Home // SECURWARE 2024, The Eighteenth International Conference on Emerging Security Information, Systems and Technologies // View article

Forensic Analysis of GAN Training and Generation: Output Artifacts Assessment of Circles and Lines

Authors:
Stefan Seidlitz
Jana Dittmann

Keywords: Media Forensic; Generative Adversarial Networks (GAN); DeepFake; Advance and Challenges.

Abstract:
Motivated by the challenges in different forensic detection tasks based on machine learning, this paper evaluates the training behavior, as well as the generation performance of images which are generated by Generative Adversarial Networks (GANs) based on simple geometric shapes using the example of circles and lines. Circles, for example, are relevant for DeepFaceFakes where eyes might be checked for inconsistencies. Therefore, we trained several StyleGAN3 models with different self-created training data sets using geometrical shapes of circles and lines. We use these models to generate fake circles and lines from a random latent vector, which we then forensically analyzed in two different ways: a visual, subjective evaluation based on an observation as well as an automated Circle-Checking approach. In both experiments, we were able to show on the example of StyleGAN3, that generative approaches have difficulties with the generation of geometric shapes: circles are often more comparable to eggs, lines are mostly not linear. Our contribution is to advance the knowledge on what kind of artifacts a Generative Adversarial Network generates. This gives a first tendency for new detection strategies to identify these artifacts, based on geometrical shapes in DeepFake images.

Pages: 141 to 147

Copyright: Copyright (c) IARIA, 2024

Publication date: November 3, 2024

Published in: conference

ISSN: 2162-2116

ISBN: 978-1-68558-206-7

Location: Nice, France

Dates: from November 3, 2024 to November 7, 2024