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

Yichao Tang, Mingyang Li, Di Miao, Sheng Li, Zhenxing Qian, Xinpeng ZhangarXiv 2025 · DOI · arXiv

Abstract

SimuFreeMark is a novel image watermarking framework that achieves state-of-the-art robustness against both conventional signal processing and AI-generated semantic edits without relying on traditional noise simulation during training. By embedding watermarks into the deep feature space of an image's inherently stable low-frequency components using a pre-trained VAE, the method generalizes better to unforeseen attacks while maintaining exceptional visual quality.

Introduction

The proliferation of Artificial Intelligence Generated Content (AIGC) has revolutionized image editing, enabling seamless object replacement, inpainting, and style transfer. However, these tools also provide malicious actors with powerful means to remove or destroy copyright watermarks. Conventional deep learning-based watermarking methods attempt to counter these threats using "noise simulation layers" during training. By applying simulated attacks (e.g., JPEG compression, blurring) within the training loop, models learn to resist specific distortions.

This paradigm suffers from a significant limitation: it is impossible to simulate every potential attack, especially as AIGC techniques evolve. Consequently, these models often overfit to known noises and fail against novel semantic distortions. Furthermore, complex noise layers increase computational overhead and complicate the training process.

Methodology

SimuFreeMark addresses these challenges by abandoning noise simulation entirely. Instead, it leverages the inherent physical and semantic stability of an image's low-frequency components.

Framework Architecture

The SimuFreeMark pipeline consists of several key stages:

  1. Frequency Decomposition: The input image $I_{ori}$ is decomposed using a 2D Fast Fourier Transform (FFT) to isolate low-frequency components ($I_{low}$), typically the central 20% of the spectrum.
  2. Latent Space Embedding: $I_{low}$ is processed by a frozen, pre-trained Stable Diffusion Variational Autoencoder (SD-VAE). The watermark $w$ is embedded into this semantic latent space via a multi-stage feature fusion network.
  3. Reconstruction: Watermarked latent features are decoded back to the spatial domain and combined with the original high-frequency components ($I_{high}$) to produce the watermarked image $I^w$.
  4. Extraction: To retrieve the watermark, the attacked image is decomposed via FFT, and a dedicated decoder extracts the bit sequence $w_{ext}$ from the low-frequency spectrum.
Detailed architecture of the SimuFreeMark framework, highlighting the frequency decomposition and VAE-based embedding process.

Training Strategy

The model employs a two-stage training process without any attack simulations. The first stage focuses on maximizing extraction accuracy, while the second stage fine-tunes the system for visual fidelity using a weighted combination of $L_2$, SSIM, and Just Noticeable Difference (JND) perceptual losses.

Experimental Setup

The model was trained on 20,000 images from the COCO dataset (resized to 512×512). Evaluation of AIGC semantic robustness utilized 1,000 images from the UltraEdit dataset, which provides standardized source images, editing instructions, and masks for reproducible testing.

Results and Analysis

SimuFreeMark was compared against several state-of-the-art (SOTA) baselines, including HiDDeN, PIMoG, RoSteALS, FIN, EditGuard, and Robust-Wide.

Quantitative Performance

As shown in the table below, SimuFreeMark achieves superior visual quality with a PSNR of 39.19 dB and an SSIM of 0.990. It maintains exceptional robustness across a wide range of attacks. Notably, it achieves over 99% bit accuracy against Gaussian and salt-and-pepper noise and remains highly resilient to JPEG compression even at low quality factors ($q=10$).

Quantitative comparison of SimuFreeMark against SOTA methods across various metrics and attack scenarios.

Semantic Robustness and Generalization

A key insight of this work is that low-frequency components are naturally shielded from AIGC edits. Because generative models must maintain global structural coherence to produce plausible images, they tend to preserve low-frequency data. By binding the watermark to deep structural features in the VAE latent space, SimuFreeMark survives complex semantic rewrites—such as diffusion-based inpainting—that typically destroy pixel-level watermarks. Ablation studies confirmed that adding noise simulation actually degrades performance, proving that relying on inherent stability is a more robust paradigm.

Conclusion

SimuFreeMark demonstrates that noise-simulation-free training can achieve superior robustness and visual quality by leveraging the stability of low-frequency semantic features. While the current framework is sensitive to geometric transformations like rotation or cropping due to its reliance on FFT alignment, future work will explore geometry-invariant representations to further enhance its real-world utility.

Summary updated

BibTeX
@misc{tang2025simufreemark,
  author = {Yichao Tang and Mingyang Li and Di Miao and Sheng Li and Zhenxing Qian and Xinpeng Zhang},
  title = {{SimuFreeMark: A Noise-Simulation-Free Robust Watermarking Against Image Editing}},
  year = {2025},
  eprint = {2511.11295},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.11295},
}
Figure