Adversarial Attack and Defense in Multi-Agent Deep Reinforcement Learning for Active Distribution Network Dispatching

Multi-agent Deep Reinforcement Learning-based (MADRL) optimal dispatching for active distribution networks (ADNs) represents a future trend, necessitating robust security measures due to the sensitivity of MADRL. However, limited research addresses this issue. This paper introduces Multi-Agent Directed Shift Attack (MADSA), a powerful attack strategy for MADRL in ADNs, which targets single and full agents. The single-agent attack maximizes the deviation of a single agent’s strategy before and after the attack, leading to ADN degradation. And the full-agent attack unifies the action direction of all agents to achieve maximum degradation. To counter MADSA, we propose an adversarial defense method named Multi-Agent Gradient Leveling Defense (MAGLD) with Gradient Leveling Regularization, which enhances the robustness of the defense strategy. Case studies show MADSA can degrade the ADN under continuous and single-step attacks, with even the small attack amplitudes significantly increasing power losses and causing overvoltage, surpassing existing methods, such as fast gradient sign method and noise attacks. The proposed defense strategy effectively mitigates these attacks, offering forward-looking considerations for the security of artificial intelligence based control in ADNs.