Abstract:Federated learning (FL) offers significant potential for preserving user data privacy, and differential privacy (DP) provides rigorous privacy guarantees. However, under non-independent and identically distributed (Non-IID) settings, the interaction between DP-induced noise and data heterogeneity often leads to substantial model performance degradation. Existing approaches typically employ uniform privacy strategies, neglecting heterogeneity differences among clients and failing to balance privacy protection with model utility effectively. To address this limitation, a heterogeneity-oriented adaptive differentially private federated learning (HA-DPFL) method is proposed. This method quantifies both client-level and global data heterogeneity using Jensen-Shannon (JS) divergence and introduces a dual adaptive mechanism: (1) the client sampling rate is adjusted according to global heterogeneity to accommodate varying distribution characteristics; (2) a personalized gradient clipping threshold is assigned to each client based on its individual heterogeneity, thereby reducing performance degradation while maintaining privacy guarantees. Experimental results on MNIST and CIFAR-10 datasets under various Non-IID settings demonstrate that HA-DPFL significantly improves model accuracy and achieves a better privacy-utility trade-off under equivalent DP guarantees.