FedECP: Enhancing global collaboration and local personalization for personalized federated learning

Le résumé fourni par la source

Personalized federated learning (PFL) has received a lot of attention, owing to its significant advantages in addressing the statistical heterogeneity problem in federated learning (FL). Existing PFL methods typically partition model parameters by layers into two parts: shared parameters, which participate in global collaboration for learning shared knowledge among clients, and personalized parameters, which are retained locally to facilitate local personalization. However, during local training, parameters inevitably absorb both personalized and shared knowledge, preventing the shared and personalized parameters from effectively fulfilling their intended roles, weakening the effectiveness of global collaboration and local personalization. To address this issue, we propose a new PFL method called FedECP. FedECP stores global and personalized knowledge in separate models, preventing the interference and achieving a clearer separation of knowledge. Furthermore, we optimize the model learning strategy at both the feature representation and model parameter levels so that shared parameters learn shared knowledge, and the personalized parameters learn client-specific personalized knowledge. We conduct extensive experiments on four benchmark datasets, comparing FedECP with twelve state-of-the-art methods. The results demonstrate that FedECP performs well in various heterogeneous scenarios.

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