FedDefense: An Integrated Robust Defense Against Backdoor Attacks in Federated Learning

Federated learning (FL) is a matter of concern on account of its privacy-protecting nature. In the context of FL, the collaborative server is unable to ascertain the specific data distribution of each client and the manner in which the model is trained, rendering it susceptible to backdoor attacks. Many existing defense methods lose their effectiveness in the presence of data heterogeneity with low sample poisoning rates from malicious clients. In this paper, we propose a new integrated backdoor defense method, FedDefense. It consists of three components: 1) filtering a subset of benign updates based on multiple metrics; 2) dynamic aggregation based on a set of key parameters; and 3) adding adaptive noise to further improve global model robustness. We compare FedDefense with state-of-the-art backdoor attack defenses in FL on three real datasets, and experimental results show that FedDefense is more resilient to backdoor attacks and has slight impact on the accuracy of the global model. For example, other methods have achieved backdoor accuracy of upwards of 90% on the CIFAR-10 dataset in the case of centralized backdoor attacks, yet FedDefense achieves the lowest backdoor accuracy of 7.93%, demonstrating significant superiority over previous defenses.