Abstract
The rise of sophisticated AI colorization models has introduced new risks for copyright infringement, particularly for creators of monochrome content like manga and classic films. This paper presents "Uncolorable Examples," a novel defensive paradigm that utilizes Perception-Aware Chroma-Restrictive Perturbation (PAChroma) to thwart unauthorized AI colorization. By embedding imperceptible, structure-aware noise into grayscale images, PAChroma forces generative models to produce near-monochrome outputs. Our results indicate that PAChroma achieves high effectiveness in color suppression while preserving the visual integrity of the original work, even under common image transformations.
Introduction
AI-based colorization technology can now transform grayscale images into vibrant color versions with minimal effort. While beneficial for restoration, this capability allows malicious actors to colorize and resell copyrighted monochrome works without authorization. Existing adversarial defenses, primarily designed for classification tasks, often fail to protect generative outputs or introduce visible artifacts that degrade the user experience.
To address this, we propose "Uncolorable Examples"—grayscale images that appear normal to humans but remain "uncolorable" for AI models. This defense treats the protection as an adversarial attack on the colorization model's objective, steering the AI toward producing a grayscale result rather than a colorful one.
Methodology: The PAChroma Framework
The core of our defense is the PAChroma (Perception-Aware Chroma-Restrictive Perturbation) method. Unlike standard adversarial attacks that target class labels, PAChroma optimizes perturbations to minimize the "colorfulness" of the AI's output.
Perception-Aware Masking
To ensure the perturbation remains invisible, we employ a Continuous Laplacian Mask. This filter identifies high-frequency details and edges where human perception is less sensitive to noise. By concentrating the perturbation in these areas, we protect smooth regions—such as faces or speech bubbles in manga—from visible distortion.
Robustness and Transferability
To ensure the defense survives real-world scenarios, we incorporate Structure-Invariant Transformations during the optimization process. By applying random shifts, flips, and jittering, the generated perturbation becomes resilient to common image processing like resizing and JPEG compression.
Experimental Evaluation
We evaluated PAChroma using two primary datasets: ImageNet (100 natural image samples) and Danbooru (4,367 manga-tagged images). The defense was tested against a variety of state-of-the-art colorization architectures, including CNN-based (DeOldify), GAN-based (BigColor), and Transformer-based (DDColor) models.
Key Results
PAChroma demonstrated significant success in both color suppression and image fidelity. In white-box settings, the method reduced the colorfulness of AI outputs by 78% to 82%.
As shown in Figure 3, PAChroma maintains a high Structural Similarity Index (SSIM) of 0.95, significantly outperforming random noise or standard adversarial (NA) baselines. Furthermore, the defense proved robust against JPEG compression at 50% quality, maintaining its protective properties where other methods failed. Even in cross-model scenarios, PAChroma degraded colorization quality more effectively than previous methods.
Discussion and Future Work
While PAChroma effectively "locks" grayscale content against automatic colorizers, challenges remain. Transferability in the manga domain is more difficult due to the structural simplicity of the medium. Future research will explore defenses against "human-in-the-loop" AI tools that utilize color hints and aim to reduce the computational overhead of generating these perturbations for large-scale libraries.
Conclusion
Uncolorable Examples provide a vital tool for creators to safeguard their monochrome intellectual property. By leveraging perception-aware noise, PAChroma offers a practical balance between image quality and robust copyright protection in the age of generative AI.
Summary updated