Gragonballer / SortingHat

Implementation of SortingHat: Efficient Private Decision Tree Evaluation via Homomorphic Encryption and Transciphering.

Private decision tree evaluation via Homomorphic Encryption and Transciphering

This is an implementation of the private decision tree evaluation (PDTE) algorithm from the paper SortingHat: Efficient Private Decision Tree Evaluation via Homomorphic Encryption and Transciphering, by Kelong Cong, Debajyoti Das, Jeongeun Park, and Hilder Vitor Lima Pereira, which was published in ACM CCS 2022.

WARNING:

This is proof-of-concept implementation. It may contain bugs and security issues. Please do not use in production systems.

Installation

The key dependencies of this project are Concrete version 1.0.0-beta and FINAL. For Concrete, cargo will take care of most of the installation except the FFTW dependency (see below). Also, we already include FINAL's code here. However, the user still need to manually install the dependencies of Concrete and FINAL, which are:

• FFTW -- See their installation instructions for details.
• GNU GMP.
• NTL.

For convinience, we prepared Bash scripts to download and install these three dependencies to /usr/local/bin. So, if you want to use them, you can simply run

./install_third_party_libs.sh

Running our PDTE

• cd src/rust_pdte
• Run tests: RUSTFLAGS="-C target-cpu=native" cargo test --release
• Run micro benchmark: RUSTFLAGS="-C target-cpu=native" cargo bench

Using the CLI

The CLI can be build using RUSTFLAGS="-C target-cpu=native" cargo build --release. cargo usually puts the executable in target/release/homdte.

Without any command line arguments, the CLI will simulate decision tree evaluation using complete binary trees of various depths. When given a command line argument of a data directory such as data/heart or data/spam, the CLI will evaluation a real model trained using the training script located in script/train.py. Some models are included in the data directory. The detailed options are available from ./homdte --help.

For convenience, a script is given under script/run_all_datasets.sh to run private decision tree evaluation on all datasets.

Optional

A script exists under script/train.py which perform the training using Concrete ML version 0.2.0. This is only necessary if new models need to be trained. Running the script is not necessary to use the CLI (described above).

Running our PDTE with transciphering

• cd src/cpp_pdte_transciphering
• make
• ./test_pdte_transciphering