Property Existence Inference against Generative Models

Generative models have served as the backbone of versatile tools with a wide range of applications across various fields in recent years. However, it has been demonstrated that privacy concerns, such as membership information leakage of the training dataset, exist for generative models. In this paper, we perform property existence inference against generative models as a new type of information leakage, which aims to infer whether any samples with a given property are contained in the training set. For example, to infer if any images (i.e., samples) of a specific brand of cars (i.e., property) are used to train the target model. We focus on the leakage of existence information of properties with very low proportions in the training set, which has been overlooked in previous works. We leverage the feature-level consistency of the generated data with the training data to launch inferences and validate the property existence information leakage across diverse architectures of generative models. We have examined various factors influencing the property existence inference and investigated how generated samples leak property existence information. In our conclusion, most generative models are vulnerable to property existence inferences. Additionally, we have validated our attack in Stable Diffusion which is a large-scale open-source generative model in real-world scenarios, and demonstrated its risk of property existence information leakage. The source code is available at https://github.com/wljLlla/PEI_Code.