Privacy-Preserving Synthetic Data for HIV Risk Prediction in People Who Inject Drugs in Kenya: Benchmarking Differentially Private Generative Models

People who inject drugs (PWIDs) in Kenya face HIV prevalence three to four times the general population rate, yet criminalization, stigma, and hard-to-reach sampling produce small, class-imbalanced datasets that are ethically difficult to share across institutional boundaries. This tension whereby the epidemiological need to predict risk meets the moral and legal need to protect subjects from re-identification, is a long-standing AI ethics problem that synthetic data is increasingly proposed to solve. Using a de-identified dataset of 6,142 PWID records from Nairobi and Coastal Kenya (17.8% HIV-positive), the study benchmarked eight augmentation strategies, comprising four deep generative methods (VAE, Tabular GAN, CTGAN, and PATE-GAN at ε = 1.0) and four classical baselines, across four classifiers, yielding 32 configurations. PATE-GAN paired with Random Forest achieved the highest precision-recall AUC (0.9165), matching or exceeding every non-private method, while providing a formal (ε = 1.0, δ = 10⁻⁵)-DP guarantee. Every method cleared a distance-to-closest-real privacy threshold. Formal DP carries near-zero utility cost in low-dimensional socio-behavioral tabular data. Remaining risks (group-level harms, surveillance creep, and consent boundaries for synthetic data derived from vulnerable populations) are discussed, and five governance-layer deployment commitments are proposed