Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
41 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
41 tokens/sec
o3 Pro
7 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

VGMShield: Mitigating Misuse of Video Generative Models (2402.13126v1)

Published 20 Feb 2024 in cs.CR, cs.AI, cs.CV, cs.LG, and eess.IV

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.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (3)
  1. Yan Pang (21 papers)
  2. Yang Zhang (1129 papers)
  3. Tianhao Wang (98 papers)
Citations (3)
Reddit Logo Streamline Icon: https://streamlinehq.com