Xuhao Zhao, Yanmin Zhu, Wenze Ma, Qihao Luo, Chenhao Zhai, Chunyang Wang, Zhaobo Wang, Jiadi Yu, Feilong Tang · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840868
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Federated recommender systems (FedRec) aim to preserve user privacy by keeping sensitive data on client devices and sharing only model parameters with a central server. However, FedRec is still vulnerable to attribute inference attacks (AIAs), where server-side adversaries exploit uploaded parameters to infer users' private attributes. Existing FedRec approaches face a suboptimal privacy-performance trade-off. Privacy-focused methods mask attribute-related features in representations to protect sensitive information, but degrade recommendation accuracy. In contrast, performance-focused methods preserve accuracy by retaining these features but risk privacy leakage through uploaded representations. To balance privacy and performance, we propose PARD, a privacy-aware representation decoupling framework that explicitly decomposes representations into privacy-relevant and privacy-irrelevant components. Only the privacy-irrelevant part is uploaded to the server, and the privacy-relevant part is retained locally. We leverage Mutual Information (MI) objectives to facilitate this decoupling: (1) suppressing sensitive leakage by minimizing the MI between privacy-irrelevant representations and attributes, and (2) retaining attribute signals by maximizing the MI for privacy-relevant representations. Since exact MI computation is intractable, we derive variational bounds and estimate them using privacy estimators under adversarial and cooperative training paradigms. Experimental results demonstrate that PARD outperforms state-of-the-art methods in both recommendation accuracy and privacy preservation. The code is available at https://github.com/XuHao-bit/PARD.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.