Data-driven FDI Attacks: A Stealthy Approach to Subvert SVM Detectors in Power System

False Data Injection (FDI) attacks can mislead the grid operator’s state estimation without being detected by the Chi-2 detector in bad data detection. While machine learning methods have been used to detect FDI attacks, existing FDI construction methods only consider the Chi-2 detector and not machine learning (ML) detectors. This paper investigates the limitation of a support vector machine (SVM) detector in detecting FDI attacks. The study proposes a novel data-driven FDI (DFDI) attack that aims to remain undetected by the SVM detector and have a significant impact on the power system operation. The DFDI attack uses the minimum-area ellipse enclosing (MEE) method to obtain the boundary of the historical measurements’ cluster, which is used as a constraint to ensure that the compromised measurements are projected into the cluster. Numerical results on the IEEE 14-bus system demonstrate the limitations of the SVM detector and drawbacks of traditional FDI (TFDI) attacks, and further verify the stealthiness of the proposed attacks against the SVM detector.