Federated Learning Overview

What is Federated Learning?

Federated Learning is a machine learning setting where many clients (e.g. mobile devices or whole organizations) collaboratively train a model under the orchestration of a central server (e.g. service provider), while keeping the training data decentralized.

Personal contribution

This repository contains our implementation of a federated learning system using the CIFAR-100 and Shakespeare datasets. The system includes both centralized and federated training approaches, with support for IID and non-IID data sharding. We have also implemented client selection strategies based on skewed probabilities and a novel method based on entropy of data. A detailed report of the project can be found in the report.pdf file.

Repo Structure

• notebooks/: Contains Jupyter notebooks to run the experiments.

• src/cifar100/: Contains the implementation for the CIFAR-100 dataset.

  • centralized.py: Centralized training script.
  • server.py: Federated learning server script.
  • client.py: Client-side training script.
  • dataset.py: Dataset handling and sharding.
  • models.py: Model definitions.
  • utils.py: Utility functions for checkpointing and plotting.
  • contrib-dataset.py: Dataset handling and sharding for our novel method based on entropy of data.
  • contrib-server.py: Federated learning server script for our novel method based on entropy of data.

• src/shakespeare/: Contains the implementation for the Shakespeare dataset.

  • centralized.py: Centralized training script.
  • server.py: Federated learning server script.
  • client.py: Client-side training script.
  • dataset.py: Dataset handling and sharding.
  • model.py: Model definitions.
  • utils.py: Utility functions for checkpointing and plotting.

Usage

1. Clone the repository:

   git clone https://github.com/maxfra01/federated-learning.git
   cd federated-learning

2. Install the required dependencies:

   pip install -r requirements.txt

3. Run the experiments either using the Jupyter notebooks in the notebooks/ directory or by running source code directly from the src/ directory.

Contributors

• Massimo Francios
• Nicolò Bonincontro
• Sergio Lampidecchia