Semantic Watermarking Reinvented: Enhancing Robustness and Generation Quality with Fourier Integrity

Sung Ju Lee, Nam Ik Cho
Seoul National University, South Korea
Accepted at ICCV 2025
Code arXiv (Coming Soon)
Teaser image

Summary of watermarking performance across different semantic watermarking methods, following the merged-in-generation scheme with no additional processing time. The proposed approaches achieve the best balance between detection robustness and image fidelity.

Abstract

Semantic watermarking techniques for latent diffusion models (LDMs) are robust against regeneration attacks, but often suffer from detection performance degradation due to the loss of frequency integrity. To tackle this problem, we propose a novel embedding method called Hermitian Symmetric Fourier Watermarking (SFW), which maintains frequency integrity by enforcing Hermitian symmetry. Additionally, we introduce a center-aware embedding strategy that reduces the vulnerability of semantic watermarking due to cropping attacks by ensuring robust information retention. To validate our approach, we apply these techniques to existing semantic watermarking schemes, enhancing their frequency-domain structures for better robustness and retrieval accuracy. Extensive experiments demonstrate that our methods achieve state-of-the-art verification and identification performance, surpassing previous approaches across various attack scenarios. Ablation studies confirm the impact of SFW on detection capabilities, the effectiveness of the center-aware embedding against cropping, and how message capacity influences identification accuracy. Notably, our method achieves the highest detection accuracy while maintaining superior image fidelity, as evidenced by FID and CLIP scores. Conclusively, our proposed SFW is shown to be an effective framework for balancing robustness and image fidelity, addressing the inherent trade-offs in semantic watermarking. Code available at github.com/thomas11809/SFWMark.

Motivation

The rise of latent diffusion models has made image generation widely accessible, but it also introduces challenges in content attribution and copyright. Semantic watermarking in the latent space offers a robust solution, especially due to its resilience against regeneration attacks, which often break pixel-level watermarks.
However, existing methods often discard the imaginary part in the Fourier domain, breaking the frequency integrity required to maintain the statistical structure of latent noise. In particular, this violates the Hermitian symmetry necessary for real-valued latent signals, leading to distorted frequency patterns that simultaneously weaken detection robustness and degrade generative quality.
To address this issue, we propose a new approach that preserves frequency integrity while enabling reliable and high-quality watermarking.

Weight distance correlation Missing imaginary patterns reveal frequency loss in baselines.

Our Approach

We introduce three key techniques to overcome the limitations of existing semantic watermarking methods.

Our Approach
  1. Hermitian Symmetric Fourier Watermarking (SFW):
    By enforcing Hermitian symmetry in the latent Fourier domain, our method preserves frequency integrity and fully utilizes both real and imaginary components. This leads to stronger detection performance without sacrificing image quality.
  2. Center-Aware Embedding Strategy:
    We embed watermarks only in the central region of the latent space, which remains stable under spatial transformations. This design greatly improves robustness against cropping attacks.
  3. HSQR: Hermitian Symmetric QR Code:
    We extend SFW to structured binary watermarks by splitting a QR code across the real and imaginary parts of the Fourier domain. This approach ensures high detection accuracy and message capacity, while preserving generative quality.

Experiments

We evaluate our methods across a wide range of distortion scenarios, including signal, compression, regeneration, and cropping attacks.

  1. Detection Performance
    Our methods (HSTR and HSQR) consistently outperform prior techniques in both verification and identification tasks, demonstrating enhanced detection robustness against various attack types.
  2. Generative Quality
    Unlike high-energy methods like RingID, which introduce visible artifacts, our approach maintains high image fidelity thanks to frequency integrity preservation.
  3. Ablation Study
    1. Without Hermitian symmetry, detection performance drops significantly, confirming the core role of SFW.
    2. Our center-aware embedding improves robustness to cropping, maintaining high accuracy under spatial distortions.
    3. As capacity increases, most methods begin to degrade, yet HSQR remains consistently accurate and reliable.

Additional Resources

More Qualitative Results

More Qualitative Results

BibTeX

To Be Updated Soon.