Generalized PTR: User-Friendly Recipes for Data-Adaptive Algorithms with Differential Privacy

TL;DR

This paper extends the classic PTR framework to a more general, data-adaptive setting for differential privacy, enabling its application to a broader range of algorithms and addressing complex privacy challenges.

Contribution

It introduces a generalized PTR framework that privately tests data-dependent privacy losses, broadening applicability beyond standard mechanisms and solving an open problem in private model release.

Findings

Successfully applied to private linear regression.

Privately released entire models in PATE setting.

Enhanced flexibility in data-adaptive differential privacy.

Abstract

The ''Propose-Test-Release'' (PTR) framework is a classic recipe for designing differentially private (DP) algorithms that are data-adaptive, i.e. those that add less noise when the input dataset is nice. We extend PTR to a more general setting by privately testing data-dependent privacy losses rather than local sensitivity, hence making it applicable beyond the standard noise-adding mechanisms, e.g. to queries with unbounded or undefined sensitivity. We demonstrate the versatility of generalized PTR using private linear regression as a case study. Additionally, we apply our algorithm to solve an open problem from ''Private Aggregation of Teacher Ensembles (PATE)'' -- privately releasing the entire model with a delicate data-dependent analysis.