This repo includes the authors' Pytorch implementation of the paper:
Computer Vision and Pattern Recognition (CVPR) 2023
[arxiv]
In this work, we take a deep look into the generalization ability of binary classifiers for the task of deepfake detection. Specifically,
-
We discover that deepfake detection models supervised only by binary labels are very sensitive to the identity information of the images, which is termed as the Implicit Identity Leakage in the paper.
-
Based on our analyses, we propose a simple yet effective method termed as the ID-unaware Deepfake Detection Model to reduce the influence of the ID representation, successfully outperforming other state-of-the-art methods.
- [03/2023] release the training and test code for our model
- [03/2023] release the pretrained weight
- Python 3 >= 3.6
- Pytorch >= 1.6.0
- OpenCV >= 4.4.0
- Scipy >= 1.4.1
- NumPy >= 1.19.5
- Download the dataset from FF++ and put them under the ./data.
.
└── data
└── FaceForensics++
├── original_sequences
│ └── youtube
│ └── raw
│ └── videos
│ └── *.mp4
├── manipulated_sequences
│ ├── Deepfakes
│ └── raw
│ └── videos
│ └── *.mp4
│ ├── Face2Face
│ ...
│ ├── FaceSwap
│ ...
│ ├── NeuralTextures
│ ...
│ ├── FaceShifter
│ ...
-
Download the landmark detector from here and put it in the folder ./lib.
-
Run the code to extract frames from FF++ videos and save them under the ./train_images or ./test_images based on the division in the original dataset.
python3 lib/extract_frames_ldm_ff++.py
You can download pretrained weights here.
To evaluate the model performance, please run:
python3 test.py --cfg ./configs/caddm_test.cfg
Our model achieved the following performance on:
| Training Data | Backbone | FF++ | Celeb-DF | DFDC |
|---|---|---|---|---|
| FF++ | ResNet-34 | 99.70% | 91.15% | 71.49% |
| FF++ | EfficientNet-b3 | 99.78% | 93.08% | 73.34% |
| FF++ | EfficientNet-b4 | 99.79% | 93.88% | 73.85% |
Note: the metric is video-level AUC.
To train our model from scratch, please run :
python3 train.py --cfg ./configs/caddm_train.cfg
Coming soon
If you have any questions, please feel free to contact us via [email protected].