Robust ConvLSTM Model With Deep Reinforcement Learning for Stealth Attack Detection in Smart Grids

The advent of modern electricity distribution systems, comprising digital communication technologies and principles, has triggered a new era of smart grids, in which advanced metering infrastructure plays a crucial role in functions, such as digital monitoring and billing. However, this advancement gave rise to vulnerabilities, mainly in the form of energy theft and adversarial attacks to falsify energy consumption data. In response, anomaly detection models have been tested and evaluated against machine-generated adversarial attacks, such as the fast gradient sign method (FGSM) and Carlini and Wagner (C&W). However, these types of attacks are mainly designed to prevent anomaly detection without considering a possible reduction in the reported energy consumption, thus overlooking the energy theft problem. Furthermore, the lack of generalization of adversarial attacks evaluated to other models remains a concern. In fact, conventional anomaly detection methods do not detect new adversarial attacks generated by artificial intelligence. Thus, this article introduces a state-of-the-art deep reinforcement learning (DRL) through a deep deterministic policy gradient, which produces novel adversarial samples that dramatically reduce reported energy consumption while evading the anomaly detection mechanism. The ConvLSTM uses the adversarial data to evaluate and enhance its resilience against such adversarial attacks. Experimental results show that conventional models experience substantial degradation in detecting such advanced attacks, achieving 17% accuracy under the evaluated adversarial attacks. In benchmarking, the proposed DRL framework against adversarial attacks generated through functions, such as FGSM and C&W. As a defensive strategy, the ConvLSTM trained with DRL-generated adversarial samples increases the model robustness against AI-driven threats in smart grids by improving its detection capabilities, achieving a 95.12% detection rate on adversarial samples. This work showcases adaptive anomaly detection methodologies and the implementation of AI-driven approaches to enhance cybersecurity in modern digital power systems.