Dear all, We have developed an algorithm for fighting spaming in publish/subscribe systems that has direct applicability to p2p file sharing systems. The rational behind this algorithm is: -You rank the content and not the users(therefore there is no need for persistent user identity) -You have positive votes only (therefore the fact that a user shares a file can be considered as a vote for this file, moreover there is no way to affect negatively the reputation of a file) -The more "similar" files a user share the less weight his vote has (It has been found that in p2p file sharing networks malicious users pollute them by advertising the same fake file with multiple fake but similar tags)
With simulations we found that in order for a single attack to take place (i.e., a malicious file to be distributed) there should be so many malicious users in the system as the number of the legitimate users that share the most popular version of a file You can find the related papers here http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5722360 and here http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5534464 (direct access to the pdf files from here http://pages.cs.aueb.gr/~fotiou/papers/inforanking.pdf and here http://pages.cs.aueb.gr/~fotiou/papers/ngi.pdf respectively) Best, Nikos On Thu, Jul 19, 2012 at 12:25 AM, James A. Donald <[email protected]> wrote: > On 2012-07-18 3:51 AM, Tony Arcieri wrote: >> >> The trust function is a collaborative filtering algorithm such as Slope >> One >> or Singular Value Decomposition. The inputs to the trust function are >> BitCoin style peer-specific "long chain" files of a peer's transfer >> history >> that a given peer has collected through direct interaction. We might look >> for the following: success/failure and transfer rate. >> >> The first thing we do is build a sparse matrix of the similarity of all >> peers to all other peers in the system (let's assume Slope One is the >> algorithm for now). This is what we'll use for computing trust. >> >> Now we actually do the collaborative filtering calculation: we look for >> peers that are similar to ourself, by inputting our own transfer history >> and the sparse matrix we calculated in the previous step. >> >> The output should be peers similar to ourself: namely ones which >> experience >> a similar history of success/failures and similar transfer rates. With >> enough information, this should begin to reveal this like which peers are >> "closest" to us on the network (i.e. least bottlenecked by the network >> relative to us, not geographical closeness or closeness in the DHT) > > > What we want is peers that are trusted by entities like ourselves, and/or > have engaged in transactions that are beneficial to entities like ourselves, > not those that allegedly trust entities that we trust and have allegedly > engaged in transactions like those that we have engaged in. > > Slope One and Singular Value Decomposition does not trivially give us that, > and it is not immediately obvious to me how to fix them to give us that. I > expect it can be done, just do not quite see how. > > _______________________________________________ > p2p-hackers mailing list > [email protected] > http://lists.zooko.com/mailman/listinfo/p2p-hackers _______________________________________________ p2p-hackers mailing list [email protected] http://lists.zooko.com/mailman/listinfo/p2p-hackers
