Inter-attribute Semantic Correlation-Guided Federated Recommender System Against Attribute Inference Attacks

Federated recommender systems (FedRSs) mitigate direct privacy leakage by keeping user data local and sharing only model updates with the server. Nevertheless, FedRSs remain vulnerable to Attribute Inference Attacks (AIAs), which leverage user embeddings to infer sensitive attributes. To defend against AIAs, recent studies have introduced adversarial learning into FedRSs. However, existing methods adopt single-task learning attackers that focus on single attributes in isolation, neglecting the semantic correlations among multiple attributes. This leads to suboptimal attack performance, thereby limiting the model’s overall privacy protection during adversarial learning. To address this issue, we propose the Multi-Attribute Collaborative Privacy-Preserving Federated Recommender System (MACPP-FedRS). MACPP-FedRS introduces a multi-task learning attacker that leverages a shared feature extractor to exploit inter-attribute semantic correlations, thereby building stronger attackers and helping the recommendation model learn more privacy-preserving representations. In addition, imbalances in attack performance across different attributes often arise in multi-task learning, with some attributes dominating the optimization process. To promote balanced privacy protection, we design an uncertainty-weighted module that adaptively adjusts the learning emphasis across attributes. Experiments on three real-world datasets demonstrate that MACPP-FedRS reduces privacy leakage under AIAs while maintaining high-quality recommendations. Our code is available at https://github.com/Loretz1/MACPP.