AML.T0043

MITRE ATLAS / Craft Adversarial Data

Adversarial data are inputs modified so that they cause the adversary's intended effect in the target model, such as misclassification or a specific chosen output, while often appearing unremarkable to a person.

Think of it likeAltering a signature by a hairsbreadth so the machine reads it as someone else and the eye sees no difference.

In plain English

Small, deliberate perturbations can flip a model's answer. The input still looks normal to a reviewer, which is what makes this hard to catch by inspection.