DeepFakeDetection: Experiments for DFDC
This repository contains the codebase for the Lab41 submissions to the DFDC. Fig1 illsutrates the overall design for the system submitted to the competion. Ultimately, the algorithm combined the predictios of three individually trained deepfake detectors that processed audio, video and the spatial power spectrum of indivual frames. The predictions from each detector was fed through a multilayer perceptron (MLP) and trained to produce the optimum set of predictions.
Fig1: Schematic illutstration of overall deepfake detection system design |
The package as well as the necessary requirements can be installed via
virtualenv -p /usr/local/bin/python3 venv
source venv/bin/activate
python setup.py install
or
virtualenv -p /usr/local/bin/python3 venv
source venv/bin/activate
pip install -e .
Fig2 illustrates the workflow for offline data preprocessing of the dfdc data.
Fig2: Preprocessing workflow for multimodal deepfake detection challenge |
We have included two preprocessing scripts for audio or video, respectively. Hyperparameters for the preprocessing can be controlled via a yaml file included in the './config' directory.
For details on the frame preprocessing:
python preprocess.py --help
For details on the audio preprocessing:
python preprocess_audio.py --help
We have provded a sample script used to train the ConvLSTM model on video frames. Hyperparameters for the training are controlled via a yaml file. For details regarding the training script:
python train.py --help
Copyright (c) 2020 IQT Labs LLC