1. A Reputation-Based Approach for Choosing Reliable Resources in Peer-to-Peer Networks E. Damiani S. De Capitani di Vimercati S. Paraboschi P. Samarati F. Violante http://seclab.dti.unimi.it/p2prep/

2. Outline 1. 1. P2P Networks & Gnutella protocol [1] 2. 2. XRep protocol [2] 3. 3. Security considerations on XRep 4. 4. Conclusions 5. 5. References

3. P2P Networks All the nodes offer the same services and follow the same behavior. Nodes act both as servers and as clients. There are no nodes with a special responsibility to monitor or supervise the network behavior. P2P networks for file sharing involves two phases: 1 Search of the peers where the requested file resides. 2 Download from the exporting peers the requested information.

4. Gnutella Gnutella is a protocol for distributed search. Nodes are called servents.(server + client) Each servent has a servent_id Servents communicate by exchanging descriptors. Ping, Pong – are used to discover servents Query, QueryHit – Searching files in the P2P network Push – allows a firewalled servent to contribute files to the P2P network The Gnutella Protocol Specification v0.4 www9.limewire.com/developer/gnutella_protocol_0.4.pdf

5. Descriptor routing 1 A servent P requiring a resource broadcasts a Query out. A servent S will respond with a QueryHit if a match is found against its local database. And S will forward incoming Query descriptors to all of its directly connected servents, except the one that delivered the incoming Query. QueryHit descriptors may only be sent along the same path that carried the incoming Query descriptor.

6. Descriptor routing 2 A servent will decrement a descriptor header’s TTL field, and increment its Hops field, before it forwards the descriptor to any directly connected servent. A servent will decrement a descriptor header’s TTL field, and increment its Hops field, before it forwards the descriptor to any directly connected servent. If, after decrementing the header’s TTL field, the TTL field is found to be zero, the descriptor is not forwarded along any connection. If, after decrementing the header’s TTL field, the TTL field is found to be zero, the descriptor is not forwarded along any connection. A servent receiving a descriptor with the same Payload Descriptor and Descriptor ID as one it has received before, should attempt to avoid forwarding the descriptor to any connected servent.

7. Gnutella1 2 3 6 5 7 8 1: initiator 2&7: responders Query QueryHit Not a descriptor 4 Match A servent requiring a file broadcasts a Query out. Servents will forward incoming Query descriptors to all of its directly connected servents, except the one that delivered the incoming Query. A servent receiving a descriptor which has received before will not forward the descriptor A servent will respond with a QueryHit if a match is found against its local database. QueryHit descriptors may only be sent along the same path that carried the incoming Query descriptor9 If the TTL field is found to be zero, the descriptor is not forwarded along any connection.

8. Structure of Gnutella descriptor Every descriptor has two parts: 1. Header Descriptor IDPayload DescriptorTTLHopsPayload Length Minimum SpeedSearch criteria Number of Hits PortIPSpeedResult SetTrailer Servent ID File Index File Size File Name file1 file2 file3 Query: sent by initiator QueryHit: sent by responder 2. Payload +

9. Motivation of Reputation systems Most P2P systems protect peers’ anonymity. Anonymity opens the door to possible misuses and abuses. No way to verify the source or content of files --Bad service, low quality files --The content of a file is different than the title --Trojan horses and viruses e.g. Mandragore – a Gnutella worm --Act as a servent and answer all Queries. --Provides a renamed copy of itself for downloading. Peer review process: the peers’ opinions is used to establish a reputation for peers and files.

10. XRep: Basic Assumptions Each servent maintains information on its own experience on files and other servents and share such experience with others upon request Each servent has a servent_id which is a digest of a public key obtained using a secure hash function Servent reputations are associated with the servent_id Each file has a resource_id which is a digest computed by applying a secure hash function to the file content File reputations are coupled to resource_id

11. Reputations Storage & votes calculation 1 Each servent maintains a resource_repository & a servent_repository that store its opinions about files and servents it had experiences A servent votes on files and servents with which it had experiences.Votes are its opinion on files and servents Votes are expressed on the basis of information available in the resource_repository & servent_repository.

