A Malicious Client Defense Scheme in Federated Learning for Large‐Scale Edge Nodes

ABSTRACT In federated learning for large‐scale edge nodes, the problem of malicious clients submitting anomalous parameters is becoming increasingly prominent. This seriously affects the accuracy and reliability of the model. For the problem of malicious client defense in a large number of clients with decentralized distribution, to reduce the detection time of malicious clients in federated learning and improve the accuracy of model training, a Blockchain‐based Grouped Federated Learning malicious client defense Scheme (BGFLS) is proposed. Specifically, to detect malicious clients quickly and accurately, a grouped federated learning architecture is proposed, which applies blockchain technology to each grouping. In addition, an algorithm is designed to detect anomalous parameters, and a block structure that supports backtracking of malicious clients is proposed. Theoretical analysis and experiments show that the BGFLS scheme has improved accuracy compared with the GeoMed scheme and Krum scheme, and its backtracking efficiency is better than that of traditional blockchain implementations. Therefore, the BGFLS scheme can quickly detect malicious clients and protect shared parameters. This study provides a practical and high‐performance solution for detecting malicious clients using federated learning in large‐scale edge computing environments, with excellent technical specifications and high operational efficiency, effectively optimizing the overall system performance and stability.