Beyond Invisibility: Learning Robust Visible Watermarks for Stronger Copyright Protection

Tianci Liu, Tong Yang, Quan Zhang, Qi Lei

Introduction

The rapid ascent of generative AI has presented a dual-edged sword for the creative community. While these tools offer unprecedented creative possibilities, they also facilitate large-scale copyright infringement through automated scraping and "style plagiarism." In the latter case, techniques like DreamBooth allow users to fine-tune models on a handful of an artist's works, effectively automating the production of derivative art that mimics a specific aesthetic.

To date, the primary defense against such exploitation has been the use of invisible adversarial perturbations—microscopic changes to pixels designed to "break" the AI's understanding of an image. However, these methods suffer from two critical flaws: they are easily removed by standard image processing (like JPEG compression) and they often fail to generalize across different AI architectures. This paper introduces HARVIM (Hard-to-Remove Visible Watermark), a framework that shifts the paradigm back toward visible watermarks, but with a sophisticated, AI-resistant optimization strategy.

The HARVIM Framework: Optimizing for Inpaint-Resistance

The central philosophy of HARVIM is that not all parts of an image are equally easy for an AI to reconstruct. If a watermark is placed over a simple, predictable texture, an AI "inpainting" tool can easily guess what was underneath. However, if the watermark covers a region of high complexity or unique structural detail, the AI will struggle to "hallucinate" the missing data accurately.

HARVIM formalizes this as a bi-level optimization problem—a mathematical game between two opposing forces:

  1. The Reconstruction Player: An AI model attempting to remove the watermark and restore the original image.
  2. The Watermark Player (HARVIM): An optimization process that selects a watermark mask $m$ to maximize the reconstruction error of the first player.

Technical Implementation

To make this optimization tractable, the authors employ several key innovations:

  • Generative Priors: HARVIM utilizes a pre-trained Normalizing Flow model (specifically RealNVP) to establish a baseline for what "natural" images look like. This allows the system to identify regions where the image's distribution is most difficult to recover.
  • Differentiable Masking: Because selecting specific pixels is a discrete (non-differentiable) operation, the authors use a sigmoid-based "soft" mask. This allows the use of standard gradient-based optimization to find the ideal watermark shape and placement.
  • Meta-Learning: The complex bi-level math is solved efficiently by "unrolling" the reconstruction steps during the watermark's optimization phase, a technique borrowed from meta-learning.

Experimental Validation and Generalization

The researchers evaluated HARVIM across three distinct datasets to test its versatility: CelebA (faces), ImageNet (diverse objects), and a curated set of Cartoon Images representing the actual creative work artists seek to protect.

Resistance to AI Removal

Against a "worst-case" scenario where the attacker knows exactly where the watermark is (Flow-R), HARVIM proved remarkably resilient. On the CelebA dataset, the Peak Signal-to-Noise Ratio (PSNR)—a standard metric for reconstruction quality—dropped from 13.02 (with random watermarking) to 7.57 (with HARVIM). A lower PSNR indicates a more distorted, failed reconstruction.

Furthermore, modern "blind" watermark removers like SLBR and DeNet, which are designed to automatically detect and erase watermarks, failed almost entirely. These tools were often unable to even identify the HARVIM watermark as a foreign element, leaving the protected image intact.

The "Backdoor Effect"

One of the most compelling findings is the impact on style-mimicking models. When an artist's portfolio is protected by HARVIM, models like DreamBooth inadvertently learn the watermark as an essential component of the artist's "style." Consequently, any images generated by the plagiarizing model are marred by distorted, broken versions of the watermark, effectively rendering the stolen style commercially useless.

Robustness and Limitations

Unlike invisible watermarks, HARVIM is inherently robust to common image transformations. Because the protection relies on the visible obstruction of critical image data, simple operations like JPEG compression or Gaussian blurring do not "wash away" the defense.

However, the approach involves certain trade-offs:

  • Aesthetics: There is a fundamental tension between the strength of protection and the visual clarity of the image. While HARVIM uses regularizers to prevent the watermark from being overly intrusive, it remains a visible modification.
  • Computation: The bi-level optimization process is currently computationally intensive, making it slower than traditional watermarking methods.

Conclusion

HARVIM represents a significant step forward in copyright protection by identifying "hard-to-reconstruct" regions as an intrinsic property of images. By leveraging these properties, HARVIM creates a defense that is not only robust against current AI tools but likely to remain effective against future, more powerful generative models. Future work will focus on improving the computational efficiency of the framework and establishing theoretical bounds for watermark unremovability.

Summary updated

Abstract

As AI advances, copyrighted content faces growing risk of unauthorized use, whether through model training or direct misuse. Building upon invisible adversarial perturbation, recent works developed copyright protections against specific AI techniques such as unauthorized personalization through DreamBooth that are misused. However, these methods offer only short-term security, as they require retraining whenever the underlying model architectures change. To establish long-term protection aiming at better robustness, we go beyond invisible perturbation, and propose a universal approach that embeds \textit{visible} watermarks that are \textit{hard-to-remove} into images. Grounded in a new probabilistic and inverse problem-based formulation, our framework maximizes the discrepancy between the \textit{optimal} reconstruction and the original content. We develop an effective and efficient approximation algorithm to circumvent a intractable bi-level optimization. Experimental results demonstrate superiority of our approach across diverse scenarios.
BibTeX
@inproceedings{liu2025beyond,
  author = {Tianci Liu and Tong Yang and Quan Zhang and Qi Lei},
  title = {{Beyond Invisibility: Learning Robust Visible Watermarks for Stronger Copyright Protection}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=3uK3fLpjtD},
}
Figure