Thwarting Passive Privacy Attacks in Collaborative Filtering

Abstract. While recommender systems based on collaborative filtering have be-come an essential tool to help users access items of interest, it has been indicated that collaborative filtering enables an adversary to perform passive privacy at-tacks, a type of the most damaging and easy-to-perform privacy attacks. In a passive privacy attack, the dynamic nature of a recommender system allows an adversary with a moderate amount of background knowledge to infer a user’s transaction through temporal changes in the public related-item lists (RILs). Un-like the traditional solutions that manipulate the underlying user-item rating ma-trix, in this paper, we respond to passive privacy attacks by directly anonymizing the RILs, which are the real outputs rendered to an adversary. This fundamen-tal switch allows us to provide a novel rigorous inference-proof privacy guaran-tee, known as δ-bound, with desirable data utility and scalability. We propose anonymization algorithms based on suppression and a novel mechanism, permu-tation, tailored to our problem. Experiments on real-life data demonstrate that our solutions are both effective and efficient.