Comparative Analysis between Supervised and Anomaly Detectors Against Electricity Theft Zero-Day Attacks

Smart power grids are vulnerable to electricity theft cyber-attacks, where malicious consumers hack their smart meters (SMs) to down-scale electricity usage readings before reporting them to electric utility companies (EUCs). This serious problem causes billions of dollars in losses to the EUCs worldwide and threatens the power grid's stability. Several machine learning (ML)-based solutions have been advised in the literature for electricity theft detection. However, most existing works propose supervised detection approaches and only a few works propose anomaly detection approaches. Therefore, in this paper, we investigate the effectiveness of the two approaches in detecting electricity theft utilizing a dataset of real electricity usage readings. Specifically, to the best of our knowledge, this work represents the first attempt to implement a comparative analysis between the performance of supervised deep learning (DL) models and anomaly detection models against electricity theft zero-day cyber-attacks. Our experimental results indicate that while the supervised detectors have proven successful against known attacks, they fail to detect new attacks. Moreover, our results demonstrate the superior performance of the anomaly detectors compared to the supervised detectors in defending against electricity theft zero-day attacks.