Defense Scheme of Federated Learning Based on GAN

• DOI: 10.3390/electronics14030406
• Publication Date: 2025-01-21T09:19:51Z
• Authors: Qing Zhang; Ping Zhang; Wenlong Lu; Xiaoyu Zhou; An Bao

Federated learning (FL), as a distributed mechanism, can have model training completed without directly uploading original data, effectively reducing the risk of privacy leakage. However, through shared gradient information, research shows that adversaries may reconstruct data. To further protect federated learning, defense scheme is proposed based on generative adversarial networks (GAN), which combined with adaptive differential privacy. Firstly, real data distribution features are learned GAN, and replaceable pseudo generated. Then, added noise. Finally, generated by in used to replace so cannot obtain user After simulation experiments carried out MNIST dataset, algorithm verified using attack method. The experimental results show superior accuracy. Compared FedAvg algorithm, only 0.48% accuracy lost. Therefore, it achieves good balance between