Definition
Plain language
An input that has been tweaked just enough to make a machine give the wrong answer, while looking normal to a person.
As stated in the literature
An input crafted, usually via gradient-based optimization, to lie on the wrong side of a model's decision boundary despite being imperceptibly or semantically unchanged to humans.
Also called: adversarial examples
Why it matters: It shows that a system can be highly accurate on ordinary inputs and still be steered to a wrong answer by someone who knows how to nudge it, which matters wherever the input comes from an untrusted source.
For example, changing a handful of pixels in a photo of a panda can make an image classifier confidently call it a gibbon, even though the picture still looks like a panda to us.