Fedra - Federated Learning Framework

General Description

Fedra is an innovative federated learning framework that enables decentralized training across distributed nodes. It leverages peer-to-peer (P2P) communication to facilitate model training on edge devices, allowing nodes to directly exchange model updates and perform local aggregation. This decentralized approach enhances privacy and security, as no raw data is shared and no central server is required. Fedra is particularly suitable for scenarios where data privacy is paramount.

Table of Contents

Contents

• General Description
• Table of Contents
• Detailed Description
  • Overview
• Configuration
• Supported Models
• Federated Learning Process
• Repository Structure
• Generating Documentation
• Versioning
• License
• Contributing
• About the Project

Detailed Description

Overview

Fedra revolutionizes federated learning by employing a decentralized model. Instead of relying on a central aggregator, each node independently trains a model on its data and engages in mutual exchange of model updates with other nodes. Local aggregation is then performed on each node, integrating insights from across the network to refine the model. This process iterates until the collective model converges or a specified number of rounds is complete.

Configuration

Setting up Fedra is streamlined and user-friendly. The framework requires node-specific configurations, managed via a simple configuration file (node.conf). This file encapsulates essential parameters such as model details, training rounds, and P2P network settings, ensuring a seamless initiation into the federated learning process.

Supported Models

The framework is model-agnostic, allowing for the integration of various neural network models. The default implementation includes a dense neural network specified in models/net.py, which can be easily replaced or extended based on the application requirements.

Federated Learning Process

1. Initialization: The global model is initialized, and configuration parameters are distributed to all participating nodes.
2. Local Training: Each node trains the model locally with its dataset.
3. Model Update Exchange: Post-training, nodes serialize and exchange model updates amongst themselves via P2P communication.
4. Local Aggregation: Each node deserializes received updates and performs local federated averaging to refine their model.
5. Convergence and Repetition: The iterative process continues, with each round of training and aggregation progressively enhancing the model's accuracy.

Repository Structure

The fedra project is structured as follows:

• fedra/main.py: Entry point for the federated learning process.
• fedra/models/: Contains the neural network models.
• fedra/network/: Implements the P2P communication.
• fedra/utils/: Provides utility functions for data loading and weight manipulation.
• data/: Hosts the dataset used for training.
• tests/: Contains unit tests for various components of the project.
• docs/: Documentation for the project.

Generating Documentation

To generate the documentation for Fedra, you first need to install the necessary dependencies:

pip3 install -r requirements.txt

Once the dependencies are installed, you can compile the documentation using the following command:

tox -e docs

This will generate HTML documentation in the docs/_build/html/ directory. Ensure that you have tox installed in your environment to use this command.

Versioning

The current version of the project is maintained in the VERSION file. Ensure to update this file as the project evolves.

License

fedra is licensed under the MIT License. See the LICENSE file for more details.

Contributing

Contributions to fedra are welcome! Please refer to the CONTRIBUTING.md file for guidelines on how to contribute to this project.

About the Project

This work was partially support by the "Trustworthy And Resilient Decentralised Intelligence For Edge Systems (TaRDIS)" Project, funded by EU HORIZON EUROPE program, under grant agreement No 101093006