This repository contains the official PyTorch implementation of the following paper at ICLR 2024 (Spotlight):
Poisoned Forgery Face: Towards Backdoor Attacks on Face Forgery Detection
Jiawei Liang, Siyuan Liang, Aishan Liu, Xiaojun Jia, Junhao Kuang, Xiaochun Cao
https://openreview.net/pdf?id=8iTpB4RNvP
Download FaceForensics++ and DeepFakeDetection and Celeb-DF-v2 datasets and place them in ./data/ folder.
Please refer to ./data/datasets.md.
(1) Basic environment setup
conda create -n dfdbd python=3.8 && conda activate dfdbd
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2
sudo apt-get install -y cmake && pip install dlib
pip install pandas opencv-python tqdm imutils cmake easydict zipp imgaug efficientnet_pytorch imagecorruptions flask(2) Build retinaface from source
git clone https://github.com/ternaus/retinaface.git
# Replace the content in the requirements.txt file with the following:
albumentations>=1.0.0
torch>=1.9.0
torchvision>=0.10.0
cd retinaface && pip install -v -e .(3) Clone the PFF code
git clone https://github.com/JWLiang007/PFF.git(1) Download landmark detector (shape_predictor_81_face_landmarks.dat) from here and place it in ./src/preprocess/ folder.
(2) Download code for landmark augmentation:
mkdir src/utils/library
git clone https://github.com/AlgoHunt/Face-Xray.git src/utils/library(3) Following SBI, we extract faces from videos.
python src/preprocess/crop_dlib_ff.py -n 32 -d Original -c c23 -p train
python src/preprocess/crop_dlib_ff.py -n 32 -d Original -c c23 -p val
python src/preprocess/crop_dlib_ff.py -n 32 -d all -c c23 -p test
python src/preprocess/crop_retina_ff.py -n 32 -d Original -c c23 -p train -gid 0
python src/preprocess/crop_retina_ff.py -n 32 -d Original -c c23 -p val -gid 0
python src/preprocess/crop_retina_ff.py -n 32 -d all -c c23 -p test -gid 0 cd src/utils/trigger_gen && python training_texture.py --kernel_size 5
# Cache the backdoor trigger using the trained trigger generator
mode=sharpen_5
cd src/utils/bd_utils && python patch_gen_server.py --gen-mode $mode --bs 1 --gid 0python src/train_sbi_bd.py src/configs/sbi/base.json -n sbi_advt_gen_sharpen_5 --yaml-path src/configs/bd/attack/advt_gen/sbi_advt_gen_sharpen_5.yaml -c c23 --gpu-id 0 python src/utils/bd_utils/test_pool.py --weight $CHECKPOINT --wl --yaml src/configs/bd/attack/advt_gen/sbi_advt_gen_sharpen_5.yaml --gid 0 -c c23 --dataset all --cache-clean --abdOur code is based on the project SBI and BackdoorBench.