Privacy-Preserving Federated Learning for Forced Displacement Broader Than Conflict in Uganda: Robustness Under Distribution Shift and Sparse Labels

Abstract

Forced displacement in Uganda extends well beyond conflict-driven flight, encompassing disaster-induced mobility, climate pressures and economic relocation, yet machine learning systems designed to support displaced populations typically assume stable data distributions and abundant labelled examples. We introduce a federated optimisation objective regularised by both differential privacy noise and a distributionally robust penalty, and we prove convergence guarantees under non-identically distributed client data when labels are scarce. The central contribution is a client-weighted aggregation rule that estimates distributional uncertainty without requiring centralised access to raw data, together with a semi-supervised pseudo-labelling mechanism that provably bounds the error introduced by unreliable labels. Theoretical analysis establishes that the proposed method attains sub-linear regret under mild convexity assumptions, while the privacy guarantee degrades gracefully as the number of participating clients grows. The framework is positioned as a methodological foundation for future empirical deployment, with explicit attention to the infrastructural constraints and ethical obligations of humanitarian computing in Uganda.