Glossary

Model extraction

Building a functionally similar copy of a model, or recovering its parameters, by querying it at scale and training on the answers (MITRE ATLAS AML.T0024.002). Rate limits, query monitoring and returning less detail, such as no raw probabilities, make it costlier.

Also called: model stealing
Defined in: MITRE ATLAS

Covered in

Guide

Assess extraction, membership and inference risks

Define what an adversary can observe and measure privacy leakage without confusing a model’s output with proof about its training data.