Covert Cyber Assault Detection in Smart Grid Networks Utilizing Feature Selection and Euclidean Distance-Based Machine Learning

Communications technologies are an integral part of efficient monitoring and reliable control in smart grids, but enhanced reliance on these technologies heightens the risk of cyber assaults. Recently, a new type of stealth, or covert, assault in smart grid networks has been discovered, which cannot be ascertained by legacy bad-data detectors using state estimation. Due to the delay-sensitive nature of smart grid networks, swift detection of abnormal changes is immensely desired. In this paper, we propose two Euclidean distance-based anomaly detection schemes for covert cyber-assault detection in smart grid communications networks. The first scheme utilizes unsupervised-learning over unlabeled data to detect outliers or deviations in the measurements. The second scheme employs supervised-learning over labeled data to detect the deviations in the measurements. Unlike the classic detection test, the proposed schemes tackle an unknown sample with low computational complexity, leading to a shorter decision time. To improve detection accuracy and further reduce the computational complexity and the associated time delay, we employ a genetic algorithm-based feature selection method to choose the distinguishing optimal feature data subset as input to both of the proposed schemes. The evaluation is carried out through the standard IEEE 14-bus, 39-bus, 57-bus and 118-bus test systems. Simulation results show that compared to the existing feature extraction-based detection schemes, the proposed schemes show significant improvement in covert cyber deception assault-detection accuracy.