Notice
Update 2024: CLRNet, ShallowNet, MesoInception4, and Xception weights are now available to download from the Google Drive link below.
update 2022: CLRNet Files and weights are temporarily removed. Contact the authors via email for access.
Overview
Title: One Detector to Rule Them All: Towards a General Deepfake Attack Detection Framework (WWW '21) (arXiv)
Citation
If you find our work useful for your research, please consider citing the following papers :)
@inproceedings{tariq2021web,
title={One Detector to Rule Them All: Towards a General Deepfake Attack Detection Framework},
author={Tariq, Shahroz and Lee, Sangyup and Woo, Simon S},
booktitle={Proceedings of The Web Conference 2021},
year={2021},
url = {https://doi.org/10.1145/3442381.3449809},
doi = {10.1145/3442381.3449809}
}Pretrained weights
The following link contains the weights for the models (CLRNet [CLR], ShallowNetV3 [SNV3], MesoInception4 [M14], and Xception [XCE]) used in our experiments
https://drive.google.com/drive/folders/1CE-HzZh76ejAsrIFSlbaEGmQHyzoj9EQ?usp=sharing
Additional Results
Updated in-domain attack results including DFDC dataset
- Note that CLRNet performs the best for DFDC dataset among all the test baselines.
Updated out-of-domain attack results (before using our defense strategy)
- Note that results from Table 5 demonstrates that models trained on DFDC, which is a quite generic and diverse dataset, can still fail to detect out-of-domain attack (see Table 5).
- See Table 6 in our paper, for defense performance against out-of-domain attack.
Dataset used for Evaluation
- Facial Reenactment
- Identity Swap
- Unknown
- DeepFake in the Wild (DFW) [Sample Links] [Source1] [Source2] [Source3]