In our CVPR'19 paper, we study the relationship between adversarial robustness and generalization in the context of the underlying data manifold. Here, we explicitly distinguish between regular adversarial examples (i.e., unconstrained) and adversarial examples constrained to the manifold, so-called on-manifold adversarial examples. For those who did not attend CVPR'19, or missed our poster, it can be downloaded below; a smaller version of the poster was also presented at the ICML'19 Workshop on Uncertainty and Robustness in Deep Learning (UDL).

CVPR'19 PosterICML'19 UDL Poster