DeepShield: Fortifying Deepfake Video Detection with Local and Global Forgery Analysis

This is a PyTorch implementation of "DeepShield: Fortifying Deepfake Video Detection with Local and Global Forgery Analysis" accepted by ICCV 2025.

---

🛠️ Environment Setup

Install the required dependencies by running the following command:

pip install -r requirements.txt

---

📊 Data Preparation

1. Face Detection & Preprocessing

Step 1: Download Datasets

Please download the following datasets and place them in the ./datasets/ directory:

• FaceForensics++ (FF++)
• Celeb-DF (v2)
• DFD
• DFDCP
• DFDC

Step 2: Extract Facial Images

We use RetinaFace to extract facial images from videos. Please refer to the RetinaFace GitHub repository for installation. Run the provided script to process the videos:

python ./preprocess/detect_faces.py

Step 3: Prepare Data Text Files

Create data annotation files in ./datasets/FaceForensics++/data_txt following this format:

[path] [start_frame] [end_frame] [label]

Example (train.txt for FF++):

original_sequences/youtube/c23/frames/071 0 452 0
original_sequences/youtube/c23/frames/054 0 367 0
manipulated_sequences/Deepfakes/c23/frames/071_054 0 452 1
manipulated_sequences/Face2Face/c23/frames/071_054 0 367 1
manipulated_sequences/FaceSwap/c23/frames/071_054 0 367 1
manipulated_sequences/NeuralTextures/c23/frames/071_054 0 367 1

2. Facial Landmark Detection

To analyze local forgery details, detect 81 facial landmarks using dlib:

1. Download shape_predictor_81_face_landmarks.dat and place it in the ./preprocess/ directory.
2. Run the landmark detection script:

python ./preprocess/detect_lands.py

3. Directory Structure

Ensure your data directory is structured as follows:

datasets
└── FaceForensics++
    ├── manipulated_sequences
    │   └── Deepfakes
    │       └── c23
    │           └── frames
    │               └── 000_003
    │                   └── 0000.png  <-- Face frames
    └── original_sequences
        └── youtube
            └── c23
                ├── frames
                │   └── 000
                │       └── 0000.png
                └── landmarks
                    └── 000
                        └── 0000.npy   <-- Landmark files

---

🚀 Training

Multi-GPU Training (DDP)

Run the following command for distributed training:

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --nnodes=1 --master_port 16677 main.py \
 --dataset ffpp_combine --val_dataset ffpp --save_all_ckpt \
 --input-size 224 --num_clips 4 --output_dir [your_output_dir] --opt adam \
 --lr 0.0003 --momentum 0.9 --weight-decay 0.0005 --epochs 80 --sched cosine \
 --duration 12 --batch-size 2 --disable_scaleup --cutout True \
 --warmup-epochs 0 --model vitb_st \
 2>&1 | tee ./output/train_ffpp_`date +'%m_%d-%H_%M'`.log

Single-GPU Training

CUDA_VISIBLE_DEVICES=0 python main.py \
 --dataset ffpp_combine --val_dataset ffpp --save_all_ckpt \
 --input-size 224 --num_clips 4 --output_dir [your_output_dir] --opt adam \
 --lr 0.0003 --momentum 0.9 --weight-decay 0.0005 --epochs 80 --sched cosine \
 --duration 12 --batch-size 2 --disable_scaleup --cutout True \
 --warmup-epochs 0 --model vitb_st \
 2>&1 | tee ./output/train_ffpp_`date +'%m_%d-%H_%M'`.log

---

🧪 Testing

To evaluate a trained model, use the following script:

CUDA_VISIBLE_DEVICES=0 python test.py \
 --dataset cdfv2 --input-size 224 --num_clips 4 \
 --duration 12 --batch-size 4 --disable_scaleup --model vitb_st \
 --test_ckpt_dir [your_checkpoint_dir] \
 2>&1 | tee ./output/test_cdf_`date +'%m_%d-%H_%M'`.log

---

📝 Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{cai2025deepshield,
  title={DeepShield: Fortifying Deepfake Video Detection with Local and Global Forgery Analysis},
  author={Cai, Yinqi and Li, Jichang and Li, Zhaolun and Chen, Weikai and Lan, Rushi and Xie, Xi and Luo, Xiaonan and Li, Guanbin},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages={12524--12534},
  year={2025}
}

Contact

Please feel free to contact the first author, namely Li Jichang, with an Email address li.jichang@foxmail.com, if you have any questions.