ALIOTH: An Efficient and Secure Weight-of-Evidence Framework for Privacy-Preserving Data Processing

Introduction

ALIOTH is a two-party framework for WoE-based dataset transformation for horizontal and vertical data partitions. The eprint version is available on Cryptology-ePrint ALIOTH.

This repository is built upon SecretFlow-SPU with optimizations from OpenBumbleBee, which provides a prototype implementation of ALIOTH, intended solely for experimental testing purposes and SHOULD NOT be used in production environments.

Build

TL;DR

git clone https://github.com/AntCPLab/OpenBumbleBee.git
cd OpenBumbleBee
python setup.py bdist_wheel

# modify psi/bazel/repositories.bzl line 129:143
# def _com_github_intel_ipp():
#     maybe(
#         http_archive,
#         name = "com_github_intel_ipp",
#         sha256 = "d70f42832337775edb022ca8ac1ac418f272e791ec147778ef7942aede414cdc",
#         strip_prefix = "cryptography-primitives-ippcp_2021.8",
#         build_file = "@psi//bazel:ipp.BUILD",
#         patch_args = ["-p1"],
#         patches = [
#             "@psi//bazel:patches/ippcp.patch",
#         ],
#         urls = [
#             "https://github.com/intel/ipp-crypto/archive/refs/tags/ippcp_2021.8.tar.gz",
#         ],
#     )

pip install dist/*.whl --force-reinstall

WOE Example

1. Launch SPU backend runtime

```sh
# OpenBumbleBee
bazel run -c opt //examples/python/utils:nodectl --jobs 16 -- --config `pwd`/examples/python/conf/2pc_alioth.json up
# alioth-impl
python utils/nodectl.py --config `pwd`/conf/2pc_alioth.json up
```

2. Run woe example

```sh
# OpenBumbleBee
bazel run -c opt //examples/python/stats:woe --jobs 16 -- --config `pwd`/examples/python/conf/2pc_alioth.json
# alioth-impl
python woe.py -c `pwd`/conf/2pc_alioth.json
```

3. Experiments

```sh
python -u exp.py -s woe -m vp hp_cat hp_num -t 1 -H 1000 10000 100000 -W 10 50 100 -K 5 10 20 2>&1 | tee output/LANLOG
python -u exp.py -s woe -m mbm_appqua -t 1 -H 1000 10000 100000 -W 1 2>&1 | tee output/LANLOG
python -u exp.py -s woe -m mbm_appqua -t 1 -H 10000 -W 1 -K 100 1000 10000 2>&1 | tee output/LANLOG
python -u exp.py -s woe -m mbm_transformation mbm_naive_transformation -t 1 -H 10000 -W 100 -K 5 10 20 2>&1 | tee output/LANLOG

#### GCD, VP
python -u exp.py -s woe -m vp -t 1 -H 800 -W 20 -K 5 2>&1 | tee output/LANLOG_GCD_VP

#### GCD, HP
python -u exp.py -s woe -m hp_cat -t 1 -H 800 -W 13 -K 5 2>&1 | tee output/LANLOG_GCD_HP_CAT
python -u exp.py -s woe -m hp_num -t 1 -H 800 -W 7 -K 5 2>&1 | tee output/LANLOG_GCD_HP_NUM

#### HCDR, VP
python -u exp.py -s woe -m vp -t 1 -H 307511 -W 51 -K 5 2>&1 | tee output/LANLOG_HCDR_VP_CAT
python -u exp.py -s woe -m vp -t 1 -H 307511 -W 69 -K 10 2>&1 | tee output/LANLOG_HCDR_VP_NUM

#### HCDR, HP
python -u exp.py -s woe -m hp_cat -t 1 -H 307511 -W 51 -K 5 2>&1 | tee output/LANLOG_HCDR_HP_CAT
python -u exp.py -s woe -m hp_num -t 1 -H 307511 -W 69 -K 10 2>&1 | tee output/LANLOG_HCDR_HP_NUM
python -u exp.py -s woe -m iv_gcd_vp iv_gcd_hp iv_hcdr_vp iv_hcdr_hp 2>&1 | tee output/LAN_LOG_IV

#### HCDR, VP
python -u exp.py -s woe -m vp -t 1 -H 307511 -W 69 -K 10 2>&1 | tee output/WANLOG_HCDR_VP_NUM

#### HCDR, HP
python -u exp.py -s woe -m hp_num -t 1 -H 307511 -W 69 -K 10 2>&1 | tee output/WANLOG_HCDR_HP_NUM

#### IV
python -u exp.py -s woe -m iv_gcd_vp iv_gcd_hp iv_hcdr_vp iv_hcdr_hp 2>&1 | tee output/WANLOG_IV

#### WOE HP_NUM
python -u exp.py -s woe -m hp_num -t 1 -H 1000 10000 100000 -W 10 50 100 -K 5 10 20 2>&1 | tee output/WANLOG_HP_NUM

#### APPROXIMATE QUANTILES
python -u exp.py -s woe -m mbm_appqua -t 1 -H 1000 10000 100000 -W 1 2>&1 | tee output/WANLOG_APPQUA

#### LR regression training with WoE
python lr.py -m gcd_woe | tee LOG_GCD_WOE
python lr.py -m gcd_normal | tee LOG_GCD_NORMAL
python lr.py -m hcdr_woe | tee LOG_HCDR_WOE
python lr.py -m hcdr_normal | tee LOG_HCDR_NORMAL
```

Citing

@misc{cryptoeprint:2025/2188,
      author = {Ye Dong and Xiangfu Song and W.j Lu and Xudong Chen and Yaxi Yang and Ruonan Chen and Tianwei Zhang and Jin-Song Dong},
      title = {{ALIOTH}: An Efficient and Secure Weight-of-Evidence Framework for Privacy-Preserving Data Processing},
      howpublished = {Cryptology {ePrint} Archive, Paper 2025/2188},
      year = {2025},
      url = {https://eprint.iacr.org/2025/2188}
}