Privacy-Preserving Federated Learning for Cross-Border Data Flows in Uganda: Robustness Under Distribution Shift and Sparse Labels

Cross-border data flows are central to Uganda's digital economy, yet their governance is complicated by competing pressures: the need for international integration and the legal obligation to protect citizen privacy. Federated learning offers a promising architecture for training machine learning models across distributed data custodians without centralising raw data. However, the practical viability of federated systems in this context is threatened by two statistical phenomena: distribution shift between participating nodes and extreme label sparsity. We introduce a generalised notion of client heterogeneity and derive convergence bounds for a FedAvg-style algorithm with differential privacy guarantees, showing that convergence degrades gracefully with increasing heterogeneity and privacy budget. We then propose a semi-supervised augmentation strategy that leverages the structure of unlabelled data to mitigate label sparsity, and we analyse its interaction with privacy mechanisms. The analysis reveals that the dominant design constraint is not communication cost but the joint effect of privacy noise and distribution skew on model bias. We conclude with design principles for federated systems operating in data-governance environments such as Uganda's, where legal fragmentation and infrastructural asymmetry are defining features.