AlexConnat / MPC-Aggreg

MPC-Aggreg

Requirements

To install the gmpy2 on Linux with pip, you must already have installed libgmp, libmpfr and libmpc (Tested on Ubuntu 18.04).

sudo apt install libgmp-dev libmpfr-dev libmpc-dev

Then install the only 3 python requirements, numpy, gmpy2 and mpyc.

pip3 install numpy gmpy2 git+git://github.com/lschoe/mpyc.git@4642a3cc65b62234491b6392b2f106fd8ae6457a

Notes

The code is in main.py and it only calls one external file called utils.py containing some helper functions.

For the purpose of the PoC, this code requires to be ran in this directory structure - namely having a directory DATA/ containing at least mnist250/ and svhn250/ - and having a (potentially empty) directory BENCHMARK/.

MPC-Aggreg/
| .gitignore
| main.py
| utils.py
|
| DATA/
| | mnist250/
| | | votes_client_i.npy
| | | […]
| | svhn250/
| | | votes_client_i.npy
| | | […]
|
| BENCHMARK/
| | mnist250_640s_5c.csv
| | […]
|
| runAll.sh

To generate these votes votes_client_i.npy, where i is the cliend ID, we wrote a little script votes_compilator.py that you find in the DATA/directory (more information in the README file of the DATA/ directory).

The script runAll.sh was used to generate the CSV files used in the benchmarking of this PoC.
What it does, is running the code main.py with both datasets "mnist" and "svhn" and different number of clients ranging from 2 (3-party MPC) to 250 (251-party MPC).

How to run on an Amazon EC2 Ubuntu (18.04) instance:

1. Log in to your instance via ssh
   ssh -i key_file ubuntu@IP_ADDRESS_OF_INSTANCE

2. Clone this repository
   git clone https://github.com/AlexConnat/MPC-Aggreg

3. In this repository, run the setup script
   cd MPC-Aggreg; ./setup.sh

This will:

1. Update the sources lists for apt
   sudo apt update

2. Install python and pip
   sudo apt install python3 python3-pip

3. Install the required system libraries
   sudo apt install libgmp-dev libmpfr-dev libmpc-dev

4. Install the required python libraries
   pip3 install numpy gmpy2 git+https://github.com/lschoe/mpyc

---

You can then run the runAll script
./runAll.sh

You can tune the parameters in the runAll script. Results could be found in the BENCHMARK directory, under the format {name of dataset}_{number of clients}_{number of samples}_{timestamp}.csv
(e.g: mnist250_8c_640s_1566203440)

NB: These CSV files contain at each line: the sample ID, the label it was assigned, and the time it took to do so.