A single gradient step finds adversarial examples on random two-layers neural networks

Open in new window