A Hybrid Federated Learning Framework with Differential Privacy for Healthcare Information Systems: Performance Analysis and Security Evaluation

Abstract

Healthcare information systems increasingly demand sophisticated machine
learning capabilities while maintaining strict patient privacy requirements.
This paper introduces a novel hybrid federated learning framework that
integrates differential privacy mechanisms specifically designed for healthcare
information systems. Our approach addresses the dual challenges of maintaining
model accuracy and ensuring robust privacy protection in distributed healthcare
environments. We propose a multi-tier architecture that combines local
differential privacy with global privacy budgeting, enabling healthcare
institutions to collaboratively train machine learning models without
compromising sensitive patient data. The framework incorporates adaptive
noise calibration based on data sensitivity levels and implements a dynamic
participant selection mechanism to optimize both privacy and performance.
Through extensive experiments on real-world healthcare datasets, our method
demonstrates superior performance compared to existing approaches, achieving
94.7% accuracy on disease prediction tasks while maintaining -differential
privacy with  = 0.5. The security evaluation reveals robust resistance against
various privacy attacks, including membership inference and model inversion
attacks. Performance analysis shows a 23% improvement in communication
efficiency and 18% reduction in training time compared to traditional federated
learning approaches. Our framework provides a practical solution for healthcare
organizations seeking to leverage collaborative machine learning while
adhering to stringent privacy regulations such as HIPAA and GDPR.