12. resource_repository: set of pairs (resource_id, value) value=0 or 1 servent_repository: set of triples (servent_id, num_plus, num_minus) num_plus, num_minus are positive integers Vote = 0 or 1 Vote = 0 or 1 Vote of servent =1, if Vote of servent =1, if num_plus>>num_minus Vote of file = Vote of file = value Reputations Storage & votes calculation 2

13. XRep: Polling Protocol Phase 1: Resource searching. p sends a Query message for searching files, and servents matching the request respond with a QueryHit Query ( Min_Speed, Search_criteria ) QueryHit (num_hit,IP,port,speed,Result_set,trailer,servent_id) Trailer: resource_ids of files in result set Initiator p Servent s

14. Phase 2: Vote polling P selects a file r that best seems to satisfy its request. Such selection may be guided by the user’s preference p polls its peers about the reputation of a file and the set of servents that offer it. Servents wishing to respond vote on the resource_id and servent_ids and send back a PollReply Initiator p Servent s Poll (resource_id, {servent_id 1… servent_id n }, PKpoll) PollReply ({(IP,port,Votes)} PKpoll )

15. Phase 3: Vote evaluation & reliability check 1. p decrypts PollReply, discards tampered ones. 2. p clusters Voter’s IP and weight the votes within a cluster --Reducing the effect of a clique 3. p selects a set of voters in each cluster, contacts them directly, and expects back confirmation messages. If not enough responses, then p repeats step 3. Initiator p Servent s TrueVote ( Vote ) TrueVoteReply ( responses )

16. Phase 4: Best servent check p cannot always download file from best servent p contacts the best servent S to check the fact that it exports file which p wants to download Preventing ID stealth attack. Initiator p Servent s AreYou (servent_id s, resource_id) AreYouRepley({response} SKs, PK s )

17. Phase 5: Resource download p selects a servent s, downloads the file r and checks it against its digest Update its resource_repository & servent_repository Initiator pServent s download ( r ) resource ( r )

18. Query (Min_Speed, Search_criteria) QueryHit (num_hit,IP,port,speed,Result,trailer,servent_id) Initiator p Servent s Poll (resource_id, {servent_id 1… servent_id n }, PKpoll) PollRepley ({(IP,port,Votes)} PKpoll ) TrueVote ( Vote ) TrueVoteRepley ( responses ) AreYou (servent_id s, resource_id) AreYouRepley({response} SKs, PKs) download ( r ) resource ( r ) { { { { { Vote polling Resource Searching Vote evaluation download Best servent check

19. XRep: Security Considerations XRep allows to protect P2P against following attacks   Self replication   Man in the middle   ID Stealth   Pseudospoofing   Shilling

20. Self replication: A malicious servent could answer all Queries and provide doctored content. Even honest peers, unaware of the malicious content, could share it and contributing to its diffusion. e.g. Mandragore – a Gnutella worm   Bad reputations of file -- Worms slightly modifying themselves  Cannot collect positive recommendations  Check reputation of the servent

21. Man in the middle: A broadcasts a Query. B responds a QueryHit. E replaces IP and Port of the QueryHit with E’s IP and Port, sends it back to A. A may download from E.   The fake content provided by E will not match the digest of the legitimate file, then be discarded. (Phase 5) A B E Query QueryHit Modified QueryHit Download

22. ID Stealth: A malicious peer answers with two QueryHit s, carrying the digest of a doctored file and one of them carrying the ID of a reputable servent   Xrep checks whether the best servent is offering that file (Phase 4).

23. Psedospoofing & Shilling: 1 Attackers create and control multiple servents. They give positive votes to the attacker. Four cases:   Multiple servents have same IP address   IP cluster (phase 3)   Servents have different but faked IP address   TrueVote/TrueVoteRepley (phase 3) These two cases are called Psedospoofing

24. Psedospoofing & Shilling: 2   Servents have different real IP address. And those IP addresses have same net_id.   IP cluster may reduce the effect (phase 3)   Servents have different real IP address. And those IP addresses have different net_id.   To ensure a high number of voters. These two cases are call Shilling.

25. Distribution of Servent & Resource An important aspect for the applicabilty of this approach frequent files are more frequently searched => the number of votes will be high few servents offering many files => these servents will probably well know Cold-start problem

26. Conclusions XRep is a reputation management protocol for anonymous P2P environments It prevents malicious behaviors in P2P network Future work: --reputation mechanism with supernodes --performance optimization

27. References [1] The Gnutella Protocol Specification v0.4 www9.limewire.com/developer/gnutella_protocol_0.4.pdf [2] “ A Reputation-based Approach for Choosing Reliable Resources in Peer-to-Peer Networks," E. Damiani, etc. [3] http://www.limewire.com/