DiffusionShield: A Watermarking Approach to Safeguarding Video Integrity Against Stable Diffusion

Zuzanna Gawrysiak, Mateusz Gabor, Tomasz Hawro

Keywords: Forgery detection, Video watermarking, AI tampering, Diffusion models, Media forensics

Abstract: To address the critical safety and reliability challenges posed by the rapid advancement of Deep Generative Models (DGMs), we propose a principled video watermarking strategy designed to withstand diffusion-based tampering. While traditional forensic methods fail against the semantic reconstruction capabilities of modern DGMs, our approach bridges this gap by integrating a computationally efficient architecture with a differentiable attack simulation layer. We demonstrate the efficacy of this attention-guided framework through extensive experiments, achieving state-of-the-art fidelity (38.9 dB PSNR) and precise localization of Stable Diffusion inpainting (F1 0.85). This work provides a robust, scalable mechanism for ensuring media integrity, directly supporting the responsible deployment of generative AI.

Submission Number: 16