Authors: S Jayashree Ananth, Naveen V S
Abstract: The digitization of the healthcare industry has resulted in massive collection of personal health information among hospitals, clinics, and research institutions. But strict privacy laws (HIPAA, GDPR), along with other institutional obstacles, hinder data collection in a centralized manner, resulting in data silos that prevent the construction of efficient machine learning models for predicting diseases, estimating treatment, and managing public health issues. In this paper, we introduce a framework for privacy-preserving federated learning (PPFL) in healthcare. Our proposed framework includes three techniques: (1) Federated Averaging with differential privacy (DP-FedAvg) for model privacy, (2) Secure Multi-Party Computation (SMPC) for private aggregation of gradients, and (3) Homomorphic Encryption (HE) for performing computations on encrypted data. Our PPFL framework is evaluated on three real-life datasets of healthcare applications (mortality prediction from ICU records, diabetic retinopathy classification, and diagnosing COVID-19 patients) and outperforms federated learning with centralization in terms of model accuracy (within 3.2%) and provides differential privacy guarantees with ε=1.0 and δ=10⁻⁵.
DOI: https://doi.org/10.5281/zenodo.20415139
