Adaptive FDIA in MTD-Enabled Smart Grids: A Data-Driven Deep Reinforcement Learning Approach

False Data Injection Attacks (FDIAs) pose a significant threat to the stability of smart grids. It enables adversaries to manipulate the state of the system while remaining undetected. Moving Target Defense (MTD) strategies alter grid parameters to invalidate attackers' knowledge, thereby increasing the detection rate. However, existing MTD methods assume attackers' inability to adapt to these hidden perturbations in real time. We challenge this assumption by proposing an adaptive FDIA framework that combines Long Short-Term Memory (LSTM) and Deep Reinforcement Learning (DRL) networks. The LSTM predicts MTD-induced reactance perturbations, while the DDPG agent strategically optimizes attack vectors to maximize stealth and impact. We evaluate our approach against state-of-the-art robust MTD on the IEEE 14-bus power system. Our approach reduces detection probability from over 98% to 52.05%, degrading MTD efficacy by nearly 46%, while inducing a 10.42% state deviation. These results expose critical vulnerabilities in MTD under adaptive adversaries, urging the development of dynamic, learning-resistant defense mechanisms.