Privacy-preserving federated learning with differential privacy for healthcare AI: a convergence and utility analysis

Abstract

Federated Learning (FL) enables collaborative model training across distributed healthcare institutions without sharing raw patient data, offering a paradigm shift for privacy-sensitive medical AI. However, FL remains vulnerable to gradient inversion attacks and model poisoning, necessitating formal privacy guarantees. This paper presents a comprehensive analysis of Differential Privacy (DP)-augmented Federated Learning for healthcare AI applications, specifically Electronic Health Record (EHR) classification. We evaluate three aggregation strategies FedAvg, FedProx, and the proposed FedNova-DP across simulated environments with 10, 25, and 50 heterogeneous clients under both IID and non-IID data distributions. The proposed FedNova-DP framework achieves 93.8% accuracy on the MIMIC-III-derived benchmark dataset under non-IID conditions with a differential privacy budget of ε = 0.5, representing a 4.4% improvement over FedAvg-DP (89.4%) under equivalent conditions. Convergence analysis demonstrates that FedNova-DP reaches target accuracy 31% faster (in communication rounds) than FedAvg. A detailed privacy-utility trade-off analysis across ε ∈ [0.1, 10] reveals that the proposed framework maintains competitive utility at strong privacy regimes (ε = 0.5, accuracy = 89.3%) compared to non-private centralized training (97.4%). These findings establish FedNova-DP as a practical, deployable solution for privacy-preserving healthcare AI at scale.