An Adaptive Differential Privacy Method Based on Federated Learning

Abstract

Differential privacy addresses privacy protection problem in federated learning. Uniformly allocating the privacy budget across all training rounds can lead to a degradation in model accuracy. Existing methods for adjusting privacy budgets often fail to adequately consider influencing factors and boundary conditions, resulting in unreasonable allocations. Therefore, we proposed an adaptive differential privacy method based on federated learning. The method determines the adjustment coefficient and scoring function based on accuracy, loss, training rounds, dataset size, and the number of clients, and uses these factors to dynamically adjust the privacy budget. Then the local model update is adjusted according to the scaling factor and the noise. Finally, the server aggregates the noised local model update and distributes the noised global model. The range of parameters and the privacy of the method are analyzed. Experimental results show that the method reduces the privacy budget by approximately 16% while maintaining comparable accuracy.