perturb

Overview ¶

Package perturb applies imperceptible pixel-level noise to break the statistical fingerprints left by AI image generators.

Every major AI generator (DALL-E, Midjourney, Stable Diffusion, etc.) produces images with characteristic frequency-domain patterns in the pixel data. Detectors like Hive Moderation and AI-or-Not exploit these patterns.

A ±1-2 value perturbation per channel is sufficient to disrupt those patterns while remaining completely invisible to the human eye — the change is well below the just-noticeable-difference (JND) threshold of ~2-3 values.