Security in Collaborative Filtering Systems

In this paper, we provide a detailed discussion of security in collaborative ltering systems, shared services that provide personalized recommendations to users based on other users with similar tastes. We show that collaborative ltering systems, as they are deployed today, are vulnerable to a wide range of individual and group attacks whose consequences include bogus, highly biased predictions, severe loss of privacy, and more. A signiicant fraction of this paper is devoted to mapping the space of attacks out as completely as possible. For each possible attack, we examine the nature of attack, its consequences, and propose countermeasures whose goals are twofold: (1) to preserve the system's ability to make accurate predictions in the face of attacks and (2) to preserve user privacy. Overall, we nd that although the space of attacks is quite rich, by using a combination of existing cryptographic techniques and some non-cryptographic ones, collaborative ltering systems can be made considerably more robust to attacks than they are today.