COPML

COPML provides a framework of a collaborative machine learning where multiple data-owners wish to jointly train a logistic regression model, while keeping their individual datasets private from the other parties.

For details of our work, please see our paper "A Scalable Approach for Privacy-Preserving Collaborative Machine Learning," accepted to NeurIPS 2020.

Usage

There are six main codes to run the experiments on the Amazon EC2 cloud. Therefore, before using the code, an Amazon EC2 cluster needs to be created first with N+1 machine instances. This can be done by using the starcluster tool, http://star.mit.edu/cluster/. In the experiments of the paper, m3.xlarge machine instances were used.

MPI4Py package should be installed at each machine

The detailed usage of the six main codes are as follows:

1. Masterless_COPML_CIFAR.py

   • Goal : get the raw data of training time of COPML with 50 iterations
   • Command : mpiexec -n (N+1) python2 Masterless_COPML_CIFAR.py (N)
   • input : ./datasets/cifar-10-batches-py/*
   • output : CPML_CIFAR_(N)

2. Masterless_BH_CIFAR.py

   • Goal : get the raw data of training time of MPC based on BH with 50 iterations
   • Command : mpiexec -n (N+1) python2 Masterless_BH_CIFAR.py (N)
   • input : ./datasets/cifar-10-batches-py/*
   • output : BH_CIFAR_(N)

3. Masterless_BGW_CIFAR.py

   • Goal : get the raw data of training time of MPC based on BGW with 50 iterations
   • Command : mpiexec -n (N+1) python2 Masterless_BGW_CIFAR.py (N)
   • input : ./datasets/cifar-10-batches-py/*
   • output : BGW_CIFAR_(N)

4. Masterless_COPML_GISETTE.py

   • Goal : get the raw data of training time of COPML with 50 iterations
   • Command : mpiexec -n (N+1) python2 Masterless_CodedPrivateML_GISETTE.py (N)
   • input : ./datasets/Gisette/*
   • output : CPML_GISETTE_(N)

5. Masterless_BH_GISETTE.py

   • Goal : get the raw data of training time of MPC based on BH with 50 iterations
   • Command : mpiexec -n (N+1) python2 Masterless_BH_GISETTE.py (N)
   • input : ./datasets/Gisette/*
   • output : BH_GISETTE_(N)

6. Masterless_BGW_GISETTE.py

   • Goal : get the raw data of training time of MPC based on BGW with 50 iterations
   • Command : mpiexec -n (N+1) python2 Masterless_BGW_GISETTE.py (N)
   • input : ./datasets/Gisette/*
   • output : BGW_GISETTE_(N)

• Download CIFAR dataset from https://www.cs.toronto.edu/~kriz/cifar.html and unzip it into ./datasets/cifar-10-batches-py/
• Download GISETTE dataset from https://archive.ics.uci.edu/ml/datasets/Gisette and unzip it into ./datasets/Gisette/

Citation

Will be updated.

Contact

The corresponding author is:

Jinhyun So

jinhyuns@usc.edu or jinhyun.soh@gmail.com