Privacy-Preserving Detection of Power Theft in Smart Grid Change and Transmit (CAT) Advanced Metering Infrastructure

For energy management and billing purposes, advanced metering infrastructure (AMI) requires periodic transmission of consumer power consumption readings by smart meters to the electric utility (EU). An efficient way for collecting readings is the change-and-transmit approach (CAT AMI) whereby readings are only transmitted if there is an enough change in consumption readings. CAT AMI, however, is plagued by malicious consumers who hack their smart meters to illegally lower their electricity bills by falsifying their readings. These attacks on the AMI could have bad economic consequences and impair the performance of the power grid if these readings are used for managing the grids. Machine learning models can be used to detect false readings but this requires disclosing consumers' CAT readings to the EU to evalaute the model. However, disclosing the consumers' readings jeopardizes consumers' privacy due to the fact that these readings can reveal sensitive information about consumers' lifestyles, e.g., their presence or absence, the appliances they use, etc. The problems of detecting power theft while protecting the consumers' privacy in CAT AMI is investigated in this paper. First, a dataset of actual readings to generate a benign dataset is developped followed by proposing new cyber-attacks tailored for CAT AMI to generate malicious samples. Then, two deep-learning detectors using a baseline model (CNN) and a CNN-GRU model are trained to detect power thefts in CAT AMI. To preserve consumers' privacy, the paper develops an approach to enable the EU to evaluate the detector using encrypted data without being able to learn the readings. Extensive experiments were carried out to assess our proposal, and the results indicate that our proposal is capable of accurately identifying malicious consumers with acceptable overhead while preserving the privacy of the consumers. Specifically, comparing to CNN model, our CNN-GRU model increases the detection rate from 93.85% to 97.14% and $HD$ from to 87.7% to 94.28%, respectively.