PaperlayerView exact revision on arXiv

Source-linked AI summary

PatchCraft: Exploring Texture Patch for Efficient AI-generated Image Detection

Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, Xinpeng Zhang

arXiv:2311.12397v3cs.CV

TL;DR

Detectors can lose performance on images from unseen generative models, limiting their generalization across sources. PatchCraft extracts inter-pixel correlation contrast between rich and poor texture regions, and reports that it outperforms a state-of-the-art detector with a clear margin.

  • Problem

    Detectors trained on one generator can degrade on unseen sources, although open-world images may come from varied or unknown generators.

    Source: §1 · p. 2

  • Method

    Smash&Reconstruction suppresses global semantics by sorting and rebuilding rich- and poor-texture patches, whose inter-pixel correlation contrast forms the detector’s fingerprint.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    The approach outperforms the state-of-the-art detector with a clear margin in the reported experiments.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The authors report that rich-versus-poor texture correlation contrast is a fingerprint shared across GAN- and diffusion-generated images, supporting detector generalization across generative models.

    Source: §1 · p. 3, §5 · p. 9

Abstract

from arXiv · show

Recent generative models show impressive performance in generating photographic images. Humans can hardly distinguish such incredibly realistic-looking AI-generated images from real ones. AI-generated images may lead to ubiquitous disinformation dissemination. Therefore, it is of utmost urgency to develop a detector to identify AI generated images. Most existing detectors suffer from sharp performance drops over unseen generative models. In this paper, we propose a novel AI-generated image detector capable of identifying fake images created by a wide range of generative models. We observe that the texture patches of images tend to reveal more traces left by generative models compared to the global semantic information of the images. A novel Smash&Reconstruction preprocessing is proposed to erase the global semantic information and enhance texture patches. Furthermore, pixels in rich texture regions exhibit more significant fluctuations than those in poor texture regions. Synthesizing realistic rich texture regions proves to be more challenging for existing generative models. Based on this principle, we leverage the inter-pixel correlation contrast between rich and poor texture regions within an image to further boost the detection performance. In addition, we build a comprehensive AI-generated image detection benchmark, which includes 17 kinds of prevalent generative models, to evaluate the effectiveness of existing baselines and our approach. Our benchmark provides a leaderboard for follow-up studies. Extensive experimental results show that our approach outperforms state-of-the-art baselines by a significant margin. Our project: https://fdmas.github.io/AIGCDetect

1. Introduction

Detectors trained on one generator can falter on unseen sources, while realistic AI-generated images raise concerns about disinformation. The paper proposes a detector based on texture-patch inter-pixel correlations and reports that it outperforms a state-of-the-art detector.

1. Introduction

  • Classifiers trained on a particular generator can perform poorly on unseen sources, motivating detectors that generalize across the varied sources found in open-world applications.The introduction also notes the societal concern that accessible fake images may contribute to disinformation dissemination.Source: §1 · p. 2
  • Smash&Reconstruction breaks global semantic information and enhances texture patches so the classifier uses their inter-pixel correlations rather than image semantics.The paper argues that semantic-based detection degrades on realistic-looking fake images.Source: §1 · p. 2
  • The approach outperforms the state-of-the-art detector with a clear margin, according to the reported experimental results.The paper also presents a benchmark intended to support comparisons under the same conditions.Source: §1 · p. 3
  • The proposed universal fingerprint is the inter-pixel correlation contrast between rich and poor texture regions, intended to generalize across GAN- and diffusion-based images.The authors attribute the contrast to generative models’ difficulty synthesizing rich texture regions.Source: §1 · pp. 2–3

2. Related Work

AI-generated image detection is framed as a binary classification task whose performance can degrade on unseen generators. Prior work explores fingerprints from frequency artifacts, image texture, data augmentation, and diffusion-model reconstruction.

2. Related Work

  • A classifier can perform well when training and test fakes come from the same generator, yet degrade significantly on unseen generators.Prior work seeks fingerprints that generalize across different generative models.Source: §2 · p. 3
  • Prior approaches use frequency anomalies [11], frequency effects attributed to upsampling [10], frequency noise patterns, texture Gram matrices [28], and data augmentation.A diffusion-specific approach uses reconstruction error from a pretrained diffusion model [48].Source: §2 · p. 3

3. Proposed Method

The method targets generator traces in inter-pixel correlations rather than global semantics, using rich-versus-poor texture contrast as a fingerprint. It extracts those features from reconstructed texture patches, then trains a binary classifier on them.

