% freeloaders 2.25% honest peers 3.5% servers 4.5% newcomers

16. Simulations: Average History

17. Simulations: Bandwidth Usage

18. BitTorrent? You give to me, I give to you… NOW! The 2 peers need to have what each one needs (usually part of the same file) Used to offload work from the originator and increase bandwidth for large files Applicable to our case?

19. Future work Reputation system is a must: need to take action on misbehavior. It's orthogonal to the "economic" one. Reputations should be used when calculating credit similar approach can be used Receipts should be inspected (lazily?) for violations. This data is distributed and peers holding it are interested in detecting misbehavior. Receipts exposing violations have similar time-frames. Analysis and simulations of collusions is needed File-spreading simulations