Identifying Provenance of Generative Text-to-Image Models

Fine-tuning provides a fast and cheap way to produce new text-to-image models that are often indistinguishable from ones trained from scratch. Unfortunately, misrepresentation of fine-tuned models creates problems for AI companies and users alike, by disincentivizing competition and misleading users on model quality and ethics of its training process.

In this paper, we propose a model provenance system that identifies models produced by fine-tuning on existing base text-to-image models, using only black-box query access to the models. Our design is informed by analysis showing that one can quantify the feature space difference between textto-image models by analyzing their responses to detailed prompts. Given a target model, our system analyzes its output, extracts visual features using a generic feature extractor, and compares the distribution against those derived from a pool of base models using Jensen-Shannon divergence. We then apply statistical hypothesis testing to determine if the target model is trained from scratch or fine-tuned, and if the latter, the likely base (parent) model. We evaluate our system across seven popular diffusion models and numerous fine-tuned variants. Our results show high accuracy in attributing model lineage, even under adversarial conditions such as image postprocessing or weight perturbations. Finally, we demonstrate real-world efficacy of our system by tracing provenance of in-the-wild models from popular online platforms.

Model trainer fine-tunes and claims ownership of the fine-tuned model.