A forensic evaluation method of stable diffusion-generated images using feature-based likelihood ratio by deep learning features

With the increasing realism of artificial intelligence (AI)-generated images from Stable Diffusion and similar models, forensic practitioners face significant challenges in image authenticity verification. This study proposes a feature-based likelihood ratio model based on deep learning features to assist forensic experts in identifying stable diffusion-generated images, with the aim of enhancing the judicial admissibility of such authentication conclusions. First, we built the development set and validation set comprising authentic images from ImageNet and generated images from Stable Diffusion v1.4. Subsequently, we trained the Swin-transformer algorithm to map the images to the feature space and determined classification through feature values analysis. The method achieved 99.4% detection accuracy on a test set. After that, we constructed a feature likelihood ratio model using deep learning features from the development set and evaluated its performance with the validation set through four assessment metrics: Tippett plot, equal error rate (EER), cost of log likelihood ratio (Cllr) values, and empirical cross-entropy (ECE) curves. The results demonstrate that the proposed feature-based likelihood ratio model with deep learning features exhibits superior performance.