A Unified Traceability Framework for Diffusion Model Generated Images

Abstract Nowadays, diffusion models have become dominant generative tools capable of producing highly realistic images, while triggering new challenges for image provenance verification. Traditional content-based and metadata-based traceability methods fail to reliably distinguish synthetic from authentic images, while existing learning-based approaches face scalability and robustness limitations. To bridge the gap, we propose a novel unified traceability framework that integrates two complementary strategies: Arcits and Aswomts. Specifically, Arcits targets image-level provenance by exploiting reconstruction residuals as image fingerprints and combining them with lightweight model fingerprints extracted from the U-Net architecture. Aswomts extends to model-level provenance by encoding structural and weight information of diffusion models via graph neural networks and weighted similarity modeling, enabling the quantification of model similarity. Comprehensive experiments on multiple datasets demonstrate that Arcits outperforms baseline methods such as PRNU, KNN, ResNet50, and VGG16 in accuracy and robustness against perturbation attacks, while Aswomts achieves reliable model attribution, surpassing traditional similarity metrics like cosine similarity and SSIM. Together, these strategies provide a scalable and robust solution for diffusion-model traceability, advancing both image- and model-level provenance analysis.