3.1. Motivation

  • Real images have similar inter-pixel correlations across rich and poor texture regions, whereas generative models leave different artifacts in those regions; the paper uses their contrast as a fingerprint.The motivation is that rich texture is more diverse and harder for generators to approximate.Source: §3.1 · p. 4

3.2. Fingerprint Feature Extraction

  • Smash&Reconstruction sorts randomly cropped patches by texture diversity and reconstructs separate rich- and poor-texture images, suppressing their global semantic information.Texture diversity is measured from pixel fluctuations and residuals in four directions; high-pass filters then extract noise patterns, and a learnable block measures the contrast between their outputs.Source: §3.2 · pp. 4–5
  • The fingerprint’s dimension depends on patch size and patch count, not input-image size, avoiding downsampling that can erase weak generator artifacts.The paper identifies downsampling-related artifact loss as a source of detector performance degradation.Source: §3.2 · p. 5
  • High-pass filters sweep across patch boundaries to capture variation trends in patch diversity, which the authors say helps detection accuracy.The approach retains these boundaries rather than restricting filtering to each patch.Source: §3.2 · p. 5

3.3. AI-generated Fake Image Detection

  • A binary classifier receives each image’s fingerprint and predicts whether the image is real or fake.The training objective updates the classifier and the learnable convolution block in the fingerprint module; at inference, an arbitrary-size image is first converted to its fingerprint.Source: §3.3 · p. 5

4. Experimental Results

The evaluation tests generalization across 17 generative models after training on ProGAN fakes. The approach outperforms baselines overall, remains strongest in most distorted-image cases, and benefits from each tested component.

4.1. Experimental Setup

  • Evaluation covers 17 generative models, with detectors trained on 360k real LSUN images and 360k ProGAN fakes for fair generalization comparisons.The test set includes commercial APIs such as Midjourney; the study targets AI-synthesized images rather than AI-manipulated images.Source: §4.1 · pp. 5–6

4.2. Detection Effectiveness

  • Our approach outperforms baselines by a clear margin; LNP is closest but trails by 4% in average detection accuracy.Performance degrades for CycleGAN and GauGAN images.Source: §4.2 · pp. 7–8, Table 2 · p. 7
  • Distortions lower every detector’s performance, but our approach performs best in most cases and its fingerprint is more robust than baselines.The evaluated distortions are JPEG compression, Gaussian blur, and downsampling.Source: §4.2 · p. 8, Table 3 · p. 8

4.3. Ablation Studies

  • Smash&Reconstruction erases image semantics and exposes rich and poor texture regions; removing it causes a marked performance drop, especially for diffusion fakes.Boundary-crossing high-pass filters support accuracy; restricting them to patches lowers accuracy by about 3.26%, while removing texture contrast performs worse than the full approach.Source: §4.3 · pp. 8–9, Table 4 · p. 8

5. Conclusions

The detector uses a universal fingerprint based on inter-pixel correlation contrast between rich and poor texture regions, which the authors attribute to generative models’ greater difficulty synthesizing rich textures. The paper also reports a benchmark and experiments showing the approach outperforms existing baselines.

5. Conclusions

  • Inter-pixel correlation contrast between rich and poor texture regions provides the detector’s proposed universal fingerprint across generative models.The authors attribute this to the greater complexity of rich texture distributions, which makes them more difficult for generative models to synthesize from random noise.Source: §5 · p. 9
  • The paper builds a benchmark spanning generative models and detectors to facilitate comparison in follow-up studies.Source: §5 · p. 9
  • Experiments show the proposed approach outperforms existing baselines.Source: §5 · p. 9

Supplement

The supplement documents the filter and classifier configurations, evaluation criteria and results, complementary ablations, and visualizations of generated images.

High-pass Filters

  • Thirty high-pass filters are constructed by rotating kernel variants across specified directions to capture diverse high-frequency signals.Source: High-pass Filters · p. 12

Classifier Structure

  • A cascaded CNN classifier is used because the proposed fingerprint allows a simple network to achieve satisfactory detection performance.Source: Classifier Structure · p. 12

Performance Evaluation

  • Detection performance is evaluated using accuracy and average precision, with detailed results reported in Tables 7–14.Source: Performance Evaluation · p. 12

Complementary Ablation Studies

AI-generated fake image visualization

Loading 2311.12397v3…