Stealing machine learning models via prediction APIs¶
Authors: Florian Tramèr; Fan Zhang; Ari Juels; Michael K. Reiter; Thomas Ristenpart
Year: 2016
Venue: arXiv preprint
type: attack collection: sok-references status: auto
Links: arXiv:1609.02943
Summary¶
Auto-extracted summary (unreviewed)
The paper studies model extraction attacks in which a black-box adversary with only query access to an ML-as-a-service prediction API reconstructs a functionally equivalent model, exploiting API features such as confidence scores and partial-feature queries. It presents an equation-solving attack for models with a logistic output layer and a path-finding attack for decision trees, extracts logistic regression, neural network and decision tree models with near-perfect fidelity, demonstrates the attacks against BigML and Amazon Machine Learning, and shows that withholding confidence values still leaves models vulnerable to less effective retraining attacks.
Abstract¶
Machine learning (ML) models may be deemed confidential due to their sensitive training data, commercial value, or use in security applications. Increasingly often, confidential ML models are being deployed with publicly accessible query interfaces. ML-as-a-service ("predictive analytics") systems are an example: Some allow users to train models on potentially sensitive data and charge others for access on a pay-per-query basis. The tension between model confidentiality and public access motivates our investigation of model extraction attacks. In such attacks, an adversary with black-box access, but no prior knowledge of an ML model's parameters or training data, aims to duplicate the functionality of (i.e., "steal") the model. Unlike in classical learning theory settings, ML-as-a-service offerings may accept partial feature vectors as inputs and include confidence values with predictions. Given these practices, we show simple, efficient attacks that extract target ML models with near-perfect fidelity for popular model classes including logistic regression, neural networks, and decision trees. We demonstrate these attacks against the online services of BigML and Amazon Machine Learning. We further show that the natural countermeasure of omitting confidence values from model outputs still admits potentially harmful model extraction attacks. Our results highlight the need for careful ML model deployment and new model extraction countermeasures.
Source: arxiv
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | equation-solving model extraction attack for models with a logistic output layer | llm | 0.85 |
| contribution | path-finding extraction algorithm for decision trees | llm | 0.85 |
| contribution | demonstration of attacks against live BigML and Amazon Machine Learning services | llm | 0.85 |
| contribution | analysis of countermeasures: label-only APIs still admit retraining attacks | llm | 0.85 |
| models | logistic regression | llm | 0.85 |
| models | neural networks | llm | 0.85 |
| models | decision trees | llm | 0.85 |
| tasks | model extraction / model stealing | llm | 0.85 |
| tasks | ML-as-a-service prediction APIs | llm | 0.85 |
| threat model | black-box adversary with query access to a prediction API, no knowledge of model parameters or training data | llm | 0.85 |
| code url | https://github.com/ftramer/Steal-ML (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/ftramer/Steal-ML | llm | 0.85 |
| limitations | label-only retraining attacks are less effective than attacks using confidence values | llm | 0.85 |
| limitations | evaluation of refined countermeasures on production ML services left for future work | llm | 0.85 |
| key result | extraction fidelity: near-perfect (close to 100% agreement) (logistic regression, neural networks and decision trees via BigML and Amazon ML APIs) | llm | 0.85 |
SoK reference [52]¶
F. Tram`er, F. Zhang, A. Juels, M. K. Reiter, and T. Ristenpart, “Stealing machine learning models via prediction APIs,” in USENIX Security 2016: 25th USENIX Security Symposium, T. Holz and S. Savage, Eds. Austin, TX, USA: USENIX Association, Aug. 10–12, 2016, pp. 601– 618. [Online]. Available: https://www.usenix.org/conference/ usenixsecurity16/technical-sessions/presentation/tramer
Listed on the reference page of the site.
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: Differential privacy 6 [p.14], Leveled / somewhat HE 2 [p.4]
- model: SVM 35 [p.1], Decision tree 27 [p.1], Logistic regression 20 [p.1], MLP 16 [p.6], Linear regression 1 [p.2]
- dataset: Credit card / German credit 5 [p.2], Iris 4 [p.5], UCI repository 4 [p.5], Breast cancer (Wisconsin) 2 [p.5], Diabetes (Pima) 2 [p.5]
- hardware: GPU 1 [p.6]
- technique: Model extraction / stealing 34 [p.1], Softmax 23 [p.6], Model inversion 6 [p.4]
- task: Medical / healthcare 4 [p.5], Anomaly detection 2 [p.4]
- topic: Benchmark 2 [p.14], Open source 2 [p.1]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/sok-references/052_stealing-machine-learning-models-via-prediction-apis.pdf |
sok-references | 19 | 1.9 MB | 0140d9ab6257 |
Database key: stealing-machine-learning-models-via-prediction-apis · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set stealing-machine-learning-models-via-prediction-apis COLUMN VALUE