Trojan Horse Hunt in Deep Forecasting Models: ESA Competition Findings

arXiv·medium signal

Researchers present findings from a European Space Agency competition to detect Trojan backdoors embedded in deep learning forecasting models used for safety-critical space operations. Attackers can embed subtle triggers that activate under specific input conditions while the model performs normally otherwise, with potentially catastrophic consequences when the trigger fires in production. The paper provides detection methodology and lessons applicable to any production forecasting pipeline in safety-critical systems.