An adversarial example is an input changed very slightly, often invisibly to people, so that an AI model gets it confidently wrong.
In plain words
Researchers found that adding a tiny, carefully chosen pattern of noise to an image can make a neural network label it as something else entirely. To a person the image looks unchanged. The same trick can also fool other models trained separately. Feeding such examples back in as training data, called adversarial training, can make a model more accurate.
A quick example
Take a photo that a model labels correctly. Add a faint layer of noise, computed to push the model the wrong way. The picture looks identical to you, but the model now names a different object, and is highly confident about it.
Where you'll meet it
- Testing how robust image classifiers are.
- Security research on attacks against AI systems.
Related
Sources
- PaperIntriguing properties of neural networks, Szegedy et al. (arXiv) · read 29 Sept 2026
- PaperExplaining and Harnessing Adversarial Examples, Goodfellow, Shlens and Szegedy (arXiv) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.