FEFL: A Fair & Efficient Defense Against Poisoning Attacks via Discrete Update Space

Federated Learning (FL) provides an efficient modeling and privacy-preserving distributed machine learning paradigm. However, the inherent vulnerabilities of FL make it susceptible to data poisoning attacks. Existing defense methods against poisoning attacks are largely based on continuous update spaces and tend to ignore a crucial aspect of collaborative training among participants: contribution size. Yet few defense strategies have been developed in FL that consider both. To address this problem, we propose a fair and efficient federated learning defense method (FEFL) against poisoning attacks. The key idea of FEFL is to compute local rankings at the client side to construct a discrete update space, with the aim of reducing the model update space available to the attacker. Simultaneously, we introduce the contribution metric to achieve the fairness among clients participating in FL tasks. The experimental results demonstrate that FEFL achieves Acc 98.58% and 90.03%, and ASR 1.79% and 0.10% with b = 10% on MNIST and FashionMNIST datasets, respectively, which has significant advantages to provide protection for FL systems.