With the widespread adoption of artificial intelligence across diverse fields, issues related to data privacy and heterogeneity have become increasingly prominent. Federated learning emerges as a distributed learning framework specifically designed to address data privacy concerns. By enabling multiple clients to collaboratively train a global model without sharing their raw data, federated learning effectively safeguards privacy while managing data heterogeneity. However, due to the highly heterogeneous model and data distributions of each client, traditional aggregation methods (e.g., FedAvg) tend to face challenges of insufficient personalization, high communication cost, and unstable model convergence. In this study, a fresh framework for personalized federated learning is put forward, named DynKT-pFL. This framework is founded on knowledge transfer and contains a dynamic temperature regulation strategy based on communication rounds and real-time client performance, as well as an adaptive knowledge coefficient fusion mechanism using an MLP combined with client performance feedback. The knowledge distillation process is achieved with accurate control and fast convergence of the model. Experiments in non-iid and client heterogeneous environments demonstrate that DynKT-pFL exhibits considerable advancement in model consistency, stability, communication efficiency, and accuracy, providing an effective solution for the practical implementation of federated learning in complex heterogeneous environments.
Ziyi Guo (Wed,) studied this question.