# WildDeepfake: A Challenging Real-World Dataset for Deepfake Detection

**Authors:** Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, Yu-Gang Jiang  
**arXiv:** [2101.01456](https://arxiv.org/abs/2101.01456)  
**Date:** 2021-01-05  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2101.01456v1/

## Key points

- Deepfake technology has raised significant public concern due to its potential for misuse and manipulation
- Deepfakes are created using AI algorithms to superimpose one person's face onto another's body, leading to the creation of misleading or fabricated content
- Researchers are developing detection systems to identify and mitigate the spread of deepfakes
- WildDeepfake dataset, introduced by a team of researchers, comprises 7,314 face sequences extracted from 707 deepfake videos collected entirely from online sources
- WildDeepfake aims to provide a more challenging testbed for evaluating the performance of deepfake detectors against authentic deceptive content found on the web
- The researchers proposed novel detection networks that leverage attention masks on real and fake faces to enhance detection accuracy
- Empirical testing demonstrated the effectiveness of ADDNets in improving detection performance against sophisticated real-world deepfakes

## Summary

In recent years, the proliferation of deepfake technology has sparked significant public concern due to its potential for misuse and manipulation. <br>
 refers to the creation of realistic-looking videos using artificial intelligence algorithms that can superimpose one person's face onto another's body. This often leads to the creation of misleading or fabricated content. To address this growing issue, researchers have been developing  to identify and mitigate the spread of such deceptive media. One promising approach to combatting deepfakes is through the development of . Several datasets, such as DeepfakeDetection and FaceForensics++, have been released to facilitate the training and evaluation of these detection systems. However, existing datasets primarily consist of videos filmed with volunteer actors in controlled settings, limiting their ability to effectively detect real-world deepfakes circulating on the internet. To bridge this gap, a team of researchers led by Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang introduced a new dataset called . This dataset comprises 7,314 face sequences extracted from 707 deepfake videos collected entirely from online sources. By incorporating real-world data into the training process, WildDeepfake aims to provide a more challenging testbed for evaluating the performance of deepfake detectors against authentic deceptive content found on the web. The researchers conducted a comprehensive evaluation comparing baseline detection networks on both existing datasets and WildDeepfake. Their analysis revealed that WildDeepfake poses greater challenges for detection algorithms compared to traditional datasets, highlighting the need for improved strategies to combat evolving deepfake techniques. In response to these challenges, they proposed two novel  that leverage attention masks on real and fake faces to enhance detection accuracy. Through empirical testing on various datasets including WildDeepfake, the researchers demonstrated the effectiveness of ADDNets in improving detection performance against sophisticated real-world deepfakes. The availability of WildDeepfake dataset on GitHub provides a valuable resource for researchers and developers seeking to advance deepfake detection technologies and safeguard against malicious use cases in online environments.