HETAL: Efficient Privacy-preserving Transfer Learning with Homomorphic Encryption

Official repository for HETAL: Efficient Privacy-preserving Transfer Learning with Homomorphic Encryption (ICML'23 Oral) by Seewoo Lee^1*, Garam Lee², Jung Woo Kim², Junbum Shin², and Mun-Kyu Lee^3**.

¹University of California, Berkeley ²CryptoLab ³Inha University

* Work done at CryptoLab

** Corresponding author

You can read the paper here: OpenReview Link

This repository uses CPU version of HEaaN library. To use GPU acceleration, please contact stat@cryptolab.co.kr.

Installation

Requirements

• OS: Linux
• Python: 3.8
• Recommended Memory: 32GB (Minimum : 16GB)

First install pipenv. Then run the following shell commands. All the required packages including heaan and heaan_sdk will be installed using existing whl files (heaan_sdk-0.2.0-cp38-cp38-linux_x86_64.whl and heaan-0.1.0+cpu-cp38-cp38-linux_x86_64.whl).

pipenv --python 3.8
pipenv shell
pipenv install

How to use

See README.md files in src/hetal and src/benchmark directories for guides.

Citation

Use the following Bibtex entry for citation.

@article{lee2023hetal,
  title={HETAL: Efficient Privacy-preserving Transfer Learning with Homomorphic Encryption},
  author={Lee, Seewoo and Lee, Garam and Kim, Jung Woo and Shin, Junbum and Lee, Mun-Kyu},
  journal={ICML},
  year={2023}
}

License

This is available for the non-commercial purpose only. See LICENSE for details.

HEaaN.SDK

HETAL will be integrated to CryptoLab's HEaaN.SDK (HEaaN.stat and HEaaN.ml), the HE-based data analysis library. Current version and its documentation is avaiable in heaan.it.