Hierarchical Federated Learning with Edge Servers

This repository implements a Hierarchical Federated Learning(HFL) system where clients are distributed across multiple edge servers in a grid-based topology. The implementation uses a CNN model and supports various datasets (MNIST, CIFAR-10, CIFAR-100) for training in a federated setting.

The system implements a two-level hierarchical federated learning approach:

1. First level: Clients train local models and send updates to their assigned edge servers
2. Second level: Edge servers aggregate client models and communicate with the global server

The implementation uses a grid-based topology where clients are distributed across a defined area and assigned to the nearest edge server.

Table of Contents

• Getting Started
• Implementation Details
• Usage

Getting Started

1. create new env

conda create -n HFLD python=3.10.6
conda activate HFLD

2. clone the repository

git clone https://github.com/sunoopy/HierFed.git

3. install required version libraries

pip install -r requirments.txt

Implementation Details

Features

• Hierarchical Learning Structure

  • Two-level federated aggregation (Client → Edge Server → Global)
  • Proximity-based client-to-edge server assignment
  • FedAvg aggregation at both edge and global levels

• Flexible Dataset Support

  • MNIST (28x28x1)
  • CIFAR-10 (32x32x3)
  • CIFAR-100 (32x32x3)

• Non-IID Data Distribution

  • Dirichlet distribution-based data allocation
  • Controllable data heterogeneity via alpha parameter

• Visualization Tools

  • Client distribution visualization
  • Edge server coverage visualization
  • Training progress monitoring

Model Architecture

The implementation uses a simple CNN with the following structure:

• 3 Convolutional layers with ReLU activation
• 2 MaxPooling layers
• Dense layers for classification

Key Parameters

Parameter	Type	Values	Default	Description
-total_rounds	int	1~	10	Number of federated learning rounds
-dataset_name	string	xxx.xxx.xxx.xxx	MNIST	dataset selection ( mnist, cifar-10, cifar-100)
-num_clients	string	1~	100	Total number of clients in the system
-sample_per_client	int	1~	100	Number of samples per client
-num_edge_servers	int	1~	4	Number of edge servers
-grid_size	int	1~	10	Size of the simulation grid
-alpha	float	1~	1.0	Dirichlet distribution parameter for non-IID
-coverage_radius	float	1~	3.0	edge server coverage area radius setting
-client_repetition	boolean	True/False	True	Joint area edge server client repetition

Usage

to be modified