Abstract:The correct classification of power quality (PQ) is the key step to ensure the normal operation of smart grid. Deep neural networks have been widely used for PQ classification, but they face serious threats from adversarial attacks. At present, there are few studies on defense algorithms for adversarial attacks in PQ classification. Furthermore, detection algorithms have not yet been studied. In order to solve the above problems, we first use Convolutional Neural Network-Long Short-Term Memory Network (CNN-LSTM) to classify PQ signals. Four representative adversarial attack algorithms (i.e., FGSM, FGSM-variant, PGD and SSA) in PQ classification are compared and summarized from three aspects: average running time of attack, robustness of attack and classification accuracy of attacked model. Second, we propose a Multi-Source Feature Detector (MSFD) to detect adversarial attacks. The model architecture of MSFD is a binary classification model based on CNN. The detection function can be implemented simply by placing MSFD in front of the classification model. Third, we propose a Multi-Source Feature Adversarial Training (MSFAT) to defend against adversarial attacks. The experimental results demonstrate the effectiveness and superiority of the proposed methods. MSFD can detect the above four representative adversarial attacks with an extremely high recognition rate, and MSFAT can significantly improve the classification accuracy of the attacked model under these attacks. MSFD and MSFAT can effectively deal with adversarial attacks in different environments without any adjustment. Compared with the most advanced adversarial training defense method in PQ classification, the proposed MSFAT can overcome the drawback that adversarial training may reduce the classification accuracy of the attacked model and is significantly effective for multiple adversarial attacks in PQ classification.

Super-efficient detector and defense method for adversarial attacks in power quality classification