DeepFake Detection Using ScatterNets

Installation

To run this repository it is highly recommended to create a separate anaconda enviroment.

Create a anaconda enviroment:

conda create --name deepfake_detection

Activate the created requirement:

conda activate deepfake_detection

Run the requiremnts file to get the requirements:

pip install -r requirements.txt

This repo uses my pytorch implementation of the dtcwt: pytorch_wavelets. You can install this however just by pip installing the requirements.txt. From the command line, the following 3 commands will allow you to run the experiments:

git clone https://github.com/fbcotter/scatnet_learn
pip install -r requirements.txt
pip install .

Dataset

To download the datasets please follow the folowwing repository.

Preprocessing has to be performed on the dataset with an appropriate number of frames sampled per second. The code for this will be updated shortly.

Training the model

The minimum required argument are the following.

python train.py --model_name "ScatterNetDeepFakeDetection" --dataset "FaceForensics"

The complete list of argumenta that can be provided is as follows:

model_name: (str) Write the name of your model.
dataset: (str) The deepfake detection to train your model on. 
            Availabel choices are: 'FaceForensics', 'FaceForensics++', 'CelebDF', 'GoogleDFD'\
        ,'FaceHQ', 'DFDC', 'DeeperForensics', 'UADFV'

num_epochs: (int) Enter the number of epochs to train the network.

loss_type: (str) Enter the desired loss type.

scheduler_type: (str) Enter the scheduler type to vary the learning rate.

log_interval: (int) Enter the intervals after which to log the details.

num_classes: (int) Enter the number of classes for a given dataset.

batch_size: (int) Enter the number of batch size.

optimizer: (str) Enter the optimizer required to step during training.

lr: (float) Enter the learning rate to train the network.

augmentation: (bool) Augmentations to the training set to make the network more generalizable.

save_path: (str) Save path for models.

logdir: (str) Log dir for tensorboard.

flush_history: (bool) Flush the tensorboard log dir.

load_model: (bool) Start training the model from a checkpoint.

parallel: (bool) Activate if multiple gpus used.

val: (bool) Activate to use validation set.

return_best: (bool) Returns the best model after training.

Testing the model

Run the following command to test the model

python test.py --model_name "ScatterNetDeepFakeDetection" --dataset "FaceForensics" --frames 30