Federated Learning (FL) enables collaborative model training across decentralized healthcare institutions without requiring the sharing of Electronic Healthcare Records (EHRs), thereby ensuring data locality and reducing privacy risks. In this study, a baseline FL framework was implemented with one central server and four hospital clients, utilizing a real-time Chronic Kidney Disease (CKD) dataset. However, privacy assessments conducted using three simulated adversarial attacks, model inversion, Membership Inference Attacks (MIA), and gradient leakage, revealed significant vulnerabilities in the plain FL setup. To address these vulnerabilities, this work proposes a secure Federated Learning-Homomorphic Encryption (FL-HE) framework that integrates FL with encryption techniques using the TenSEAL library. The proposed FL-HE framework introduces a layer-wise encryption strategy, securing model parameters, bias, and feature normalization, ensuring end-to-end confidentiality. While the integration of HE introduces computational overhead, the FL-HE framework achieves a high prediction accuracy of 98.6%, nearly identical to the 98.7% achieved by the unencrypted FL model. These results underscore the strong privacy-preserving capabilities of the FL-HE framework without compromising the performance of the model, making it suitable for applications like in healthcare, where the privacy of data is of utmost importance.
M et al. (Sat,) studied this question.