# DeepFakes: a New Threat to Face Recognition? Assessment and Detection

**Authors:** Pavel Korshunov, Sebastien Marcel  
**arXiv:** [1812.08685](https://arxiv.org/abs/1812.08685)  
**Date:** 2018-12-20  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.08685v1/

## Key points

- Ease of automatically replacing a person's face in videos using GANs has raised concerns about Deepfake videos
- Deepfake videos involve swapping faces or altering facial expressions
- Public scandals have involved celebrities' faces being superimposed onto pornographic content
- The paper introduces the first publicly available set of Deepfake videos generated from the VidTIMIT database
- Training and blending parameters significantly impact the quality of Deepfake videos
- Face recognition systems relying on VGG and Facenet neural networks are vulnerable to Deepfakes, with false acceptance rates of 85.62% and 95.00% respectively
- Lip-sync inconsistency detection is ineffective in distinguishing Deepfake videos
- Visual quality metrics used in presentation attack detection domains achieved an equal error rate of 8.97% when applied to high-quality Deepfakes
- GAN-generated Deepfake videos pose significant challenges for face recognition systems and existing detection methods
- Further research and development are crucial to combat the threat posed by Deepfakes

## Summary

In recent years, the ease of automatically replacing a person's face in a video with someone else's face using pre-trained generative adversarial networks (GANs) has raised concerns about the spread of Deepfake videos. These videos involve swapping faces onto different bodies or altering facial expressions and have been implicated in public scandals such as celebrities' faces being superimposed onto pornographic content. To address this growing issue and develop methods for detecting Deepfake videos, this paper introduces the first publicly available set of Deepfake videos generated from the VidTIMIT database. The authors utilized open source software based on GANs to create these Deepfakes and emphasized that training and blending parameters significantly impact the quality of the resulting videos. To demonstrate this impact, they generated two sets of 320 videos each with low and high visual quality by tuning different parameter sets. The study revealed that state-of-the-art face recognition systems relying on VGG and Facenet neural networks are vulnerable to Deepfake videos, exhibiting false acceptance rates of 85.62% and 95.00%, respectively. This highlights the urgent need for effective methods to detect Deepfake videos. Several baseline approaches were considered to identify potential detection methods; however an audio-visual approach based on lip-sync inconsistency detection proved ineffective in distinguishing Deepfake videos. On the other hand, a method based on visual quality metrics commonly used in presentation attack detection domains achieved an equal error rate of 8.97% when applied to high-quality Deepfakes. Overall, the experiments conducted in this study demonstrate that GAN-generated Deepfake videos pose significant challenges for both face recognition systems and existing detection methods. Furthermore, as face-swapping technology continues to advance detecting these manipulated videos will become even more difficult; therefore further research and development are crucial to effectively combat the threat posed by Deepfakes to face recognition systems and society at large.