Dynamic Information Fusion for Personalized Federated Learning

Abstract

The Internet of Things generates vast amounts of decentralized data, where data heterogeneity presents a significant challenge in federated learning. Many existing personalized federated learning (pFL) approaches prioritize local model accuracy at the expense of global generalization, leading to reduced cost-effectiveness. To address this issue, a federated learning algorithm called Dynamic Information Fusion for Personalized Federated Learning (FedIF) is proposed. Two model heads carrying different information are fused by FedIF to obtain a personalized model fitting local data better. The personalization steps of FedIF are performed after local training in each round of the FedAvg algorithm. This design allows FedIF to be combined with other algorithms that improved FedAvg based algorithms. Comparative and ablation experiments are conducted between FedIF and other state-of-the-art personalized federated learning algorithms. Experimental results indicate that the FedIF achieves competitive performance in most evaluated heterogeneous settings.