Detect Any Deepfakes (DADF)

This repository is the official implementation of Detect Any Deepfakes (CCBR 2023).

We propose DADF, a novel framework that adapts the Segment Anything Model (SAM) for the task of face forgery detection and localization. By incorporating low-level forgery traces into the SAM architecture, our model achieves state-of-the-art performance in pixel-level forgery localization while maintaining the versatility to be applied to other segmentation tasks (e.g., camouflage, shadow detection).

To ensure reproducibility, we recommend using a clean Conda environment.

# Create environment
conda create -n dadf python=3.8 -y
conda activate dadf

# Install PyTorch (Adjust CUDA version based on your driver)
pip install torch==1.13.0+cu117 torchvision==0.14.0+cu117 --extra-index-url [https://download.pytorch.org/whl/cu117](https://download.pytorch.org/whl/cu117)

# Install Dependencies
pip install -r requirements.txt

Please download the pre-trained weights and put them into the pretrained/ directory.

• **default or vit_h**: ViT-H SAM model. (Recommended)
• vit_l: ViT-L SAM model.
• vit_b: ViT-B SAM model.

Please organize your datasets under the ./load/ directory.

1. Download the dataset and put it in ./load.
2. Download the pre-trained SAM (Segment Anything) and put it in ./pretrained.
3. Training:

# run.sh
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nnodes=1 --nproc_per_node=4 --master_port=29501 \
train.py --config configs/demo.yaml

⚠️ Note: The SAM model consumes significant memory. We use 4 x A100 graphics cards for training. If you encounter memory issues, please try using graphics cards with larger VRAM or reduce the batch size.

We support distributed training (DDP). Adjust nproc_per_node based on your GPU availability.

# Example: Train on 4 GPUs
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nnodes=1 --nproc_per_node=4 --master_port=29500 \
train.py --config configs/dadf_vit_h.yaml

Note: If you encounter OOM (Out of Memory) errors with ViT-H, try reducing the batch size in the config or switching to ViT-B/L backbones.

Evaluate the model performance and generate localization maps.

python test.py --config configs/dadf_vit_h.yaml --model checkpoints/dadf_best.pth

If you find this code or paper useful, please cite our work:

@article{lai2023detect,
  title={Detect Any Deepfakes: Segment Anything Meets Face Forgery Detection and Localization},
  author={Lai, Yingxin and Luo, Zhiming and Yu, Zitong},
  journal={Chinese Conference on Biometric Recognition (CCBR)},
  year={2023}
}

This project is built upon SAM-Adapter. We thank the authors for their excellent codebase.