Adversarial training-based robust model for transmission line’s insulator defect classification against cyber-attacks

Adversarial training-based robust model for transmission line’s insulator defect classification against cyber-attacks

The increased sophistication of smart grids has generated significant interest in employing unmanned aerial vehicles (UAVs) to monitor the operational condition of insulators, especially in identifying insulator defects to avoid substantial power loss, shortened lifespan of power lines, and power outages. The classification process of insulator defects remains a concern due to factors such as the small size of the objects, intricate backgrounds, and a restricted amount of available data. Moreover, although cyber-attacks can affect deep learning (DL) models to cause misclassification of insulator defects, existing literature has yet to address this issue. Hence, this work employs the YOLOv9 model, a state-of-the-art object detector, to classify defects in power transmission line insulators, including accurate identification of small defects within intricate backgrounds. Additionally, this study is the first to introduce fast gradient sign method (FGSM) and projected gradient descent (PGD)-based adversarial attacks to insulator defect classification tasks to address their severity, while proposing adversarial training as a defensive measure. The experimental results show that the YOLOv9 model achieves a mean average precision (mAP) value of 96.5%, outperforming existing YOLOv8s, YOLOv7, YOLOv5s, and Faster R-CNN by 1.4%, 1.8%, 6.7%, and 10.3%, respectively. Also, this investigation demonstrates the severe effects of adversarial attacks, where mAP values of the YOLOv9 model decrease up to 14% and 9.9% under FGSM and PGD attacks, respectively. The proposed adversarial trained YOLOv9 model ensures robustness against cyber-attacks and maintains high accuracy in normal conditions, even in the presence of Gaussian noise.

Files associated with this item cannot be displayed due to copyright restrictions.