FL-OA: A Byzantine-robust federated learning framework for smart devices using outsourced auditing

FL-OA is a Byzantine-robust federated learning framework for smart devices, in which the server works with a third-party institution holding an additional root dataset to perform outsourced auditing, achieving robust aggregation without strong assumptions. The framework introduces a gradient ascent step and a correction term in local training to mitigate divergence among benign updates, and uses a parameter importance metric to extract key parameters for auditing, mitigating the curse of dimensionality. Theoretical analysis and extensive experiments show that FL-OA is more effective than existing defense methods in resisting Byzantine attacks.