Research on the Recovery Strategy of Federated Learning Model Based on Zero-Knowledge Proof

Abstract

With the continuous advancement of intelligent algorithms, people have put forward new demands for data privacy protection, and the traditional privacy protection technology can not meet the needs of reality. To address this issue, a model safety verification method is proposed by combining European distance and cosine similarity constraint. The results showed that in the MNIST dataset and FMNIST dataset, the impact of three different attacks on the proposed research scheme showed a low value below 2%. In the face of malicious customers with different proportions, the impact of the MNIST and the FMNIST dataset are stable at a low range, and the maximum did not exceed 5%. The findings denote that the raised method has excellent performance in the defense model poisoning attack, and can effectively improve the robustness and privacy security in the federated learning system.