Privacy-Preserving Federated Learning for Crimes Against Humanity in Uganda: Robustness Under Distribution Shift and Sparse Labels

The documentation and analysis of crimes against humanity in Uganda confront a dual computational challenge: data are distributed across sensitive, institutionally siloed repositories, and the labels required for supervised learning are exceptionally sparse due to the cost and risk of expert annotation. We introduce a novel aggregation rule, the Distributionally Robust Federated Averaging (DRFA) algorithm, which integrates a Wasserstein-distance-based ambiguity set into the federated optimisation objective to bound performance degradation when client data distributions diverge from the centralised target distribution. To address sparse labels, we propose a semi-supervised consensus mechanism that propagates confidence-weighted pseudo-labels through a graph-regularised manifold, operating entirely within the privacy constraints of differential privacy. We provide formal convergence guarantees for DRFA under non-identically distributed client data and derive a generalisation bound that characterises the trade-off between privacy budget, label sparsity, and worst-case risk. The framework is evaluated theoretically through a series of propositions and proofs, establishing that the proposed method achieves sub-linear regret under conditions of bounded distribution shift. The article concludes by delineating the institutional and ethical conditions under which such a system could be responsibly deployed by Ugandan human rights bodies, emphasising that technical robustness is a necessary but insufficient condition for evidentiary credibility.