Federated learning for climate-model downscaling across data-scarce national meteorological services: a differential-privacy evaluation of forecast skill and sovereign data leakage

National meteorological services in data-scarce regions face a structural tension: statistical downscaling of coarse climate-model output requires large, pooled training corpora, yet the station records that would supply such corpora are sovereign assets whose transfer across borders is legally and politically constrained. Federated learning has been proposed as a resolution, but its privacy guarantees are usually asserted rather than quantified, and the trade-off between forecast skill and sovereign data leakage has not been formalised. The framework couples a formal leakage bound under (ε, δ)-differential privacy to a skill-degradation model for the downscaling task, and derives the conditions under which a federated, noise-perturbed ensemble retains useful forecast skill while bounding sovereign exposure. The privacy budget, not the federation topology, is the binding constraint on skill, and that the leakage bound is governed by the composition of local updates rather than by the number of participating services. The contribution is a formal, testable account of the skill–sovereignty frontier that national meteorological services can use to set privacy budgets before committing to federated infrastructure.