Federated Learning with Vision Transformers on Medical Images

This repository provides a modular implementation of Federated Learning (FL), Split Learning (SL), and Federated Split Vision Transformers with Block Sampling (FeSViBS) applied to medical imaging datasets like HAM10000 and BloodMNIST.

---

Overview

This project re-implements and simplifies the FeSViBS architecture from MICCAI 2023, using:

• Split learning with client-specific ViT heads and MLP tails
• Federated aggregation of split model parts
• IID and non-IID data distribution strategies
• Support for Differential Privacy (optional)

---

Methods Implemented

Method	Notebook/Script	Description
Centralized	centralized.ipynb	Vanilla training using full dataset
Local FL	local.py	Independent client training without parameter sharing
SLViT	SLViT.ipynb,	Split ViT training with shared backbone
SViBS	FeSViBS_Notebook.ipynb	SLViT with block sampling
FeSViBS	FeSViBS_Notebook.ipynb	Federated + Split ViT with block sampling
Data Split	organise_ham_clients.py	Prepares HAM10000 data for 6-client federated setup

---

Datasets Used

1. HAM10000
   Pigmented skin lesion dataset
   📥 Kaggle Link

2. BloodMNIST
   Blood cell images from the MedMNIST collection
   📦 Loaded via medmnist

---