Controllable Blind AC FDIA via Physics-Informed Extrapolative AVAE.

Area of Science:

• Electrical Engineering
• Cybersecurity
• Artificial Intelligence

Background:

• False data injection attacks (FDIAs) threaten power system AC state estimation, particularly without voltage measurements.
• Existing machine learning methods for FDIAs have limitations in controlling errors and generalizing beyond training data.

Purpose of the Study:

• To develop a novel method for generating controllable and stealthy false data injections for AC state estimation.
• To improve the robustness and adaptability of power system security against sophisticated cyberattacks.

Main Methods:

• Proposed a physics-informed extrapolative adversarial variational autoencoder (PI-ExAVAE).
• Incorporated physically consistent priors from AC power flow equations to ensure plausibility and stealth.
• Utilized adversarial training to generate realistic attack vectors.

Main Results:

• Achieved a 90% success rate in bypassing detection tests across different attack configurations.
• Demonstrated superior performance over existing methods like SAGAN in generating smoother, more realistic deviations.
• Showcased the model's ability to extrapolate beyond training data, targeting unseen operational scenarios.

Conclusions:

• Integrating physics knowledge into data-driven models enhances adaptability and robustness against evolving detection mechanisms.
• The PI-ExAVAE model offers a powerful tool for generating realistic and effective FDIAs for cybersecurity research.
• Physics-informed AI is crucial for developing resilient power systems against sophisticated cyber threats.