Translational symmetry in convolutions with localized kernels causes an implicit bias toward high frequency adversarial examples.

Using Learning Dynamics to Explore the Role of Implicit Regularization in Adversarial Examples.