Countering Evasion Attacks for Smart Grid Reinforcement Learning-based Detectors

Fraudulent customers in smart power grids employ cyber-attacks by manipulating their meters and reporting false consumption readings to reduce bills. To combat these attacks mitigate financial losses, various machine learning-based electricity theft detectors have been proposed. Unfortunately, are vulnerable serious cyber-attacks, specifically evasion attacks. The objective of this paper is investigate the robustness deep reinforcement learning (DRL)-based against our proposed through a series experiments. Firstly, we introduce DRL-based implemented using double Q networks (DDQN) algorithm. Secondly, propose attack model generate adversarial black box scenario. These samples generated modifying malicious reading deceive make them appear as benign samples. We leverage attractive features (RL) determine optimal actions for Our compared with an FGSM-based model. experimental results reveal significant degradation detector performance due attack, achieving success rate (ASR) ranging from 92.92% 99.96%. Thirdly, counter enhance detection robustness, hardened defense training process. This process involves retraining on achieves outstanding performance, ASR 1.80% 9.20%. Finally, address challenge whether model, which has adversarially trained samples, capable defending vice versa. conduct extensive experiments validate effectiveness models.