Differential Privacy Hierarchical Federated Learning Method based on Privacy Budget Allocation

Federated learning can effectively protect users’ personal data from being obtained by attackers, and differential privacy can enhance the privacy of federated learning and solve the privacy leakage problem caused by model training parameters. Because of the rapid increase in the number of terminals, hierarchical federated learning framework has a wide application prospect because of its advantages such as high scalability. Most of the existing researches on the combination of federated learning and differential privacy are based on the terminal-server framework, while the relevant researches based on hierarchical federated learning framework are lacking. In this paper, we propose a differential privacy hierarchical federated learning method based on privacy budget allocation, which can effectively resist member inference attacks by adding quantitative noise to model parameter update when the model parameter update is uploaded. In order to achieve the balance between data availability and privacy protection, an accurate privacy budget allocation strategy is designed to ensure the efficiency of federated learning and the availability of the model. We verify the effectiveness of the proposed method on the MNIST dataset, and the experimental results show that the accuracy of the global model can still be maintained at a high level under the premise of protecting data security.