Abstract
Data heterogeneity is a key challenge in the field of federated learning. Many existing personalized federated learning approaches focus on the performance of local models, neglecting the generalization capabilities of the global model, which may not be cost-effective. To address this issue, a federated learning algorithm called Personalized Federated Learning with Adaptive Information Fusion (FedIF) is proposed, which fuses two model heads carrying different information to obtain a personalized model fitting the local data better. The personalization steps of the FedIF are carried out after the local training in each round of the FedAvg algorithm, allowing it to be combined with other algorithms that improve upon the FedAvg. Comparative and ablation experiments between FedIF and other state-of-the-art personalized federated learning algorithms were conducted under three public datasets and two medical imaging datasets. The exceptional performance of our algorithm is attested across a wide range of experimental settings.