Federated Learning Differential Privacy Preservation Method Based on Differentiated Noise Addition
Liquan Han, Di Fan, Jinyuan Liu, Wei Du
Abstract
Differential privacy is an essential tool in federated learning privacy preservation. However, existing differential privacy-preserving techniques introduce excessive noise perturbations for privacy-preserving purposes, which makes federated learning models less usable. Under the condition of satisfying differential privacy, this paper proposes a new noise addition method, differential privacy preservation by federated learning based on differentiated noise addition (DDPFL), to reduce the loss of model usability. Specifically, the method derives importance coefficients for each model parameter by analyzing the value of the gradient update size, the magnitude of the absolute value of the weight parameter, and the relationship between the global model and the local model gradient update trend during the training process for each client. Then, noise is added to the model parameters using a differentiated noise addition mechanism according to the magnitude of the importance coefficients to achieve noise perturbation of the local model. Finally, the effectiveness of the proposed method is validated by an experimental comparison of the MNIST dataset and the Fashion-Mnist dataset. The usability of the model improves under the same differential privacy condition as the federated learning differential privacy preserving method before the improvement.