AML.T0020
MITRE ATLAS / Training Data Poisoning
Adversaries manipulate the data used to train or fine-tune a model so that the resulting model behaves the way they want. Poisoning can degrade accuracy broadly or install a targeted backdoor that only activates on a chosen trigger.
In plain English
A model is the sum of what it was shown. Corrupt a small slice of the training set and you can make the model fail on demand, often without anyone noticing that overall accuracy barely moved.