DRAG: Data Reconstruction Attack using Guided Diffusion

DRAG: Data Reconstruction Attack using Guided Diffusion, ICML 2025

🚀 Getting Started

Environment

We provided dependencies in conda for reproducing.

conda create -f environment.yml

Extra dependencies: CUDA, which is for compiling plugin for StyleGAN2-ADA, which is not a necessary part for our diffusion based attacks.

Download necessary dataset and checkpoint, which are storing in <PROJ_ROOT>/datasets and <PROJ_ROOT>/checkpoints by default.

<PROJ_ROOT>
├── checkpoints
│   ├── stylegan2-ada-pytorch
│   │   └── ffhq.pkl
│   └── stylegan-xl
│       └── imagenet256.pkl
├── datasets
│   ├── coco
│   ├── ffhq
│   └── imagenet2012
└── ...

To reproduce the results in the paper, check commands in the script run.sh.

Experiment Settings

We sample 10 images for each dataset, determined with the script roll_dice.py with random seed 0.

Dataset	Samples
MSCOCO	119,138,725,1044,1703,1919,2111,2591,4111,4497
FFHQ	337,429,1729,1917,2890,4919,6044,7532,8223,9399
IN-1K	6091,11341,16904,17849,24681,28026,36044,36293,37807,49165

🎯 Model Checkpoint

TO BE UPDATED

🤝 Acknowledgment

This work builds on and benefits from several open-source efforts:

• SIMBA: Split Inference - Metrics, Benchmarks and Algorithms
• Guidance with Spherical Gaussian Constraint for Conditional Generation
• StyleGAN2-ADA
• StyleGAN-XL

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

📄 Citation

If you find our work useful, please cite us:

@inproceedings{lei2025drag,
    title={DRAG: Data Reconstruction Attack with Guided Diffusion},
    author={Wa-Kin Lei and Jun-Cheng Chen and Shang-Tse Chen},
    booktitle={International Conference on Machine Learning (ICML)},
    year={2025}
}