A Novel Evasion Attack Against Global Electricity Theft Detectors and a Countermeasure

Abstract

The smart grid advanced metering infrastructure (AMI) is vulnerable to electricity theft cyber-attacks in which malicious smart meters report low readings to reduce the consumers’ bills. To avoid this problem, several machine-learningbased detectors have been proposed to detect electricity theft. Most of these detectors are global in the sense that they are trained on different consumption levels, including low and high consumptions, to be used for all consumers. In this paper, we introduce a novel type of evasion attacks against global detectors as follows. A malicious consumer who has high consumption level can send false readings for a low consumption profile (that resembles the profiles the detector is trained on) to evade the detector, i.e., steal electricity without being detected. We firstly conduct experiments to prove that the existing global detectors are vulnerable to this new kind of evasion attacks. To launch this attack, we train a Generative Adversarial Network (GAN) on a real dataset to generate fake low consumption readings that can evade the detector. The given results indicate that the success rate of the attack is between 82% and 97%. To thwart this attack, we divide the consumers into clusters of close electricity consumption levels and train one detector for each cluster. Therefore, if a malicious consumer in any cluster tries to imitate the consumption profiles of consumers in other clusters, he/she will be detected. On the other hand, it is not profitable to imitate the electricity consumption profiles of consumers in his/her cluster to evade detection. To prove the effectiveness of our countermeasure, extensive experiments are conducted and the results indicate that our countermeasure can successfully thwart the attack.