Federated Learning with Flower

Overview

This project implements a Federated Learning (FL) system using the Flower framework. It compares centralized and federated learning approaches on the CIFAR-10 dataset using PyTorch. Additionally, it incorporates Differential Privacy (DP) techniques to enhance security.

Features

• Federated Learning (FL): Distributed model training using multiple clients.
• Centralized Training: Standard model training on a single dataset.
• Differential Privacy (DP): Adds noise to model updates for enhanced security.
• Custom Model: Convolutional Neural Network (CNN) for image classification.
• Federated Simulation: Uses Flower's simulation framework for testing FL models.
• Performance Comparison: Measures accuracy and training time for both centralized and federated models.
• Configurable Hyperparameters: Supports easy modification of training parameters via config.json.

Project Structure

.
├── client.py            # Implements the Flower client
├── server.py            # Configures the federated learning server
├── federated_main.py    # Runs federated learning experiments
├── centralized_main.py  # Runs centralized training experiments
├── data_loader.py       # Loads and processes CIFAR-10 dataset
├── model.py             # Defines the CNN model
├── train.py             # Implements training and evaluation functions
├── config.py            # Stores device settings (CPU/GPU selection)
├── config.json          # Stores training hyperparameters
├── config_manager.py    # Handles config file updates
├── utils.py             # Helper functions for model parameter handling
├── main.py              # Compares centralized and federated training
└── README.md            # Project documentation

Installation

1. Create a Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows use: venv\Scripts\activate

2. Install Dependencies

pip install -r requirements.txt

Running the Project

To run the project, navigate to the project directory and execute the following command:

cd code
python -m main

This command runs the main.py script, which iterates over different hyperparameter configurations and executes federated learning experiments.

Configuration

Modify config.json to adjust hyperparameters:

{
    "CENTRALIZED_EPOCHS": 100,
    "NUM_CLIENTS": 2,
    "NUM_ROUNDS": 1,
    "BATCH_SIZE": 8,
    "NOISE_MULTIPLIER": 0.1,
    "EVAL_RATIO": 0.3,
    "FIT_RATIO": 1.0
}

• CENTRALIZED_EPOCHS: Number of epochs for centralized training
• NUM_CLIENTS: Number of clients in federated learning
• NUM_ROUNDS: Number of federated training rounds
• BATCH_SIZE: Training batch size
• NOISE_MULTIPLIER: Level of differential privacy noise
• EVAL_RATIO: Ratio of clients used for evaluation

Model Architecture

The CNN model (model.py) consists of convolutional, pooling, and fully connected layers designed for CIFAR-10 classification.

Logging & Evaluation

The project logs accuracy results in CSV files for further analysis:

• Centralized Training: log/centralized.csv
• Federated Learning: log/federated_acc.csv

Acknowledgments

• Flower for federated learning framework
• PyTorch for deep learning support