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

TL;DR

PAChroma advances a defense against unauthorized AI colorization by embedding imperceptible perturbations into grayscale inputs to degrade colorization quality. It combines a perception-aware chroma-restrictive perturbation with a continuous Laplacian mask and diverse input transformations to minimize a colorfulness loss across diverse colorization models. The approach satisfies four practical criteria—effectiveness, imperceptibility, transferability, and robustness—and demonstrates substantial colorization suppression on ImageNet and manga datasets while preserving the grayscale appearance. This work provides a practical framework for copyright-aware defenses in generative media and sets the stage for extensions to higher resolutions and user-guided colorization defenses.

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.

Figures (7)

• Figure 1: Overview of Uncolorable Examples. Without protection, grayscale images can be illegitimately colorized by AI colorization models. Our method, PAChroma, generates Uncolorable Examples by adding human-imperceptible perturbations to the input, effectively invalidating unauthorized colorization.
• Figure 2: Uncolorable Examples from PAChroma invalidate colorization via imperceptible perturbations. Top-left: DeOldify DeOldify2021; top-right: ACDO dakini2024animecolor; bottom-left: DDColor kang2023ddcolorphotorealisticimagecolorization; bottom-right: MC-V2 mangacolorv2_2022. Each image is shown with its CF score, SSIM between the input, or PSNR between the output.
• Figure 3: Effectiveness and Imperceptibility of random noise, Nullifying Attack (NA), and PAChroma. PAChroma preserves grayscale structure while preventing colorization, outperforming NA in imperceptibility. Each image includes CF score, SSIM between the input, or PSNR between the output.
• Figure 4: Transferability of PAChroma among DeOldify, BigColor, and DDColor. Shown with CF, SSIM between inputs, or PSNR between outputs.
• Figure 5: Robustness of NA, NA-Mask, and PAChroma to JPEG compression (Q=$X$%) and Random Resized Cropping (RRC). Each image is shown with its CF score on the bottom corner.