Privacy-Preserving Federated Learning for Biophysics Core Life in Uganda: Robustness Under Distribution Shift and Sparse Labels

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Laboratories providing biophysics core services in Uganda hold fluorescence images, electrophysiological traces and macromolecular characterisation records that cannot be pooled in a single server for legal, ethical and infrastructural reasons, yet models trained in one site routinely lose accuracy when applied in another where instruments, reagents and preparation protocols differ and where expert annotation is scarce. The framework proposes clipped federated optimisation with distributed differential privacy, domain-discrepancy regularisation and consistency-based semi-supervised learning, and it derives formal bounds that relate privacy noise, inter-client discrepancy and unlabelled-data utility to performance on unseen sites. The hypothesis predicts that explicit control of shift dominates naive increases in local sample size, and the analysis is designed to identify bandwidth-aware aggregation schedules and governance-compatible trust assumptions that would permit estimation and validation in future field deployments without requiring raw data centralisation.