Matrix-Completion-Based False Data Injection Attacks Against Machine Learning Detectors

Abstract

False data injection (FDI) attacks can manipulate power system measurements, leading to system economic losses and security issues. Although machine-learning (ML) detectors can effectively detect FDI attacks, the current methods used to construct FDI attacks do not take into account the presence of ML detectors. To tackle this problem, we propose novel convex matrix-completion-based FDI (MC-FDI) attacks on DC and AC power flow models from an attacker’s perspective, accounting for the temporal correlation between compromised and historical measurements. The proposed attacks minimize the nuclear norm of the compromised measurement matrix to make the compromised measurement consistent with the historical measurements, and also maximize the L1-norm of the incremental voltage angle to ensure a sufficient negative impact on the power system operation. Moving target defense (MTD) is proposed to detect the proposed MC-FDI attacks from the defender’s standpoint. The idea is to actively change the line impedance to corrupt the spatial and temporal correlation of the compromised measurements in the MC-FDI attacks. Numerical results on the IEEE 14-bus and IEEE 118-bus systems show the stealthiness of the proposed attacks to both the Chi-square detector and ML detectors as well as the efficacy of MTD in detecting the MC-FDI attacks.