Untitled

Introduction

This repository contains Jupyter Notebooks for experimenting with Model Inversion Attacks and various Defense methods on MRI images using WGAN-GP. Each notebook corresponds to a specific step:

• wgan-mri.ipynb: WGAN-GP for generating MRI images (no defense)
• model-inversion-attack-gmi.ipynb: Perform GMI (Generative Model-Inversion Attacks Against Deep Neural Networks) attack
• wgan-mri-autoencoder-defend.ipynb: WGAN + Autoencoder Purifier
• wgan-mri-dp-defense.ipynb: WGAN + Differential Privacy Defense
• wgan-mri-gradient-clipping.ipynb: WGAN + Gradient Clipping Defense
• Mutual_Information_Regularization_based_Defense.ipynb: WGAN + Mutual Information Regularization based Defense

---

Requirements

• GPU (>4GB VRAM)
• RAM >25GB

---

Datasets

1. MRI Dataset (for WGAN & Defense)https://www.kaggle.com/datasets/nhtphmm/mri-dataset

• Used to train WGAN-GP and the defense methods.

1. DataMRI (for GMI Attack)https://www.kaggle.com/datasets/phamquanuet/datamri

• Used to train or fine-tune the GMI model.

How to download? (Kaggle API)

kaggle datasets download -d nhtphmm/mri-dataset --unzip -p ./data/mri-dataset
kaggle datasets download -d phamquanuet/datamri --unzip -p ./data/datamri

---

Checkpoints

Each notebook requires a pretrained checkpoint, which should be placed in the root folder:

1. WGAN-GP (wgan-mri.ipynb)
   • Download checkpoint from link
2. GMI Attack (model-inversion-attack-gmi.ipynb)
   • Download checkpoint from link
3. Autoencoder Purifier (wgan-mri-autoencoder-defend.ipynb)
   • Download checkpoint from link
4. Differential Privacy (DP-inspired) (wgan-mri-dp-defense.ipynb)
   • Download checkpoint from link and save
5. Gradient Clipping (wgan-mri-gradient-clipping.ipynb)
   • Download checkpoint from link and save
6. Mutual information regularization based defense (Mutual_Information_Regularization_based_Defense.ipynb)
   • Download checkpoint from link and save

---

Quick Start

1. Clone the repository:

git 
cd REPO_NAME

RUN ALL