Adaptive Differential Privacy Algorithm for Federated Learning on Small Datasets
Lei Xia, Huanbo Yang · 2024
We address the challenge of enhancing federated learning with small datasets by proposing an adaptive differential privacy technique and a dynamic learning rate optimization algorithm. Our method adds adaptive noise to protect data privacy and uses historical momentum information to improve stability and convergence speed. Extensive experiments demonstrate that our approach significantly improves model performance, accelerates the learning process, and ensures privacy, offering a practical solution for federated learning with small datasets.