FedRDA: Federated learning with representation decoupling and divergence-aware aggregation

Personalized Federated Learning (PFL) aims to tailor personalized models for each client while enabling global knowledge sharing within the federated learning framework. However, in scenarios with highly heterogeneous data distributions, existing methods often struggle to effectively balance the tension between the global model and local personalized requirements. To address this challenge, we propose a novel personalized federated learning framework, FedRDA, which achieves efficient separation and dynamic aggregation of global and local representations through a Representation Decoupling Module (RDM) and a Divergence-Aware Aggregation Module (DAAM). During local training, FedRDA separates global and personalized representations via a filtering mechanism, ensuring the global one supports knowledge sharing while the personalized one focuses on local tasks. To enhance global representation learning, we use contrastive and mean squared error losses to align them with the global center on the server. To disentangle personalized from redundant global information, FedRDA uses vCLUB to minimize their mutual information, improving the personalized model's expressiveness. FedRDA also uses task vectors and divergence-aware aggregation to dynamically adjust client contributions, enabling better global model aggregation under heterogeneous data. Extensive experiments on five image classification benchmark datasets demonstrate the superiority of FedRDA over existing approaches in both personalization performance and global generalization.Our code is public at https://github.com/daoyingyijian/FedRDA.