Certified Backdoor Detection for Artificial Intelligence Climate Downscaling Models Under Trigger-Pattern Injection: Detection Bounds and Forecast Integrity Assurance

Statistical downscaling models that map coarse reanalysis fields to station-scale temperature are increasingly deployed in climate services, yet their learned mappings can be silently compromised by trigger-pattern backdoors injected during training. The framework rests on a distributional separation principle: the trigger induces a low-dimensional shift in the model's internal representation that is detectable with high probability whenever the trigger's activation energy exceeds a threshold determined by the model's Lipschitz constant and the natural variability of the input distribution. We derive finite-sample detection bounds that hold uniformly over trigger patterns, establish a lower bound on the trigger magnitude required for undetectability, and show that the certified detector's false-negative rate decays exponentially in the number of clean calibration samples. The framework is operationalised on a real reanalysis record for Kampala, Uganda (0.35°N, 32.58°E), comprising 312 monthly air-temperature observations from 2000 to 2023 with a mean of 22.443 °C and a warming trend of −0.1564 °C per decade. The certified detector is shown to be sound under the stated assumptions, and its guarantees are shown to degrade gracefully when the Lipschitz estimate is loose. The contribution is a detection-theoretic foundation for forecast integrity assurance in climate downscaling, with explicit bounds that a climate service can audit before deployment.