Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Color Matters: Demosaicing-Guided Color Correlation Training for Generalizable AI-Generated Image Detection

Nan Zhong, Yiran Xu, Mian Zou

TL;DR

This work tackles the generalization gap in AI-generated image detectors by exploiting camera imaging pipeline cues. It introduces DCCT, a demosaicing-guided, self-supervised pretraining framework that learns stable color-correlation features from high-frequency residuals, formalized with a bound on the 1-Wasserstein distance between photographic and AI-generated distributions. By freezing dual conditional networks and training a lightweight classifier on their outputs, DCCT achieves state-of-the-art generalization across more than 20 unseen generators and demonstrates robustness to benign post-processing. The approach highlights the value of camera-aware pretraining for reliable detection in the face of rapidly evolving generative models and lays groundwork for extending to diverse sensor designs.

Abstract

As realistic AI-generated images threaten digital authenticity, we address the generalization failure of generative artifact-based detectors by exploiting the intrinsic properties of the camera imaging pipeline. Concretely, we investigate color correlations induced by the color filter array (CFA) and demosaicing, and propose a Demosaicing-guided Color Correlation Training (DCCT) framework for AI-generated image detection. By simulating the CFA sampling pattern, we decompose each color image into a single-channel input (as the condition) and the remaining two channels as the ground-truth targets (for prediction). A self-supervised U-Net is trained to model the conditional distribution of the missing channels from the given one, parameterized via a mixture of logistic functions. Our theoretical analysis reveals that DCCT targets a provable distributional difference in color-correlation features between photographic and AI-generated images. By leveraging these distinct features to construct a binary classifier, DCCT achieves state-of-the-art generalization and robustness, significantly outperforming prior methods across over 20 unseen generators.

Color Matters: Demosaicing-Guided Color Correlation Training for Generalizable AI-Generated Image Detection

TL;DR

This work tackles the generalization gap in AI-generated image detectors by exploiting camera imaging pipeline cues. It introduces DCCT, a demosaicing-guided, self-supervised pretraining framework that learns stable color-correlation features from high-frequency residuals, formalized with a bound on the 1-Wasserstein distance between photographic and AI-generated distributions. By freezing dual conditional networks and training a lightweight classifier on their outputs, DCCT achieves state-of-the-art generalization across more than 20 unseen generators and demonstrates robustness to benign post-processing. The approach highlights the value of camera-aware pretraining for reliable detection in the face of rapidly evolving generative models and lays groundwork for extending to diverse sensor designs.

Abstract

As realistic AI-generated images threaten digital authenticity, we address the generalization failure of generative artifact-based detectors by exploiting the intrinsic properties of the camera imaging pipeline. Concretely, we investigate color correlations induced by the color filter array (CFA) and demosaicing, and propose a Demosaicing-guided Color Correlation Training (DCCT) framework for AI-generated image detection. By simulating the CFA sampling pattern, we decompose each color image into a single-channel input (as the condition) and the remaining two channels as the ground-truth targets (for prediction). A self-supervised U-Net is trained to model the conditional distribution of the missing channels from the given one, parameterized via a mixture of logistic functions. Our theoretical analysis reveals that DCCT targets a provable distributional difference in color-correlation features between photographic and AI-generated images. By leveraging these distinct features to construct a binary classifier, DCCT achieves state-of-the-art generalization and robustness, significantly outperforming prior methods across over 20 unseen generators.
Paper Structure (17 sections, 1 theorem, 13 equations, 8 figures, 5 tables, 2 algorithms)

This paper contains 17 sections, 1 theorem, 13 equations, 8 figures, 5 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Method: DCCT
  4. Preliminaries
  5. DCCT via Self-Supervised Reconstruction
  6. Theoretical Insight
  7. AIGI Detection
  8. Experiments
  9. Experimental Setups
  10. Main Results
  11. Further Analysis
  12. Ablation Studies
  13. Conclusion and Discussion
  14. Overall DCCT Pipeline and Algorithm
  15. Further Evaluation of the One-Class Detector DCCT$^\dagger$
  16. ...and 2 more sections

Key Result

Proposition 3.1

Let $\bm x'$ and $\bm y'$ denote the high-frequency input and output residuals, respectively. Assume locally Gaussian statistics for the joint distribution of $(\bm x', \bm y')$ within small patches. There exists a constant $\delta > 0$ such that, for any input $\bm x'$ possessing non-trivial energy provided that the generative model fails to perfectly replicate the spectral correlations induced b

Figures (8)

  • Figure 1: Illustration of the CFA pipeline: (a) original RGB scene; (b) RAW CFA mosaic (single‑channel), shown in grayscale with a zoomed‑in patch where a color‑coded RGGB Bayer pattern is overlaid to visualize the sampling locations of each channel rather than true RGB values; and (c) demosaiced RGB reconstruction via simple bilinear interpolation for visualization purpose.
  • Figure 2: Bayer-like masking strategy, where a single channel (representing the CFA raw input) is retained as the network input, while the remaining two color components serve as the reconstruction targets (ground-truth labels).
  • Figure 3: The system diagram of the DCCT via self-supervised reconstruction.
  • Figure 4: Distributions of the anomaly score $D$ for photographic vs. AI-generated images on the GenImage dataset zhu2023genimage. We compute $D$ by averaging the pixel-wise difference between the negative log-likelihood $\mathrm{NLL}$ and the entropy $H$ of the color feature maps predicted by DCCT (trained exclusively on photographic images). The resulting distributions demonstrate a clear separation between photographic and AI-generated content across various generators.
  • Figure 5: t-SNE embeddings of learned features for photographic and AI-generated images.
  • ...and 3 more figures

Theorems & Definitions (1)

  • Proposition 3.1