A ONE-WAY HASHING-BASED PRIVACY FRAMEWORK FOR SECURE FEDERATED CLINICAL ANALYTICS IN NIGERIA

Abstract

Machine learning is expanding rapidly across Nigerian clinical practice. However, without a data privacy framework built for the constraints of a low-resource health system, that expansion is outpacing the safeguards patients need. This study contributes a validated Privacy-Preserving Machine Learning (PPML) framework, combining SHA-256 salted one-way hashing with Federated Learning (FL), that fills a specific gap in the literature: no prior framework of this kind has been tested and aligned against Nigeria’s own NDPR/NDPA regulatory requirements rather than a generic privacy standard. The framework was tested using Design Science Research methodology on 13,240 patient records across 12 simulated federated nodes representing Nigeria’s six geopolitical zones. Key findings: SHA-256 hashing cut membership inference and linkage attack success by three-quarters or more, at an accuracy cost so small it was not statistically distinguishable from random variation; the framework’s Random Forest configuration reached 0.871 accuracy and 94.6% recall on the clinically critical deteriorating-patient class; and hashing added only 2.4% training overhead on commodity 4-core, 8 GB hospital hardware, with full compliance across all eight NDPR/NDPA requirements. The practical implication is that Nigerian hospitals do not have to choose between protecting patient privacy and running collaborative AI on their existing infrastructure. Of the configurations tested, only federated learning combined with SHA-256 hashing meets privacy, utility, and efficiency targets simultaneously, on hardware most institutions can already access.