Spread Spectrum Image Watermarking Through Latent Diffusion Model.

Abstract

The rapid development of diffusion models in image generation and processing has led to significant security concerns. Diffusion models are capable of producing highly realistic images that are indistinguishable from real ones. Although deploying a watermarking system can be a countermeasure to verify the ownership or the origin of images, the regeneration attacks arising from diffusion models can easily remove the embedded watermark from the images, without compromising their perceptual quality. Previous watermarking methods that hide watermark information in the carrier image are vulnerable to these newly emergent attacks. To address these challenges, we propose a robust and traceable watermark framework based on the latent diffusion model, where the spread-spectrum watermark is coupled with the diffusion noise to ensure its security and imperceptibility. Since the diffusion model is trained to reduce information entropy from disordered data to restore its true distribution, the transparency of the hidden watermark is guaranteed. Benefiting from the spread spectrum strategy, the decoder structure is no longer needed for watermark extraction, greatly alleviating the training overhead. Additionally, the robustness and transparency are easily controlled by a strength factor, whose operating range is studied in this work. Experimental results demonstrate that our method performs not only against common attacks, but also against regeneration attacks and semantic-based image editing.