Robust Federated Learning with Multi-Step Aggregation

In this paper we address the problem of robust federated learning, where the coordinator needs to robustly aggregate locally trained models to minimize the impact of malicious agents. We start by analyzing the state-of-the-art robust aggregation techniques and highlighting their weaknesses, which we exploit to propose a novel attack strategy. We then counteract the proposed attack with a novel multi-step aggregation technique, which cascades three robust aggregation steps. We provide theoretical results characterizing the robustness of the proposed aggregation, and showing that how it outperforms the state of the art. Additionally, we characterize the convergence of a federated algorithm employing our proposed aggregation. We corroborate our theoretical results with experiments for a classification task.