Home

Awesome

<p align="center"> <h2 align="center"><strong>Evading Forensic Classifiers with Attribute-Conditioned Adversarial Faces [CVPR 2023]</strong></h2> <p align="center"> <a href="https://fahadshamshad.github.io"><strong> Fahad Shamshad</strong></a>, <a href="https://koushiksrivats.github.io"><strong> Koushik Srivatsan</strong></a>, <a href="https://scholar.google.com/citations?user=2qx0RnEAAAAJ&hl=en"><strong> Karthik Nandakumar</strong></a> <br> <span style="font-size:4em; "><strong> MBZUAI, UAE</strong>.</span> </p> </p> <p align="center"> <a href="https://openaccess.thecvf.com/content/CVPR2023/html/Shamshad_Evading_Forensic_Classifiers_With_Attribute-Conditioned_Adversarial_Faces_CVPR_2023_paper.html" target='_blank'> <img src="https://img.shields.io/badge/CVPR-Paper-blue.svg"> </a> <a href="https://koushiksrivats.github.io/face_attribute_attack/" target='_blank'> <img src=https://img.shields.io/badge/Project-Website-87CEEB"> </a> <a href="https://www.youtube.com/watch?v=ZkPuU3lIK9U" target='_blank'> <img src="https://badges.aleen42.com/src/youtube.svg"> </a> </p> <p align="center"> <img src="./docs/static/images/twitter_gif.gif" align="center" width="100%"> </p> <br/>

Updates :loudspeaker:

<br/> <br/>

Attribute-conditioned adversarial face image generation

<p align="center"> <img src="./docs/static/images/pipeline.png" align="center" width="100%"> </p>

<a name="instructions-for-code-usage"></a>

Intructions for Code usage

Setup

git clone https://github.com/koushiksrivats/face_attribute_attack.git

cd face_attribute_attack
# use anaconda to build environment 
conda create -n faa python=3.8
conda activate faa
# install packages
pip install -r requirements.txt

Dataset and pre-trained weights

  1. Download the forensic classifier training data:
    • You can download the real FFHQ images here
    • You can download the fake (styleGAN generated) FFHQ images here
    • Re-arrange them into the following folder structure. 
      data
  |__ train
          |__ fake
          |__ real
  |__ test
          |__ fake
          |__ real
  2. Download the pre-trained StyleGAN2 weights:
    • Download the pre-trained StyleGAN2 weights from here.
    • Place the weights in the 'pretrained_models' folder.

Usage

Train forensic classifier

python classifier_training.py \
  --train_data data/train \
  --test_data data/test \
  --batch_size 128 \
  --epochs 10 \
  --classifier_name resnet50 \
  --output_path forensic_classifier_trained_models/resnet50/ \
  --wandb_project_name project_name \
  --experiment_name resnet50_forensic_classifier \
  --resume_training False

# Note: The trained model will be saved in the output_path under the name 'best_epoch.pt'

Adversarial faces with text as reference

python text_as_reference.py --config_file configs/config_text_as_reference.ini

Adversarial faces with image as reference

python image_as_reference.py --config_file configs/config_image_as_reference.ini

Adversarial transferability with meta-optimization (Uses the text-as-reference method)

python adversarial_transferability.py --config_file configs/config_adversarial_transferrability.ini

If you're using this work in your research or applications, please cite using this BibTeX:

@inproceedings{shamshad2023evading,
  title={Evading Forensic Classifiers With Attribute-Conditioned Adversarial Faces},
  author={Shamshad, Fahad and Srivatsan, Koushik and Nandakumar, Karthik},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16469--16478},
  year={2023}
}