DiffWA: Diffusion Models for Watermark Attack

Xinyu Li

TL;DR

DiffWA (Diffusion Watermark Attack) is a new framework that uses conditional diffusion models to strip invisible watermarks from images while simultaneously restoring their visual quality. By combining distance-guided sampling with an "estimator" to speed up the process, the researchers achieved high watermark destruction (Bit Error Rates near 0.5) while maintaining impressive image fidelity (PSNR > 31dB).

What Problem Does This Paper Address?

Digital watermarking is a cornerstone of copyright protection, with modern "blind" watermarking algorithms (like HiDDeN) becoming incredibly robust against traditional attacks like cropping, blurring, or JPEG compression.

The authors identify a significant gap in the current landscape: most existing watermark attack algorithms either fail to remove the watermark entirely or cause so much visual damage to the image that it becomes useless. There is currently an imbalance where watermark addition algorithms are far more advanced than the attack algorithms meant to test them. DiffWA aims to bridge this gap by treating watermark removal as an image restoration task.

Methods

The authors propose a framework based on Conditional Diffusion Models, specifically drawing inspiration from the "Palette" image-to-image framework. The core idea is to train a model on clean, unwatermarked images so it learns the distribution of "natural" data.

Key technical components include:

  • Distance Guidance: During the sampling (denoising) process, the model is guided by a distance metric (either Mean Squared Error or Structural Similarity). This ensures the generated "clean" image stays structurally similar to the original watermarked input while moving toward the manifold of unwatermarked images.
  • Estimator Acceleration: Diffusion models are notoriously slow because they require hundreds or thousands of iterations. The authors introduced a ResNet34-based estimator that predicts a noisy version of the image ($x^N$) directly from the watermarked input. This allows the model to skip the initial diffusion steps and perform only a fraction of the denoising iterations (e.g., 100 steps instead of 1000).
  • Combinatorial Method: To maximize effectiveness, the authors experimented with a two-stage approach where one model pre-processes the image to shift its distribution, and a second diffusion model performs the final watermark removal and restoration.

Datasets

The researchers evaluated their method using:

  • CIFAR-10: A standard benchmark dataset of 60,000 32x32 color images.
  • HiDDeN: This was used as the "target" watermarking scheme to be attacked. The authors trained their own HiDDeN model with various noise layers (Crop, Dropout, Gaussian blur, JPEG) to ensure it was highly robust before attempting to break it.

Key Insights and Ideas

  • Watermarks as "Noise": The fundamental "aha moment" is treating a watermark not as a hidden signal to be decoded, but as a specific type of perturbation that can be removed through the generative power of diffusion.
  • Frequency Reconstruction: The authors analyzed the HiDDeN scheme and found it embeds data across all frequency components (Low-Low to High-High). Because diffusion models naturally reconstruct images from low to high frequencies, they are uniquely suited to "overwriting" these embedded signals.
  • The Fidelity-Attack Trade-off: There is a delicate balance controlled by the gradient scale ($\eta$). If $\eta$ is too high, the image looks perfect but the watermark might remain; if it's too low, the watermark is destroyed but the image loses its original identity. The authors found that keeping $\eta$ relatively close to zero provides the best results for watermark removal.

Key Results

The proposed DiffWA model outperformed traditional attack methods across several metrics:

  • Watermark Destruction: The model achieved a Bit Error Rate (BER) of over 0.4 (where 0.5 represents a total random guess, meaning the watermark is effectively unreadable).
  • Image Quality: Despite the aggressive attack, the restored images maintained a PSNR of more than 31 and an SSIM of more than 0.97, indicating that the "attacked" images are almost indistinguishable from the originals to the human eye.
  • Efficiency: Using the Estimator Acceleration, the model achieved similar performance to the full diffusion process while significantly reducing the computational overhead.

Limitations and Future Work

  • Resolution Constraints: The experiments were conducted on CIFAR-10 (32x32 pixels). While this proves the concept, real-world watermarking often occurs on high-resolution photography. Scaling this approach to megapixels remains a challenge for diffusion models.
  • Ethical Implications: The authors acknowledge that while this research helps improve watermarking robustness, it could also be misused for copyright infringement. They suggest that future work should focus on developing watermarking techniques specifically designed to resist diffusion-based attacks.
  • Algorithm Specificity: While effective against HiDDeN, it is yet to be seen how this method performs against other state-of-the-art watermarking frameworks like ReDMark.

Summary updated

Abstract

With the rapid development of deep neural networks(DNNs), many robust blind watermarking algorithms and frameworks have been proposed and achieved good results. At present, the watermark attack algorithm can not compete with the watermark addition algorithm. And many watermark attack algorithms only care about interfering with the normal extraction of the watermark, and the watermark attack will cause great visual loss to the image. To this end, we propose DiffWA, a conditional diffusion model with distance guidance for watermark attack, which can restore the image while removing the embedded watermark. The core of our method is training an image-to-image conditional diffusion model on unwatermarked images and guiding the conditional model using a distance guidance when sampling so that the model will generate unwatermarked images which is similar to original images. We conducted experiments on CIFAR-10 using our proposed models. The results shows that the model can remove the watermark with good effect and make the bit error rate of watermark extraction higher than 0.4. At the same time, the attacked image will maintain good visual effect with PSNR more than 31 and SSIM more than 0.97 compared with the original image.
BibTeX
@misc{li2023diffwa,
  author = {Xinyu Li},
  title = {{DiffWA: Diffusion Models for Watermark Attack}},
  year = {2023},
  eprint = {2306.12790},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.12790},
}
Figure