← AI security in plain English

AML.T0059

MITRE ATLAS / Erode Dataset Integrity

Adversaries poison or manipulate portions of a dataset to reduce its usefulness and the trust placed in it, forcing the owner to spend effort finding and correcting errors.

Think of it likeSalting a few pages of a reference work so every future reader has to check all of it.

In plain English

Corrupting part of a dataset devalues the whole thing, because nobody can tell which records are clean. The cost lands on the owner as re-verification work, not on the attacker.