HDP-FL: A Hierarchical Differential Privacy Framework for Federated Learning in Healthcare

Abstract

Federated learning (FL) enables collaborative model training in Internet of Things (IoT)-enabled healthcare environments without directly sharing raw medical data. However, shared gradients and model parameters may still expose sensitive information, and uniform privacy budgets cannot reflect different medical-data sensitivities and user privacy preferences. This paper proposes HDP-FL, a hierarchical differential privacy framework that allocates privacy budgets according to data sensitivity and user-specific privacy demands. HDP-FL combines local Gaussian perturbation, server-side Gaussian perturbation, adaptive gradient clipping, and learning-rate adjustment to provide fine-grained privacy protection while maintaining model utility. Theoretical analysis shows that the proposed perturbation mechanism satisfies (ε, δ)-differential privacy under calibrated Gaussian noise. Experiments on MNIST, CheXpert, and COVID-19 datasets under IID and non-IID settings demonstrate that HDP-FL reduces gradient reconstruction success rates and preserves competitive accuracy under label-flipping and gradient inversion attacks.