Hello, you are using an old browser that's unsafe and no longer supported. Please consider updating your browser to a newer version, or downloading a modern browser.
Training Camp • Cybersecurity Glossary
A model extraction attack steals an ML model by querying its API and training a substitute model that mimics its behavior, exposing IP and enabling further attacks.
Model Extraction Attack Definition: A model extraction attack steals an ML model by querying its API and training a substitute model that mimics its behavior, exposing IP and enabling further attacks.
A model extraction (or model stealing) attack is an adversarial machine-learning technique in which an attacker repeatedly queries a target model's prediction API and uses the inputs and outputs to train a substitute model that approximates the original's behavior. This steals intellectual property, exposes the victim to query-based monetary cost, and can serve as a stepping stone for crafting transferable adversarial examples or inferring training data. Defenses include rate limiting, query monitoring, output perturbation, and watermarking.
Model Extraction Attack is one of the topics you'll master in the AIGP Boot Camp.
AIGP Boot Camp →