1. 1 Maze A Hybrid P2P file sharing system Design by Networking and distributed System lab at Peking University Presenter:Elaine

2. 2 Outline The design of Maze ( USENIX Worlds ’ 04) An Empirical Study of Free-Riding Behavior in the Maze P2P File-Sharing System (IPTPS ’ 05 ) Robust Incentives via Multilevel Tit-for-tat (IPTPS ’ 06 ) P2P search mechanism in Maze (ACM INFOSCALE '06)

3. 3 What is Maze? One of the first large scale academic P2P-file sharing system More than 2,300,000 Registers Peak moment: more than 100K users Average per day:30K users 13TB of data are exchanged daily

4. 4 Structure of Maze (5) Clients contcat to multiple replicas and perform “swarm download”

5. 5 What will central sever log? Metadata –owner ID –file name –File type –Size Each transaction information User ’ s reputation point

6. 6 Specialty in Maze Central metadata indexing and query processing Social network –Download network –Friend network Eventually need to reduce the dependencies upon the central server. These friends form the bases over which we plan to add P2P search capabilities. Maze forum Incentive model

7. 7 Maze Points system 1. New users are initialized with 4096 points. 2. Uploads: +1.5 points per MB uploaded 3. Downloads: -1 point per MB downloaded within 100MB, -0.7 per additional MB between 100MB and 400MB, -0.4/MB between 400MB and 800MB,and -0.1 per additional MB over 800. 4. Downloads requests are ordered by T = requestTime− 3 ∗ logP, where P is a user ’ s point total. 5. Users with P < 512 have a download bandwidth quota of 300Kb/s.

8. 8 An Empirical Study of Free-Riding Behavior in the Maze P2P File- Sharing System

9. 9 In-depth analysis of the effectiveness of the incentive policies and how users react to them Definition of Free-riding behavior –Peers who only download without contributing.

10. 10 Experiment Methodology The central servers log the following information per client: Unless otherwise stated, results are analyzed using logs from 9/28 to 10/28.

11. 11 Statistics suggest the existence of free-riding Results are analyzed using logs from 9/28 to 10/28. server-like users client-like users The ratio is 4.4:1. We found that client-like users are responsible for 51% downloads but only 7.5% uploads. So, a big portion free-Riders exist!

12. 12 Are they not capable to contribute ? Or not willing to contribute? –What factors contributes to the free- riding behavior?

13. 13 When a user ’ s point is low.. He can do: –Legal Contribution –Illegal Whitewashing Sybil ’ s attack

14. 14 The only way that the free-riders can survive the Maze system without cheating is through contribution –From 9/28~10/28, we found that the top 10% popular files account for more than 98.8% of total transfer traffic, and over half of which were downloads from the client-like user. –So they got lots of popular files! They can easily make back their deficits provided that 1.The Maze system can quickly direct queries to them and 2.Their contents are available

15. 15 The central index sever is indeed a factor! –Currently, new content of a peer does not make into the index until a few days later –Adding more indexing servers. –A more complete solution is to implement the P2P searching in the future releases

16. 16 Do they hide the files? –The study in [6] shows that 70% of Gnutella users do not have any files to share –The average number of shared files of client-like users is 491, versus 281 of the server-likeusers –Not the case in Maze

17. 17 So what mainly made them become free riders? Users with positive point changes have longer session time, on average 2.89 times more than those with negative point changes (218 minutes versus 75 minutes)

18. 18 Brief Summary Points system is effective in general –But can ’ t avoid free-riders using account whitewashing. Query and search mechanism, and we can accomplish it by installing P2P searching mechanism and/or increase the frequency of updating the central index. More savvy incentive policies (e.g. encourage people to increase their online session durations and punish whitewashing behavior

19. 19 Robust Incentives via Multi- level Tit-for-tat

20. 20 ;.;. Private History –Giving higher priority to peers with whom he has successfully downloaded Shared History –Non-subjective Looking at the overall contribution of a peer –subjective Reputation is given by all other peers, weighted by the reputation of the assigning peers

21. 21 Existent cheating We analyzed the complete log of all transactions in the Maze system over a one month period, more detail in [11] –Besides the expected free riding behavior and user whitewashing –User collusion Pair-wise collusion Spam account collusion –a form of whitewashing that allows whitewashed points to be collected at a single user

22. 22 The limitation of other incentive systems Private history coverage problem. –A one month download log only enforces Tit- for-Tat to only 2% of a peer ’ s upload –Peer have no opinion of the requesting peer. The blind uploads bring a lot of opportunity for free-riders.

