Federated Split Learning with Efficient File-Based Gradient Synchronization

Introduction

The project improves Split Federated Learning (SFL) by introducing a file-based gradient synchronization mechanism. The system modularizes split servers and clients as independent microservices, enabling flexible scaling and deployment tailored to different use cases.

Features

• File-based gradient synchronization for efficient model updates.
• Microservice-oriented design: split servers and clients can be deployed and scaled independently.
• Support for balanced and imbalanced (Non-IID) client data setups.

Prerequisites

• Docker Desktop (Windows)
• NVIDIA driver and NVIDIA Container Toolkit

Quick Start

Note: Services must be started in order: Database & Backend → Server services → Client services.
Starting them out of order may cause connection failures.

1. Start the database and backend:

docker-compose -f docker-compose.db-backend.yml up

2. Start the Split Server and Federated Server:

docker-compose -f docker-compose.servers.yml up

3. Start clients:

docker-compose -f docker-compose.clients.yml up

or for imbalanced (Non-IID) data:

docker-compose -f docker-compose.clients_imbalance.yml up

4. Open the FastAPI documentation at <backend_ip>:8000/docs.
   Locate the POST /federated_train_async API, provide the desired number of global rounds, and trigger the federated training process.