Deep Learning-Based Attack And Defense Method For Electric Theft In Smart Grid

With the continuous popularization of smart grids and the establishment of advanced measurement systems,a large amount of electricity consumption data has been generated in smart grids.These data can effectively help power companies to detect non-technical losses(NTL)in the power grid,especially electricity stealing users.At present,many researchers have used deep learning technology to analyze the characteristics of the collected user electricity consumption data,and based on this,they have designed detection methods for NTL.However,there are still many shortcomings in the existing detection methods to be improved.This paper proposes a deep learning-based electricity stealing detection framework by deeply excavating the characteristics of users’ electricity consumption data,and further designs a covert electricity stealing attack and its defense method that can evade the existing mainstream detection methods.The main work contents are as follows:1.A hybrid NTL detection framework based on convolutional neural networks is proposed.The current methods for NTL detection are studied,which lack the power consumption comparison of users and their communities.In response to this problem,this paper combines the width and depth convolutional neural network models with the maximum information coefficient algorithm,and extracts the characteristics of the user’s electricity consumption behavior pattern,statistical characteristics and the correlation between the user’s electricity consumption data and the community’s electricity loss.detection.Detailed experiments show that the detection framework has good robustness and accuracy in the face of various attack methods.2.A stealth electricity stealing attack strategy is proposed,and a defense method based on deep neural network is designed for this attack strategy.Existing NTL detection methods based on machine learning have detection weaknesses,especially the attack strategy of electricity stealing users manipulating adjacent electricity meters to transfer bills while maintaining their normal electricity consumption patterns.Such attacks are almost impossible to be discovered by existing detection methods.In order to deal with this problem,this paper analyzes the electricity consumption deviation of normal users and electricity stealing users from horizontal and vertical perspectives,and defines a power consumption pattern feature and a correlation feature,respectively.Then,a hybrid detection model is designed based on these two features.The experimental results show that the method has good detection performance.3.A covert real-time attack method based on conditional generative adversarial network is proposed.At present,NTL detection methods based on deep learning have achieved good results.However,these detection methods lack the consideration of the vulnerability of deep learning,so there is a huge threat when deploying NTL detection systems.This paper designs an attack method that misleads the anomaly detection system by adding disturbance to the electricity stealing data,and introduces a feature extraction module into the conditional generative adversarial network,which reduces the training time and improves the attack success rate of the generated samples.The experimental results show that the generated attack samples can evade most of the existing detection methods.