Reproduce “Privacy Profiles for Private Selection” 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
Antti Koskela, Rachel Redberg, Yu-Xiang Wang
Privacy Profiles for Private Selection develops a direct approximate-DP analysis for selecting the best result from a random number of private candidates. By working with hockey-stick divergence and the base mechanism’s full privacy profile rather than converting through RDP, the framework gives tighter accounting for Report Noisy Max, Generalized PTR, and DP-SGD hyperparameter tuning, including settings where RDP guarantees are unavailable.
Key insights
- The derivative of the candidate-count probability-generating function characterizes how repeated evaluations alter the distribution of the selected maximum, providing a common route to private-selection bounds.
- Direct privacy-profile accounting can avoid looseness introduced by converting RDP guarantees into (ε,δ)-DP, while also supporting mechanisms such as Generalized PTR that do not naturally admit RDP analysis.
- Binomial randomization offers a controllable alternative to Poisson randomization, reducing variability in the number of evaluations at comparable privacy cost.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.