<![CDATA[The rapid evolution of generative artificial intelligence (AI) models such as DALL·E, Midjourney, and Stable Diffusion has intensified the risk of visual disinformation, as synthetic images increasingly resemble real ones. Traditional binary detection methods (real vs. fake) have become insufficient, creating a growing need for provenance attribution, i.e., identifying the specific generative model responsible for producing an image. This study presents a comparative evaluation of two Convolutional Neural Network (CNN) architectures—XceptionNet and EfficientNetB0—for multi-class attribution of synthetic images. The primary objective is to evaluate and compare the effectiveness of both architectures in attributing images to four specific source classes: DALL·E, Midjourney, Stable Diffusion, and real images. Both models were trained and tested using a transfer learning approach on a balanced dataset of 2,000 samples and assessed using accuracy, precision, recall, and F1-score. Experimental results show that EfficientNetB0 outperforms XceptionNet, achieving 95.2% accuracy compared to 93.5%, while also exhibiting more stable training behavior and stronger discriminative capability for visually similar classes. The findings indicate that EfficientNetB0 offers a more reliable balance of computational efficiency and feature extraction performance, making it a suitable architecture for provenance attribution tasks involving generative AI imagery. ]]>
Copyrights © 2026
