Clustering-Based Custom Global Modules for Personalized Federated Learning

The field of federated learning faces a fundamental challenge posed by the non-independent and identically distributed (Non-IID) data among heterogeneous clients. Personalized federated learning (PFL) addresses this issue by providing customized models for each client. The existing PFL methods ignore the impact of redundant information brought by shared knowledge on processing local tasks of clients, and clients prefer to find the mapping of local data distribution from shared knowledge space. This paper proposes a personalized aggregation method, FedCGM, which customizes a clustering coordinator for each client. The method extracts core knowledge from the global model that aligns with the local data distribution and adapts it to the local client. The challenge of FedCGM lies in identifying the mapping relationship between local data distributions and shared knowledge. To address this challenge, we construct posterior probability indices of the global modules on the target data distribution among participating members. This approach establishes a mapping relationship between data distribution and shared knowledge. We conducted extensive experiments on real-world datasets and found that FedCGM achieved higher accuracy compared to the state-of-the-art PFL methods.