Preventing Sensitive Data Generation with Positive-unlabeled Diffusion Models

Abstract Diffusion models often generate sensitive data that are unwanted by users, mainly because the unlabeled training data frequently contain such sensitive data. However, labeling all sensitive data in the large-scale unlabeled training data is impractical. To solve this, my research colleagues and I propose positive-unlabeled diffusion models, which prevent the generation of sensitive data using a small amount of labeled sensitive data in addition to unlabeled training data. If we have access to clean normal data, then we can prevent the generation of sensitive data since such sensitive data are excluded from the training data. Our key idea is to approximate the training objective function of normal (negative) data using only unlabeled and sensitive (positive) data. Therefore, even without labeled normal data, we can maximize the training objective function for normal data and minimize it for labeled sensitive data, ensuring the generation of only normal data.