Turning Data Heterogeneity into a Backdoor Shield for Personalized Federated Learning

Le résumé fourni par la source

Personalized Federated Learning (PFL) has emerged as an effective paradigm capable of accommodating data heterogeneity by tailoring models to local client distributions. Nevertheless, such personalization exacerbates susceptibility to backdoor threats, as the induced model diversity and diminished shared knowledge undermine the robustness of global defenses, which opens new avenues for backdoor exploitation. Although adversarial training can mitigate backdoors in centralized settings, its applicability to PFL is fundamentally limited by the inherent data heterogeneity. To tackle this challenge, we propose Client-Centric Shielding (CCS), a novel PFL framework that leverages data heterogeneity as an inherent defense, thereby shielding clients from backdoor attacks. At the client level, CCS integrates local adversarial training with a personalized projection network, which aligns feature spaces and strengthens the learning of robust and discriminative representations. At the server level, a clustering-based defense is employed; rather than operating on the full parameter space as in prior work, our approach utilizes batch normalization statistics to construct low-dimensional client vectors, enabling efficient detection and isolation of malicious clients. Comprehensive empirical evaluations demonstrate that CCS achieves state-of-the-art (SOTA) robustness against sophisticated backdoor attacks while consistently outperforming existing defense baselines.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.