Fig. 4
From: Adversarial attack and defense in reinforcement learning-from AI security view
The first line shows dominant adversarial examples for the original map. The fist picture denotes the original map for attack, and the three columns on the right are the dominant adversarial examples of successful attacks. Meanwhile, the red dotted lines represent the perturbation band. The second line denotes the direction in which the value gradient rises the fastest. By comparison between dominant adversarial examples and the contour graph, can found that on the perturbation band, the value gradient rises fastest