Small crafted perturbations that fool models.
What is an adversarial attack on a machine learning model?