FedEMA-Distill：基于指数移动平均的知识蒸馏方法以增强联邦学习的鲁棒性

Federated learning (FL) often suffers from performance degradation when clients hold heterogeneous non-Independent and Identically Distributed (non-IID) data, particularly when some clients exhibit adversarial behavior, leading to client drift, slow convergence, and high communication overhead. This study proposes FedEMA-Distill, a server-side procedure that combines an exponential moving average (EMA) of the global model with ensemble knowledge distillation from client-uploaded prediction logits evaluated on a small public proxy dataset. Clients perform standard local training, upload only compressed logits, and may utilize different model architectures, thus requiring no changes to client-side software while still supporting model heterogeneity across devices. Experiments conducted on CIFAR-10 and CIFAR-100 datasets demonstrate that this method effectively enhances the robustness and efficiency of federated learning.