MTOR explained: AI video detection specs and 97.76% AP rate

arXiv USA

Overview

This paper introduces MTOR, a multimodal AI-generated video detector that leverages video-level multimodal semantics and a novel concept of Temporal Over-Regularity (TOR) to identify AI-generated videos with high precision. Extensive experiments across five benchmarks covering 46 video generator variants show MTOR outperforming 16 representative baselines, achieving 97.76% AP, 97.92% AUC, and 90.24% ACC. TOR specifically targets the consistent temporal persistence and reduced variability characteristic of AI-generated videos.

In Depth

Key Findings

This research presents MTOR, an advanced multimodal detector for AI-generated videos that significantly improves detection accuracy and generalization. MTOR achieves this breakthrough by integrating video-level multimodal semantics with a novel feature called Temporal Over-Regularity (TOR). Across extensive experiments on five benchmarks encompassing 46 distinct video generator variants, MTOR outperformed 16 leading baseline detectors, securing an impressive 97.76% Average Precision (AP), 97.92% Area Under the Curve (AUC), and 90.24% Accuracy (ACC). A crucial aspect of MTOR’s success is TOR’s ability to specifically identify the consistent temporal persistence and reduced variability often observed in AI-generated content compared to real-world videos.

Technical / Clinical Details

MTOR’s superior detection capabilities are derived from two primary, complementary components:

• Video-Level Multimodal Semantics: MTOR comprehensively analyzes both visual and auditory information within a video, extracting high-level semantic features. This integrated analysis allows it to assess the overall coherence and semantic consistency of a video, discerning subtle discrepancies that differentiate human-made from AI-generated content. This approach goes beyond detecting localized visual artifacts, enabling the identification of narrative inconsistencies or unrealistic scenarios.
• Temporal Over-Regularity (TOR): This is a core novel contribution of MTOR. AI-generated videos frequently exhibit a lower degree of temporal variability and a more ‘smoothed’ or ‘predictable’ flow compared to authentic footage. The TOR component quantitatively analyzes this consistent temporal persistence and reduced variability, identifying it as a unique footprint of AI generation. It specifically evaluates statistical correlations between frames and the regularity of feature evolution over time.
• High-Precision Cross-Generator Performance: The robust evaluation against 46 generator variants across five benchmarks demonstrates MTOR’s exceptional generalization capabilities. This suggests its effectiveness against a wide range of current and potentially future generative models, indicating strong real-world applicability in combating deepfake videos.

Background & Context

The rapid advancements in deepfake technology have made AI-generated videos increasingly realistic, raising significant concerns about their potential misuse in spreading misinformation and disinformation. Maliciously crafted AI-generated videos can lead to social unrest, political manipulation, and reputational damage, among other severe consequences. In this context, the development of highly accurate and automated detection techniques for AI-generated videos is crucial for safeguarding the integrity of digital media and maintaining public trust. Previous detection methods often suffered from limited generalizability, frequently focusing on specific generation models or purely visual artifacts, leaving vulnerabilities for novel generation techniques.

Strategic Significance & Outlook

The introduction of MTOR represents a significant breakthrough in the field of AI-generated video detection, pushing the frontier of anti-deepfake technology. This system will serve as a powerful tool for various organizations combating the spread of misinformation, including social media platforms, news agencies, and national security bodies. Looking forward, there is potential to adapt MTOR’s technology for real-time video stream detection and to enhance its robustness against even more sophisticated AI generation techniques. Furthermore, this research opens new avenues for understanding the intrinsic characteristics of AI-generated content, thereby contributing significantly to the future development of comprehensive content authentication technologies.

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.

• Your email and selected fields are used only to deliver the newsletter.
• We never share your information with third parties.
• You can unsubscribe anytime via the link in each email.

See our Privacy Policy for details.

Agree & Continue ▶

Takes about a minute · Unsubscribe anytime