Deep Recurrent Electricity Theft Detection in AMI Networks with Evolutionary Hyper-Parameter Tuning

Despite its prominent role in future smart power grid for monitoring and billing purposes, advanced metering infrastructure (AMI) opens the door wide for serious cyber-security threats. This paper investigates an electricity theft cyber-attack scenario where malicious customers hack their smart meters and manipulate the energy consumption readings in order to reduce their electricity bills. In this context, machine learning techniques offer an attractive means to efficiently detect such cyber-attacks. The contributions of this work are two-fold: 1) Unlike the existing works that rely on shallow and static machine learning techniques, we propose a detection mechanism based on deep recurrent neural networks that exploit the time-series nature of the energy consumption readings to improve the detection performance, and 2) While the existing works did not consider fine-tuning of the detector's hyper-parameters, we employ a multi-objective evolutionary (Non-dominated Sorting Genetic Algorithm)-based optimization approach to fine-tune the detector's hyper-parameters to further improve the detection performance. Real energy consumption data is utilized to test the proposed detection mechanism, which demonstrates a superior detection performance compared with a state-of-the-art (static and shallow) detector.