Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploiting the Final Component of Generator Architectures for AI-Generated Image Detection

Yanzhu Liu, Xiao Liu, Yuexuan Wang, Mondal Soumik

TL;DR

This paper tackles the generalization gap in AI-generated image detection by proposing a taxonomy of generators based on their final architectural component and a contamination-based training regime that uses only the final component to imprint detectable traces on real images. By constructing negative samples via φ*(E*(x)) for three representative components and training a detector on a DINOv3 backbone, the authors demonstrate strong zero-shot performance across diverse unseen generators and real-world benchmarks, even with as few as 100 samples per component. They further show that combining a sparse set of negatives via K-Medoids maintains high generalization, and that independent mini-batch sampling from real and constructed data yields faster, more stable convergence. The work yields practical implications for robust, scalable AI-generated image detection, providing a principled approach to exploit architecture-level traces without requiring access to full generator models.

Abstract

With the rapid proliferation of powerful image generators, accurate detection of AI-generated images has become essential for maintaining a trustworthy online environment. However, existing deepfake detectors often generalize poorly to images produced by unseen generators. Notably, despite being trained under vastly different paradigms, such as diffusion or autoregressive modeling, many modern image generators share common final architectural components that serve as the last stage for converting intermediate representations into images. Motivated by this insight, we propose to "contaminate" real images using the generator's final component and train a detector to distinguish them from the original real images. We further introduce a taxonomy based on generators' final components and categorize 21 widely used generators accordingly, enabling a comprehensive investigation of our method's generalization capability. Using only 100 samples from each of three representative categories, our detector-fine-tuned on the DINOv3 backbone-achieves an average accuracy of 98.83% across 22 testing sets from unseen generators.

Exploiting the Final Component of Generator Architectures for AI-Generated Image Detection

TL;DR

This paper tackles the generalization gap in AI-generated image detection by proposing a taxonomy of generators based on their final architectural component and a contamination-based training regime that uses only the final component to imprint detectable traces on real images. By constructing negative samples via φ*(E*(x)) for three representative components and training a detector on a DINOv3 backbone, the authors demonstrate strong zero-shot performance across diverse unseen generators and real-world benchmarks, even with as few as 100 samples per component. They further show that combining a sparse set of negatives via K-Medoids maintains high generalization, and that independent mini-batch sampling from real and constructed data yields faster, more stable convergence. The work yields practical implications for robust, scalable AI-generated image detection, providing a principled approach to exploit architecture-level traces without requiring access to full generator models.

Abstract

With the rapid proliferation of powerful image generators, accurate detection of AI-generated images has become essential for maintaining a trustworthy online environment. However, existing deepfake detectors often generalize poorly to images produced by unseen generators. Notably, despite being trained under vastly different paradigms, such as diffusion or autoregressive modeling, many modern image generators share common final architectural components that serve as the last stage for converting intermediate representations into images. Motivated by this insight, we propose to "contaminate" real images using the generator's final component and train a detector to distinguish them from the original real images. We further introduce a taxonomy based on generators' final components and categorize 21 widely used generators accordingly, enabling a comprehensive investigation of our method's generalization capability. Using only 100 samples from each of three representative categories, our detector-fine-tuned on the DINOv3 backbone-achieves an average accuracy of 98.83% across 22 testing sets from unseen generators.
Paper Structure (19 sections, 13 equations, 7 figures, 2 tables, 1 algorithm)

This paper contains 19 sections, 13 equations, 7 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Architecture attribution
  4. Last step inversion
  5. Zero-shot deepfake detection
  6. Method
  7. A Taxonomy of Generation Architectures based on Final Component
  8. Fake sample construction --- $\varphi\circ$ E
  9. Construction from the three representative components
  10. Sparse Sampling from Combined Constructions
  11. Detection model --- $\theta$
  12. Experiments
  13. Model implementation & baselines
  14. Results & discussions
  15. Conclusion
  16. ...and 4 more sections

Figures (7)

  • Figure 1: The proposed taxonomy of generation architectures based on final component.
  • Figure 2: The proposed framework for detecting AI-generated images
  • Figure 3: Performance comparison of our model trained on the final component of SD2
  • Figure 4: Performance comparison of our model trained on the final component of HiDream
  • Figure 5: Feature visualization of the three final components.
  • ...and 2 more figures