Glossary
Model Extraction
Model extraction reconstructs a model's underlying behavior — approximating what it would output for a given input — by systematically querying its inference endpoint and observing the pattern of responses, rather than obtaining the model's own parameters directly. Exposing a model as a queryable service, rather than distributing the model itself, was often assumed to protect the model — extraction shows that assumption is only partial: enough queries can approximate the model's behavior closely without ever touching its actual parameters.
Defending against this looks less like a standard access-control problem and more like the rate-limiting and abuse-detection problem applied to a specific goal: bounding how many queries a given caller can make, watching for query patterns characteristic of systematic extraction rather than normal use, and treating an inference endpoint as a resource whose total exposure compounds with cumulative access, not just a single request evaluated in isolation.