Impact Analysis of Adversarial Attacks on Deep Reinforcement Learning-Based Wide-Area Microgrid Control

The growing integration of deep reinforcement learning (DRL) into power system control offers unprecedented adaptability and performance benefits. However, it also introduces new vulnerabilities that can be exploited by adversarial cyber threats. Unlike traditional model-based controllers, DRL agents depend heavily on real-time observations, feedback loops, and learned policies-making them inherently susceptible to malicious manipulation. This paper investigates the cybersecurity and resilience of a DRL-based wide-area secondary controller operating within an autonomous microgrid. Specifically, we evaluate a Deep Q-Learning Network (DQN) agent responsible for maintaining voltage and frequency stability through coordinated decision-making. To assess its robustness, we simulate a comprehensive suite of cyberattack scenarios-including false data injection, strategically triggered backdoors, spoofing, reward poisoning, and coordinated multi-vector attacks targeting both observation and actuation channels of the trained DQN. We also introduce a set of evaluation metrics-including reward degradation, frequency and voltage deviation, and a composite system stability index-to quantitatively assess the impact of these attacks. Our simulation results reveal that even subtle, state-dependent, or reward-level manipulations can significantly degrade control performance while evading detection by conventional monitoring systems. In contrast, high-magnitude sensor attacks result in a complete loss of observability but are easier to detect by system operators. Our experimental results also showed that coordinated attacks that simultaneously disrupt multiple channels induce severe impacts, triggering transient instability in frequency and voltage regulation.