ADPFL: Adaptive Differential Privacy-Enhanced Federated Learning

Federated Learning (FL) provides a privacy-preserving solution as a paradigm of distributed learning by allowing clients to train models locally. However, traditional FL still faces privacy leakage risks during parameter uploading, and the existing Differential Privacy (DP)-based FL (DPFL) often degrade model performance due to improper noise injection. In this paper, an adaptive DPFL framework ADPFL is proposed, which achieves a trade-off between privacy protection and model utility through dynamic fine-grained privacy budget allocation, adaptive gradient clipping, and optimized client selection. Specifically, ADPFL first controls the noise scale by dynamically adjusting the privacy budget at the global level based on global accuracy differences across training rounds and at the local level based on the importance of CNN layers. Second, the gradient clipping threshold can be adaptively adjusted based on training progress, which also regulates the noise while minimizing information loss and suppressing outlier gradients. Third, ADPFL improves the convergence speed by selecting high-accuracy clients with diverse data distributions, while reducing privacy budget consumption. Finally, theoretical analyses demonstrate that ADPFL satisfies (ϵ, δ)-DP constraints, and provide the convergence bound. Experimental results show that ADPFL outperforms baseline methods in both privacy protection and model utility.