Training Humans for Synthetic Face Image Detection

Ramlah Sara RehmanEwald MeierMathias IbsenChristian Rathgeb^*Robert NicholsChristoph Busch

• Darmstadt University of Applied Sciences, Darmstadt, Germany

Fake identities created using highly realistic synthetic face images have become increasingly prevalent in recent years, driven by advancements in generative neural networks that are readily accessible online and easy to use. These fake identities can be exploited for malicious purposes, such as spreading misinformation or committing fraud. Given the widespread availability of online content and the ease of generating fake online identities, it is desirable that users are able to distinguish real face images from synthetic ones. Additionally, it is important to explore whether specialised training can enhance the ability of individuals to detect synthetically generated face images. In this work, we address these challenges by designing an online experiment to evaluate human detection capabilities and the impact of training on detecting synthetic face images. As part of the experiments, we recruited 184 participants divided into an experimental group and a control group, where the experimental group underwent a tailored training session halfway through the experiment. The study shows that training may moderately enhance human capabilities to detect synthetic face images. Specifically, it was found that the experimental group generally outperformed the control group after training, primarily due to improved abilities in detecting synthetic face images. However, after training, the experimental group showed increased sensitivity and misclassified also more authentic face images, as compared to the control group.