AdpFL: A Privacy-Preserving Federated Learning Framework through Adaptive Model Pruning on Non-IID Data

Le résumé fourni par la source

Federated learning (FL) has shown great potential, especially with the rise of complex foundation models and growing privacy needs. FL has experienced challenges, including high communication costs, privacy concerns of user data, and the complexities of non-independent and identically distributed (non-IID) data. This paper aims to address these challenges through our proposed FL framework called Adaptive Pruning FL (AdpFL). AdpFL conducts an exploration phase prior to the formal FL process, during which each local dataset's deviation and corresponding threshold are computed locally. These results are then utilised to perform model pruning during the formal FL training. By doing so, the method achieves full adaptivity while simultaneously addressing challenges related to communications costs and non-IID data. Meanwhile, AdpFL enhances pruning by replacing part of the parameters with zeros, thereby preventing membership inference attacks (MIA) and ensuring privacy in both local and global models. The extensive experiments on various datasets and FL tasks demonstrate that our method not only efficiently accelerates global convergence but also achieves superior accuracy and robustness compared to traditional methods. AdpFL reduces computational overhead and decreases the effective update size (i.e., the proportion of non-zero parameters) by approximately half while improving global accuracy by up to 20%, and it also lowers the accuracy of white-box MIA attacks across various models. These improvements make FL more efficient, scalable, and practical in real-world situations.

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