Reproduce “Accuracy-First Rényi Differential Privacy and Post-Processing Immunity” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ossi Räisä, Antti Koskela, Antti Honkela
Researchers at CISPA Helmholtz Center and the University of Helsinki proved that pure ex-post differential privacy is post-processing immune while its probabilistic relaxation is not, and introduced "alpha"-ex-post Rényi Differential Privacy (RDP) which guarantees post-processing immunity. The developed framework, including a data-dependent stopping rule, was applied to synthetic data generation and image classifier fine-tuning, demonstrating adaptive privacy budgeting with validated utility.
Key insights
- Pure *ex-post* privacy holds post-processing immunity, confirming its robustness, while its "delta"-probabilistic relaxation (with "delta" > 0) does not, revealing a critical vulnerability in common relaxed *ex-post* definitions.
- Making the privacy bound an explicit part of a mechanism's output allows for a rigorous definition and proof of post-processing immunity for *ex-post* differential privacy.
- The Brownian mechanism, crucial for adaptive noise, satisfies the newly proposed "alpha"-ex-post RDP and its post-processing immunity, providing a robust tool for accuracy-first DP.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.