/Adversarial-Attack-CycleGAN-and-pix2pix

Generating Adversarial Images for Image-to-Image models in Pytorch

Primary LanguagePython

Adversarial-Attack-CycleGAN-and-pix2pix

Generating Adversarial Images for Image-to-Image models in Pytorch, code implementations for "Disrupting Image-Translation-Based DeepFake Algorithms with Adversarial Attacks," paper accepted to WACV DeepPAB workshop.

We provide PyTorch implementations for adversarially attacking CycleGAN, pix2pix and pix2pixHD models.

The code was written by Chin-Yuan Yeh.

This PyTorch implementation is to be used alongside the pytorch-CycleGAN-and-pix2pix repository or the pix2pixHD repository.

dataset specifics

This work uses CelebA-HQ and CelebAMask-HQ. The datasets can be found in this google drive link provided by this repository CelebAMask-HQ. (In case google drive link fails, backup version can also be found here)

Getting Started

Installation

  • clone this repo
git clone https://github.com/jimmy-academia/Adversarial-Attack-CycleGAN-and-pix2pix.git
  • clone the CycleGAN-pix2pix or pix2pixHD repo and copy the corresponding files to the correct place.

for CycleGAN or pix2pix:

git clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix.git
cd Adversarial-Attack-CycleGAN-and-pix2pix
cp -v attack_cyc_pix.py adversaryfunc.py ../pytorch-CycleGAN-and-pix2pix

or for pix2pixHD

git clone https://github.com/NVIDIA/pix2pixHD.git
cd Adversarial-Attack-CycleGAN-and-pix2pix
cp -v attack_pixhd.py adversaryfunc.py ../pix2pixHD

Model Setup -- Prepare Your Own Model:

  • download and prepare the CelebaHQ and mask datasets
bash scripts/download_dataset.sh
python scripts/prepare_dataset.py --imagesize [default: 256] --datatype [default: Smile | Blond | Bald | Glass | Blond-pix | Blond-HD | Custom ]
  • than move resulting new directory to the correct location, ex:
mv smilehq ../pytorch-CycleGAN-and-pix2pix/datasets

refer to customizations for dataset preparation details.

  • train the model with original CycleGAN-pix2pix or pix2pixHD repository:
cd ../pytorch-CycleGAN-and-pix2pix
python train.py --dataroot  datasets/smilehq --name smile --model cycle_gan

python train.py --dataroot  datasets/black_masked --name black_masked --model pix2pix
python train.py --dataroot  datasets/masked_blond --name masked_blond --model pix2pix

(pix2pixHD)

python train.py --dataroot datasets/black_masked --name black_masked --resize_or_crop scale_width_and_crop --loadSize 572 --fineSize 512 --label_nc 0 --no_instance 

python train.py --dataroot datasets/masked_blond --name masked_blond --resize_or_crop scale_width_and_crop --loadSize 572 --fineSize 512 --label_nc 0 --no_instance

Model Setup -- Use Pretrained Weights

  • Alternatively, you can use pretrained weights by:
bash script/download_weights.sh
  • then move the corresponding folders to the correct location, ex:
mkdir -p ../pytorch-CycleGAN-and-pix2pix/checkpoints/
mv weights/smile ../pytorch-CycleGAN-and-pix2pix/checkpoints/
  • also copy the dataset directory for single running example if you didn't download the celebahq dataset, ex:
cp -vr single_dset ../pytorch-CycleGAN-and-pix2pix/datasets

Attack:

  • basic attack -- CycleGAN
python attack_cyc.py --dataroot datasets/<single_dset|smilehq|...> --name smile
  • basic attack -- pix2pix
python attack_pix.py --dataroot datasets/<single_dset|smilehq|...>
  • basic attack -- pix2pixHD
python attack_pixhd.py
  • more attack templates: in template directory
a for ...