Defense against backdoor attacks in federated learning with robust adaptive learning rates

Abstract

Federated Learning (FL) serves as a privacy-preserving paradigm that not only protects user privacy, but also improves model generalization ability and data security. However, by launching a backdoor attack, a vicious client can embed the backdoor in a the global model to deviate the direction of the model update, leading to the desired misclassification. To defend against backdoor attacks, we proposes a Robust Adaptive Learning Rate method (RALR). RALR takes into account the way of voting the gradient symbols of the clients by dimension, which means that no single client will have too much power. In addition, RALR adaptively finds the learning threshold so that the symbol voting value of each dimension reaches a certain number before it can participate in the global aggregation, and the bad influence of backdoor attackers on the global model training will be weakened as a result. In addition, the introduction of the sign gradient mechanism effectively protects the privacy of the update parameters. RALR not only ensures the performance of the main task under different experimental conditions, but also effectively eliminates the backdoor. The experimental results show that the robust adaptive learning rate method can defend against the backdoor attack very effectively. The successful rate of the attack is reduced to 1.9 % compared to the existing defense.