Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthetic Image Verification in the Era of Generative AI: What Works and What Isn't There Yet

Diangarti Tariang, Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, Luisa Verdoliva

TL;DR

This paper surveys the rapid evolution of synthetic image generation and the corresponding detection and attribution landscape. It contrasts generation paradigms—GANs and diffusion models—and discusses how visible artifacts are evolving into invisible forensic traces, such as artificial fingerprints and spectral signatures. The authors synthesize data-driven detection methods, forensic-cue based approaches, and attribution techniques, highlighting generalization, calibration, and open-set challenges. They also chart open research directions, including intent characterization, explainability, robustness to adversarial manipulation, universal detectors, and active signaling mechanisms to safeguard information integrity in the era of generative AI.

Abstract

In this work we present an overview of approaches for the detection and attribution of synthetic images and highlight their strengths and weaknesses. We also point out and discuss hot topics in this field and outline promising directions for future research.

Synthetic Image Verification in the Era of Generative AI: What Works and What Isn't There Yet

TL;DR

This paper surveys the rapid evolution of synthetic image generation and the corresponding detection and attribution landscape. It contrasts generation paradigms—GANs and diffusion models—and discusses how visible artifacts are evolving into invisible forensic traces, such as artificial fingerprints and spectral signatures. The authors synthesize data-driven detection methods, forensic-cue based approaches, and attribution techniques, highlighting generalization, calibration, and open-set challenges. They also chart open research directions, including intent characterization, explainability, robustness to adversarial manipulation, universal detectors, and active signaling mechanisms to safeguard information integrity in the era of generative AI.

Abstract

In this work we present an overview of approaches for the detection and attribution of synthetic images and highlight their strengths and weaknesses. We also point out and discuss hot topics in this field and outline promising directions for future research.
Paper Structure (37 sections, 9 figures, 1 table)

This paper contains 37 sections, 9 figures, 1 table.

Table of Contents

  1. Introduction
  2. Synthetic image generation
  3. Generative Adversarial Networks
  4. Diffusion Models
  5. Forensic artifacts
  6. Synthetic image detection
  7. Data-driven methods
  8. Methods exploiting forensic cues
  9. Low-level artifacts
  10. High-level artifacts
  11. Experimental evaluation
  12. Generalization analysis
  13. In the wild
  14. Calibration
  15. Synthetic image attribution
  16. ...and 22 more sections

Figures (9)

  • Figure 1: Top: examples of synthetic images, generated using (from left to right) Latent Diffusion, Stable Diffusion, Midjourney v5, DALL·E Mini, DALL·E 2, DALL·E 3. The prompt used for their generation is the following: a photo of the Rome Colosseum with a UFO over it, detailed, 8k. Bottom: Average Power Spectra of the artificial fingerprints for each of such model. Forensic artifacts are clearly visible as spectral peaks in the Fourier domain, stronger or weaker based on the specific model. We can observe that the first three images share very similar artifacts while the fingerprints of the three releases of DALL-E differ greatly from one another, testifying to very different generative architecturescorvi2023intriguing.
  • Figure 2: Taxonomy of synthetic image detection methods.
  • Figure 3: Synthetic image detection results in terms of AUC with and without post-processing (PP).
  • Figure 4: Synthetic image detection results in terms of AUC for three methods over various contents.
  • Figure 5: Synthetic image detection results in terms of AUC on images from X generated by DALL·E 3, Midjourney and Firefly.
  • ...and 4 more figures