Making Local Models Learn Autonomously with Global Feature Tracking and Client Drift Releasing for Federated Learning

Data heterogeneity presents a significant challenge in federated learning, leading to inconsistent optimization of local models. Specifically, the client drift resulting from this heterogeneity significantly undermines model performance. To mitigate client model drift, existing approaches either align client and server models or utilize partial variance reduction. However, these methods compel the convergence direction on the client side to align with the global model, thereby restricting local model training from fully capturing the knowledge of the local dataset. Additionally, they underutilize global information, potentially causing local overfitting. To address this problem, we propose FedTR, a novel algorithm incorporating global feature tracking and client-released strategies to empower local models to learn autonomously without constraints. Meanwhile, shared global feature centroids effectively prevent local models from overfitting. Experimental results on five real medical datasets demonstrate the significant advantages of our algorithm over existing methods in non-IID data settings, improving both local model accuracy and convergence rate. Specifically, the FedTR algorithm enhances average accuracy by 3% \(\sim \) 5%. We provide evidence of the convergence rate of our algorithm.