Fig. 2
From: Adversarial attack and defense in reinforcement learning-from AI security view
Examples for adversarial attacks on Pong policy trained with DQN(Huang et al. 2017). The first line: computing adversarial perturbations by fast gradient sign method (FGSM)(Goodfellow et al. 2014a) with an ℘∞-norm constraint. The trained agent who should have taken the “down” action took “noop” action instead under adversarial attack. The second line: authors utilized the FGSM with ℘1-norm constraint to compute the adversarial perturbations. The trained agent can not take action correctly, which should have moved up, but took “down” action after interference. Videos are available at http://r11.berkeley.edu/adversarial