This repo is a prototype implementation of privacy-preserving XGBoost (https://xgboost.readthedocs.io/en/latest/). We adopt several property-preserving encryption schemes to encrypt the XGBoost model so that the privacy-preserving model can predict an encrypted query.
An extended abstract of this work (https://arxiv.org/abs/2011.04789) appears in Privacy-preserving Machine Learning Workshop at NeurIPS 2020.
This package requires python>=3.8. Install the dependencies with
- python -m pip install .
This command installs the dependencies in a specific order.
Run the tests with:
- cd test
- python -m pytest
Install dependencies required to run the examples with:
- python -m pip install .[examples] Run the examples with:
- cd examples
- jupyter nbconvert --to notebook --execute foo.ipynb --output foo.ipynb You can clear the notebook output with
- jupyter nbconvert --clear-output --ClearMetadataPreprocessor.enabled=True foo.ipynb
This package depends on the Paillier partially homomorphic encryption scheme (https://en.wikipedia.org/wiki/Paillier_cryptosystem). It also includes source code for a modified version of Boldyreva et. al.'s order-preserving encryption scheme (https://github.com/tonyo/pyope). The source code is place in the 'third-party/ope/' directory.
See DEVELOPMENT.md
See CONTRIBUTING for more information.
This project is licensed under the Apache-2.0 License.