Enhancing Privacy and Security in Recommender Systems Through Federated Learning and Differential Privacy
Zhigang Yang, Tafadzwa Mbodza · 2024
The domain of Recommender Systems (RS) plays a pivotal role in tailoring user experiences across digital platforms, yet it is fraught with pressing privacy issues due to the aggregation of user data in centralized repositories. This study introduces a groundbreaking framework designed to bolster the privacy and security of RS by amalgamating Federated Learning (FL) and Differential Privacy (DP) through an innovative Perturbed User-Based Interaction Matrix. By leveraging FL, our framework decentralizes the model training process, ensuring that user data remains confined to local devices and is not transferred to central servers. Simultaneously, DP is employed to inject controlled noise into the model updates, thereby mitigating the risk of inadvertent data leakage during the aggregation phase. The approach further integrates Laplace noise to obscure individual user contributions, achieving an effective equilibrium between privacy and utility. Empirical evaluation using the MovieLens dataset reveals that our method preserves robust recommendation accuracy, achieving a Precision@10 score of 0.82, while maintaining a commendable privacy budget (epsilon) of 1.0. This research provides a viable pathway for the deployment of privacy-enhanced recommendation systems, successfully addressing privacy concerns without sacrificing the utility of the recommendations.