Low-Rank False Data Injection Attacks With Incomplete Network Information Against Machine-Learning Detectors
Bo Liu, Hongyu Wu · IEEE Transactions on Industrial Informatics · 2025
Machine learning (ML) methods have proven effective in detecting false data injection (FDI) attacks, which pose significant threats to the security of power systems. However, existing FDI construction methods do not adequately consider the presence of ML detectors. To address this issue, we propose convex low-rank FDI (LR-FDI) models for both dc and ac power systems. These models employ matrix completion techniques to ensure the temporal correlation consistency of compromised measurements with historical measurements. The LR-FDI attacks maximize the L1-norm of the incremental voltage, ensuring a sufficient negative impact on the power system operation with incomplete network information (INI). To maintain the spatial correlation consistency of the compromised measurements, we derive novel mathematic INI models as constraints and integrate them into the proposed attack model. The effectiveness of the LR-FDI attacks and their stealthiness to ML detectors are demonstrated through numerical results on the IEEE 118-bus system. The proposed LR-FDI attacks with INI augment the realism of an attacker's knowledge and capabilities, making them more challenging to be detected.