SFedSV: Optimized Secure Federated Learning for Balancing Training Efficiency and Model Accuracy

With the advancement of distributed computing technologies, federated learning has emerged as an innovative solution for cross-institution collaborative modeling by enabling data usage without visibility. However, recent studies have revealed that federated learning still risks compromising user privacy during the learning process. Although some secure federated learning algorithms have been proposed, they cannot balance training efficiency and model accuracy. To address this challenge, this paper proposes SFedSV, a secure federated learning algorithm based on homomorphic encryption. The method incorporates a singular value-based aggregation mechanism integrated with homomorphic encryption technology and residual compensation mechanism, effectively protecting gradient privacy while maintaining model training efficiency and accuracy. Experimental validation on the MNIST dataset demonstrates that the proposed approach not only successfully resists gradient leakage attacks but also preserves the global model’s accuracy and training efficiency.