Benchmarking FedAvg and FedCurv for Image Classification Tasks

This repository contains the code for running the experiments in the paper "Benchmarking FedAvg and FedCurv for Image Classification Tasks", Casella et al., ITADATA2022.

Installation

You need to install OpenFL, the framework for Federated Learning developed by Intel:

$ pip install openfl

For more installation options check out the online documentation.

Getting Started

OpenFL enables data scientists to set up a federated learning experiment following one of the workflows:

• Director-based Workflow: Setup long-lived components to run many experiments in series. Recommended for FL research when many changes to model, dataloader, or hyperparameters are expected

• Aggregator-based Workflow: Define an experiment and distribute it manually. All participants can verify model code and FL plan prior to execution. The federation is terminated when the experiment is finished

The quickest way to test OpenFL is to follow our tutorials.
Read the blog post explaining steps to train a model with OpenFL.
Check out the online documentation to launch your first federation.

Requirements

• Ubuntu Linux 16.04 or 18.04.
• Python 3.6+ (recommended to use with Virtualenv).

OpenFL supports training with TensorFlow 2+ or PyTorch 1.3+ which should be installed separately. User can extend the list of supported Deep Learning frameworks if needed.

Project Overview

What is Federated Learning

Federated learning is a distributed machine learning approach that enables collaboration on machine learning projects without having to share sensitive data, such as, patient records, financial data, or classified information. The minimum data movement needed across the federation is solely the model parameters and their updates.

Usage

You can test Federated Learning in the following settings:

Datasets

• CIFAR-10
• MNIST
• MEDMNIST2D (organamnist)

Partitionings:

• Uniform (IID)
• Quantity Skew (non-IID)
• Prior shift (non-IID) in three versions: labels quantity skew, pathological labels skew, Dirichlet Labels skew
• Covariate shift (non-IID)

In order to use one of the previous partitionings methods you need to overwrite numpy.py with the provided non-iidness.py

Steps:

1. ./start_director.sh
2. Start how many envoys you want with ./start_envoy.sh
3. Create your Jupyter notebook with all the settings you want!

Contributors

Bruno Casella - bruno.casella@unito.it Roberto Esposito - roberto.esposito@unito.it Carlo Cavazzoni - carlo.cavazzoni@leonardo.com Marco Aldinucci - marco.aldinucci@unito.it

Support

We welcome questions, issue reports, and suggestions:

• GitHub Issues

Citation

@inproceedings{casella2022benchmarking,
  author = {Bruno Casella and
  Roberto Esposito and
  Carlo Cavazzoni and
  Marco Aldinucci},
  editor = {Marco Anisetti and
  Angela Bonifati and
  Nicola Bena and
  Claudio Ardagna and
  Donato Malerba},
  title = {Benchmarking FedAvg and FedCurv for Image Classification Tasks},
  booktitle = {Proceedings of the 1st Italian Conference on Big Data and Data Science, {ITADATA} 2022,
  September 20-21, 2022},
  series = {{CEUR} Workshop Proceedings},
  publisher = {CEUR-WS.org},
  year = {2022},
  url = {https://iris.unito.it/bitstream/2318/1870961/1/Benchmarking_FedAvg_and_FedCurv_for_Image_Classification_Tasks.pdf}
}