Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Finding AI-Generated Faces in the Wild

Gonzalo J. Aniano Porcile, Jack Gindi, Shivansh Mundra, James R. Verbus, Hany Farid

TL;DR

This work tackles the problem of distinguishing real faces from AI-generated faces in the wild, with a focus on fake online profiles. It trains an EfficientNet-B1–based detector on a large, diverse dataset spanning GANs and diffusion models, including both face and non-face content, and evaluates across resolutions and compressions. The detector achieves strong in-engine performance and robust out-of-domain generalization for many engines, with evidence that its learned signal is semantic/structural rather than a brittle low-level artifact, as shown by explainability analyses. The study demonstrates practical potential for platform-scale defenses against synthetic profile photos, while acknowledging an ongoing arms race with evolving generative technologies and laundering techniques.

Abstract

AI-based image generation has continued to rapidly improve, producing increasingly more realistic images with fewer obvious visual flaws. AI-generated images are being used to create fake online profiles which in turn are being used for spam, fraud, and disinformation campaigns. As the general problem of detecting any type of manipulated or synthesized content is receiving increasing attention, here we focus on a more narrow task of distinguishing a real face from an AI-generated face. This is particularly applicable when tackling inauthentic online accounts with a fake user profile photo. We show that by focusing on only faces, a more resilient and general-purpose artifact can be detected that allows for the detection of AI-generated faces from a variety of GAN- and diffusion-based synthesis engines, and across image resolutions (as low as 128 x 128 pixels) and qualities.

Finding AI-Generated Faces in the Wild

TL;DR

This work tackles the problem of distinguishing real faces from AI-generated faces in the wild, with a focus on fake online profiles. It trains an EfficientNet-B1–based detector on a large, diverse dataset spanning GANs and diffusion models, including both face and non-face content, and evaluates across resolutions and compressions. The detector achieves strong in-engine performance and robust out-of-domain generalization for many engines, with evidence that its learned signal is semantic/structural rather than a brittle low-level artifact, as shown by explainability analyses. The study demonstrates practical potential for platform-scale defenses against synthetic profile photos, while acknowledging an ongoing arms race with evolving generative technologies and laundering techniques.

Abstract

AI-based image generation has continued to rapidly improve, producing increasingly more realistic images with fewer obvious visual flaws. AI-generated images are being used to create fake online profiles which in turn are being used for spam, fraud, and disinformation campaigns. As the general problem of detecting any type of manipulated or synthesized content is receiving increasing attention, here we focus on a more narrow task of distinguishing a real face from an AI-generated face. This is particularly applicable when tackling inauthentic online accounts with a fake user profile photo. We show that by focusing on only faces, a more resilient and general-purpose artifact can be detected that allows for the detection of AI-generated faces from a variety of GAN- and diffusion-based synthesis engines, and across image resolutions (as low as 128 x 128 pixels) and qualities.
Paper Structure (14 sections, 5 figures, 2 tables)

This paper contains 14 sections, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data sets
  4. Real Faces
  5. GAN Faces
  6. GAN Non-Faces
  7. Diffusion Faces
  8. Diffusion Non-Faces
  9. Training and Evaluation Data
  10. Model
  11. Results
  12. Explainability
  13. Comparison
  14. Discussion

Figures (5)

  • Figure 1: The evolution of statistical models of natural images: (a) a fractal pattern with a $1/\omega$ power spectrum; (b) a synthesized textile patternportilla2000parametric; (c) a GAN-generated face karras2021alias; and (d) a diffusion-generated scene with the prompt "a beekeeper painting a self portrait" stabilityAI.
  • Figure 2: Representative examples of AI-generated images used in our training and evaluation (see also Table \ref{['tab:datasets']}). Some synthesis engines were used to generate faces only and others were used to synthesize both faces and non-faces. In order to respect user privacy, we do not show examples of real photos.
  • Figure 3: True positive rate (TPR) for correctly classifying an AI-generated face (with a fixed FPR of $0.5\%$) as a function of: (a) resolution where the model is trained on $512 \times 512$ images and evaluated against different resolution (solid blue) and trained and evaluated on a single resolution $N \times N$ (dashed red); and (b) JPEG quality where the model is trained on uncompressed images and a range of JPEG compressed images and evaluated on JPEG qualities between $20$ (lowest) to $100$ (highest).
  • Figure 4: The Margaret Thatcher illusion thompson1980margaret: the faces in the top row are inverted versions of those on the bottom row. The eye and mouth inversion in the bottom right is evident when the face is upright, but not when it is inverted. (Credit: Rob Bogaerts/Anefo https://commons.wikimedia.org/w/index.php?curid=79649613))
  • Figure 5: Examples of AI-generated faces and their normalized integrated gradients, revealing that our model is primarily focused on facial regions: (a) an average of $100$ StyleGAN 2 faces, (b) DALL-E 2, (c) Midjourney, (d,e) Stable Diffusion 1,2.