Participation Bias Correction for Byzantine-Robust Federated Learning Under Heterogeneous Client Availability

Federated learning enables distributed clients to collaboratively train a global model without sharing their private data. However, Byzantine robustness becomes more challenging under heterogeneous client availability, since Byzantine clients with higher availability can participate more frequently than honest clients and accumulate disproportionate influence over the global model across rounds. Existing Byzantine-robust aggregation methods mainly defend against malicious updates after the participating clients have been selected, and therefore cannot directly correct the persistent participation bias caused by unequal client availability. This bias can raise the effective Byzantine weight beyond the level assumed by current-round robustness guarantees even when the population-wide Byzantine fraction is moderate. To address this availability-correlated threat, we propose a robust federated aggregation method with participation-bias correction that reduces Byzantine influence based on historical client availability rather than only current-round update information. The key idea of our method is a participation-level correction scheme that controls cross-round client influence before update-level robust aggregation. Specifically, the server first estimates each client’s availability from historical online records and converts the estimates into bounded inverse-propensity weights, assigning smaller aggregation weights to clients that participate more frequently. It then applies temporal difference clipping to limit large deviations from the previously applied server update and aggregates the clipped differences using a weighted geometric median with the corrected participation weights. Extensive experiments on MNIST and CIFAR-10 under sign-flip and Gaussian attacks demonstrate that the proposed method consistently improves robustness when Byzantine clients have higher availability while preserving clean training utility. The results further show that participation bias correction provides the main robustness gain, while temporal difference clipping offers additional stabilization under different availability gaps, Byzantine ratios, and non-IID data distributions while incurring small computation and memory overhead.