Privacy Preserving Collaborative Filtering via the Johnson-Lindenstrauss Transform

Recommendation systems have become increasingly popular as a result of the significant growth in online information. However, it requires a substantial amount of historical user data to generate accurate predictions, which raises concerns over user privacy. Differential privacy is a well acknowledged privacy notion that has become an important standard for the preservation of privacy. Unfortunately, existing privacy preservation methods based on differential privacy protect user privacy at the cost of utility, aspects of which have to be sacrificed to ensure that privacy is maintained. In this paper, we propose a Johnson Lindenstrauss privacy preserving collaborative filtering (JLCF) method. The proposed method preserves users' privacy without compromising utility. It guarantees user privacy by directly perturbing the original dataset using a transfer matrix. We prove that the proposed method achieves -differential privacy. In addition, we theoretically analyse the utility of the proposed method, and our extensive experiments show that the prediction accuracy is improved.