Fig. 7
From: Adversarial attack and defense in reinforcement learning-from AI security view
Attacking flow for enchanting attack (Lin et al. 2017). Enchanting attack from the original state st, the whole processing flow can be concluded as follow: 1) action sequence planning; 2) generating adversarial examples with target actions; 3) agent takes actions under adversarial example; 4) environment gives the next sate st+1. Meanwhile, adversary utilizes the prediction model to attack the target agent with initial state st