| Title |
Distinguishing real from synthetic: machine learning approaches for distinguishing ai-created images from photographs / |
| Translation of Title |
Atskyrimas tarp tikro ir dirbtinio: mašininio mokymo metodai dirbtinio intelekto sukurtų vaizdų atskyrimui nuo nuotraukų. |
| Authors |
Zabotka, Mindaugas |
| Full Text |
|
| Pages |
42 |
| Keywords [eng] |
neuroniniai tinklai, GAN, VAE, difuzijos modeliai, sugeneruoti vaizdai neural networks, GAN, VAE, diffusion model, generated images |
| Abstract [eng] |
This work addresses the growing challenge of distinguishing between AI-generated images and real photographs, a critical issue in maintaining authenticity and trust in digital environments. With advancements in generative models such as GANs, VAEs, and diffusion models, AI-generated images have become increasingly realistic, raising ethical, social, and security concerns. The research investigates the feasibility of creating a universal neural network model to classify AI-generated and real images effectively. A diverse dataset of 2,060 images was created. Several neural networks, using the Vision Transformer (ViT) architecture, were trained on this dataset and evaluated for their ability to generalize to images from previously unseen generative models. Key findings reveal that while the models achieved higher accuracy on known and diffusion-based unseen generators, they struggled significantly with data from generators such as Recraft. This emphasises the inherent difficulty of developing a universal model due to the absence of consistent features across generative models and the rapid evolution of these technologies. The study contributes valuable insights into the limitations of current detection methodologies. |
| Dissertation Institution |
Vilniaus universitetas. |
| Type |
Master thesis |
| Language |
English |
| Publication date |
2025 |
