Privacy-Preserving Collaborative Filtering Using Fully Homomorphic Encryption

Protecting the privacy of users is one of the most important issues in recommender systems, where new items, e.g., books, movies, and friends in online social networking service/sites, are recommended to target users. To identify recommended items, encryption-based privacy-preserving collaborative filtering is widely used to generate recommendations. However, existing solutions are either slow or not scalable. To tackle this issue, in this paper, we first propose a privacy-preserving user-based CF protocol using the BGV fully homomorphic encryption scheme, which is named BGV-CF. By reducing interactions and the amount of communication traffic among users and recommendation servers, the proposed BGV-CF protocol significantly facilitates the recommendation process. Then, we propose an optimized BGV-CF (OBGV-CF) protocol where some computations are offloaded to users during the recommendation process. The security of the proposed schemes is qualitatively analyzed and quantitative analyses of the computation and communication costs are performed. In addition, provable security analysis using random oracles is provided. The BGV-CF and OBGV-CF protocols are implemented using C++, and testbeds using the MovieLens dataset are conducted. Experimental results demonstrate that the proposed BGV-CF and OBGV-CF successfully achieve their design goals.