Paper derives exact privacy-utility bounds for diffusion denoisers
Original title: Diffusion Sensitivity Geometry and Strong Duality for Optimal Privacy-Utility Tradeoffs
Exact two-record sensitivity and an SDP replace loose diameter bounds and isotropic Gaussian noise.
A cs.CR paper argues that differential privacy for diffusion models often uses loose sensitivity bounds and isotropic Gaussian noise. It characterizes worst-case sensitivity of ideal diffusion denoisers using only two records, reducing analysis for a domain of size m to at most m(m-1) constraints. It also gives exact guarantees for several privacy notions and proves strong duality for an SDP that computes minimum weighted mean-squared error.
- Worst-case sensitivity depends on only two records, independent of dataset size.
- Guarantees cover Rényi, approximate, Gaussian, and mutual-information privacy.
- Global optimality is certified by strong duality of a semidefinite program.
From the original article
1 min read at arxiv.orgExisting approaches to differential privacy in diffusion models rely on loose sensitivity bounds and isotropic Gaussian noise, leading to unnecessary estimation error. We develop a framework for the optimal privacy-utility tradeoff of ideal diffusion denoisers, replacing conservative bounds with exact sensitivity characterizations and provably optimal noise calibration.
By Tue Do · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.