LAUSR

False data injection (FDI) attacks against power system state estimation through manipulating measurements can result in economic losses and grid operating security issues. FDI attacks are stealthy to the traditional bad data detector. However, existing FDI construction methods fail to consider the stealthiness of attacks against machine-learning (ML) detectors. Since the historical measurement patterns are generally utilized by ML detectors, we apply the tensor completion (TC) technique in the FDI construction to manipulate compromised measurements matching the historical measurement patterns. We propose a novel convex TC-based FDI (TC-FDI) attack algorithm that 1) minimizes the nuclear norm of the compromised measurement tensor to make the compromised measurements consistent with the historical ones and 2) maximizes the L1-norm of the incremental voltage to ensure a sufficient negative impact on the power system operation. Further, the reactance perturbation strategy (RPS) is utilized to detect the TC-FDI attacks by breaking the spatial and temporal correlation of the compromised measurements. Numerical results on the IEEE 14-bus system show the stealthiness of the proposed attacks to the statistic-based detectors and ML detectors. The efficacy of the RPS in detecting TC-FDI attacks is also demonstrated.