Ewald Meier successfully defended his Master thesis on Human Detection of Synthetically generated Face Images

The rapid development of generative AI, especially in the area of image
generation, brings many benefits to the general public. But all these benefits
come at a price. As the generated images become more realistic and
thus harder to distinguish from the real ones, they also become more widely
available through many free and easy-to-use online tools. This increases the
danger these fake images pose to the general public. This creates the need
for a countermeasure to mitigate these dangers. This thesis seeks to provide
such a countermeasure in the form of a comprehensive guideline that can
be easily applied by the general public. In addition, a diverse dataset of fake
images needs to be created to provide a foundation for this study and to
allow for further studies in this area.
This led to the following two research questions, which will be answered
in the course of this thesis. „What should a dataset of fake images look like
in order to test the performance of humans in detecting them, considering
different easy-to-use tools and diversity of the dataset?“ and „How does a
guideline for humans look like to improve their performance in detecting
fake images?“.
In order to answer these questions, the landscape of currently available online
tools that allow the creation of fake images is first assessed using a total
of 7 defined criteria. Then, 4 of the best performing tools are selected to create
the dataset. To ensure a diverse dataset, the images created are evenly
distributed across 2 genders, 4 age groups, and 4 ethnicities. Real images
are also introduced into the dataset for comparison to and thus allow the
dataset to be used in research. Finally, all images are processed to have the
same style to shift the focus to the content of the images alone. This dataset
is then analyzed to find any anomalies that indicate a fake image. This is
used to create the guidelines.
The resulting dataset confirmed the potential dangers that these online tools
pose to the general public, as these images are in large part indistinguishable
from real images at first glance. The results suggest that some of the fake images
are even preferred by people over real images, which is consistent with
the current state of research. However, the guidelines created in this thesis
provide a viable countermeasure to this danger, as they offer a classification
accuracy of over 90% and are easy to apply, as they guide the user through
the entire image analysis process without requiring any prior knowledge.