The paper proposes using the backward pass of backprop algorithm to compute the adversarial example. This leads to computational time saving because instead of two backward passes. Overall, a good paper. The rebuttal was generally satisfying. It's encouraged to further consider the comments in the final version.