Privacy-Preserving Medical Data Collaborative Modeling: A Differential Privacy Enhanced Federated Learning Framework

Abstract

This paper proposes a novel privacy-preserving federated learning framework enhanced with adaptive differential privacy for secure medical data collaboration. The framework addresses critical challenges in protecting patient privacy while enabling effective collaborative model training across healthcare institutions. We introduce a dual-layer privacy protection mechanism that combines local and central differential privacy, dynamically adjusting privacy budget allocation based on training progress and data sensitivity. The framework implements a hierarchical architecture with edge servers performing preliminary aggregation to reduce communication overhead and enhance privacy protection. A novel adaptive privacy budget allocation strategy is developed to optimize the privacy-utility trade-off throughout the training process. The framework incorporates robust aggregation mechanisms to handle data heterogeneity while maintaining privacy guarantees. Theoretical analysis establishes convergence properties and privacy bounds under various operating conditions. Experimental evaluation of real-world medical datasets demonstrates that our framework achieves 92.5% accuracy while reducing privacy loss by 85% compared to baseline methods. The framework shows strong resistance to various privacy attacks, with membership inference attack success rates reduced by 87%. The results validate the framework's effectiveness in enabling secure and efficient collaborative learning in healthcare settings while maintaining strict privacy protection for sensitive medical data.