Utilizing Additional Personalized Representations for Personalized Federated Learning

Shulan Yin, Yingxun Fu, ma li

Abstract:

Federated learning (FL) allows multiple devices to collaboratively train models while preserving data privacy. Recently, personalized federated learning (pFL) has attracted significant attentions, due to its ability to effectively address the issue of statistical heterogeneity in FL. However, in heterogeneous environments, local feature extractors often produce data-specific drift, causing that clients are hard to balance local personalization with effective global collaboration. To address this problem, we propose a pFL method called FedAPR, which allows each client to learn global information while preserving personalized information by utilizing additional representations locally. To
evaluate the effectiveness of our method, we simulate various heterogeneous scenarios across multiple benchmark datasets,
and conduct extensive comparisons with nine state-of-the-art methods. The experiment results show that our method achieves the best performance among the tested methods.