jax_privacy.auditing

Library for empirical privacy auditing/estimation.

This library provides functions for estimating the privacy of a model, based on attack scores of held-in and held-out canaries

Classes

Bonferroni()	Use Bonferroni correction across all possible thresholds.
BootstrapParams([num_samples, quantiles, ...])	Parameters for bootstrapping.
CanaryScoreAuditor(in_canary_scores, ...)	Class for auditing privacy based on attack scores.
Explicit(threshold)	Use a specific threshold value.
MultiSplit([num_samples, ...])	Splits data multiple times with significance correction.
Split([threshold_estimation_frac, seed])	Split data to choose threshold and then compute the bound.
ThresholdStrategy()	Base class for threshold selection strategies.