Structural Deep Learning for Guaranteed Attacks with Zero System Knowledge

The present outlook in power systems is rapidly changing due to the introduction of (1) new active devices into the grid, such as photovoltaic (PV) panels, wind generators, and energy storage devices, and (2) new data from sensing and control devices. While this abundant data improves situational awareness and enhances control schemes, it can make the power grid more vulnerable than ever to cyber-attacks with dire consequences. Cyberattack withdraws much attention due to its potential impact, its financial losses, and its implications for national security. To understand the risks, this work looks into the operation of the electric grids, e.g., how to solve power flow equations. Specifically, this work investigates the good and the bad parts of existing methods and proposes to have a stochastic solution for power flow analysis for robustness. The finding is that no matter how the solution method is improved, system information is crucial to securely analyzing the grid. This gives utilities a false sense of security by hiding such information. For example, in a false data injection attack (FDIA), an attacker must know system information and measurements. If system information is hidden, the grid seems impossible to attack successfully, e.g., passing the Chi-square test based on system information. This dissertation shows that a carefully designed system can not only attack successfully but also with a strong performance guarantee.