Generative machine learning models have enabled users to specify text prompts and obtain corresponding images. While generative models may be made available for use by cloud service providers, ensuring that users utilize the models in compliance with terms of service can be difficult. In particular, determining whether a particular image was generated by a particular model is difficult. Image watermarks, manual inspection, or machine learning techniques to inspect images do not produce conclusive evidence in this regard. This disclosure describes techniques to embed information about the generative model that generates an image and an associated token ID and session information into the media generated by the model, with minimal impact to the media. The embedded information can be utilized to determine whether the image was generated/modified by a particular model, and if so, the prompt used for such modification. The described techniques can improve authentication of media generated using generative models and can provide unambiguous evidence.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.