FedDefense: A Defense Mechanism for Dishonest Client Attacks in Federated Learning

Abstract Federated Learning (FL), which allows multiple participants to co-train machine learning models, enhances privacy-preserving by avoiding exposing local data. In recent years, FL has been considered a promising paradigm. However, during the FL process, individual clients may fall out on the client’s side, or a particular client may engage in dishonest behavior such as uploading malicious data, thereby hindering the training of the global model. Most of the existing defense methods are considered only from the perspective of data filtering or model weighting, which have the disadvantages of poor robustness and high computational cost. Therefore, we propose a novel security FL (FedDefense) scheme based on client selection and adaptive rewards to defend against dishonest client attacks. First, to reduce the likelihood of poisoned clients participating in aggregation, we design a randomized subset method for client contribution evaluation via Kullback–Leibler (KL) divergence. Second, we reduce the server’s dependence on clients through a dynamic reward strategy to ensure healthy model training. Numerical analysis and performance evaluation show that the proposed technique prevents the threat of dishonest clients during FL processing. Compared with existing methods, our approach has significant advantages in terms of efficiency and performance.