Enhancing Generalization in Federated Learning via Adaptive Gradient Transformation

Abstract

Federated learning (FL) has garnered significant attention in the consumer electronics sector due to its capability to optimize model performance on edge devices while ensuring user data privacy. In practice, however, FL systems face considerable challenges, primarily stemming from the statistical heterogeneity introduced by non-independently and identically distributed (non-IID) user data, which often leads to training instability. Current mitigation strategies predominantly address the statistical bias caused by non-IID data, yet they frequently overlook critical issues related to convergence stability and model overfitting. To address this research gap, this paper introduces two novel algorithms: Federated averaging with adaptive gradient transformation (FedAvg-AGT) and its variant integrated with sharpness-aware minimization, FedSAM-AGT. These methods facilitate client models in exploring flatter loss landscapes, thereby preventing convergence toward sharp minima and enhancing generalization performance on complex consumer electronics data. Theoretical analysis validates the convergence properties of the proposed algorithms, while experimental results on large-scale consumer electronics datasets demonstrate that both FedAvg-AGT and FedSAM-AGT achieve superior robustness and accuracy compared to existing state-of-the-art baselines.