← AI security in plain English

AML.T0014

MITRE ATLAS / Discover AI Model Family

Adversaries determine the general family of a model, such as the architecture or the base model it was fine-tuned from, which narrows the set of known attacks likely to work against it.

Think of it likeIdentifying the make of a lock from its keyhole, because that tells you which picks to bring.

In plain English

Attacks transfer within a model family. Once the adversary knows what the system is built on, published research against that architecture becomes directly applicable.