Guarding the Grid: Data Poisoning Attacks and Countermeasures in Power System Stability Assessment AI Model

Artificial Intelligence (AI) algorithms have become a critical tool for addressing the transient stability assessment issue in modern power systems. By leveraging historical or simulation data to establish correlations between features and transient stability, AI eliminates the need for modeling and analyzing intricate physical mechanisms, thereby reducing problem complexity and enhancing computational efficiency. However, the black-box nature of AI introduces significant security risks. Malicious attackers can exploit vulnerabilities during the training process, embedding backdoors in the models to manipulate their outputs and potentially disrupt power system operations. This paper explores the feasibility of embedding backdoors in AI models for power systems transient stability classifier and proposes a data poisoning-based backdoor attack method tailored to these systems. A backdoor trigger is designed to exploit the difficulty of accessing system nodes, causing AI models to misclassify specific scenario samples. To counter such threats, detection schemes are developed at both the model and sample levels. Finally, the effectiveness of the proposed attack and detection methods is validated through a case study on a AI-driven transient stability classifier.