Skip to content
MesaTEE GBDT-RS : a fast and secure GBDT library, supporting TEEs such as Intel SGX and ARM TrustZone
Rust C Python Makefile C++ Shell
Branch: master
Clone or download

Latest commit

Fetching latest commit…
Cannot retrieve the latest commit at this time.


Type Name Latest commit message Commit time
Failed to load latest commit information.
ci Fix coverage script Mar 30, 2020
dataset First Release May 15, 2019
examples Fix unused warning. May 15, 2019
sgx Fix rand-sgx dep Mar 30, 2020
xgb-data First Release May 15, 2019
.drone.yml Upgrade drone script for sgx 1.1.1 Mar 30, 2020
.gitignore First Release May 15, 2019
Cargo.toml Upgrade SGX SDK to v1.1.0 Dec 18, 2019
LICENSE First Release May 15, 2019 Update May 30, 2019
gbdt.pdf Add paper info and pdf May 16, 2019

MesaTEE GBDT-RS : a fast and secure GBDT library, supporting TEEs such as Intel SGX and ARM TrustZone

Build Status codecov

MesaTEE GBDT-RS is a gradient boost decision tree library written in Safe Rust. There is no unsafe rust code in the library.

MesaTEE GBDT-RS provides the training and inference capabilities. And it can use the models trained by xgboost to do inference tasks.

New! The MesaTEE GBDT-RS paper has been accepted by IEEE S&P'19!

Supported Task

Supppoted task for both training and inference

  1. Linear regression: use SquaredError and LAD loss types
  2. Binary classification (labeled with 1 and -1): use LogLikelyhood loss type

Compatibility with xgboost

At this time, MesaTEE GBDT-RS support to use model trained by xgboost to do inference. The model should be trained by xgboost with following configruation:

  1. booster: gbtree
  2. objective: "reg:linear", "reg:logistic", "binary:logistic", "binary:logitraw", "multi:softprob", "multi:softmax" or "rank:pairwise".

We have tested that MesaTEE GBDT-RS is compatible with xgboost 0.81 and 0.82

Quick Start

Training Steps

  1. Set configuration
  2. Load training data
  3. Train the model
  4. (optional) Save the model

Inference Steps

  1. Load the model
  2. Load the test data
  3. Inference the test data


    use gbdt::config::Config;
    use gbdt::decision_tree::{DataVec, PredVec};
    use gbdt::gradient_boost::GBDT;
    use gbdt::input::{InputFormat, load};

    let mut cfg = Config::new();

    // load data
    let train_file = "dataset/agaricus-lepiota/train.txt";
    let test_file = "dataset/agaricus-lepiota/test.txt";

    let mut input_format = InputFormat::csv_format();
    let mut train_dv: DataVec = load(train_file, input_format).expect("failed to load training data");
    let test_dv: DataVec = load(test_file, input_format).expect("failed to load test data");

    // train and save model
    let mut gbdt = GBDT::new(&cfg); train_dv);
    gbdt.save_model("gbdt.model").expect("failed to save the model");

    // load model and do inference
    let model = GBDT::load_model("gbdt.model").expect("failed to load the model");
    let predicted: PredVec = model.predict(&test_dv);

Example code

  • Linear regression: examples/
  • Binary classification: examples/

Use models trained by xgboost


  1. Use xgboost to train a model
  2. Use examples/ to convert the model
    • Usage: python xgboost_model_path objective output_path
    • Note depends on xgboost python libraries. The converted model can be used on machines without xgboost
  3. In rust code, call GBDT::load_from_xgboost(model_path, objective) to load the model
  4. Do inference
  5. (optional) Call GBDT::save_model to save the model to MesaTEE GBDT-RS native format.

Example code

  • "reg:linear": examples/
  • "reg:logistic": examples/
  • "binary:logistic": examples/
  • "binary:logitraw": examples/
  • "multi:softprob": examples/
  • "multi:softmax": examples/
  • "rank:pairwise": examples/



At this time, training in MesaTEE GBDT-RS is single-threaded.


The related inference functions are single-threaded. But they are thread-safe. We provide an inference example using multi threads in example/

SGX usage

Because MesaTEE GBDT-RS is written in pure rust, with the help of rust-sgx-sdk, it can be used in sgx enclave easily as:

gbdt_sgx = { git = "" }

This would import a crate named gbdt_sgx. If you prefer gbdt as normal:

gbdt = { package = "gbdt_sgx", git = "" }

For more information and concret examples, please look at directory sgx/gbdt-sgx-test.


Apache 2.0


Tianyi Li @n0b0dyCN

Tongxin Li @litongxin1991

Yu Ding @dingelish

Steering Committee

Tao Wei, Yulong Zhang


Thanks to @qiyiping for his/her great previous work gbdt. We read his/her code before starting this project.

You can’t perform that action at this time.