Adversarial Attack

Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary

A technique that makes small, deliberate changes to input data to push an AI model into wrong predictions – often changes too subtle for a person to notice.

What is an Adversarial Attack?

An adversarial attack deceives a machine learning model by making small, deliberate changes to its input. The changes are often too subtle for a person to notice, yet they push the model into a confidently wrong prediction.

The classic demonstrations come from image recognition, where adjusting a handful of pixels is enough to make a classifier mislabel an image, but the same idea applies to text, audio, and any other input a model consumes.

Types of Adversarial Attacks

  1. Evasion attacks change the input at prediction time so the model gets it wrong, such as altering a stop sign so an autonomous car’s vision system reads it as a yield sign.
  2. Poisoning attacks corrupt the training data itself, planting weaknesses during model training that an attacker can exploit after the model is deployed.
  3. Model extraction attacks query a model repeatedly and use its answers to reconstruct a working copy, stealing the model’s behavior without ever seeing its weights.

Applications of Adversarial Attacks

Example of an Adversarial Attack

The best-known example comes from a 2014 paper by Ian Goodfellow and colleagues. They took a photo that a classifier correctly labeled as a panda, added a layer of noise invisible to the human eye, and the classifier labeled the result a gibbon – with 99% confidence.

The image looks identical to a person. That gap between what humans see and what models compute is exactly what adversarial attacks exploit, and why models deployed in security-sensitive settings need defenses against them.

Related AI terms: Robustness · Security · Model Training · Image Recognition · Facial Recognition

Did you like the Adversarial Attack gist?

Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.

Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About

Read the Governor's Letter

Stay ahead with Governor's Letter, the newsletter delivering expert insights, AI updates, and curated knowledge directly to your inbox.

By subscribing to the Governor's Letter, you consent to receive emails from AI Guv.
We respect your privacy - read our Privacy Policy to learn how we protect your information.

Browse All AI Terms A–Z

Every term in the dictionary, in alphabetical order. Jump to a letter or scroll the full list.

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

X

Y

Z