SimuFreeMark: A Noise-Simulation-Free Robust Watermarking Against Image Editing

TL;DR

SimuFreeMark addresses the need for robust image watermarking under both conventional processing and AIGC semantic edits by abandoning hand-crafted attack simulations. It relies on the inherent stability of low-frequency image components and binds the watermark to deep semantic features via a frozen SD-VAE, enabling a simulation-free training paradigm. The approach is validated with extensive experiments showing superior robustness across a range of attacks and high visual fidelity, outperforming state-of-the-art methods. This simulation-free framework offers a generalizable path for watermarking in dynamic editing environments and suggests extending stability-based ideas to geometry-invariant representations in the future.

Abstract

The advancement of artificial intelligence generated content (AIGC) has created a pressing need for robust image watermarking that can withstand both conventional signal processing and novel semantic editing attacks. Current deep learning-based methods rely on training with hand-crafted noise simulation layers, which inherently limit their generalization to unforeseen distortions. In this work, we propose $\textbf{SimuFreeMark}$, a noise-$\underline{\text{simu}}$lation-$\underline{\text{free}}$ water$\underline{\text{mark}}$ing framework that circumvents this limitation by exploiting the inherent stability of image low-frequency components. We first systematically establish that low-frequency components exhibit significant robustness against a wide range of attacks. Building on this foundation, SimuFreeMark embeds watermarks directly into the deep feature space of the low-frequency components, leveraging a pre-trained variational autoencoder (VAE) to bind the watermark with structurally stable image representations. This design completely eliminates the need for noise simulation during training. Extensive experiments demonstrate that SimuFreeMark outperforms state-of-the-art methods across a wide range of conventional and semantic attacks, while maintaining superior visual quality.

Figures (4)

• Figure 1: The network comparison of our framework and other frameworks.
• Figure 2: Stability comparison of frequency components.
• Figure 3: The flowchart of SimuFreeMark.
• Figure 4: Visual quality comparison with SOTA methods.