XAI-Guided Adversarial Attacks on DRL in UAV-Enabled MEC

Explainable AI (XAI) enhances transparency in AI-driven applications but can also be exploited for malicious purposes. This paper investigates adversarial attacks that leverage XAI techniques to undermine Deep Reinforcement Learning (DRL)-based service provisioning in Unmanned Aerial Vehicle (UAV)-enabled Multi-access Edge Computing (MEC). Focusing on latency-sensitive services in dynamic, resource-constrained UAV networks, we propose two XAI-guided attacks: Salient Nodes Overload (SNO) and Subtle Nodes Drift (SND). These attacks exploit DRL explanations to degrade performance and mislead the agent. Experiments on UAV-enabled MEC frameworks using Double-Dueling Deep Q-Network (D3QN) and Deep Q-Network (DQN) show SNO causes immediate degradation, reducing job completion by about 40% in D3QN and 20% in DQN. The stealthier SND causes a delayed decline, as D3QN’s resilience defers deterioration for nearly twice as many episodes, complicating early detection. Both attacks reduce UAV hovering time, with SND causing up to 30% greater loss in D3QN.