Personalization‐Based Adaptation for Privacy Federated Recommendation

ABSTRACT The advantages of federated learning in collaborative computing of deep learning make it a crucial approach for distributed architectures in recommender systems. However, existing federated recommender systems typically share unified item embeddings across all clients, which fails to capture user‐specific characteristics of items. How to adaptively retain both the commonality and individuality meanings of item embeddings in the recommendation becomes a critical challenge, while simultaneously preventing personalized information leakage. Therefore, this paper proposes a novel federated recommendation method (named 2P‐FedRec) that constructs a privacy‐preserving personalized recommender system in an adaptive manner. Specifically, this method employs an adaptive attention module to generate item representation containing global item embeddings (capturing cross‐user commonalities) and personalized embeddings (capturing user‐specific preferences), and utilizes two regularizers to guide the optimization of independence between these two embeddings. To protect user privacy, we also apply local differential privacy (LDP) with noise injection to the uploaded parameters, preventing the reconstruction of sensitive data. Extensive experiments on Epinions and Yelp datasets demonstrate that 2P‐FedRec outperforms the state‐of‐the‐art baselines while maintaining privacy.