NeurIPS 2020

Backpropagating Linearly Improves Transferability of Adversarial Examples

Meta Review

The paper proposes to improve the transferability of adversarial examples by using linear approximations of DNNs. Although there are still some requests from reviewers to improve the presentation (please take their advice into account, it will increase the impact of your paper), there is a consensus that the experimental evaluations of the method are complete and sufficiently robust to convince of the soundness of the methods and its working hypotheses.