Efficient Deep Learning Based Detector for Electricity Theft Generation System Attacks in Smart Grid

This thesis investigates the problem of electricity theft attacks in the generation domain. In this attack, the adversaries manipulate readings to claim higher energy injected into the grid for overcharging utility companies by hacking smart meters monitoring renewable-based Distributed Generation (DG). In prior research, deep learning (DL) based detectors were developed to detect such behavior, though they relied on different data sources and overlooked the critical impact of small perturbations which an attacker could integrate into its reported energy. This thesis takes advantage of addressing this gap by proposing an efficient DL-based detector that can offer higher accuracy and detection rate using only a single source of data by adding two features to enhance the performance. Subsequently, the proposed detector is further extended to deal with the small perturbations that attackers can add. We carry out extensive simulation designing two different detectors, one for solar DG units electricity theft issue, and the other for multiple fuel types (i.e., solar, and wind). We use a realistic dataset, and the results show that the proposed models detect the adversaries with higher rate detection even with small perturbations.