Abstract The increasing need for secure and accurate analysis of sensitive healthcare data across distributed sources such as blood banks has amplified the adoption of Federated Learning (FL), which allows collaborative training without the distribution of raw data. While FL helps address data silos and ensures a degree of privacy, studies reveal that gradient updates still leak sensitive information, posing risks of data reconstruction. Although differential privacy techniques have been introduced to mitigate such threats, uniform noise injection often compromises learning performance. To overcome these limitations, this research proposes a novel privacy-adaptive framework titled Self-learning Heterogeneous-based privacy-Enhanced end-to-end Learning Design for Federated Learning (SHELD-FL). The framework integrates self-learning privacy budgeting to dynamically allocate privacy budgets based on gradient sensitivity, and heterogeneous differential privacy to vary noise levels per client according to data sensitivity, achieving a better privacy-utility balance. A gradient boosting classifier is used at the client side to enhance classification under non-IID conditions, and the Builder Optimization Algorithm (BOA) is employed at the server to optimize noise regulation during aggregation. Experimental results on electronic health records from simulated blood banks demonstrate that the proposed SHELD-FL framework achieves a high classification accuracy of 98.39% under a strong privacy setting (ε = 10), outperforming baseline approaches by 3–5%. Moreover, the framework reduces communication latency by approximately 40%, indicating its efficiency and scalability in real-world federated environments. These findings confirm that SHELD-FL offers a reliable, adaptive, and privacy-preserving solution for secure and collaborative healthcare data analysis across distributed institutions.
Shankar et al. (Sun,) studied this question.