Privacy-Preserving Federated Learning for Conflict Resolution Peacebuilding in Uganda: Robustness Under Distribution Shift and Sparse Labels

Federated learning offers a compelling architecture for peacebuilding analytics in conflict-affected regions, where data sovereignty, institutional sensitivity, and infrastructural fragility preclude centralised data aggregation. We introduce a federated empirical risk minimisation objective regularised by a Wasserstein-distance penalty that anchors local models to a robust global prior, and we prove a generalisation bound that degrades gracefully with both participation heterogeneity and label scarcity. The framework is analysed theoretically rather than evaluated empirically, and we derive conditions under which the proposed regulariser dominates unregularised FedAvg in expected risk. We further propose a privacy-accounting mechanism compatible with local differential privacy that preserves the utility guarantees of the central result.