Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

S^2F-Net:A Robust Spatial-Spectral Fusion Framework for Cross-Model AIGC Detection

Xiangyu Hu, Yicheng Hong, Hongchuang Zheng, Wenjun Zeng, Bingyao Liu

TL;DR

S^2F-Net addresses the poor cross-model generalization of AIGC detectors by exploiting universal frequency-domain artifacts arising from upsampling. It combines a Spatial Smash & Residual Learning pipeline with a Learnable FFT-based Frequency Attention module to adaptively emphasize texture-rich high-frequency and texture-poor low-frequency cues, followed by a Cascaded Discriminative Module for decision making. The approach demonstrates strong cross-model generalization, achieving an average accuracy of 90.49% across 17 generative models on the AIGCDetectBenchmark and showing robustness to common distortions via structured ablations. This dual-domain framework and the learnable spectral weighting offer a practical path toward model-agnostic detection of AI-generated content in real-world settings.

Abstract

The rapid development of generative models has imposed an urgent demand for detection schemes with strong generalization capabilities. However, existing detection methods generally suffer from overfitting to specific source models, leading to significant performance degradation when confronted with unseen generative architectures. To address these challenges, this paper proposes a cross-model detection framework called S 2 F-Net, whose core lies in exploring and leveraging the inherent spectral discrepancies between real and synthetic textures. Considering that upsampling operations leave unique and distinguishable frequency fingerprints in both texture-poor and texture-rich regions, we focus our research on the detection of frequency-domain artifacts, aiming to fundamentally improve the generalization performance of the model. Specifically, we introduce a learnable frequency attention module that adaptively weights and enhances discriminative frequency bands by synergizing spatial texture analysis and spectral dependencies.On the AIGCDetectBenchmark, which includes 17 categories of generative models, S 2 F-Net achieves a detection accuracy of 90.49%, significantly outperforming various existing baseline methods in cross-domain detection scenarios.

S^2F-Net:A Robust Spatial-Spectral Fusion Framework for Cross-Model AIGC Detection

TL;DR

S^2F-Net addresses the poor cross-model generalization of AIGC detectors by exploiting universal frequency-domain artifacts arising from upsampling. It combines a Spatial Smash & Residual Learning pipeline with a Learnable FFT-based Frequency Attention module to adaptively emphasize texture-rich high-frequency and texture-poor low-frequency cues, followed by a Cascaded Discriminative Module for decision making. The approach demonstrates strong cross-model generalization, achieving an average accuracy of 90.49% across 17 generative models on the AIGCDetectBenchmark and showing robustness to common distortions via structured ablations. This dual-domain framework and the learnable spectral weighting offer a practical path toward model-agnostic detection of AI-generated content in real-world settings.

Abstract

The rapid development of generative models has imposed an urgent demand for detection schemes with strong generalization capabilities. However, existing detection methods generally suffer from overfitting to specific source models, leading to significant performance degradation when confronted with unseen generative architectures. To address these challenges, this paper proposes a cross-model detection framework called S 2 F-Net, whose core lies in exploring and leveraging the inherent spectral discrepancies between real and synthetic textures. Considering that upsampling operations leave unique and distinguishable frequency fingerprints in both texture-poor and texture-rich regions, we focus our research on the detection of frequency-domain artifacts, aiming to fundamentally improve the generalization performance of the model. Specifically, we introduce a learnable frequency attention module that adaptively weights and enhances discriminative frequency bands by synergizing spatial texture analysis and spectral dependencies.On the AIGCDetectBenchmark, which includes 17 categories of generative models, S 2 F-Net achieves a detection accuracy of 90.49%, significantly outperforming various existing baseline methods in cross-domain detection scenarios.
Paper Structure (29 sections, 8 equations, 9 figures, 9 tables)

This paper contains 29 sections, 8 equations, 9 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. GANs Series Models
  4. Diffusion Models
  5. Related Detection Methods
  6. Our work
  7. Design Motivation
  8. Spatial Domain Feature Extraction: Semantic Smash and Residual Learning
  9. Semantic Smash & Physical Reconstruction
  10. Residual Mapping and Compression
  11. FFT-based Learnable Frequency Attention Module
  12. Spectral Transformation & Geometrically Constrained Masking
  13. Group-wise Energy Aggregation & Spatial Recalibration
  14. Cascaded Discriminative Module
  15. experimental results
  16. ...and 14 more sections

Figures (9)

  • Figure 1: Energy Comparison after Fourier Centralization
  • Figure 2: Entropy Distribution Contrast (a) and (c) , Density Difference (b) and Histogram Difference (d) illustrates the differences in entropy region distribution between images generated by GANs and real images & images generated by Stable Diffusion and real images.
  • Figure 3: Kernel Density Estimation (KDE) of texture richness distribution between real images and synthetic images from advanced generative models
  • Figure 4: Average Discrete Cosine Transform (DCT) heatmap comparison between GAN and Diffusion-generated images
  • Figure 5: S$^2$F-Net's overall architecture: integration of spatial feature extraction and learnable frequency attention mechanism
  • ...and 4 more figures