The context: One of the greatest unsolved flaws of deep learning is its vulnerability to so-called adversarial attacks. When added to the input of an AI system, these perturbations, seemingly random or undetectable to the human eye, can make things go completely awry.
