New Method Infers Dataset Usage Without Shadow Models or Held-out Data

Determining the extent to which a specific dataset was used to train a machine learning model, known as Dataset Usage Inference (DUI), is a critical challenge. Existing DUI methods are often impractical because they demand expensive shadow models and access to both known training samples and a confirmed held-out dataset. These requirements are rarely met, especially for large, modern models and real-world data ownership disputes. A new, practical DUI framework has been developed to overcome these limitations. This method eliminates the need for shadow models and actual held-out data. Instead, it generates synthetic non-member samples, extracts various membership signals, and frames DUI as a mixture proportion estimation problem. Experiments with large image generative models demonstrate that this approach reliably quantifies dataset usage, providing a valuable tool for data owners to ascertain how their data was utilized in model training.