GIFD: A Generative Gradient Inversion Method with Feature Domain Optimization

A PyTorch official implementation for GIFD: A Generative Gradient Inversion Method with Feature Domain Optimization, accepted to ICCV-2023.

Results

Setup

We provide the environment configuration file exported by Anaconda, which can help you build up conveniently.

conda env create -f environment.yml
conda activate GIFD 

Dataset and model file

Download the ImageNet and FFHQ and provide their paths in the yml file.

While the model weights of BigGAN are downloaded automatically, StyleGAN2 weights require downloaded manually as follows.

gdown --id 1c1qtz3MVTAvJpYvsMIR5MoSvdiwN2DGb (shape predictor, placed in the root directory)

gdown --id 1JCBiKY_yUixTa6F1eflABL88T4cii2GR (stylegan pre-trained checkpoint, placed in the inversefed\genmodels\stylegan2_io)

Quick start

We prepare three configuration files for performing gradient inversion attacks, including the BigGAN-based, the StyleGAN2-based, and the GAN-free methods, where we give detailed descriptions of every parameter. Feel free to contact me at fang-h23@mails.tsinghua.edu.cn if you have any concerns. You can simply start by specifying the path of the config file.

python rec_mult.py --config $CONFIG_PATH

Citation

@inproceedings{fang2023gifd,
  title={GIFD: A Generative Gradient Inversion Method with Feature Domain Optimization},
  author={Fang, Hao and Chen, Bin and Wang, Xuan and Wang, Zhi and Xia, Shu-Tao},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={4967--4976},
  year={2023}
}

Acknowledgement

Our code is based on invertingGradients and ILO and we are grateful for their great devotion.

For BigGAN, we use PyTorch official implementation and weights.

For StyleGAN2, we adapt this Pytorch implementation, which is based on the official Tensorflow code.

We express great gratitude for their contribution to our community!