EFLSS: An Efficient Federated Learning Security Scheme in Byzantine Attack Scenarios

Federated Learning (FL), a distributed machine learning approach, aims to protect user privacy by performing model training without centralized data collection. However, FL systems are susceptible to malicious participants uploading incorrect or harmful updates in the face of Byzantine Attacks, leading to model performance degradation or even failure. To address this problem, this paper proposes a new approach that combines an confusion factor and a parameter screening mechanism to securely and efficiently counter Byzantine Attacks. Local gradients are added with confusion factors before uploading to protect the security of local data. In addtion, we propose an improved anomaly detection algorithm that can identify and filter Byzantine nodes during the gradient aggregation phase to ensure the accuracy of the global model. Experimental results show that compared with the traditional FL mechanism, the global model obtained in this scheme has higher accuracy.