Security architecture of smart grid knowledge question-answering systems based on large language models
Yun Fu, Junrong Liu, Qiucheng Ban, Linyan Zhou · Cyber-Physical Systems · 2025
With the rapid advancement of large language models (LLMs), their application in smart grid knowledge services is increasingly promising, yet challenged by data sensitivity and output unpredictability. This paper proposes GRASP, a secure response architecture that integrates trusted execution environments, adaptive input–output auditing, and knowledge-grounded verification to enhance LLM trustworthiness in power systems. GRASP isolates inference processes, filters malicious inputs, constrains risky outputs, and reinforces factual consistency through a domain-specific knowledge graph. Experimental results demonstrate that GRASP significantly improves the safety, accuracy, and reliability of smart grid question-answering systems.