SAFELearning: Enabling Backdoor Detection in Federated Learning with Secure Aggregation
SAFELearning: Enable Backdoor Detectability In Federated Learning With Secure Aggregation
AI summary
SAFELearning enables backdoor detection under federated learning secure aggregation through Oblivious Random Grouping (ORG) and Partial Parameter Disclosure (PPD). ORG partitions participants into one-time random subgroups unknown to them, and PPD discloses aggregated subgroup models for anomaly detection without leaking individual model privacy. Experiments show that it significantly reduces backdoored model accuracy under common backdoor strategies without harming main task accuracy, with lower communication and computation overhead than state-of-the-art secure aggregation protocols.