Identity-Focused Inference and Extraction Attacks on Diffusion Models

These models have been widely adopted across industries such as healthcare [Wolleb et al.(2022)] and the creative arts [Saharia et al.(2022)] due to their ability to generate high-fidelity synthetic content. However, with the access to personal images from social media and other online data stores, concerns regarding the inclusion of sensitive data, particularly facial images [Kim et al.(2023)] [Huang et al.(2023)], without the knowledge or consent of the data owners have become increasingly prevalent. This issue raises significant challenges related to privacy, intellectual property, and the ethical use of personal data in AI systems. A central challenge in this context is determining whether data related to a specific individual's identity was used to train these models. In this paper, we introduce the concept of identity inference, which holds model owners accountable for the potential unauthorized use of personal data. Unlike traditional membership inference, which seeks to determine whether a particular data point was part of the training set, identity inference focuses on detecting whether any known or unknown data point related to the individual's identity was used. As diffusion models become more prominent, especially in domains involving sensitive data like facial images, the risk of training on unauthorized data becomes a growing concern [Miernicki and Ng(2021)].