Toward Communication-Efficient Adversarial Federated Learning for Robust Edge Intelligence

Abstract

Federated learning (FL) enables decentralized model training without sharing raw data but remains vulnerable to adversarial attacks and often suffers from performance degradation under non-IID data distributions. By analyzing vanilla knowledge distillation (VKD) and adversarial knowledge distillation (AKD), we observe that neither approach alone can effectively transfer both clean and robust knowledge from teacher to student models. To address this limitation, we first propose Pre-trained Model-guided Adversarial Federated Learning (PM-AFL), which exploits the complementary strengths of VKD and AKD to improve both clean accuracy and adversarial robustness. Furthermore, we extend PM-AFL to PM-AFL++, which leverages locally augmented data to enrich knowledge diversity, thereby enhancing the effectiveness of both VKD and AKD and improving model generalization. To further alleviate the discrepancy between local and global models under non-IID settings, PM-AFL++ introduces a consistency regularization term that aligns local adversarial predictions with the corresponding global clean outputs. As a result, local models can effectively absorb diverse knowledge from the teacher while remaining consistent with the global model, leading to improved global generalization. Extensive experiments on multiple benchmark datasets demonstrate that the proposed framework consistently achieves superior clean accuracy and adversarial robustness while reducing per-round communication by 23- $73\times $ and requiring fewer communication rounds than the baselines.