International Joint Conference on Neural Networks (IJCNN 2025) – VERIMEDIA Workshop
The paper introduces WILD, a new dataset designed for synthetic image source attribution, addressing the challenge of identifying which generative model created a given synthetic image. The dataset is divided into two parts:
-
Closed set: Contains 10,000 images from 10 popular commercial and open-source text-to-image generators (e.g., Adobe Firefly, DALL-E 3, Midjourney), with 1,000 images per generator. Images are generated using fixed, randomized prompts to ensure diversity and minimize bias.
-
Open set: Includes 10,000 images from 10 additional generators (e.g., StyleGAN variants, DALL-E 1) to simulate real-world scenarios where unknown models may be encountered.
Half of the images in both sets are post-processed (e.g., compressed, cropped, resized) to test model robustness. We evaluated seven baseline methods (e.g., CLIP-based classifiers, CNNs, Vision Transformers) on WILD. Key findings:
-
CNNs excel on plain images but degrade with post-processing.
-
Vision Transformers (ViT) show greater robustness to post-processing.
-
Open-set attribution remains challenging, with performance dropping significantly for post-processed images.
Thanks to all the forensic researchers from various Italian institutions who contributed to this study! 
IPLAB Authors: Mirko Casu, Orazio Pontorno, Claudio Ragaglia, Luca Guarnera, Sebastiano Battiato