AVoiD-DF

This repository contains code and models for our paper

AVoiD-DF: Audio-Visual Joint Learning for Detecting Deepfake

In this paper, we propose an Audio-Visual Joint Learning for Detecting Deepfake (AVoiD-DF), which exploits audio-visual inconsistency for multi-modal forgery detection.

Clone the repo:

git clone https://github.com/SYSU-DISG/AVoiD-DF.git

Download the datasets from the following link:

https://pan.baidu.com/s/1MckHs-H57jTma5v0o6XYMA

(Sorry, due to protocol restrictions, the datasets and weights are not publicly available now.)

We use Python 3.6. Install requirements by running:

pip install -r requirements.txt

We create our own dataloader found in 'data_processing/my_dataset', you can modify it to train with your own data.

Training

You can do training on your own data by a simple command:

python train.py

Note

Unfortunately, due to protocol restrictions we cannot release the complete source code and models. We have open sourced some of the modules and our training codes are based on the released code of ViT.

Acknowledgements

Our work is based on the official version of AVoiD-DF, and some of our codes refer to ViT (Vision Transformer). Thanks for sharing!

Citation

If you find our repo helpful to your research, please cite our paper, thanks.