AML.T0015
MITRE ATLAS / Evade AI Model
Adversaries craft adversarial data or generate deepfakes so that a model misclassifies input or accepts fabricated content as authentic. Evasion can be used to slip past AI-based detection or to gain access to a system that gates on model output.
In plain English
When a model is the gate, fooling the model is the way in. Small, deliberate changes to an input can flip its verdict while a person looking at the same input notices nothing odd.