SECUREREC: Privacy-Preserving Recommendation with Distributed Matrix Factorization

Recommender systems have received much attention recently because of their abilities to capture the interests of users. A standard solution is to collect and analyze users' historical behavior data, which might raise privacy concerns, e.g., Facebook-Cambridge Analytica data scandal. Collaborative filtering has been widely used in recommender systems for its simplicity. However, it suffers from an efficiency issue owing to a large amount of data and time-consuming operations. Therefore, an interesting question arises: how to provide recommendation services and protect users' privacy at the same time based on distributed matrix factorization? The paradox is that sharing inaccurate information about user data makes it difficult for the recommender to infer personal preference. In this paper, we propose an item recommender system named SECUREREC. We formulate the notion of (alpha, beta)-accuracy. We prove that SECUREREC is (alpha, beta)-accurate and epsilon-differentially private. Experimental results on three real-world datasets show that SecureRec achieves comparable precision to non-private item recommendation methods while offering privacy guarantees to users.