Model extraction
Also known as: model stealing · API extraction
Model extraction is an attack where an adversary queries a model (usually via an API) enough times to reconstruct a close copy of its behavior or weights.
- Query the API, learn the input→output mapping.
- Train a surrogate to mimic it.
- The goal: a “good enough” copy of the model you paid for.
The defense: rate limiting, output quantization, and anomaly detection on query patterns all raise the cost of extraction. Assume an API is a surface to be probed.