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

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.

💡 Research Summary

The paper tackles the pressing problem of detecting AI‑generated images (AIGIs) that are increasingly indistinguishable from genuine photographs. Existing detection methods fall into two camps: artifact‑based approaches that target generator‑specific fingerprints, and generic representation methods that rely on large‑scale pretrained models (e.g., CLIP). Both suffer from poor generalization when confronted with unseen generators or when image post‑processing is applied.

To overcome these limitations, the authors propose a novel “Demosaicing‑guided Color Correlation Training” (DCCT) framework that exploits a fundamental difference between camera‑captured photos and AI‑generated images: the physical imaging pipeline. Real cameras acquire light through a Color Filter Array (CFA), most commonly a Bayer pattern, which samples only one color per pixel. The raw sensor output is a single‑channel mosaic that must be demosaiced—interpolated—to reconstruct the missing two color channels. This process introduces characteristic high‑frequency aliasing and inter‑channel correlations that are absent in purely digital synthesis, where the generator directly outputs full RGB values without any CFA‑induced modulation.

DCCT mimics the CFA sampling by masking each RGB image according to the Bayer pattern, producing a single‑channel observation (x) (the simulated raw) and a two‑channel target (y) (the missing colors). Both (x) and (y) are high‑pass filtered to emphasize subtle residuals. A U‑Net is trained in a self‑supervised manner to model the conditional distribution (p_{\theta}(y’|x’)) using a mixture of discretized logistic components, similar to PixelCNN++. The training objective is the negative log‑likelihood over photographic images only; no AI‑generated data are needed for pretraining.

The authors provide a theoretical analysis showing that, under locally Gaussian assumptions for the joint statistics of ((x’,y’)), the 1‑Wasserstein distance between the conditional distributions of photographs and AI‑generated images is lower‑bounded by a positive constant (\delta). Intuitively, the CFA‑induced linear mapping (T_{\text{CFA}}) creates a distinct mean relationship between (x’) and (y’) that a digital generator cannot replicate perfectly, leading to a persistent distributional gap in the high‑frequency color‑correlation space.

After pretraining, the U‑Net’s intermediate representations—high‑frequency color‑correlation features—are extracted for any image (photographic or AI‑generated). These features are fed to a lightweight binary classifier (e.g., a two‑layer MLP) that learns to separate the two classes. Because the features are rooted in a physical imaging process, they remain stable across a wide range of generators and survive common post‑processing operations such as JPEG compression, additive noise, and color jitter.

Extensive experiments were conducted on multiple benchmark datasets covering more than 20 unseen generators, including recent diffusion models, text‑to‑image systems, and various GAN architectures. DCCT consistently outperformed state‑of‑the‑art artifact‑based detectors and generic representation baselines, achieving higher accuracy, AUC, and robustness to degradations. Ablation studies confirmed the importance of the high‑pass filtering, the mixture‑of‑logistics output, and the CFA‑aligned masking strategy.

In summary, the paper introduces a camera‑aware self‑supervised pretraining paradigm that leverages the immutable CFA‑demosaicing pipeline to learn discriminative color‑correlation features. By grounding detection in a physical process that AI generators cannot emulate, DCCT offers a scalable, generalizable, and robust solution to AI‑generated image detection, marking a significant advance in digital forensics.