Source Attribution for Images Generated by Diffusion-Based Text-to-Image Models: Exploring the Forensics Approach
Xinqi Jiang, Jinyu Tian · 2024
As the image generation technology continues to advance, images generated by artificial intelligence have become ubiquitous on the internet, leading to a plethora of controversies regarding image copyright. Addressing this issue, tracing back to the source model of generated images has become a crucial approach to resolving it, aiding in establishing the legitimacy and ownership of images. Most existing approaches primarily focus on discerning image authenticity while overlooking the issue of source attribution for generated images. Few traditional image attribution methods are constrained by temporal limitations, only applicable to attributing past generative models such as GANs, while their effectiveness is limited for newer methods like diffusion generative models that have emerged in recent years. This paper proposes a novel method capable of attributing images generated by text-to-image diffusion models and maintaining effectiveness even for untrained models. We construct a semantic feature extraction network and train a feature-difference recognition model using a siamese network approach based on the semantic similarity of generated images. We collect and construct a dataset for training and testing, validating the outstanding performance of our method. Our method enhances compatibility with black-box models, ensuring effective source identification for images generated through text input. Through rigorous experimental validation, our method demonstrates significant progress, providing effective forensics approach for attributing generated images.