G-Channel Removed Reconstruction Error (GRRE)

• GRRE is a forensic method that exploits the distinctive reconstruction error gap between natural and AI-generated images when the green channel is removed.
• It utilizes a diffusion-based reconstructor to restore the G-channel, resulting in markedly higher errors for real images versus lower errors in synthetic outputs.
• Empirical evaluations demonstrate GRRE's superior accuracy across diverse datasets and its robustness under compression, noise, and cross-dataset challenges.

G-Channel Removed Reconstruction Error (GRRE) is a forensic method for robust detection of AI-generated images, particularly those synthesized by diffusion models and GANs. The approach exploits the distinctive reconstruction behavior that arises when the green (G) channel of an RGB image is removed and then restored using a learned diffusion-based reconstructor. A pronounced discrepancy in reconstruction error is observed between natural and synthetic images: for real images, the removal of the G-channel leads to significant reconstruction errors, whereas AI-generated images typically yield lower errors due to their proximity to the diffusion model’s learned manifold. This phenomenon forms the basis of a highly generalizable and robust forensic signature for distinguishing synthetic from real photographic content (He et al., 6 Jan 2026).

1. Motivation and Theoretical Rationale

Empirical analysis in image processing indicates that the G-channel contains the greatest proportion of luminance, high-frequency texture, and structural information in natural scenes. When the G-channel is zeroed in a real photograph and a diffusion model attempts to hallucinate the missing data, the structural deficit results in relatively high pixel-wise reconstruction errors. In contrast, AI-generated images, often constructed to be statistically consistent with the model's training distribution, are less perturbed by channel removal—their G-channel content can be more accurately inferred by the same model. As a result, GRRE leverages this reconstruction fidelity gap (real ≫ fake) as a forensic feature that is robust to post-processing and generalizes across generative model families.

2. Formalization and Mathematical Framework

Let $x \in \mathbb{R}^{H \times W \times 3}$ denote an RGB image. The G-channel removal is effected by the mask $m = [1,\, 0,\, 1] \in \mathbb{R}^3$, producing

$$x_{noG}(u, v) = x(u, v) \odot m = [\,R(u, v),\,0,\,B(u, v)\,].$$

The masked image $x_{noG}$ is input into a diffusion-based reconstructor $\mathcal{R}_\theta$, yielding

$$x_{recon} = \mathcal{R}_\theta(x_{noG}) \in \mathbb{R}^{H \times W \times 3}.$$

The resulting pixel-wise absolute error map is

$$E(u,v) = |x(u,v) - x_{recon}(u,v)| \in \mathbb{R}^3.$$

Global reconstruction error metrics include, for instance,

$$\mathrm{GRRE}_{L1}(x) = \frac{1}{HW} \sum_{u,v} \|E(u,v)\|_{1},\ \mathrm{GRRE}_{MSE}(x) = \frac{1}{HW} \sum_{u,v} \|E(u,v)\|_2^2.$$

These metrics compactly summarize the discrepancy induced by G-channel removal and subsequent restoration.

3. Network Architecture and Learning Schemes

The reconstruction module adopts a standard Denoising Diffusion Probabilistic Model (DDPM) U-Net backbone with encoder–decoder topology, four down/up-sampling stages, skip connections, and GroupNorm + SiLU nonlinearity. Cross-block time conditioning is delivered via a four-layer sinusoidal MLP. Training uses the standard diffusion denoising loss:

$$\mathcal{L}_{diff} = \mathbb{E}_{x_0,\, t,\, \varepsilon} \left[ \|\varepsilon - \varepsilon_{\theta}(x_t, t)\|_2^2 \right],$$

where $$x_t = \sqrt{\bar{\alpha}_t} x_0 + \sqrt{1 - \bar{\alpha}_t} \varepsilon$$. During inference, a reverse diffusion process (T steps descending to 0) is executed from the corrupted $x_{noG}$ to recover $x_{recon}$.