23. 23 Share history –The Eigentrust Algorithm –Terminology Local trust value: c ij. The opinion that peer i has of peer j, based on past experience. Global trust value: t i. The trust that the entire system places in peer i. P1 P3 P7 P4 P2 P5 P8 P1 P5 P3

24. 24 The Math

25. 25 The Eigentrust Algorithm Fortunately, if n is large, the trust vector ->ti will converge to the same vector for every peer i.

26. 26 EigenTrust problem Also EigenTrust helps to punish colluders in Maze But suffers from both false negatives and false positives

27. 27 False negatives Leg hugger –High reputation peers randomly download from a colluder Example –Larry is a spam account colluder that we detected., e.g. Most of Larry ’ s 30GBs uploads are to other colluders –200MB uploads to some reputable peers boosts his rank nearly 100 times Another 734KB upload to one super peer further promotes its rank to 6.6*10 -5. –Larry gets a higher rank than most of peers who upload more than 30GB to legitimate users

28. 28 False positive Satellite networks –Encourage peers in satellite networks download from outsede peers

29. 29 So, each of the above incentives has their own weakness.. Private history –Coverage Shared history –Non-subjective Solve above,but Collusion –Subjective Solve above,but False positive False negative The multi-level Tit-for-tat!

30. 30 Designing Multi-Trust incentive M be a N X N matrix that defines a one step Rank among peers, i.e., M i,j is peer j ’ s rank from i ’ s perspective. The two step rank matrix (one level indirect trust) can be expressed as M 2. The entry (M 2 ) i;j aggregates other peers one step rank to yield the rank of j from i ’ s perspective. Imposes service differentiation by looking at which tier j falls into when its downloading request arrives at i. – The smaller level it belongs to, the higher priority it is given. – Within the same tier, two peers will be ranked according to their values in the matrix of that tier.

31. 31 Coverage experiment Measurement Metrics: –Trust upload ratio which is the volume of traffic that are served by either a M friend, or a friend that is either a {M or M 2 } friend, over the total traffic in the system –Using one month traffic log, simulated with incentive mechanism of Multi-Trust and Tit-for- tat. a.k.a Tit-for-tat

32. 32 Comparism to Eigentrust First, we use the completed transactions in our one month traffic log to calculate a set of ranking values for each peer. Next, the evaluator node extracts download requests for the following two week period, and statistically sorts the requesters into a service queue according to their local subjective ranking. Peers with lower queue positions are served first, i.e., higher ranked. Metric –peers ’ queue positions Expected results –True colluders will have queue positions no earlier than EigenTrust, whereas local distributors will move up

33. 33 Colluder punishment

34. 34 Leg-hugger Larry downloads from 73 peers in the two weeks following our one month traffic log It gets high rank in the EigenTrust because of his high rank friend. In multitrust his high reputable friend still helps him getting into 16 (22%) peers ’ trust list, but overall he is punished in all the other peers.

35. 35 Satellite cluster Wayne is the local distributor in one cluster. It downloads from 14 peers in the next two weeks, among them there are 11 external peers and 3 internal peers (peer 9, 10 and 13).

36. 36 Brief summary Multi-trust performs no worse than EigenTrust in punishing pairwise or spam account colluders. Meanwhile, it solves the two problems brought by EigenTrust –dropping the high rank of leg-hugger peer –rising the low rank of local distributor inside its own satellite cluster