Enhancing the Performance of Cyber-Attacks Detection and Localization in Smart Grids Using Deep Transfer Learning

Nowadays, the detection and localization of cyber-attacks have become increasingly important. Artificial intelligence methods have significantly improved effectiveness in this field. For instance, one such technique involves neural networks, which were previously not appropriately adapted to the dataset. In this study, ResNet-18 and LeNet convolutional neural networks have been employed for the detection and localization of cyber-attacks. By utilizing pre-trained networks, such as ResNet-18 and LeNet, combined with a deep transfer learning approach, an approximate 10% increase in accuracy has been achieved. This improvement is attributed to the attack localization process, which is based on extracting features and recognizing patterns in both safe and attacked data. The dataset, created by Oak Ridge National Laboratories (ORNL), includes three subsets containing measurements related to normal, disturbance, control, and cyberattack behaviors in the electric transmission system. The transfer learning approach in these convolutional neural networks focuses on leveraging the simplest structures within their architectures to enhance performance.