Efficient and Privacy-Preserving ConvLSTM-Based Detection of Electricity Theft Cyber-Attacks in Smart Grids

In Advanced Metering Infrastructure (AMI) networks, Smart Meters (SMs), are installed at consumers' houses, provide electric utilities with fine-grained power consumption data necessary for accurate billing, load monitoring, and energy management. However, utility companies are still subjected to electricity theft cyber-attacks in which fraudulent consumers may manipulate their reported readings and hence reduce their bills. Several ML-based electricity theft detectors have been proposed in the literature, however, they either do not capture well the deeper periodicity and temporal features in energy consumption data or violate consumers' privacy by running these models over unencrypted power consumption data. To address these challenges, we propose in this paper a Conv-LSTM-based detector that integrates a 2-D Convolutional Neural Network (CNN) model with a Long Short-Term Memory (LSTM) network to significantly improve the model's functionality and detection accuracy, specifically addressing the inherent periodicity and temporal dependencies in electricity consumption data. Moreover, to run the proposed model over encrypted 2D data and preserve consumers' privacy, we designed a novel lightweight Inner Product Functional Encryption (IPFE) scheme that allows SMs to send their encrypted power consumption data to the Electric Utility (EU) which can securely compute the first feature map of the first convolutional layer of the Conv-LSTM detector while preserving consumer privacy. Our analysis and experiments demonstrate that our scheme is secure and efficiently detecting fraudulent consumers with minimal overhead. In specific, our model achieves a Detection Rate (DR) of 92.95%, a False Alarm Rate (FAR) of 3.68%, and a High Detection (HD) rate of 89.27%, resulting in an overall Accuracy (ACC) of 94.65%. Moreover, our scheme achieves high Precision (PR) at 98.80% and a robust Area Under the Curve (AUC) value of 98.50%. These results highlight the effectiveness of our approach in enhancing both detection accuracy and reliability all while protecting consumers' privacy.