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

Abstract

We propose PAChroma, a method that embeds imperceptible perturbations into grayscale images to prevent unauthorized AI colorization. The resulting Uncolorable Examples resist AI colorization while maintaining visual quality, transferability, and robustness, protecting monochrome content such as manga and films from unauthorized colorization and resale.

Futa Waseda

Project Assistant Professor | Trustworthy AI: Robustness, Reliability, and VLM Defense