This is a personal implementation of F3-Net , so there are lots of difference compared to the official version. To learn more details about F3-Net, please check the paper here.
Model is tested on FaceForensics++ LQ data and reports AUC.
| Model | Paper | Valid(Mine) | Test(Mine) |
|---|---|---|---|
| Baseline | 89.3 | 92.0 | 89.6 |
| FAD | 90.7 | 91.3 | 89.5 |
| LFS | 88.9 | 87.5 | 84.7 |
| Both | 92.8 | 91.0 | 88.6 |
| Mix | 93.3 | \ | \ |
Obviously, there's something wrong with the implementation of LFS branch and I'm working on it now.
Hyperparameters are in train.py.
| Variable name | Description |
|---|---|
| dataset_path | The path of dataset, support FF++ only. |
| pretrained_path | The path of pretrained Xception model. |
| batch_size | 128 in paper. |
| max_epoch | how many epochs to train the model. |
| loss_freq | print loss after how many iterations |
| mode | mode of the network, see details below. |
Download Xception model trained on ImageNet (through this link) or use your own pretrained Xception.
Then modify the pretrained_path variable.
The dataset related function is designed for FaceForensics++ dataset. Check this github repo or paper for more details of the dataset.
After preprocessing, the data should be organized as following:
|-- dataset
| |-- train
| | |-- real
| | | |-- 000
| | | | |-- frame0.jpg
| | | | |-- frame1.jpg
| | | | |-- ...
| | | |-- 001
| | | |-- ...
| | |-- fake
| | |-- Deepfakes
| | | |-- 000_167
| | | | |-- frame0.jpg
| | | | |-- frame1.jpg
| | | | |-- ...
| | | |-- 001_892
| | | |-- ...
| | |-- Face2Face
| | | |-- ...
| | |-- FaceSwap
| | |-- NeuralTextures
| |-- valid
| | |-- real
| | | |-- ...
| | |-- fake
| | |-- ...
| |-- test
| | |-- ...
There are four modes supported in F3-Net.
| Mode(string) | |
|---|---|
| 'FAD' | Use FAD branch only. |
| 'LFS' | Use LFS branch only. |
| 'Both' | Use both of branches and concate before classification. |
| 'Mix'(unavailable) | Use both of branches and MixBlock. |
Note:
Mode 'Mix' is unavailable yet. If you're interested in this part, check 'class Mixblock' in models.py.
Environment:
Pytorch, torchvision, numpy, sklearn, pillow are needed.
To train the model
python train.py
Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. arXiv preprint arXiv:2007.09355, 2020