Repository for the official implementation of our paper "Adv-Inversion: Stealthy Adversarial Attacks via GAN-Inversion for Facial Privacy Protection".
This repository contains the official implementation of Adv-Inversion, a novel framework for generating identity-preserving adversarial face samples via GAN inversion, designed to enhance facial privacy protection against unauthorized recognition systems.
1.Create a new conda environment
conda create -n advinversion python=3.10
conda activate advinversion
pip install -r requirements.txt2.Download Checkpoints
We use IR152, IRSE50, FaceNet, and MobileFace model checkpoints provided by AMT-GAN. The Google Drive link to these checkpoints is here. After downloading, create the pretrained_models folder:
cd Adv-Inversion-main
mkdir pretrained_models Place the .pth files in the pretrained_models folder (you can unzip the files from assets.zip to get them).
Additional StyleGAN and ArcFace backbone weights are necessary. Please download them from the provided Google Drive link, unzip them, and place them in the pretrained_models folder.
Download the enhanced FSE model weights from the given Google Drive link and place them in the pretrained_models folder.
3.Download Datasets
For our experiments, we use FFHQ and CelebA-HQ datasets for evaluation. Due to ownership restrictions, we cannot provide direct access to these datasets. Please download them manually. We only provide the target face for your use. Ensure that the source face is placed in the dataset/Celeba/source or dataset/FFHQ/source directory.
To perform the adversarial attack, use the following command (Place the source face in the dataset/Celeba/source directory, and adjust the batch size when the available GPU memory is insufficient.):
python attack_IDF.pyTo evaluate the attack results using FID metric:
python eval_fid_metric.pyTo evaluate the attack results using SSIM and PSNR metrics:
python eval_ssim_psnr_metric.pyTo be added after paper acceptance.
If you have any questions, please contact [wanghb69@mail2.sysu.edu.cn]