Model inversion
Also known as: model inversion attack
Model inversion is an attack that reconstructs sensitive training inputs from a model’s behavior.
- The model “memorizes” some data; inversion coaxes it out.
- A route from weights back to private inputs.
- A reason to audit memorization before release.
The exposure: a model trained on sensitive data can, in principle, give some of it back. If your training data is sensitive, assume memorization is possible and audit accordingly.