A1486
Title: Byzantine-robust distributed one-step estimation
Authors: Chuhan Wang - Beijing Normal University (China) [presenting]
Abstract: A Robust One-Step Estimator (ROSE) is proposed that solves the Byzantine failure problem in distributed M-estimation with only a single iteration, even when a moderate fraction of worker nodes exhibit Byzantine behavior. This estimator achieves higher asymptotic relative efficiency than conventional median-based estimators, while maintaining the same computational complexity. It is also robust to outliers or missing data that may occur during centralized aggregation. We establish the asymptotic normality of the estimator as the parameter dimension p increases with the sample size, and under mild assumptions, we derive its convergence rate.
