FL-OA Boosts Byzantine Robustness in Federated Learning.

Federated Learning (FL) enables collaborative model training across multiple devices without sharing raw data, but its distributed nature makes it highly vulnerable to Byzantine attacks from malicious participants. Existing defense mechanisms often rely on restrictive assumptions, such as a minority of attackers or the server possessing a matching root dataset. Furthermore, these methods struggle with the inherent divergence among benign updates and the computational complexity of comparing high-dimensional model updates. To overcome these limitations, a new framework called FL-OA (Federated Learning with Outsourced Auditing) has been proposed. FL-OA introduces a novel approach where the central server partners with a third-party organization that holds an additional root dataset. This third party performs outsourced auditing, allowing the server to achieve robust aggregation without needing strong assumptions about the attacker's proportion. Additionally, FL-OA incorporates a gradient ascent step and a correction term during local training to reduce the divergence among benign updates. It also designs a parameter importance indicator to focus auditing on critical parameters, thereby alleviating the "curse of dimensionality" in update comparisons. Extensive experiments and theoretical analysis demonstrate FL-OA's superior performance against Byzantine attacks compared to existing methods.