A Trusted Execution Environment for Secure Reasoning on Large-Scale Models in the Power Industry

The growing deployment of large-scale models in the power industry improves grid operation and decision-making. However, it also introduces security concerns, such as adversarial reasoning attacks and data manipulation. To address these challenges, this paper proposes a Trusted Execution Environment (TEE)-based secure reasoning framework enhanced with Explainable Artificial Intelligence (XAI). XAI methods are integrated to identify critical input features and detect anomalies by analyzing abnormal feature attribution patterns. The framework ensures that sensitive data and reasoning processes are securely executed within hardware-isolated environments while maintaining interpretability and operational transparency. Experimental results in simulated power grid scenarios demonstrate that the proposed approach significantly improves both the security and explainability of large model reasoning, offering a reliable defense mechanism for critical power infrastructures.