Defending against Backdoors in Federated Learning with Robust Learning Rate

Federated learning (FL) allows a set of agents to collaboratively train model without sharing their potentially sensitive data. This makes FL suitable for privacy-preserving applications. At the same time, is susceptible adversarial attacks due decentralized and unvetted One important line against backdoor attacks. In attack, an adversary tries embed functionality during training that can later be activated cause desired misclassification. To prevent attacks, we propose lightweight defense requires minimal change protocol. high level, our based on carefully adjusting aggregation server's rate, per dimension round, sign information agents' updates. We first conjecture necessary steps carry successful attack in setting, then, explicitly formulate conjecture. Through experiments, provide empirical evidence supports conjecture, test under different settings. observe either completely eliminated, or its accuracy significantly reduced. Overall, experiments suggest outperforms some recently proposed defenses literature. achieve this by having influence over trained models. addition, also convergence rate analysis scheme.