A Differential Privacy Federated Learning Scheme with Improved Noise Perturbation

Federated learning solves the privacy problem in deep learning by uploading model parameters at the edge. However, through the existing research, it is shown that through the analysis of model parameters and server-side aggregation parameters, there is still a risk of privacy leakage of private information. In order to solve this problem, differential privacy technology provides lightweight privacy protection by adding noise disturbance in the model training process, but it cannot fully consider the heterogeneity and diversity of model parameters, indiscriminately adding excessive or insufficient noise can negatively affect model performance. In this paper, we propose a differential privacy federated learning framework IDP-FL with improved noise perturbation. It leverages the benefits of gradient descent to enhance gradient perturbation techniques, dynamically adjusts the gradient clipping bound for various training stages, and incrementally tightens the gradient norm boundary as the algorithm progresses. we conduct a series of extensive experiments on public datasets MNIST and FEMNIST, and the results indicate that our IDP-FL algorithm improves performance by an average of 3.55% and 2.32% respectively compared to the DP-FL and NbAFL algorithms.