Towards Personalized Federated Learning via Contrastive-Augmented Local Memorization Retrieval

Abstract

Federated learning (FL) enables clients to collaboratively train statistical models while maintaining the privacy of their local data. However, traditional FL methods often suffer performance degradation due to data heterogeneity across clients. To mitigate this issue, we propose an efficient personalized federated learning (pFL) framework, FedCLR. In this framework. we first propose contrastive-augmented representation learning at the local level to build a global model capable of providing more discriminative feature representations. During inference, FedCLR utilizes this global model in conjunction with local memorization retrieval mechanism to achieve better personalization performance. Comprehensive experiments conducted on public datasets and various heterogeneous data environments demonstrate that FedCLR significantly outperforms recent state-of-the-art pFL methods in terms of accuracy.