Awesome
Exploring Spatial-Temporal Features for Deepfake Detection and Localization
An implementation code for Deepfake Detection & Localization and a new dataset for evaluation.
Table of Contents
Background
A new Spatial-Temporal Deepfake Detection and Localization (ST-DDL) network with the newly proposed Anchor-Mesh Motion (AMM) algorithm and Fusion Attention (FA) module. In addition, a new public forgery dataset, ManualFake, with over 16,000 videos, including the new properties such as the generation of commercial software, the transmission of online social networks, namely the Facebook (Post), Facebook (Messenger), Wechat (Mobile), Wechat (PC), YouTube, Whatsapp, and Tiktok.
<p align='center'> <img src='https://github.com/paper1765/paper1765/blob/master/imgs/framework.jpg?raw=true' width='870'/> </p> <p align='center'> <em>The framework of ST-DDL.</em> </p>Usage
For pre-process the videos:
cd preprocess
python preprocess.py
For testing:
python test.py
Then the model will detect the videos in the data/ManualFake_preview/ and save the results in the res/ directory.
Note: The pre-trained weight can be downloaded from here.
<p align='center'> <img src='https://github.com/paper1765/paper1765/blob/master/imgs/result.jpg?raw=true' width='480'/> </p> <p align='center'> <em>Demo results of localization.</em> </p>Dataset
We construct a new Deepfake dataset, namely ManualFake, that innovatively features commercial Deepfake software, online social network scenarios, and multi-source videos (e.g., online interviews). The full dataset can be downloaded from Google Drive or Baidu Pan. If you would like to use our ManualFake dataset, please fill out this Google Form and, once accepted, we will send you the unzip code of our dataset.
<p align='center'> <img src='https://github.com/paper1765/paper1765/blob/master/imgs/dataset.jpg?raw=true' width='480'/> </p> <p align='center'> <em>The overview of ManualFake.</em> </p>License
The ManualFake dataset is for non-commercial research and educational purposes only. Before we provide your access to download, please put your ACADEMIC email address in Google Form and the download link will be sent to you once the form is accepted.