What is Model Extraction?
Model extraction aims to reconstruct, approximate, or replicate a model’s functionality from its inputs and outputs. The target can be behavior, decision boundaries, or a substitute model. Unlike Model Inversion, it primarily concerns exposed capability and intellectual property, not information about training data.
Model extraction security risks should be assessed in the context of the system, data flows, identities, integrations, and decision consequences. A precise boundary helps owners evaluate the exposure without overstating what one control can achieve.
What is Model Extraction used for?
It can create a competing substitute, evade service intent, or drive costly query activity. Use authentication, authorization, rate limits, anomaly monitoring, and response minimization where appropriate. Assess the value of the exposed capability and accessibility. Rate limiting changes cost and observability; it does not independently prevent extraction.
