Federated learning secure aggregation

A simple Python implementation of a secure aggregation protocole for federated learning.

Introduction

This project is an implementation of a practical secure aggregation for privacy-preserving Machine Learning as it is described in this paper.

It will train a neural network using federated learning. The neural network is created in order to predict human activity based on accelerometer and gyroscope data. The field of recognize human activity is beyond the scope of this work.

It is devided into two parts:

• Server: which will centralize information and orchestrate multiple clients to process new weights and biais of a shared machine learning neural network.
• Client: which will train the shared neural network with a part of the training dataset

Please read the following to understand how to run the project.

Run the project

Server

The server project uses docker-compose to run severals services (containers):

• web: this service is a web API coded with python and Flask. This service is accessible using http://localhost:3000
• db: this service is based on the official postgres docker image to run a database. This service is accessible using postgresql://localhost:5432 with username: postgres and password: root. The default database schema is federated.
• pgadmin: this service is a web application to navigate throw the postgres database. This service is accessible using http://localhost:5050

To run the complete stack, use the following command:

docker-compose up

If you want to export the docker image of this API, use the following command:

docker build . -t <IMAGE_TAG>:<VERSION>

Client

The client is a simple python script to interact with the server. The bash script run.sh is used to run several client in the same time. Run the following command when the web service is up and running:

./run.sh

Python lib used

• Cryptography: https://cryptography.io/en/latest/
• SSlib: https://github.com/jqueiroz/python-sslib

Références

https://dl.acm.org/doi/pdf/10.1145/3133956.3133982

http://doi.acm.org/10.1145/3302505.3310068