Abstract
The growing deployment of inverter-based resources (IBRs) in modern power systems introduces critical cyber-physical vulnerabilities. To address this gap, a novel spatio-temporal multi-head self-attention convolutional network (STAC) is proposed for the cyber defense of modern power systems with high IBR penetration. The STAC architecture effectively detects various cyberattacks targeting IBRs, including sparse measurement injection, sparse input hijacking, load deviation, switching, and local controller attacks. The proposed approach provides a model-agnostic framework that: (1) leverages STAC's multi-head attention mechanisms to identify temporal patterns distinctive to different attack vectors; (2) integrates adversarial training to enhance robustness against sophisticated attacks; and (3) validates performance across high IBR penetration scenarios using the modified 68-bus test system. Experimental results demonstrate detection within 6 ms of attack onset with no false positives during benign grid events such as load changes and line trips. The STAC model accurately classifies all five distinct attack types with an overall 93% accuracy, outperforming conventional methods. The STAC approach requires no internal inverter model knowledge, enabling straightforward deployment on existing systems.