Federated matrix factorization for privacy-preserving recommender systems

Recommender systems recommend contents or services via collecting and analyzing numerous user data, which may raise serious privacy concerns when the recommender is untrusted. Inspired by federated learning, a user-level distributed matrix factorization framework has been proposed where the model can be learned via collecting gradient information from users (instead of the raw data). This approach focuses on protecting model-privacy and value-privacy from untrusted recommender but has limited consideration on existence-privacy. To address this issue, we enhance the aforementioned framework with Homomorphic Encryption and randomized response. Extensive experiments demonstrate that our method can provide more secure protection for users’ privacy with less performance degradation and smaller computational burden.