Skip to content

wj23027/FedTree

 
 

Repository files navigation

Documentation

Overview

FedTree is a federated learning system for tree-based models. It is designed to be highly efficient, effective, and secure. It has the following features currently.

  • Federated training of gradient boosting decision trees.
  • Parallel computing on multi-core CPUs and GPUs.
  • Supporting homomorphic encryption, secure aggregation and differential privacy.
  • Supporting classification and regression.

The overall architecture of FedTree is shown below. FedTree_archi

Getting Started

You can refer to our primary documentation here.

Prerequisites

  • CMake 3.15 or above
  • GMP
  • NTL
  • gRPC 1.50.0 (required for distributed version)

You can follow the following commands to install NTL library.

wget https://libntl.org/ntl-11.5.1.tar.gz
tar -xvf ntl-11.5.1.tar.gz
cd ntl-11.5.1/src
./configure SHARED=on
make
make check
sudo make install

If you install the NTL library at another location, please pass the location to the NTL_PATH when building the library (e.g., cmake .. -DNTL_PATH="PATH_TO_NTL").

For gRPC, please remember to add the local bin folder to your path variable after installation, e.g.,

export PATH="$gRPC_INSTALL_DIR/bin:$PATH"

If your gRPC version is not 1.50.0, you need to go to src/FedTree/grpc directory and run

protoc -I ./ --grpc_out=. --plugin=protoc-gen-grpc=`which grpc_cpp_plugin` ./fedtree.proto
protoc -I ./ --cpp_out=. ./fedtree.proto

Clone and Install submodules

git clone https://github.com/Xtra-Computing/FedTree.git
cd FedTree
git submodule init
git submodule update

Standalone Simulation

Build on Linux

# under the directory of FedTree
mkdir build && cd build 
cmake .. -DDISTRIBUTED=OFF
make -j

Build on MacOS

Build with Apple Clang

You need to install libomp for MacOS.

brew install libomp

Install FedTree:

# under the directory of FedTree
mkdir build
cd build
cmake -DOpenMP_C_FLAGS="-Xpreprocessor -fopenmp -I/usr/local/opt/libomp/include" \
  -DOpenMP_C_LIB_NAMES=omp \
  -DOpenMP_CXX_FLAGS="-Xpreprocessor -fopenmp -I/usr/local/opt/libomp/include" \
  -DOpenMP_CXX_LIB_NAMES=omp \
  -DOpenMP_omp_LIBRARY=/usr/local/opt/libomp/lib/libomp.dylib \
  ..
make -j

Run training

# under 'FedTree' directory
./build/bin/FedTree-train ./examples/vertical_example.conf

Distributed Setting

For each machine that participates in FL, it needs to build the library first.

mkdir build && cd build
cmake .. -DDISTRIBUTED=ON
make -j

Then, write your configuration file where you should specify the ip address of the server machine (ip_address=xxx). Run FedTree-distributed-server in the server machine and run FedTree-distributed-party in the party machines. Here are two examples for horizontal FedTree and vertical FedTree.

Distributed Horizontal FedTree

# under 'FedTree' directory
# under server machine
./build/bin/FedTree-distributed-server ./examples/adult/a9a_horizontal_server.conf
# under party machine 0
./build/bin/FedTree-distributed-party ./examples/adult/a9a_horizontal_p0.conf 0
# under party machine 1
./build/bin/FedTree-distributed-party ./examples/adult/a9a_horizontal_p1.conf 1

Distributed Vertical FedTree

# under 'FedTree' directory
# under server (i.e., the party with label) machine 0
./build/bin/FedTree-distributed-server ./examples/credit/credit_vertical_p0_withlabel.conf
# open a new terminal
./build/bin/FedTree-distributed-party ./examples/credit/credit_vertical_p0_withlabel.conf 0
# Under party machine 1
./build/bin/FedTree-distributed-party ./examples/credit/credit_vertical_p1.conf 1

Other information

FedTree is built based on ThunderGBM, which is a fast GBDTs and Radom Forests training system on GPUs.

Citation

Please cite our paper if you use FedTree in your work.

@misc{fedtree,
  title = {FedTree: A Fast, Effective, and Secure Tree-based Federated Learning System},
  author={Li, Qinbin and Cai, Yanzheng and Han, Yuxuan and Yung, Ching Man and Fu, Tianyuan and He, Bingsheng},
  howpublished = {\url{https://github.com/Xtra-Computing/FedTree/blob/main/FedTree_draft_paper.pdf}},
  year={2022}
}

Releases

No releases published

Packages

No packages published

Languages

  • C++ 94.6%
  • Cuda 2.6%
  • Python 2.0%
  • CMake 0.8%