PAG-FL: Personalization via Adaptive Gradient Filtering in Federated Learning

Personalized Federated Learning (PFL) has emerged as a key paradigm for delivering personalized AI services across diverse clients. While prior research has primarily focused on handling data heterogeneity, recent studies have additionally targeted noise and communication efficiency. However, most approaches exchange knowledge only at the logit or prototype level, which limits representation sharing and struggles to mitigate both data heterogeneity and noise. Moreover, communication overhead accumulated over hundreds of training rounds still imposes substantial bandwidth costs, and fixed personalization ratios hinder effective optimization to non-IID data. To address these challenges, we propose Personalization via Adaptive Gradient Filtering in Federated Learning (PAG-FL), a communication-efficient and robust PFL framework. PAG-FL reduces communication costs by collecting only class-wise average embeddings and local prediction heads from clients. It further alleviates embedding bias through L2 normalization and a projection layer, while optimizing personalization ratios using thresholds derived from the interquartile range (IQR) of gradients. Extensive experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 demonstrate that PAG-FL consistently outperforms state-of-the-art methods. Specifically, under label noise and non-IID conditions, PAG-FL achieves accuracy improvements of up to 19% and F1 Score gains of up to 17% compared to the best baselines. In fixed and subset user scenarios, it maintains this robustness, yielding substantial accuracy gains of approximately 19%. Furthermore, PAG-FL significantly accelerates convergence, reducing communication rounds by up to 96% and communication overhead by up to 1,700×.