Transfer Learning-Based Classification of Cyber Attacks Against Power Grids

Power grids serve as the backbone of critical infrastructures, enabling the efficient control and distribution of electricity. Subsequently, the number of cyber attacks against power grids continues to grow, especially during times of international conflict and uncertainty. Thus, we must continue to further the research that secures and ensures the stability of power grids. Related research has made progress to this end through the application of artificial intelligence-based systems, but such works suffer from the following limitations: (a) they strongly focus on attack detection, neglecting attack classification, (b) they produce complex systems requiring large amounts of computation without concern for reducing complexity, and (c) they lack concern for how the systems can be adapted when new attacks arise. These limitations motivate our work, where we develop efficient and high-performing classification systems that are adaptable to new attacks. Specifically, we develop diverse benign and attack datasets consisting of cyber and physical layer data using our power system testbed. Additionally, we adopt SHapley Additive exPlanations to reduce the total number of features required to accurately classify attacks by 83% while maintaining a superior accuracy of 97%. Lastly, we use transfer learning to enhance fine-tuning and adapt our classification system to classify new cyber attacks in power grid environments.