Watermarking Via Gaussian Noise Modulation in Diffusion Models

Abstract

The rapid advancement of AI-generated image technologies has posed significant challenges to digital content copyright protection. While watermarking remains a key solution, conventional post-generation watermarking methods often compromise image quality and create visible artifacts, negatively impacting user experience. To address these issues, we propose a novel plug-and-play watermarking framework that seamlessly integrates with diffusion-based generative models through Gaussian noise modulation.Our framework consists of three core components: a watermark embedder that encodes copyright information into the latent space while preserving the statistical properties of Gaussian noise, an image generator that encodes latent variables into watermarked images, and a robust extractor that utilizes diffusion model inversion to recover watermarks directly from the noise domain rather than the image domain.By leveraging Gaussian noise modulation and KL divergence optimization, the embedded watermark maintains the original Gaussian distribution characteristics, enabling high-fidelity image generation. Additionally, the integration of the Bidirectional Explicit Linear Multi-step (BELM) sampler significantly reduces inversion errors, further enhancing the stability and accuracy of both embedding and extraction processes.Extensive experiments demonstrate its superior accuracy and robustness against common attacks, including geometric transformations, photometric adjustments and quality degradation, making it a highly effective solution for copyright protection in AI-generated content.

Keywords: Watermarking, Diffusion model, Copyright Protection