← AI Terminology

Model Stealing

Model stealing (extraction) attacks query a victim model to train a substitute that imitates its predictions or functionality, threatening IP and safety controls.

A core ML security risk for APIs.
Why It Matters in AI
APIs expose behaviour that can be distilled. Attackers clone capabilities or probe safety edges. Defences (rate limits, output perturbation, watermarking, legal terms) matter for commercial AI.
Key Points
Aspect Description
Eval Agreement fidelity of stolen model
Attack Query → labeled set → train clone
Impact IP theft; safety bypass via local clone
Related Knowledge distillation, membership inference
Targets Cloud ML APIs, open endpoints
Defences Auth, throttling, detection, watermarking
Simple Analogy
Interviewing a secretive expert with clever questions until you can hire a stand-in who gives the same answers.
Common Usage Examples
  • Paper demos of extraction on vision APIs
  • Monitor anomalous query patterns
  • Watermark proprietary model outputs
  • Legal ToS + technical rate limits
Summary
In short: Model stealing extracts a surrogate via queries — an IP and security threat against hosted AI APIs.