A GAN-Based Real-Time Covert Energy Theft Attack Against Data-Driven Detectors

The advanced metering infrastructure (AMI) system has been rapidly established around the world, effectively improving the communication capability of the power system. Problematically, it turns out malicious users can easily commit energy theft by tampering with smart meters. Thus, many data-driven methods have been proposed to detect energy theft in AMI. However, existing detection schemes lack consideration for well-planned covert attacks, making them vulnerable. This paper proposes a real-time covert attack model based on conditional generative adversarial network (CGAN). In particular, based on the transferability of adversarial samples, we first extract the data features that the malicious detection model focuses on during the detection process. Then, we utilize these extracted features and a generator to generate adversarial perturbations that can mislead malicious detection models. Finally, to make the generated perturbations more stealthy, a discriminator is used to simulate malicious detection models to correct them. Extensive experiments demonstrate that our proposed attack method can evade most current detection methods.