Intermediate features matter in prototype-guided personalized federated learning

The recent rise of Federated Learning (FL) as a privacy-preserving distributed learning paradigm has attracted significant attention in both research and application. Among the emerging topics of FL, personalized FL (pFL) has emerged as a focal point, with the primary challenge being the development of efficient, customized solutions for heterogeneous data environments. Recent efforts integrating prototype learning into FL have shown promise, yet they often neglect the utilization of intermediate features. We are thus motivated to address this gap by proposing a novel approach named FedPSC. This method first employs an embedding scheme to learn global category prototypes that are used to align local training processes across different clients. Most importantly, it explores the potential of multi-level category prototypes by leveraging intermediate features, thereby further aligning local feature learning at different hierarchical levels. Additionally, FedPSC incorporates supervised contrastive learning with a simple yet effective modification, extending it to the intermediate level as well, which complements the category prototypes and enhances model learning. Our comprehensive experiments on public benchmark datasets indicate that FedPSC outperforms recent FL methods in multiple aspects, particularly in terms of accuracy.