基于动态调整的差分隐私联邦学习优化方法

urrently, federated learning has become a focus of attention in both academia and industry due to its advantage of processing data locally. However, federated learning faces challenges of data heterogeneity and privacy leakage in practical applications. Existing research has clearly indicated that when federated learning models are trained under the dual constraints of data heterogeneity and differential privacy mechanisms, their utility often suffers significant losses. The optimization methods proposed in existing work have issues such as poor portability and limited improvement effects. Therefore, this paper proposes an optimized federated learning method with differential privacy based on dynamic adjustment. In this method, dynamic clipping based on gradient magnitude is applied to calculate the corresponding clipping threshold according to gradient information, addressing the issue of local model direction deviation. Additionally, local noise reduction based on dynamic aggregation is implemented to determine the optimal aggregation parameters using generated data, tackling the noise pollution caused by federated aggregation. Experimental results on public datasets demonstrate that the proposed algorithm can effectively improve the model prediction accuracy.