Definition
Model extraction is a type of adversarial attack against machine learning models where an attacker attempts to steal the model's parameters or reconstruct its functionality. This is achieved by querying the target model with carefully crafted inputs and observing the outputs to infer its internal workings. Understanding model extraction is crucial for protecting proprietary AI assets and ensuring the integrity of AI driven systems. For instance, a malicious actor might probe a fraud detection model to learn how it identifies suspicious transactions, enabling them to bypass future checks. This differs from data poisoning, which aims to corrupt the training data rather than steal the model itself. Knowledge of model extraction is relevant for cybersecurity professionals and appears in advanced security certifications.