[Paper Review] Practical Deepfake Detection: Vulnerabilities in Global Contexts

Yang A. Chuming, Daniel J. Wu+1 more|arXiv (Cornell University)|Jun 20, 2022

Generative Adversarial Networks and Image Synthesis4 citations

TL;DR

This paper proposes a video corruption pipeline to simulate real-world quality degradation in low-bandwidth environments, demonstrating that state-of-the-art deepfake detection models fail on corrupted videos—even when the original video is authentic. The key finding is that models confidently classify legitimate videos as fake after applying CRF50 or datamoshing, revealing critical vulnerabilities in global deployment contexts.

ABSTRACT

Recent advances in deep learning have enabled realistic digital alterations to videos, known as deepfakes. This technology raises important societal concerns regarding disinformation and authenticity, galvanizing the development of numerous deepfake detection algorithms. At the same time, there are significant differences between training data and in-the-wild video data, which may undermine their practical efficacy. We simulate data corruption techniques and examine the performance of a state-of-the-art deepfake detection algorithm on corrupted variants of the FaceForensics++ dataset. While deepfake detection models are robust against video corruptions that align with training-time augmentations, we find that they remain vulnerable to video corruptions that simulate decreases in video quality. Indeed, in the controversial case of the video of Gabonese President Bongo's new year address, the algorithm, which confidently authenticates the original video, judges highly corrupted variants of the video to be fake. Our work opens up both technical and ethical avenues of exploration into practical deepfake detection in global contexts.

Motivation & Objective

• To investigate the practical robustness of deepfake detection models under real-world video quality degradation common in low-bandwidth regions.
• To simulate video corruptions such as bitrate reduction, resolution downscaling, CRF adjustment, and datamoshing to mimic real-world video quality loss.
• To evaluate the performance of a state-of-the-art deepfake detection model on corrupted versions of the FaceForensics++ dataset.
• To explore ethical risks of model failure in regions with limited technological infrastructure and lower digital literacy.
• To use the Gabon presidential address case as a real-world example of how corruption can cause false detection of authentic videos as fake.

Proposed method

• Developed a video corruption pipeline applying five distinct video degradation techniques: bitrate reduction (1.0 and 0.5 Mbps), resolution downscaling (720p, 480p, 240p), constant rate factor (CRF) adjustment (30, 40, 50), and datamoshing.
• Applied these corruptions independently and in combination to both real and fake videos from the FaceForensics++ dataset.
• Used the top-performing model from the 2020 Kaggle Deepfake Detection Challenge—an ensemble of 7 EfficientNet-B7 networks—trained on a different dataset to ensure out-of-domain evaluation.
• Evaluated model performance via classification accuracy on corrupted videos, measuring the percentage of deepfakes correctly identified as fake.
• Conducted ablation studies to isolate the impact of individual corruptions, particularly CRF50 and datamoshing, on detection confidence.
• Used the Gabonese presidential address video as a case study, applying the same corruptions to assess model behavior on a real-world authentic video.

Experimental results

Research questions

• RQ1 How do common video quality degradations—such as low bitrate, low resolution, high CRF, and datamoshing—affect the performance of state-of-the-art deepfake detection models?
• RQ2 To what extent do video corruptions that do not align with training-time data augmentations undermine model reliability?
• RQ3 Can a legitimate video be misclassified as fake by a deepfake detection model after applying real-world-like video corruptions?
• RQ4 What are the sociopolitical risks of model failure in regions with poor internet infrastructure and lower technological literacy?
• RQ5 How do inorganic corruptions like datamoshing and high CRF affect both model confidence and human perceptual similarity to authentic content?

Key findings

• The deepfake detection model achieved 95.5% accuracy on uncorrupted deepfakes from FaceForensics++, confirming its strong baseline performance.
• After downscaling to 240p and applying CRF50, the model’s accuracy dropped to 58.6% on deepfakes, indicating a significant failure under quality degradation.
• When combined with datamoshing, the model’s accuracy on corrupted deepfakes dropped to 69.0%, showing that complex corruptions amplify model vulnerability.
• The Gabonese presidential address video—originally classified as authentic with 98.8% confidence—was misclassified as fake with 56.5% confidence after 240p + CRF50 corruption.
• The model’s confidence in authentic videos decreased dramatically under high-CRF and datamoshing corruptions, even though the videos remained semantically recognizable to humans.
• Corruptions such as CRF50 and datamoshing, which lack direct analogs in standard data augmentation, caused the most severe performance degradation, indicating a critical gap in model robustness.