Federated Hybrid Training and Self-Adversarial Distillation: Towards Robust Edge Networks

Federated learning (FL) is a distributed training paradigm that enhances data privacy in mobile edge networks by enabling data owners to collaboratively train models without sharing raw data. However, data heterogeneity and adversarial attacks pose significant challenges in developing an unbiased and robust global model for edge deployment. In this paper, we observe that most existing federated adversarial training frameworks either struggle to balance the trade-off between accuracy and robustness, or suffer from compromised data privacy and heavy computational overhead. To mitigate this problem, we propose Federated hyBrid Adversarial training and self-adversarial disTillation (FedBAT), a novel framework designed to improve both robustness and generalization in FL. FedBAT achieves this by integrating local FL-based hybrid adversarial training (FHAT) with global augmentation-invariant self-distillation (AISD), leveraging both data augmentation and feature distillation perspectives. From data augmentation perspective, FHAT is employed to defend against adversarial attacks by balancing accuracy and robustness through a weighted combination of standard and adversarial training. From feature distillation perspective, AISD aligns local adversarial features of augmented images with their corresponding unbiased global clean features. This feature-level alignment mitigates bias caused by data heterogeneity and further enhances the global model's robustness and generalization. Moreover, we interpret the performance benefits of FedBAT through an analysis of the learned feature distributions, providing further insight into its effectiveness. Extensive experimental results across multiple datasets demonstrate that FedBAT yields comparable or superior performance gains in improving robustness while maintaining clean accuracy compared to several baselines.