Trust, Risk & Safety advanced

Adversarial Example

An input perturbed just enough to fool a model, while looking unchanged to a person.

Adversarial examples exploit the fact that a model’s decision boundaries do not match human perception. Tiny, targeted changes flip the output with high confidence. They are a reminder that high benchmark accuracy says nothing about behaviour under attack.

In practice: A stop sign with a few stickers that a vision model reads as a speed limit sign.