Private collaborative filtering under untrusted recommender server

Abstract

Recommender systems play an increasingly vital role in modern E-commerce. However, exploiting users’ preferences with recommender algorithms leads to serious privacy risks, especially when recommender service providers are unreliable. To deal with the problem, this paper proposes a Client/Server framework to create a private recommender system (PrivateRS). The system assumes that the Server side is untrustworthy. On the Client side, each user firstly rates the items and randomizes the ratings with a differential privacy mechanism. The ratings are further substituted by private symbols which are autonomously defined by each user to hide the ordinal meaning of the ratings. Using those symbols, the Server applies a private collaborative filtering algorithm to predict the ratings of items for the user. During this process, new similarity metrics are provided to search the nearest neighbours for users or items without knowing the real meanings of those symbols. Experimental results demonstrate that even though the ordinal meaning of the rating is significantly obfuscated, the proposed algorithms can still generate accurate recommendations with acceptable loss.