FedFR-ADP: Adaptive differential privacy with feedback regulation for robust model performance in federated learning

Abstract

Privacy preservation is a critical concern in Federated Learning (FL). However, traditional Local Differential Privacy (LDP) methods face challenges in balancing FL model accuracy with noise strength. To address this, we propose a novel adaptive differential privacy method with feedback regulation, FedFR-ADP. First, we employ Earth Mover’s Distance (EMD) to measure the data heterogeneity of each client and adaptively apply Gaussian noise based on the degree of heterogeneity, making the noise addition more targeted and effective. Second, we introduce a feedback regulation mechanism to dynamically tune the privacy budget according to the global model’s error feedback, further enhancing model performance. Finally, we validate our approach through experiments on two commonly used image classification datasets. The experimental results demonstrate that FedFR-ADP outperforms three benchmark algorithms, including DP-FedAvg, in terms of model training accuracy and Mean Squared Error (MSE) under varying degrees of heterogeneity. Compared to these benchmarks, FedFR-ADP achieves at least a 3.05% and 1.76% improvement in training accuracy across both datasets, with significantly reduced MSE fluctuations. This not only boosts model accuracy but also provides more stable noise control.