Task-aware Kernel Stein Certificates for Renyi Differential Privacy

Forum topic · 小凯 · 2026-10-20

Summary

A paper by Leoni Carla Wirth, Tara Trautwein, and Gesine Reinert (arXiv:2610.23798) introduces kernel Stein discrepancy-based certificates for Renyi differential privacy. Renyi divergence has become a standard basis for privacy guarantees, but it is difficult to compute outside the Gaussian setting, especially for task-specific frameworks such as Renyi Pufferfish privacy. The authors use kernel Stein discrepancies constructed from escort distributions to certify Renyi differential privacy, and by designing kernels that focus on the relevant privacy directions, they derive task-aware privacy certificates. The paper provides exact identities for the Gaussian and Laplace mechanisms and upper bounds for exponential mechanisms, along with separate task-aware guarantees. A real-world dataset example demonstrates how the protected comparison set and the released subspace jointly determine the resulting certificate.

Paper Overview

Field: Machine Learning Authors: Leoni Carla Wirth, Tara Trautwein, Gesine Reinert Published: 2026-10-16 arXiv: 2610.23798

Summary

Renyi differential privacy, based on Renyi divergence, has become a notion of choice for privacy guarantees. However, outside the Gaussian setting, Renyi divergence can be difficult to compute. In particular, Renyi differential privacy guarantees that exploit task-specific information, as in Renyi Pufferfish privacy, can be challenging.

This paper uses kernel Stein discrepancies based on escort distributions to provide certificates for Renyi differential privacy. By designing the kernel to focus on the relevant privacy directions, the authors also derive task-aware privacy certificates.

Contributions

Kernel Stein discrepancy certificates for Renyi differential privacy using escort distributions.
Task-aware privacy certificates via kernels focused on relevant privacy directions.
Exact identities for the Gaussian and Laplace mechanisms.
Upper bounds for exponential mechanisms, with separate task-aware guarantees.
A real dataset example illustrating how the protected comparison set and the released subspace jointly determine the final certificate.

---

*Automatically collected on 2026-10-20*

Task-aware Kernel Stein Certificates for Renyi Differential Privacy

Summary

Paper Overview

Summary

Contributions

Tags