Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Present and Future Generalization of Synthetic Image Detectors

Pablo Bernabeu-Perez, Enrique Lopez-Cuena, Dario Garcia-Gasulla

TL;DR

This work conducts a systematic analysis and uses its insights to develop practical guidelines for training robust synthetic image detectors, and identifies critical flaws in detectors and workarounds to enable the deployment of real-world detector applications enhancing accuracy, reliability and robustness beyond the limitations of current systems.

Abstract

The continued release of increasingly realistic image generation models creates a demand for synthetic image detectors. To build effective detectors we must first understand how factors like data source diversity, training methodologies and image alterations affect their generalization capabilities. This work conducts a systematic analysis and uses its insights to develop practical guidelines for training robust synthetic image detectors. Model generalization capabilities are evaluated across different setups (e.g. scale, sources, transformations) including real-world deployment conditions. Through an extensive benchmarking of state-of-the-art detectors across diverse and recent datasets, we show that while current approaches excel in specific scenarios, no single detector achieves universal effectiveness. Critical flaws are identified in detectors, and workarounds are proposed to enable the deployment of real-world detector applications enhancing accuracy, reliability and robustness beyond the limitations of current systems.

Present and Future Generalization of Synthetic Image Detectors

TL;DR

This work conducts a systematic analysis and uses its insights to develop practical guidelines for training robust synthetic image detectors, and identifies critical flaws in detectors and workarounds to enable the deployment of real-world detector applications enhancing accuracy, reliability and robustness beyond the limitations of current systems.

Abstract

The continued release of increasingly realistic image generation models creates a demand for synthetic image detectors. To build effective detectors we must first understand how factors like data source diversity, training methodologies and image alterations affect their generalization capabilities. This work conducts a systematic analysis and uses its insights to develop practical guidelines for training robust synthetic image detectors. Model generalization capabilities are evaluated across different setups (e.g. scale, sources, transformations) including real-world deployment conditions. Through an extensive benchmarking of state-of-the-art detectors across diverse and recent datasets, we show that while current approaches excel in specific scenarios, no single detector achieves universal effectiveness. Critical flaws are identified in detectors, and workarounds are proposed to enable the deployment of real-world detector applications enhancing accuracy, reliability and robustness beyond the limitations of current systems.
Paper Structure (29 sections, 13 figures, 8 tables)

This paper contains 29 sections, 13 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Architecture
  5. Train Datasets
  6. Benchmarking Datasets
  7. Train Experiments
  8. Single-class Models
  9. Multi-class Models
  10. Image Alteration Methods
  11. Deployment Experiments
  12. Generalization to Source
  13. Image Decision Boundary
  14. Scale Generalization
  15. Human Evaluators
  16. ...and 14 more sections

Figures (13)

  • Figure 1: Examples of the In-the-wild dataset.
  • Figure 2: Recall of SuSy on authentic and synthetic evaluation datasets, under different scaling factors.
  • Figure 3: Recall of SID on authentic and synthetic evaluation datasets, under different scaling factors.
  • Figure 4: Detector architecture used, based on a ResNet-18 from lopez2023sr, including ResNet blocks (blue), bottlenecks (red), adaptative average pooling 2D (orange), concatenation (yellow) and an MLP (green).
  • Figure 5: Scalability analysis showing images processed per second using single-crop (green) and 5-crop (blue) approaches across different hardware configurations. Note the axis break highlighting the GPU acceleration gain.
  • ...and 8 more figures