Home

Awesome

GBDF: Gender Balanced DeepFake Dataset

<img src="Fig1.png" width="400" align="right" >

A publicly available gender-balanced annotated deepfake dataset, GBDF, from FaceForensics++ (FF++), Celeb-DF, and Deeper Forensics-1.0 consisting of 10,000 live and fake videos generated using different identity and expression swapping deepfake generation techniques. The dataset consist of 10,000 videos with 5000 each for males and females with 1:4 real to fake ratio.

Version 1.0 (07.17.2022)

Table of Contents

Abstract

Facial forgery by deepfakes has raised severe societal concerns. Several solutions have been proposed by the vision community to effectively combat the misinformation on the internet via automated deepfake detection systems. Recent studies have demonstrated that facialanalysis based deep learning models can discriminate based on protected attributes such as gender and race. For the commercial adoption and massive roll-out of the deepfake detection technology, it is vital to evaluate and understand the fairness (the absence of any prejudice or favoritism) of deepfake detectors across demographic variations (protected attributes). As the performance differential of deepfake detectors between demographic sub-groups would impact millions of people of the deprived sub-group. This paper aims to evaluate the fairness of the deepfake detectors across males and females. However, existing deepfake datasets are not annotated with demographic labels to facilitate fairness analysis. To this aim, we manually annotated existing popular deepfake datasets with gender labels and evaluated the performance differential of current deepfake detectors across gender. Our analysis on the gender-labeled version of the datasets suggests (a) current deepfake datasets have skewed distribution across gender, and (b) commonly adopted deepfake detectors obtain unequal performance across gender with mostly males outperforming females. Finally, we contributed a gender-balanced and annotated deepfake dataset, GBDF, to mitigate the performance differential and to promote research and development towards fairness-aware deep fake detectors For more details, please take a look at the Research Paper.

Database Properties

The GBDF dataset is created using FF++(c23 version), Celeb-DF, DeeperForensics-1.0 and consist of 10,000 videos with 5000 each for males and females.As none of these existing deepfake datasets contain demographic information, we manually annotated ground truth gender labels for these datasets. The real/fake videos from these deepfakes datasets are merged to create GBDF dataset.Deepfakes in the GBDF dataset are created using different Identity Swapping (i.e., FaceSwap, FaceSwap-Kowalski, FaceShifter, Encoder-decoder style and End- to-end Face Swapping techniques) and Expression swapping (i.e., Face2Face and NeuralTextures) deepfake generation techniques. The majority of the videos in GBDF are from Caucasians. The ratio of real to fake videos in the GBDF dataset is 1 : 4. The GBDF is further divided into gender-balanced and subject independent training and testing subsets in the ratio of 70 : 30.

<img src="Fig2.png" width="700" >

Download

GBDF provides gender annotations for deepfakes of the FaceForesincs++ database,Celeb-DF database,Deeper Forensics-1.0 database.

Grad-CAM Results

GRAD-CAM visualization of our best performing EfficientNet V2-L based deepfake detector for live and fake images for males and females is shown. The detector was trained on GBDF dataset, For real images highly activated region is the cheek for females and the ocular region for males. For fake images, the mouth and cheek region for males and the complete face region for females are the most activated region. These results were consistent across the datasets depending on the deepfake generation technique. <img src="Fig3.png" width="700" >

Citing

If you use this work, please cite the following papers.

@misc{https://doi.org/10.48550/arxiv.2207.10246,
  doi = {10.48550/ARXIV.2207.10246},
  
  url = {https://arxiv.org/abs/2207.10246},
  
  author = {Nadimpalli, Aakash Varma and Rattani, Ajita},
  
  keywords = {Computer Vision and Pattern Recognition (cs.CV), Artificial Intelligence (cs.AI), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {GBDF: Gender Balanced DeepFake Dataset Towards Fair DeepFake Detection},
  
  publisher = {arXiv},
  
  year = {2022},
  
  copyright = {arXiv.org perpetual, non-exclusive license}
}


@inproceedings{roessler2019faceforensicspp,
	author = {Andreas R\"ossler and Davide Cozzolino and Luisa Verdoliva and Christian Riess and Justus Thies and Matthias Nie{\ss}ner},
	title = {Face{F}orensics++: Learning to Detect Manipulated Facial Images},
	booktitle= {International Conference on Computer Vision (ICCV)},
	year = {2019}
}



@inproceedings{jiang2020deeperforensics1,
  title={{DeeperForensics-1.0}: A Large-Scale Dataset for Real-World Face Forgery Detection},
  author={Jiang, Liming and Li, Ren and Wu, Wayne and Qian, Chen and Loy, Chen Change},
  booktitle={CVPR},
  year={2020}
}

@inproceedings{Celeb_DF_cvpr20,
   author = {Yuezun Li, Xin Yang, Pu Sun, Honggang Qi and Siwei Lyu},
   title = {Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics},
   booktitle= {IEEE Conference on Computer Vision and Patten Recognition (CVPR)},
   year = {2020}
}

Acknowledgment

This work is supported in part from National Science Foundation (NSF) award no. 2129173. The research infrastructure used in this study is supported in part from a grant no. 13106715 from the Defense University Research Instrumentation Program (DURIP) from Air Force Office of Scientific Research.

License

This project is licensed under the terms of the Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license. All images used in this project belongs to the FF++,celeb-DF,DeeperForensics-1.0. The copyright of the images remains with the original owners. The copyright of the annotations remains with the VCBSL:Visual Computing and Biometric Security Lab, Wichita State University 2022.