AI Research Portal

This paper introduces a pixel-level mapping pre-processing step for generalized AI-generated image detection. The core idea is to disrupt the pixel value distribution of images, thereby breaking non-essential semantic patterns that detectors often exploit as shortcuts. This forces the detector to focus on more fundamental and generalizable high-frequency traces inherent to the image generation process. The paper empirically claims that this approach significantly boosts the cross-generator performance of state-of-the-art detectors across various generative models, including GANs and diffusion-based models.

Existing AI-generated image detectors face several bottlenecks:

Overfitting to Semantic Cues: Current detectors often overfit to source-specific semantic cues, failing to learn universal generative artifacts. This limits their generalization to images from unseen generative models.
Semantic Bias Discrepancies: The issue primarily stems from semantic bias discrepancies caused by imperfect fitting to training data in generative models, manifesting as visual artifacts like blurring and texture anomalies. Different models exhibit specific semantic biases.

The paper empirically demonstrates that simple high-pass filtering and image patch shuffling fail to effectively eliminate semantic information, making them inadequate for reducing the impact of semantic bias on generalization in detection tasks. These findings reveal critical limitations in existing approaches and highlight the need for more sophisticated semantic suppression methods.
A novel pixel-level mapping method is introduced that attenuates semantic bias during classifier training. By transforming low-frequency information while amplifying high-frequency artifacts, the method significantly reduces classifier reliance on biased semantic features, addressing the core generalization challenge in synthetic image detection.
Extensive validation across multiple benchmarks confirms the effectiveness of the method. The results show consistent performance gains when detecting images from unseen generative architectures, demonstrating the generalization capability of the proposed method.

The proposed method introduces a pixel-level mapping module as a pre-processing step before feeding images into a classification network. This module aims to reduce semantic bias by transforming pixel values, thereby converting low-frequency information into high-frequency components and amplifying inter-pixel differences.

Framework Pipeline: An input image (Ix) first passes through the pixel-level mapping module to produce a transformed image (I'x), which is then sent to a classification network to determine if it is real or generated.

Fixed Pixel-level Mapping Module: Given an input RGB image I ∈ RH×W×3 with pixel values ranging from 0 to 255, a mapping function Φε is applied to each pixel value v ∈ [0,256). This function amplifies differences between adjacent pixel values and normalizes pixel intensities. Key Equation: Φε(ν) = v - round(v/256, 2) × 256, where round represents the round operation (NumPy's np.round with decimals=2). Setting decimals=2 disrupts monotonic arrangements and normalizes transformed pixel values to the approximate range of [-1.28, 1.28].

Random Pixel-level Mapping Module: This variant applies different random mapping rules to each input sample and each channel within that sample. For each input image I ∈ RH×W×3, a per-channel mapping table T_c ∈ R256 is constructed, where each entry follows an independent and identically distributed (i.i.d.) uniform distribution over 256 dimensions. Key Equation: T_c ~ U(−1,1)256, c∈ {0,1,2}. The transformed image I' ∈ RH×W×3 is then generated by: I'[x, y] = T_c [I_c[x,y]], ∀x∈ [0, H), y ∈ [0, W), where I_c[x, y] ∈ {0,..., 255} denotes the original pixel value at location (x, y) in channel c.

Objective: The classifier f_θ: X → [0,1] is trained to minimize the cross-entropy loss: L = -1/N Σ[y_i log f_θ(x_i) + (1 - y_i) log(1 - f_θ(x_i))], where y_i = 1 if x_i ~ P_gen(x/G_j), G_j ∈ G, and y_i = 0 if x_i ~ P_real(x).

Architecture: The detector network is implemented with a ResNet-50 architecture.
Implementation: The method is implemented using PyTorch.
Hardware: 8 Nvidia 3090 GPUs.
Image Pre-processing: During training, images are randomly cropped to size 128×128. During testing, images are center-cropped.
Optimizer: Adam optimizer.
Learning Rate: Initial learning rate of 2×10-4.
Moment Decay Rates: First-order moment decay rate is 0.9, second-order moment decay rate is 0.999.
Weight Decay: 2 × 10-4.
Training Epochs: 200 epochs.
Batch Size: 128.
Classification Threshold: A uniform classification threshold of 0.5 is maintained on all evaluation benchmarks.

Datasets:
- Training: ForenSynths (20 semantic classes, 4 used: car, cat, chair, horse), GenImage (SDv1.4 as generative model), LSUN, ImageNet.
- Test: Self-Synthesis (9 GAN models), GenImage (state-of-the-art diffusion models, including MidJourney and WuKong), UniversalFakeDetect.
Metrics: Classification accuracy (Acc) and average precision (AP).
Baselines: UniFD (Ojha, Li, and Lee 2023), NPR (Tan et al. 2024b), C2P-CLIP (Tan et al. 2025), DRCT (Chen et al. 2024), BSA (Zheng et al. 2024), ResNet-50, high-frequency filtering, patch shuffling (size=8, size=2).
Numeric Results:
- Self-Synthesis GAN dataset (Table 1):
  - Fixed-mapping: 99.6%/100.0% (Acc/AP) mean across 9 GANs.
  - Random-mapping: 99.4%/100.0% (Acc/AP) mean across 9 GANs.
  - Outperforms UniFD by 20.3% in classification accuracy and NPR by 4.7%.
- GenImage dataset (Table 2):
  - Fixed-mapping: 98.4% mean accuracy.
  - Random-mapping: 97.5% mean accuracy.
  - Exceeds baseline UniFD by 9.6% and C2P-CLIP by 2.6%.
- Comparison with semantic reduction methods on GenImage (Table 3):
  - Fixed-mapping: 98.4 mAcc, 99.8 mAP.
  - Random-mapping: 97.5 mAcc, 99.6 mAP.
  - NPR (spectrum-based): 88.6 mAcc, 93.7 mAP.
  - BSA (shuffle-based): 74.7 mAcc, 85.6 mAP.
  - ResNet-50 (baseline): 67.0 mAcc, 76.9 mAP.
Runtime/Training Cost: Not stated in the paper.

The random mapping variant exhibits slightly inferior performance compared to fixed mapping, suggesting that consistent transformation patterns benefit training convergence despite equivalent semantic suppression.

Method evidence: "To this end, we propose a pixel-level mapping approach that converts monotonically ordered pixel values (0-255) into a new set of pixel values, thereby altering the tight spatial arrangement between neighboring pixels." (section: Pixel-level Mapping for Semantic Bias Reduction)
Result evidence: "Our pixel-level mapping method achieves a new state-of-the-art result with an average accuracy of 98.4%, exceeding baseline UniFD and state-of-the-art C2P-CLIP by 9.6% and 2.6% respectively." (section: Evaluation on GenImage dataset)
Limitation evidence: "Notably, the random mapping variant exhibits slightly inferior performance compared to fixed mapping, suggesting that consistent transformation patterns benefit training convergence despite equivalent semantic suppression." (section: Comparison with different semantic reduction methods)

Beyond Semantic Features: Pixel-level Mapping for Generalized AI-Generated Image Detection

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Evidence

Additional Figures