Model Extraction

Model extraction involves creating a functional copy of a machine learning model, typically accessed via an API, by querying it with strategically chosen inputs. Current research focuses on improving the efficiency and accuracy of extraction techniques, particularly in data-free scenarios, using methods like generative models and derivative-free optimization to create surrogate models. This area is crucial for understanding and mitigating security risks in machine learning-as-a-service platforms and protecting intellectual property in deployed models, with ongoing work exploring both improved attack methods and novel defenses like watermarking and query unlearning.