PAC Privacy Preserving Diffusion Models

arXiv:2312.01201v2 Announce Type: replace
Abstract: Data privacy protection is garnering increased attention among researchers. Diffusion models (DMs), particularly with strict differential privacy, can potentially produce images with both high privacy and visual quality. However, challenges arise such as in ensuring robust protection in privatizing specific data attributes, areas where current models often fall short. To address these challenges, we introduce the PAC Privacy Preserving Diffusion Model, a model leverages diffusion principles and ensure Probably Approximately Correct (PAC) privacy. We enhance …

abstract arxiv attention challenges cs.ai cs.lg current data data privacy differential differential privacy diffusion diffusion models however images privacy privacy preserving protection quality researchers robust type visual