MAID

Datasets

Our experiments are based on the open-source datasets DiffusionForensics, Artifact and GenImage.

Installation

1. Clone this repository and navigate to the MAID folder:

   git clone https://github.com/Zhu-Luyu/MAID.git
   cd MAID

2. Install the required packages:

   conda env create -f environment.yaml -n maid
   conda activate maid

3. Download pre-trained diffusion models:

   • DDIM
   • The repository for Stable Diffusion v1.5 has been removed, but you can use Stable Diffusion v1.4 as a substitute.
   • DiT
   • DeepFloyd IF-I-M-v1.0
   • LDM

Training & Evaluation

Modify the script parameters as needed to run training and evaluation:

sh train.sh
sh eval.sh

To run DMA extraction before training or evaluation:

cd dma
# DDIM
python compute_dma.py --diffusion_name "ddim" --diffusion_path path/to/your/ddim/checkpoint_file.ckpt --dataroot path/to/img_dataset --postfix "_ddim" --batch_size 100
# IF. The usage of LDM, SD, and DiT is similar
python compute_dma.py --diffusion_name "if" --diffusion_path path/to/your/if/model_folder --dataroot path/to/img_dataset --postfix "_if" --batch_size 100

Model Classes

We selected the following model classes for the experiment:

1. DiffusionForensics (LSUN bedroom subset)

Framework	Classes
GAN	StyleGAN
Diffusion Model	ADM, IDDPM, PNDM
-	Real

2. Artifact

Framework	Classes
GAN	BiqGAN, CIPS, CycleGAN, Denoising Diffusion GAN, Diffusion GAN, Gansformer, GauGAN, Lama, ProGAN, ProjectedGAN, StarGAN, StyleGAN, Taming Transformer, Generative Inpainting
Diffusion Model	Latent Diffusion, Stable Diffusion, VQ Diffusion, Glide, Palette, Mat
-	Real

3. GenImage

Framework	Classes
GAN	BigGAN
Diffusion Model	ADM, Glide, Midjourney, SDv1.5, VQDM, wukong
-	Real

Acknowledgments

Our code is based on the frameworks provided by CNNDetection and DNF. We greatly appreciate their contributions and code.

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{zhu2025maid,
  title={MAID: Model Attribution via Inverse Diffusion},
  author={Luyu Zhu and Kai Ye and Jiayu Yao and Chenxi Li and Luwen Zhao and Yuxin Cao and Derui Wang and Jie Hao},
  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  year={2025}
}