Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TrueFake: A Real World Case Dataset of Last Generation Fake Images also Shared on Social Networks

Stefano Dell'Anna, Andrea Montibeller, Giulia Boato

TL;DR

The paper tackles the gap between laboratory fake-image detection and real-world robustness when images are disseminated and processed by social networks. It introduces TrueFake, a 600k-image dataset combining real, DM-generated, and GAN-generated content, with 180k images shared across Facebook, X, and Telegram to simulate in-the-wild conditions. The study evaluates five state-of-the-art detectors and a proposed R50-E2P approach under non-shared and shared scenarios, examining generalization to unseen generators and resilience to social-network processing. Findings show substantial performance degradation due to social processing, with CLIP-D providing a strong baseline and R50-E2P suggesting a viable direction using pretrained feature encoders for improved generalization. The dataset offers a practical benchmark to drive development of robust forensic detectors in real-world settings.

Abstract

AI-generated synthetic media are increasingly used in real-world scenarios, often with the purpose of spreading misinformation and propaganda through social media platforms, where compression and other processing can degrade fake detection cues. Currently, many forensic tools fail to account for these in-the-wild challenges. In this work, we introduce TrueFake, a large-scale benchmarking dataset of 600,000 images including top notch generative techniques and sharing via three different social networks. This dataset allows for rigorous evaluation of state-of-the-art fake image detectors under very realistic and challenging conditions. Through extensive experimentation, we analyze how social media sharing impacts detection performance, and identify current most effective detection and training strategies. Our findings highlight the need for evaluating forensic models in conditions that mirror real-world use.

TrueFake: A Real World Case Dataset of Last Generation Fake Images also Shared on Social Networks

TL;DR

The paper tackles the gap between laboratory fake-image detection and real-world robustness when images are disseminated and processed by social networks. It introduces TrueFake, a 600k-image dataset combining real, DM-generated, and GAN-generated content, with 180k images shared across Facebook, X, and Telegram to simulate in-the-wild conditions. The study evaluates five state-of-the-art detectors and a proposed R50-E2P approach under non-shared and shared scenarios, examining generalization to unseen generators and resilience to social-network processing. Findings show substantial performance degradation due to social processing, with CLIP-D providing a strong baseline and R50-E2P suggesting a viable direction using pretrained feature encoders for improved generalization. The dataset offers a practical benchmark to drive development of robust forensic detectors in real-world settings.

Abstract

AI-generated synthetic media are increasingly used in real-world scenarios, often with the purpose of spreading misinformation and propaganda through social media platforms, where compression and other processing can degrade fake detection cues. Currently, many forensic tools fail to account for these in-the-wild challenges. In this work, we introduce TrueFake, a large-scale benchmarking dataset of 600,000 images including top notch generative techniques and sharing via three different social networks. This dataset allows for rigorous evaluation of state-of-the-art fake image detectors under very realistic and challenging conditions. Through extensive experimentation, we analyze how social media sharing impacts detection performance, and identify current most effective detection and training strategies. Our findings highlight the need for evaluating forensic models in conditions that mirror real-world use.
Paper Structure (18 sections, 1 equation, 1 figure, 4 tables)

This paper contains 18 sections, 1 equation, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Dataset
  4. Real Images Collection
  5. GAN-generated Images
  6. DM-generated Images
  7. Social Network Sharing
  8. Detection methods
  9. R50-ND
  10. MISLNet
  11. NPR
  12. CLIP-D
  13. P2G
  14. ResNet50 as Feature Encoder to Prototypes (R50-E2P)
  15. Experimental results
  16. ...and 3 more sections

Figures (1)

  • Figure 1: The TrueFake dataset contains a total of 600,000 images: a) 100,000 real images sourced from the FORLAB iuliani2021leak and FFHQ ffhq-dataset datasets, b) 200,000 images generated using five DM-based techniques (Stable Diffusion (versions 1.5 to 3), Stable Diffusion XL, and FLUX.1), c) 120,000 images produced with GANs (StyleGAN 1, 2, and 3), and d) 180,000 of them shared via social media platforms. Indeed, a subset of 20,000 real, 25,000 DM-generated, and 15,000 GAN-generated images were also shared on three social networks (Facebook, X, and Telegram) by using the respective APIs.