Mitigating Label Skewness in Robust Federated Learning via Feature Synthesis

Robust federated learning (RFL) has become a promising paradigm, in which edge devices collaboratively perform adversarial training (AT) under the coordination of a central server to produce a robust global model, thereby enhancing resistance to adversarial attacks. However, recent studies have shown that the inherent label skewness nature of traditional FL introduces additional challenges in adversarial settings. This not only exacerbates model vulnerability but also highlights the limited effectiveness of directly applying conventional AT methods to FL. To address this, we shift our focus to the data level, the key issue in label skewness, and propose a simple yet remarkably effective feature-level class-conditional data augmentation method, dubbed FedCCFA. We first approximate the feature distribution of each class using global statistics and synthesize feature embeddings by mimicking class-wise Gaussian distributions. These synthetic features are then injected into the local AT process, forming a class-conditional augmentation strategy that improves robustness against both data heterogeneity and adversarial attacks. Moreover, due to its lightweight design that relies merely on global statistics, FedCCFA incurs minimal communication and computational overhead, making it expected to be highly compatible with a wide range of existing methods. Extensive experiments across multiple datasets demonstrate the effectiveness of FedCCFA, showing performance comparable to or better than that of baseline methods.