Reproduce “Identification, Amplification and Measurement: A bridge to Gaussian Differential Privacy” 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
Yi Liu, Ke Sun, Linglong Kong, Bei Jiang
A new framework enables the formal identification, precise measurement, and practical amplification of Gaussian Differential Privacy (GDP) properties for algorithms previously characterized by other differential privacy notions. This allows for tighter privacy composition guarantees and enhanced interpretability of privacy mechanisms.
Key insights
- An algorithm is μ-GDP if and only if the tail condition (as ε→∞) of its privacy profile δ_A(ε) vanishes at a rate of O(e^(-ε^2)), allowing for objective identification.
- The Gaussian Differential Privacy Transformation (GDPT) acts as an order-preserving mapping from privacy profiles to μ values, simplifying the verification and comparison of GDP guarantees.
- The partial derivative of μ_GDP(ε, δ) with respect to ε is uniformly bounded, which enables the construction of efficient staircase approximations for precise numerical measurement of μ.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.