Privacy-enhanced Lossless Federated Recommendation with client-selection
Fang Ting, Wang Yong Li, Sha Xue Qi · 2023
With the implementation of privacy protection legislation such as GDPR, corporations are finding it increasingly difficult to legally gather users’ data. The fundamental principle behind federated learning is to train a machine learning model without knowing user’s local raw private data and although some current works protect interaction information by using virtual scoring, the uploaded data remains unencrypted real gradients that could betray some private information. In this paper, we propose privacy-enhanced lossless federated collaborative recommendation framework via homomorphic encryption and client selection (PriFedRec++). Our PriFedRec++ can secure the privacy of two separate types of content of users, namely, rating values and rating behaviors, without compromising recommendation performance. Extensive investigations clearly demonstrate the efficacy of our PriFedRec++ in providing accurate and privacy-aware recommendations and there are also experiments which indicate the effectiveness of client selection strategies in accelerating model convergence.