minseok-ryu / dpfl-iadmm-classification

Privacy-preserving algorithms for a distributed empirical risk minimization model with a multiclass logistic regression function

Differentially Private Inexact ADMM for a Federated Learning model

In this open-source code, we implement inexact differentially private alternating direction method of multipliers (DP-IADMM) algorithms for solving a distributed empirical risk minimization problem with the multi-class logistic regression function. In specific, the following three algorithms are implemented:

• "ObjT": DP-IADMM with a trust-region of solutions incorporated with the objective perturbation method.
• "ObjP": DP-IADMM with a proximal function incorporated with the objective perturbation method.
• "OutP": DP-IADMM with a proximal function incorporated with the output perturbation method.

Install and Run

git clone https://github.com/minseok-ryu/DP-IADMM-Multiclass-Logistic.git

After downloading the code, open the terminal and go to the directory where "run_1.py" exists.

1. Do the followings:

conda create -n DPFL	
conda activate DPFL	
conda install numpy	
conda install cupy
pip install GPUInfo
pip install mlxtend

Packages cupy and GPUInfo are required for GPU computation. mlxtend is required for MNIST dataset.

2. Run:

python run_1.py

3. Go to "Outputs" directory to see the results

Important Note

To use GPU, change from "import numpy as np" to "import cupy as np" located in the first line of Algorithms.py, Models.py, Read.py, and Functions.py.

Acknowledgements

This material is based upon work supported by the U.S. Department of Energy, Office of Science, under contract number DE-AC02-06CH11357.