HIV-PrivFed: Privacy–Utility–Fairness-Aware Federated Learning for HIV Risk Stratification and Testing Prioritization Under Non-IID Clinical Heterogeneity

Background/Objectives: HIV screening from routinely collected behavioral data could help to prioritize testing across clinics that cannot centralize sensitive records; however, the utility, privacy, and client-level fairness of federated models on real non-IID data are poorly characterized. Methods: Using Sialon-II, a cohort of men who have sex with men in 13 European cities (13 natural clients; horizontal federated learning), we benchmarked ten conventional classifiers and a centralized residual MLP, compared local-only, FedAvg, and FedProx baselines, and evaluated HIV-PrivFed, a differentially private FedProx-based algorithm with client-by-client Rényi accounting and client-level fairness auditing (fairness is audited, not enforced); secure aggregation overhead was benchmarked in simulation. Results: The best conventional model reached an AUROC of 0.696 (AUPRC 0.235), statistically indistinguishable from logistic regression on the original partition. Federated learning exceeded local-only learning (AUROC 0.607 versus 0.565). Achieved worst-client privacy budgets (ε = 22.1–224.1) were far looser than the nominal targets and give no meaningful guarantee; accuracy was retained at these budgets (AUROC 0.629–0.650) but fell to 0.553 and 0.524 at ε=8 and 4. Client-level estimates were highly uncertain. Conclusions: Useful accuracy and a meaningful privacy guarantee did not coexist in this small-cohort setting; larger cohorts and external validation are needed before deployment.