Adaptive Iteration Differential Private Federated Learning with Gradient-Guide Synthetic Data

Motivated by model utility enhancement, communication round optimization, and robustness improvement against data heterogeneity in federated learning with differential privacy, we introduce the Adaptive Iteration Differential Private Federated Learning with Synthetic Data(AIS-DPFL) framework. Our approach employs privacy-preserved gradient guidance to address non-IID challenges, generating high-quality synthetic data while maintaining privacy. In addition, we introduce an adaptive local iteration strategy that optimizes training efficiency and convergence, complemented by modifications of the local objective function to improve accuracy. Finally, extensive experiments are conducted to validate the superiority of AIS-DPFL in accuracy and convergence compared to existing methods.