A binary classifier—typically a ResNet-50 modified to receive a three-channel error map—operates either on $E$ directly or its scalar projection. The final single logit is interpreted as the posterior for the “AI-generated” class. The classifier is trained with the binary cross-entropy loss.

4. Detection Workflow

Inference proceeds as follows:

Input: test image x; pretrained diffusion reconstructor ℛ_θ; classifier f_φ; threshold τ 1. x_noG ← x ⊙ [1,0,1] 2. x_recon ← ℛ_θ(x_noG) # run reverse diffusion steps 3. E ← |x - x_recon| # pixel-wise absolute error 4. logit z ← f_φ(E) 5. p ← sigmoid(z) 6. if p ≥ τ then label← “AI-generated” else label← “real”

The operating threshold $\tau$ is selected on a validation set to maximize detection accuracy or $F_1$; in practice, $\tau=0.5$ suffices when the output is well calibrated.

5. Empirical Evaluation and Comparative Performance

Experimental validation spans diverse domains and generative methods, using the DiffusionForensics suite (CelebA-HQ, ImageNet, LSUN-Bedroom; each with real and multiple GAN/diffusion outputs). Baseline detectors include DIRE, FIRE, AEROBLADE, and FakeInversion. Metrics include accuracy (Acc), AUC, and average precision (AP). Both intra-dataset cross-model (train/test using different generators) and cross-dataset (train/test on different domains) settings are reported.

Condensed results (Accuracy/AUC in %) are as follows:

Method	CelebA-HQ→Others	LSUN→Others	Average
AEROBLADE	73.5 / 72.3	80.2 / 84.1	76.9 / 78.2
FakeInversion	67.9 / 61.8	62.2 / 59.7	65.0 / 60.8
DIRE	99.6 / 100.0	93.0 / 98.9	96.3 / 99.5
FIRE	86.1 / 99.7	70.4 / 69.9	78.3 / 84.8
GRRE (Ours)	100.0/100.0	98.4/99.8	99.2/99.9

On LSUN-Bedroom (11 generators), GRRE attains 98.4% accuracy and 99.8% AUC. Cross-dataset transfer further demonstrates GRRE's robustness: trained on CelebA-HQ (SD-v2), tested on ImageNet (ADM, SD-v1), GRRE achieves 91.6% accuracy and 98.6% AUC versus FIRE at approximately 53%/55%.

6. Robustness, Ablation, and Generalization

GRRE maintains robust detection under a suite of perturbations: under heavy JPEG compression (Q=40), AUC ≥ 74%; under Gaussian noise ($\sigma=0.15$), AUC ≥ 92%. In every stress scenario, GRRE outperforms alternatives by 10–30 AUC points. Ablation studies on channel removal confirm the centrality of the G-channel; removal of R (RRRE) and B (BRRE) yields substantially lower average AUC (81.8% and 87.6%, respectively) compared to GRRE (98.4%). This finding underscores the unique forensic value of the green channel's information density.

Variant	Average AUC / AP
RRRE	81.8% / 95.1%
BRRE	87.6% / 99.6%
GRRE	98.4% / 99.8%

7. Limitations and Prospective Developments

GRRE's accuracy presumes a sufficiently powerful diffusion reconstructor; undertrained or biased models can degrade detection efficacy. The computational cost of running a full $T$-step (e.g., 1000-step) diffusion process for each image may impede real-time or large-scale deployment. Possible extensions include leveraging simultaneous multi-channel removal (tri-stream detectors), training-free application via out-of-the-box diffusion reconstructors coupled with adaptive thresholding, and adaptation to multi-frame scenarios such as AI-generated video detection.

A plausible implication is that channel-removal-based approaches, exemplified by GRRE, constitute a generalizable and scalable paradigm with the potential for robust forensic defense in increasingly hostile generative environments (He et al., 6 Jan 2026).