D5 · Seeing
A Panda, One Nudge, a Gibbon
Build an adversarial example, explain why it works, and say why defenses are hard.
In 2013 a team including Christian Szegedy and Ian Goodfellow found something nobody was looking for. Add a tiny, carefully chosen pattern to a photo a network gets right, and it changes its answer, with confidence. To a person, the two photos look the same.