Abstract
The rapid advancement of generative models has led to the creation of highly realisticvisuals, raising concerns about the spread of misinformation and fake news in digital
contexts. This thesis investigates the human ability to distinguish between authentic
and AI-generated images, comparing it to the performance of AI-based detection models. The comparison focuses on commonly used performance metrics such as accuracy,
precision, and recall.
The thesis begins by examining the broader context of generative modelling and
reviewing state-of-the-art AI image detectors. In addition, it provides an overview of
existing research on the perception of AI-generated content, highlighting the associated
challenges and risks.
The main part of the thesis is concerned with the evaluation of human performance in
distinguishing AI-generated images from authentic-human ones. A baseline data set was
created to facilitate a direct comparison between human detection ability and AI-based
detector models. A human baseline study was conducted to establish a benchmark,
also providing insight into human decision-making processes. Subsequently, detector
models were trained on a large-scale data set to assess whether classifiers can achieve
competitive performance with limited resources.
Two architectures were developed and evaluated against the human benchmark: a
custom model and a transfer learning model. The results indicate that, while humans
slightly outperform the models in accuracy (56%) and precision (54.72%), the models
excel in recall (74%), identifying a greater proportion of AI-generated images. Both
architectures performed similarly across all metrics, demonstrating that it is feasible to
train models with limited resources that approximate human performance. The findings
suggest that AI models, particularly with their higher recall, are more effective for comprehensive detection of AI-generated imagery, especially in the context of digital media.
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Ulrich Bodenhofer (Supervisor) |