| Abstract: |
The explosion of healthcare data in distributed medical facilities provided an unprecedented opportunity to create reliable artificial intelligent models, yet the trade-off for privacy has been unavoidable. In this paper, we present Adaptive-DP-FL, a federated learning framework tailored for healthcare applications integrating adaptive differential privacy methods. Our framework tackles three key challenges: (1) achieving clinically acceptable model accuracy under privacy constraints; (2) controlling communication overhead in healthcare networks that have limited bandwidth; and (3) reaching fair levels of performance across a diverse collection hospital data distributions. We introduce a dynamic budget allocation where the magnitude of noise injection scales according to gradient sensitivity and training convergence measures, resulting in 34% accuracy improvement over fixed- budget differential privacy strategies at matched privacy guarantees (ε=1.0). We also present a hierarchical privacy-preserving aggregation protocol and TopK-DP gradient compression, which decrease the communication overhead by 87.3% without degradation to model quality. Comprehensive experimental results on five real-world healthcare datasets with 511,893 patient records show that Adaptive-DP-FL achieves AUC-ROC of 0.923 for the task of mortality prediction, which reflects a performance degradation by only 2.1% compared to its centralized baselines; and guarantees formal (ε,δ)-differential privacy. |
