Federated Membership Inference on Federboost

This package implements 'Federboost' from Tian et al Which is a horizontally federated Gradient Boosted Decision Tree algorithm. With the regularisation terms of XGBoost. This is not a 1 to 1 conversion from XGBoost to python, the models perform remarkably close when regularisation parameters are used, but they differ more when it is not used.

This is a multi-class classification implementation that uses softmax. Also different types of Membership Inference Attack can be tested, both using federated information or without.

Get the paper here

TODO:

• Clear up for other researchers
  • code cleanup & explanation
• Change naming from old SFXGBoost to FXGBoost

---

Usage: install conda environment:

conda env create -f environment.yml

Activate your environment:

conda activate fedxgboost_mpi

build and run the experiments using:

pip install . && mpiexec -np 3 python tests/main.py (experiment number 1 to 3)

here -np 3 implies 1 server and two participants in the federated network. for this you will have to change the POSSIBLE_PATHS parameter for the healthcare dataset.

or use "run.sh":

./run.sh myexperimentname (experiment number 1 to 3)

---

Tian, Z., Zhang, R., Hou, X., Liu, J., & Ren, K. (2020). Federboost: Private federated learning for gbdt. arXiv preprint arXiv:2011.02796.

---

Made in collaboration with the RIVM (Dutch National Institute for Public Health and the Environment) for my master thesis at the University of Twente.

check https://jaap-meerhof.github.io for my contact information