Adversarial Example

What is an adversarial example?

An adversarial example is an input deliberately tweaked to fool an AI. The classic case: a photo of a panda gets a sprinkle of nearly invisible noise. To a human it's still a panda. To the AI, it's confidently a "gibbon." A tiny change in the input flips the model's answer completely.

Why does it fool the AI?

AI sees images by hunting for features in the pixels — and those features aren't the same as the edges and colors a human notices. An attacker just nudges the pixels along the model's weak spots, making it overconfident about the wrong features. The result is an answer that makes no sense, and a reminder that the model's judgment is brittle.

Where do adversarial examples show up?

Image recognition
Making a self-driving car read a "stop" sign as a "speed limit" sign — potentially deadly in the real world.
Speech recognition
Hiding commands in audio that human ears can't hear, so a voice assistant executes them.
Text
Using specific wording or characters to push a model into wrong or harmful output.

Why does it matter?

Adversarial examples show that AI's capability and reliability are two different things. A model that's 99% right on a test set can be beaten by 1% of clever noise. Studying them helps us harden AI — and think through the risks before it lands in self-driving cars, medicine and security.

Bottom line: an adversarial example is a drug for AI — invisible to humans, but enough to make the model get it totally wrong.