This is really great research and interesting: (very roughly) how to compute a very small mask which, when applied to any image, makes the neural network misclassify it, whereas humans would notice no essential difference.
Quite remarkable.
Quite remarkable.