Black-box attacks and defense for DNN-based power quality classification in smart grid

Abstract

Machine learning (ML) models are widely used in smart grid, but they are vulnerable to adversarial examples that are maliciously crafted using the input data. Therefore, the use of these models in smart grid can cause significant damage. Although there are many studies on ML models attack and defense, there are few studies based on smart grid environment. In this paper, we first utilize deep neural network (DNN) to classify power quality (PQ) signals. Then an algorithm based on the Projected Gradient Descent (PGD) is proposed to attack the classification model, which employs multiple iterations to find the optimal perturbation. Finally, a defense model based on Multi-Level-Denoising Autoencoder (MLD-AE) is proposed to defend against the black-box attacks. The experimental results show that, compared with the Fast Gradient Sign Method (FGSM), the proposed black-box attack algorithm can significantly reduce the classification ability of the model with a smaller disturbance. Furthermore, compared with adversarial training defense methods, the proposed defense model can improve the power system’s ability to resist black-box attacks without reducing the ability of the classification model to classify clean PQ signals, and has excellent scalability.

Keywords

• Smart grid
• Adversarial examples
• DNN
• Black-box attacks
• PGD
• MLD-AE