Positive-Unlabeled Diffusion Models for Preventing Sensitive Data Generation

2025-03-05

AI Summary

Diffusion models often inadvertently generate sensitive content implicitly present in training data, yet large-scale unlabeled datasets are prohibitively difficult to comprehensively sanitize. Method: We propose the first positive–unlabeled (PU) learning framework tailored for diffusion models, requiring only a small number of labeled sensitive samples and no labels for normal samples to selectively suppress sensitive generation. Our approach integrates PU learning with variational inference by reparameterizing the reverse diffusion process and optimizing a lower bound on the evidence lower bound (ELBO) for normal data—thereby enabling provably falsifiable suppression of sensitive generation. Results: Evaluated on multiple benchmark datasets, our method significantly reduces the rate of sensitive image generation while preserving generation fidelity, with negligible degradation in Fréchet Inception Distance (ΔFID < 0.5), thus achieving a robust trade-off between safety and quality.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Security and Privacy: Data transparency and provenance

Abstract

Diffusion models are powerful generative models but often generate sensitive data that are unwanted by users, mainly because the unlabeled training data frequently contain such sensitive data. Since labeling all sensitive data in the large-scale unlabeled training data is impractical, we address this problem by using a small amount of labeled sensitive data. In this paper, we propose positive-unlabeled diffusion models, which prevent the generation of sensitive data using unlabeled and sensitive data. Our approach can approximate the evidence lower bound (ELBO) for normal (negative) data using only unlabeled and sensitive (positive) data. Therefore, even without labeled normal data, we can maximize the ELBO for normal data and minimize it for labeled sensitive data, ensuring the generation of only normal data. Through experiments across various datasets and settings, we demonstrated that our approach can prevent the generation of sensitive images without compromising image quality.

Problem

• Prevent generation of sensitive data in diffusion models
• Use limited labeled sensitive data to train models
• Maintain image quality while avoiding sensitive outputs

Innovation

• Positive-unlabeled diffusion models prevent sensitive data generation.
• Approximates ELBO for normal data using unlabeled and sensitive data.
• Ensures only normal data generation without labeled normal data.