Privacy Raven Has Left the Nest

Summary: Trail of Bits released PrivacyRaven, a Python library that simulates privacy attacks on deep learning systems, covering model extraction, membership inference, and model inversion. The tool operates under a label-only black-box threat model, where the adversary receives only labels from an API querying the model, and model inversion is still in incubation.
What you should know: The source notes that even a medical diagnosis model returning only a terse "Yes" or "No" can be attacked to compromise training data confidentiality, and that weaker threat models exposing confidence scores or model explanations substantially increase attack effectiveness.

Read the original at Trail of Bits Blog ↗