Update (May 19, 2025): We have added a link to the full dataset in the Dataset section below.

Introduction

This is the official repo of the paper "DiffusionGuard: A Robust Defense Against Malicious Diffusion-based Image Editing" (ICLR 2025).

Contents

• Introduction
• Installation
  • Python environments
• Protecting images against inpainting
• Configuration
• Dataset

Installation

Python environments

conda create -n diffusionguard python=3.10
conda activate diffusionguard
pip install -r requirements.txt # We recommend using torch version 2.1.1 and CUDA version 12.2 for best compatibility.

Protecting images against inpainting

You can protect images from diffusion-based image inpainting by generating adversarial perturbations using main.py. All images must be placed in the assets folder.

To run DiffusionGuard on image keanu.png for mask image keanu_mask.png in assets folder, please run the following code:

python main.py --config-name diffusionguard orig_image_name='keanu.png' mask_image_names='["keanu_mask.png"]'

In this example, orig_image_name refers to the original image you wish to protect, and mask_image_names contains the list of masks applied for protection.

If multiple mask images are given, the code will automatically merge them by taking the union of all masks.

Configuration

• method: Protection method. Defaults to diffusionguard.

• orig_image_name: Filename of the image to be protected. Should be placed in assets folder.

• mask_image_names: List of filenames for the masks to be used. Should be placed in assets folder.

• model: Specifies the models used.

  • inpainting: Inpainting model. Defaults to runwayml/stable-diffusion-inpainting.

• training:

  • size: Image resolution of the image to be protected. Image should be square. Defaults to 512.

  • iters: Number of PGD optimization iterations.

  • grad_reps: Number of reps to accumulate gradients.

  • batch_size: Batch size for PGD iteration.

  • eps: Linf threshold value for the adversarial noise. Defaults to 16/255.

  • step_size: PGD step size. Defaults to 1/255.

  • num_inference_steps: Number of inference steps for the inpainting pipeline.

  • mask: Mask-related settings.

    • generation_method: Mask generation method. Choose from single, global, and contour_shrink.

    • contour_strength: Shrink strength for contour_shrink method. Defaults to 1.0.

    • contour_iters: Shrink iterations for contour_shrink method. Defaults to 15.

    • contour_smoothness: Contour smoothing parameter for contour_shrink method. Defaults to 0.1.

Dataset

Our dataset is available at this Google Drive link.