Reproduce “Stealing Machine Learning Models via Prediction APIs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Florian Tramèr, Fan Zhang, Ari Juels, Michael K. Reiter, Thomas Ristenpart
The authors show how prediction APIs can be used to copy models’ input-to-output behavior, using confidence scores to recover parameters and adaptive queries, including incomplete inputs, to trace decision-tree paths. In experiments, the BigML German Credit tree was extracted exactly with 1,150 queries in 631 seconds. Label-only attacks are possible but require more queries and were evaluated on model classes.
Key insights
- Information-rich API outputs (high-precision confidence values, acceptance of partial feature vectors) significantly enhance the efficiency and fidelity of model extraction attacks across various model types.
- Successful model extraction converts black-box access into effective white-box access, substantially accelerating subsequent adversarial goals like model inversion attacks or the creation of evasion examples.
- The cost of extracting a model can be considerably lower than the cost of sustained legitimate usage, creating a strong economic incentive for adversaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.