The Limits of Provable Security Against Model Extraction

ePrint Report: The Limits of Provable Security Against Model Extraction

Ari Karchmer

Can we hope to provide provable security against model extraction attacks? As a step towards a theoretical study of this question, we unify and abstract a wide range of "observational" model extraction defense mechanisms -- roughly, those that attempt to detect model extraction using a statistical analysis conducted on the distribution over the adversary's queries. To accompany the abstract observational model extraction defense, which we call OMED for …

eprint report provable security security