VGMShield: Mitigating Misuse of Video Generative Models

Yan Pang, Yang Zhang, Tianhao WangarXiv 2024 · DOI · OpenReview

Abstract

The rapid advancement of Video Generative Models (VGMs) has introduced significant societal risks, including the proliferation of photorealistic misinformation and deepfakes. This paper introduces VGMShield, a multi-layered defense framework designed to detect, trace, and prevent the misuse of AI-generated video content. Unlike traditional image-based detection, VGMShield leverages spatial-temporal dynamics—such as motion trajectories and frame-to-frame consistency—as a unique "fingerprint" for identification. Our evaluation across multiple open-source and commercial models demonstrates that VGMShield achieves over 90% accuracy in both detection and source attribution, while providing a proactive defense mechanism against unauthorized image-to-video synthesis.

Introduction

The evolution of video generation, from early research prototypes to sophisticated systems like OpenAI’s Sora, has democratized high-quality video production. However, this capability is a double-edged sword, enabling the creation of deceptive content ranging from fabricated news events to non-consensual deepfakes. Current forensic tools often fail to account for the temporal dimension, focusing instead on static artifacts that are increasingly minimized by modern generators. VGMShield addresses this by treating video security as a lifecycle challenge, encompassing the detection of fakes, attribution to specific models, and proactive prevention.

Methodology

VGMShield treats detection and tracing as classification tasks powered by advanced video recognition backbones. The framework evaluates three primary architectures: I3D (3D CNN), X-CLIP (cross-frame attention), and Video Masked Autoencoders (MAE). MAE proved most effective due to its reconstruction-based training, which necessitates a deep understanding of complex temporal relationships rather than just memorizing static textures.

The VGMShield framework utilizes a video detection model (W) to distinguish between real and fake videos generated by model (G), leveraging key frame selection and captioning.

The system identifies three core spatial-temporal weaknesses in current VGMs that serve as detectable fingerprints:

  • Attention Shifts: Inconsistencies in spatial focus between adjacent frames leading to flickering.
  • Motion Variations: Deviations from natural physical trajectories and struggles with maintaining "stillness."
  • Frequency Fluctuations: Irregular high-frequency noise patterns that differ significantly from the natural noise of real camera sensors.

Results and Analysis

The framework was tested against a diverse set of models using the OpenVid-1M and InternVid datasets. A key finding is the existence of distinct "motion fingerprints" for different generative architectures, allowing for precise model attribution.

Motion PCA comparison showing distinct clusters for different generative models (e.g., Hunyuan, VideoCrafter, VGen) compared to real video, illustrating unique spatial-temporal fingerprints.

VGMShield achieved over 90% accuracy in targeted detection. Even in data-agnostic scenarios, the MAE-based model maintained high performance in tracing videos back to their parent models. Furthermore, the prevention module—utilizing adversarial perturbations—successfully disrupted video synthesis in commercial models like Gen-2. By adding invisible noise to source images, the defense successfully immobilized objects or created logical inconsistencies, rendering the generated video useless for misinformation purposes.

Conclusion and Future Work

VGMShield represents a significant step toward robust video forensics by prioritizing temporal coherence. While the industry transition toward Diffusion Transformers (DiT) presents new challenges for fingerprinting, the focus on spatial-temporal dynamics remains a promising avenue for future-proofing AI defense systems. Future research will focus on scaling these detection models to handle the massive volume of real-time social media uploads and improving resistance to "purification" attacks like re-encoding.

Summary updated

Abstract

With the rapid advancement in video generation, people can conveniently utilize video generation models to create videos tailored to their specific desires. Nevertheless, there are also growing concerns about their potential misuse in creating and disseminating false information. In this work, we introduce VGMShield: a set of three straightforward but pioneering mitigations through the lifecycle of fake video generation. We start from \textit{fake video detection} trying to understand whether there is uniqueness in generated videos and whether we can differentiate them from real videos; then, we investigate the \textit{tracing} problem, which maps a fake video back to a model that generates it. Towards these, we propose to leverage pre-trained models that focus on {\it spatial-temporal dynamics} as the backbone to identify inconsistencies in videos. Through experiments on seven state-of-the-art open-source models, we demonstrate that current models still cannot perfectly handle spatial-temporal relationships, and thus, we can accomplish detection and tracing with nearly perfect accuracy. Furthermore, anticipating future generative model improvements, we propose a {\it prevention} method that adds invisible perturbations to images to make the generated videos look unreal. Together with fake video detection and tracing, our multi-faceted set of solutions can effectively mitigate misuse of video generative models.
BibTeX
@misc{pang2024vgmshield,
  author = {Yan Pang and Yang Zhang and Tianhao Wang},
  title = {{VGMShield: Mitigating Misuse of Video Generative Models}},
  year = {2024},
  eprint = {2402.13126},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.13126},
}
Figure