A Continual Deepfake Detection Benchmark: Dataset, Methods, and Essentials

Pytorch implementation for WACV 2023 paper "A Continual Deepfake Detection Benchmark: Dataset, Methods, and Essentials". We propose a continual deepfake detection benchmark (CDDB) over a new collection of deepfakes from both known and unknown generative models.

A Continual Deepfake Detection Benchmark: Dataset, Methods, and Essentials
Chuqiao Li, Zhiwu Huang, Danda Pani Paudel, Yabin Wang, Mohamad Shahbazi, Xiaopeng Hong, Luc Van Gool
[Project Page] [Paper]

Dependencies

Run pip install -r requirements.txt to install required dependencies.

Datasets

Our proposed CDDB benchmark dataset contains mixed deepfake sources as follows:

We collected the benchmark dataset from pervious opensource deepfake datasets:

ProGAN,StyleGAN, BigGAN, CycleGAN, GauGAN, CRN,IMLE, SAN, FaceForensics++ ,WhichFaceReal from [CNNDetection]
GLOW, StarGAN from [GanFake Detection]
WildDeepfake from [WildDeepFake]

Download the following datasets to datasets/

CDDB Dataset
[ProGAN pretrained weights]

Organize the dataset folder as:

Project
├── datasets
|   ├── CDDB
|   ├── checkpoints

Training:

DyTox

Using DyTox, train the model with binary/multi labels, with different sequences:

cd DyTox
bash train.sh 0,1 \
    --options configs/data/ganfake_easy.yaml configs/data/ganfake_easy_order.yaml configs/model/ganfake_pretrain_dytox.yaml \
    --name dytox_easy_m1500_sumblog0.1 \
    --data-path ./datasets/CDDB/  \
    --output-basedir ./datasets/checkpoints  \
    --memory-size 1500 \
    --binary_loss sum_b_log \
    --binary_weight 0.1

LUCIR:

Using LUCIR, train the model with binary/multi labels, with different sequences:

cd LUCIR
python lucir_main.py --name lucir_easy_m1500_sumasig0.1 --checkpoints_dir ./datasets/checkpoints  --dataroot ./datasets/CDDB/ --task_name gaugan,biggan,cyclegan,imle,deepfake,crn,wild --multiclass  0 0 1 0 0 0 0 --batch_size 32 --num_epochs 40 --binary_loss sum_a_sig --binary_weight 0.1

iCaRL:

Using iCaRL, train the model with binary/multi labels, with different sequences:

cd iCaRL
python main_icarl_CNND.py --name icarl_easy_m1500_sumasig0.1 --checkpoints_dir ./datasets/checkpoints --model_weights ./datasets/checkpoints/no_aug/model_epoch_best.pth --dataroot ./datasets/CDDB/ --task_name gaugan,biggan,cyclegan,imle,deepfake,crn,wild --multiclass  0 0 1 0 0 0 0  --batch_size 32 --num_epochs 30 --schedule 10 20 30 --add_binary --binary_weight 0.1 --binary_loss sum_a_sig

Loss Term:

As described in the paper, we implement diffent kinds of binary loss term

SumLogit

--binary_loss sum_a_sig

SumFeat

--binary_loss sum_b_sig

SumLog

--binary_loss sum_b_log

Max

--binary_loss max

Citation

When using the code/figures/data/etc., please cite our work

@inproceedings{li2022continual,
  title={A Continual Deepfake Detection Benchmark: Dataset, Methods, and Essentials},
  author={Li, Chuqiao and Huang, Zhiwu and Paudel, Danda Pani and Wang, Yabin and Shahbazi, Mohamad and Hong, Xiaopeng and Van Gool, Luc},
  booktitle={Winter Conference on Applications of Computer Vision (WACV)},
  year={2023}
}

Coral79 / CDDB