Uncolorable Examples: Preventing Unauthorized AI Colorization via Perception-Aware Chroma-Restrictive Perturbation

Yuki Nii, Futa Waseda, Ching-Chun Chang, Isao Echizen

TL;DR

This paper introduces "Uncolorable Examples," a novel defensive paradigm designed to protect grayscale media, such as heritage films and manga, from unauthorized AI colorization. By embedding human-imperceptible perturbations into grayscale images using a method called PAChroma, the authors effectively trick AI colorization models into outputting flat grayscale results while preserving the original visual quality for human viewers.

What Problem Does This Paper Address?

Recent advances in deep learning have made automatic image colorization remarkably convincing. Tools like DeOldify and DDColor can transform black-and-white images into vibrant, realistic color versions with a single click. However, this technology opens the door to widespread copyright infringement. Malicious actors can easily colorize monochrome manga, classic animations, or historic films without the creator's consent and resell them illegally.

Until now, content creators had no technical means to protect their grayscale work from being exploited by generative AI. This paper tackles this gap by framing the defense as an adversarial attack problem: how can we modify a grayscale image so that it looks completely normal to a human, but completely paralyzes an AI colorization model?

Overview of the Uncolorable Examples framework.

Methods

The authors propose PAChroma (Perception-Aware Chroma-Restrictive Perturbation), an optimization-based method that crafts adversarial perturbations tailored for grayscale protection.

Instead of trying to force the AI model to generate ugly or unnatural colors—which is unreliable due to the sheer diversity of colorization architectures—PAChroma actively steers the model's output back toward grayscale. It achieves this by minimizing a specialized Colorfulness Loss ($L_{CF}$), which mathematically quantifies visual vividness.

To make this defense practical for the real world, PAChroma introduces two key innovations:

  1. Continuous Laplacian Mask: Standard adversarial noise often creates visible artifacts on smooth surfaces (like a character's face or a manga speech bubble). PAChroma applies a Laplacian filter to map out the high-frequency regions (edges) of the image. It then concentrates the adversarial perturbations along these edges, exploiting a quirk of human vision called contrast masking to keep the changes completely imperceptible.
  2. Structure-Preserving Input Transformations: To ensure the defense works even if a pirate resizes, crops, or compresses the image, PAChroma applies a suite of random, block-wise transformations (geometric shifts, frequency-domain filtering, intensity jitters) during the optimization loop. This forces the perturbation to generalize across different model architectures (black-box transferability) and survive common image processing pipelines (robustness).

Datasets

The authors evaluated their approach on two distinct image domains:

  • ImageNet: A standard dataset of natural images. The authors sampled 100 images from the validation set, resized to $256 \times 256$ pixels.
  • Danbooru: To test the method on artwork, they collected 4,367 manga-tagged illustrations, resizing them to $576 \times 576$ pixels and randomly sampling subsets for evaluation.

Key Insights and Ideas

  • Grayscale as the Ultimate Defense Target: Forcing an AI to output a grayscale image is a much more robust defense than forcing it to output "bad" colors. Because colorization models are trained to find plausible color distributions, pushing them toward zero colorfulness exploits a shared vulnerability across CNN, GAN, and Transformer-based architectures.
  • The Manga Challenge: Manga is uniquely difficult to protect because it contains massive, flat regions with zero texture. Standard adversarial noise stands out immediately in these areas. Edge-weighting via the Laplacian mask is the "aha moment" that allows the defense to work on line art without ruining the reading experience.
Qualitative comparison showing how PAChroma preserves image structure compared to standard noise and basic nullifying attacks.

Key Results

PAChroma was tested against several state-of-the-art colorization models, including DeOldify, BigColor, and DDColor.

  • Effectiveness: In white-box scenarios, PAChroma successfully suppressed colorization, reducing the output colorfulness score by 78.40% to 82.80%.
  • Imperceptibility: It maintained excellent visual fidelity, achieving high structural similarity (SSIM) scores of 0.91 to 0.95 relative to the original unprotected images, vastly outperforming baseline attacks which left heavy, pixelated distortion.
  • Transferability & Robustness: In black-box settings (where the defender doesn't know what model the attacker is using), PAChroma still achieved a modest but perceptibly degrading color suppression. Crucially, the defense remained effective even after the protected images were subjected to JPEG 50% compression and random resizing.

Limitations and Future Work

While PAChroma marks a significant first step, the authors note a few limitations:

  • Manga Domain Transferability: Black-box transferability remains limited on manga images compared to natural images. The lack of texture in manga restricts where the algorithm can safely hide adversarial noise, making it easier for a different AI model to bypass the defense.
  • User-Guided Colorization: The current method only targets fully automatic colorization pipelines. It does not yet account for interactive, user-guided tools where a human provides color hints or text prompts to guide the AI.
  • Computational Overhead: Generating these examples requires iterative optimization, which may be too slow for high-resolution video processing or real-time deployment without further optimization.

Summary updated

Abstract

AI-based colorization has shown remarkable capability in generating realistic color images from grayscale inputs. However, it poses risks of copyright infringement -- for example, the unauthorized colorization and resale of monochrome manga and films. Despite these concerns, no effective method currently exists to prevent such misuse. To address this, we introduce the first defensive paradigm, Uncolorable Examples, which embed imperceptible perturbations into grayscale images to invalidate unauthorized colorization. To ensure real-world applicability, we establish four criteria: effectiveness, imperceptibility, transferability, and robustness. Our method, Perception-Aware Chroma-Restrictive Perturbation (PAChroma), generates Uncolorable Examples that meet these four criteria by optimizing imperceptible perturbations with a Laplacian filter to preserve perceptual quality, and applying diverse input transformations during optimization to enhance transferability across models and robustness against common post-processing (e.g., compression). Experiments on ImageNet and Danbooru datasets demonstrate that PAChroma effectively degrades colorization quality while maintaining the visual appearance. This work marks the first step toward protecting visual content from illegitimate AI colorization, paving the way for copyright-aware defenses in generative media.
BibTeX
@misc{nii2025uncolorable,
  author = {Yuki Nii and Futa Waseda and Ching-Chun Chang and Isao Echizen},
  title = {{Uncolorable Examples: Preventing Unauthorized AI Colorization via Perception-Aware Chroma-Restrictive Perturbation}},
  year = {2025},
  eprint = {2510.08979},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.08979},
}
Figure