Iclr 2022

Byzantine-Robust Learning on Heterogeneous Datasets via Bucketing

EPFL researchers show that existing Byzantine-robust aggregation rules (Krum, coordinate-wise median, RFA) fail catastrophically on non-iid data, then fix the problem with a one-step bucketing scheme that randomly groups worker updates before aggregation. The first result with provable convergence guarantees for Byzantine robustness under realistic data heterogeneity.