Leveraging Pre-Trained Visual Models for AI-Generated Video Detection

View on arXiv ← Back to list

Authors: Keerthi Veeramachaneni, Praveen Tirupattur, Amrit Singh Bedi, Mubarak Shah

Published: 2025-07-17 15:36:39+00:00

AI Summary

This paper addresses the critical challenge of detecting generic AI-generated videos, extending beyond traditional deepfake detection which primarily focuses on human faces. The authors propose an approach leveraging features extracted from pre-trained visual models to distinguish between real and AI-generated video content. This method achieves high detection performance, which can be further improved by training a simple linear classification layer.

Abstract

Recent advances in Generative AI (GenAI) have led to significant improvements in the quality of generated visual content. As AI-generated visual content becomes increasingly indistinguishable from real content, the challenge of detecting the generated content becomes critical in combating misinformation, ensuring privacy, and preventing security threats. Although there has been substantial progress in detecting AI-generated images, current methods for video detection are largely focused on deepfakes, which primarily involve human faces. However, the field of video generation has advanced beyond DeepFakes, creating an urgent need for methods capable of detecting AI-generated videos with generic content. To address this gap, we propose a novel approach that leverages pre-trained visual models to distinguish between real and generated videos. The features extracted from these pre-trained models, which have been trained on extensive real visual content, contain inherent signals that can help distinguish real from generated videos. Using these extracted features, we achieve high detection performance without requiring additional model training, and we further improve performance by training a simple linear classification layer on top of the extracted features. We validated our method on a dataset we compiled (VID-AID), which includes around 10,000 AI-generated videos produced by 9 different text-to-video models, along with 4,000 real videos, totaling over 7 hours of video content. Our evaluation shows that our approach achieves high detection accuracy, above 90% on average, underscoring its effectiveness. Upon acceptance, we plan to publicly release the code, the pre-trained models, and our dataset to support ongoing research in this critical area.

Key findings
The proposed method achieved high detection accuracy, averaging above 90%, for generic AI-generated video detection across both open-source and closed-source models. However, a reduced accuracy was observed when detecting videos from the latest closed-source generation models, highlighting the increasing realism and advanced capabilities of these generative technologies.
Approach
The approach involves extracting features from pre-trained visual models (SigLIP and VideoMAE) that have learned the distribution of real visual content. These features are then utilized in either a training-free, distance-based classification method, or to train a parameter-efficient linear classification layer for improved detection performance.
Datasets
VID-AID (a newly compiled dataset including 10,000 AI-generated videos from 9 text-to-video models: ModelScope, Text2Video-Zero, Zeroscope, Latte, OpenSora, Sora, Veo, DreamMachine, Video Poet, and 4,000 real videos from YouTube-VOS). GenVideo was also used for comparative evaluation in the supplementary material.
Model(s)
UNKNOWN
Author countries
United States
← Previous
Next →