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.
You can refer to our primary documentation here.
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
git clone https://github.com/Xtra-Computing/FedTree.git
cd FedTree
git submodule init
git submodule update
# under the directory of FedTree
mkdir build && cd build
cmake .. -DDISTRIBUTED=OFF
make -j
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
# under 'FedTree' directory
./build/bin/FedTree-train ./examples/vertical_example.conf
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.
# 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
# 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
FedTree is built based on ThunderGBM, which is a fast GBDTs and Radom Forests training system on GPUs.
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}
}