[Paper Review] DeepFakes: a New Threat to Face Recognition? Assessment and Detection

Pavel Korshunov, Sébastien Marcel|arXiv (Cornell University)|Dec 20, 2018

Digital Media Forensic DetectionComputer Science24 references490 citations

TL;DR

The paper creates a public Deepfake VidTIMIT-based dataset using GANs, demonstrates vulnerability of VGG/Facenet face recognition to swapped faces, and evaluates detection methods showing IQM+SVM best among baselines, with lip-sync methods failing.

ABSTRACT

It is becoming increasingly easy to automatically replace a face of one person in a video with the face of another person by using a pre-trained generative adversarial network (GAN). Recent public scandals, e.g., the faces of celebrities being swapped onto pornographic videos, call for automated ways to detect these Deepfake videos. To help developing such methods, in this paper, we present the first publicly available set of Deepfake videos generated from videos of VidTIMIT database. We used open source software based on GANs to create the Deepfakes, and we emphasize that training and blending parameters can significantly impact the quality of the resulted videos. To demonstrate this impact, we generated videos with low and high visual quality (320 videos each) using differently tuned parameter sets. We showed that the state of the art face recognition systems based on VGG and Facenet neural networks are vulnerable to Deepfake videos, with 85.62% and 95.00% false acceptance rates respectively, which means methods for detecting Deepfake videos are necessary. By considering several baseline approaches, we found that audio-visual approach based on lip-sync inconsistency detection was not able to distinguish Deepfake videos. The best performing method, which is based on visual quality metrics and is often used in presentation attack detection domain, resulted in 8.97% equal error rate on high quality Deepfakes. Our experiments demonstrate that GAN-generated Deepfake videos are challenging for both face recognition systems and existing detection methods, and the further development of face swapping technology will make it even more so.

Motivation & Objective

• Provide a publicly available dataset of GAN-based Deepfake videos for face swapping.
• Assess the vulnerability of state-of-the-art face recognition systems to Deepfakes.
• Evaluate baseline Deepfake detection methods and identify strengths and limitations.

Proposed method

• Generate Deepfakes using GAN-based face swapping on VidTIMIT to produce LQ (64x64) and HQ (128x128) videos.
• Evaluate VGG and Facenet face recognition performance on original and Deepfake videos.
• Test lip-sync inconsistency as an audio-visual detection baseline and several image-based quality/ML baselines (PCA/LDA, IQM, SVM).
• Provide open-source dataset and implementation for reproducibility.

Experimental results

Research questions

• RQ1 How vulnerable are current face recognition systems (VGG and Facenet) to GAN-based Deepfake face swaps?
• RQ2 Can existing detection approaches reliably distinguish Deepfake videos from genuine footage, and which features are most effective?
• RQ3 Does lip-sync based detection outperform image-based Quality of Information measures for detecting Deepfakes?
• RQ4 What is the impact of video quality (LQ vs HQ) on both recognition vulnerability and detection accuracy?

Key findings

• VGG and Facenet face recognition show high vulnerability to Deepfake videos, with FAR of 85.62% (HQ) and 88.75% (LQ) for VGG, and 95.00% (HQ) and 94.38% (LQ) for Facenet at the EER threshold from licit data.
• Lip-sync based detection fails to distinguish Deepfakes from originals, indicating limitations of audio-visual inconsistency approaches.
• IQM-based features with an SVM classifier detect HQ Deepfakes with an 8.97% EER and 9.05% FRR@FAR10%, while HQ performance is notably better than LQ.
• Image quality metric (IQM) based methods with PCA/LDA show varying effectiveness, e.g., IQM+PCA+LDA yields 20.52% EER (LQ) and 66.67% FRR@FAR10%.
• Overall, baseline detectors are less effective against HQ Deepfakes, underscoring the need for more robust detection methods and datasets.