Personalized Federated Learning for Assessing Characteristic Client Data

Recently, Federated Learning (FL) has garnered significant as a powerful approach for learning from distributed data. Although FL provides notable advantages such as enhanced privacy protection, it also introduces challenges, particularly in characterizing the data distribution of individual clients, which is often non-i.i.d. (non-independent and identically distributed), given that client data are not centralized on the server. In this study, we propose a federated learning framework utilizing autonomous client models, each comprising two distinct local models: a “general” model, which captures shared data patterns across clients, and an “individual” model, which focuses on clientspecific data. By leveraging both the general and individual models, the proposed method facilitates the efficient extraction of distinctive data from each client. Preliminary empirical results demonstrate the effectiveness of the proposed method, particularly in the extraction and analysis of characteristic data. This study provides a novel approach that supports data characterization, which is critical in distributed learning systems.