[Paper Review] AVTENet: A Human-Cognition-Inspired Audio-Visual Transformer-Based Ensemble Network for Video Deepfake Detection

Ammarah Hashmi, Sahibzada Adil Shahzad+3 more|arXiv (Cornell University)|Oct 19, 2023

Digital Media Forensic Detection5 citations

TL;DR

This paper proposes AVTENet, a novel audio-visual transformer-based ensemble network that leverages multiple pure Transformer models to detect deepfake videos by analyzing visual, audio, and audio-visual cues in tandem. It achieves state-of-the-art performance with 99% accuracy on the FakeAVCeleb Testset-II, outperforming existing uni-modal, multi-modal, and ensemble methods by effectively fusing modality-specific representations through ensemble learning.

ABSTRACT

The recent proliferation of hyper-realistic deepfake videos has drawn attention to the threat of audio and visual forgeries. Most previous studies on detecting artificial intelligence-generated fake videos only utilize visual modality or audio modality. While some methods exploit audio and visual modalities to detect forged videos, they have not been comprehensively evaluated on multimodal datasets of deepfake videos involving acoustic and visual manipulations, and are mostly based on convolutional neural networks with low detection accuracy. Considering that human cognition instinctively integrates multisensory information including audio and visual cues to perceive and interpret content and the success of transformer in various fields, this study introduces the audio-visual transformer-based ensemble network (AVTENet). This innovative framework tackles the complexities of deepfake technology by integrating both acoustic and visual manipulations to enhance the accuracy of video forgery detection. Specifically, the proposed model integrates several purely transformer-based variants that capture video, audio, and audio-visual salient cues to reach a consensus in prediction. For evaluation, we use the recently released benchmark multimodal audio-video FakeAVCeleb dataset. For a detailed analysis, we evaluate AVTENet, its variants, and several existing methods on multiple test sets of the FakeAVCeleb dataset. Experimental results show that the proposed model outperforms all existing methods and achieves state-of-the-art performance on Testset-I and Testset-II of the FakeAVCeleb dataset. We also compare AVTENet against humans in detecting video forgery. The results show that AVTENet significantly outperforms humans.

Motivation & Objective

• To address the growing threat of hyper-realistic audio-visual deepfakes that manipulate both voice and face in videos.
• To overcome the limitations of existing uni-modal deepfake detection methods that fail to detect cross-modal manipulations.
• To develop a multi-modal, Transformer-based ensemble framework capable of detecting both visual and acoustic forgeries simultaneously.
• To evaluate the model on a comprehensive benchmark dataset that includes diverse manipulation techniques across audio and video streams.
• To establish a scalable, self-supervised learning-compatible architecture for future deepfake detection research.

Proposed method

• The framework integrates three distinct Transformer-based models: an audio-specific model (AST-based AN), a video-specific model (ViViT-based VN), and an audio-visual fusion model (AV-HuBERT-based AVN) to extract modality-specific representations.
• Each component model is pre-trained on large-scale self-supervised datasets to capture discriminative spatio-temporal and spectral features from audio and video streams.
• The ensemble strategy combines predictions from the three models via early fusion and consensus voting to improve robustness and detection accuracy.
• The AVN component uses AV-HuBERT to extract joint audio-visual embeddings, though the authors note it underperforms due to limited visual context capture outside the lip region.
• The model is trained and evaluated on the FakeAVCeleb dataset, which contains diverse deepfake manipulations including face swapping, fsgan, and RTVC.
• Hyperparameter tuning and ablation studies are conducted to analyze the contribution of each component and optimize the ensemble performance.

Figure 1: An example of audio-visual deepfakes. Cyber-attackers aim to create convincing counterfeit videos by manipulating audio and visual streams. Sharing such malicious content can have harmful consequences. Therefore, our proposed method exploits manipulations in both audio and visual streams t

Experimental results

Research questions

• RQ1 Can a multi-modal Transformer-based ensemble model outperform existing uni-modal and multi-modal deepfake detection methods on complex, real-world audio-visual deepfake datasets?
• RQ2 How does the integration of purely Transformer-based models for audio, video, and audio-visual fusion improve detection accuracy compared to CNN-based baselines?
• RQ3 To what extent do different pre-trained models (e.g., AST, ViViT, AV-HuBERT) contribute to detecting various types of audio-visual manipulations such as face swapping and voice cloning?
• RQ4 Why does the AV-HuBERT-based AVN component underperform on visually manipulated fake videos, and how can this be mitigated?
• RQ5 Can an ensemble of specialized Transformer models achieve superior generalization and robustness across diverse deepfake generation techniques?

Key findings

• AVTENet achieves a state-of-the-art detection accuracy of 99% on Testset-II of the FakeAVCeleb dataset, significantly outperforming existing methods.
• The model outperforms all baseline uni-modal and multi-modal approaches, including VGG16 (81%), Xception (76%), and Lip Forensics (76%), on the same benchmark.
• The ensemble model with AST, ViViT, and AV-HuBERT components (AVTENet_ff) achieves the highest performance, demonstrating the effectiveness of multi-expert learning.
• The AVN component based on AV-HuBERT underperforms on visually manipulated videos (e.g., faceswap) because it fails to capture visual cues outside the lip region.
• Replacing AV-HuBERT with AST and ViViT in the AVN component leads to worse performance, indicating redundancy and lack of complementary information in the fused model.
• The ablation study confirms that the ensemble strategy enhances robustness and generalization across diverse manipulation types, including audio-only and visual-only forgeries.

Figure 2: Illustrations of the audio and visual streams of some samples selected from the FakeAVCeleb Dataset, where real streams are indicated in green, and fake ones are indicated in red.