Federated Learning Under Non-IID Patient Distributions for Malaria Diagnosis in Low-Connectivity Clinics: A Robustness Evaluation of Differential Privacy Budgets

Federated learning promises to train malaria diagnostic models across clinics without centralising patient images, but two design choices interact in ways that current evaluations do not characterise: the non-IID partition of patient data across sites, and the differential privacy budget spent on each round of aggregation. the present analysis treats the privacy budget as a design parameter whose admissible range depends on the heterogeneity of the site-level label and feature distributions, and it derives the robustness consequences of that dependence. The analysis is methodological and theoretical rather than empirical: no patient images were collected, and no federated model was trained. Instead, the paper formalises the non-IID federated objective, defines a robustness criterion based on the worst-site sensitivity of the aggregated model, and derives a bound linking the privacy budget, the number of participating clinics, and the between-site divergence of the diagnostic distribution. The bound is then used to construct an evaluation protocol that a low-connectivity clinic network could execute with its own data. The central result is that the privacy budget cannot be selected independently of the non-IID structure: for a fixed robustness target, the admissible budget contracts as between-site divergence grows, and the contraction is steeper than the standard composition analysis predicts.