Module selection 

Differentially private selection of hyperparameter configurations.

The defect this replaces

HyperparameterNoiseMechanism was stored on PrivateHPOConfig and never matched on anywhere in the crate: no exponential mechanism, no report-noisy-max, no noise on the choice at any point. PrivateResultsAggregator::aggregate_results sorted the evaluations exactly and returned the exact top five, and optimize() tracked the exact argmax.

Private hyperparameter optimization is entirely about privatising the selection step (Liu & Talwar, STOC 2019; Chaudhuri, Monteleoni & Sarwate, JMLR 2011), so an exact argmax over utilities computed from private data leaks the selection and provides no guarantee for the chosen configuration.

What is implemented

• The exponential mechanism (McSherry & Talwar, FOCS 2007): index i is returned with probability proportional to exp(epsilon * u_i / (2 * Delta_u)). The weighting itself is delegated to the crate’s audited crate::privacy::noise_mechanisms::ExponentialMechanism rather than reimplemented here.
• Report-noisy-max with Gumbel noise, which is equivalent in distribution to the exponential mechanism – asserted by a test in this module – and with Laplace noise at scale 2 Delta_u / epsilon.
• A Gaussian argmax at scale sqrt(2 ln(1.25/delta)) * 2 Delta_u / epsilon, which requires a delta and errors when none is configured.
• SparseVector selection is refused: the sparse vector technique answers a stream of threshold queries and is not a one-shot selection primitive. Use crate::privacy::noise_mechanisms::SparseVectorMechanism.

Every selection reports the epsilon it consumed so the caller can charge it.

Structs

NoisySummary

A differentially private summary, together with what it actually cost.

SelectionOutcome

Outcome of one private selection.

Constants

OBJECTIVE_SENSITIVITY_KEY

Key under which the objective’s global sensitivity is looked up in SensitivityBounds::global_sensitivity.

SUMMARY_QUANTILES

Quantiles released by noisy_summary_statistics, in order.

Functions

exponential_mechanism_index

Select an index with the exponential mechanism.

exponential_mechanism_probabilities

The exponential mechanism’s selection probabilities, in closed form.

gaussian_sample

One Gaussian sample with standard deviation sigma, by Box-Muller.

gaussian_sigma

Analytic Gaussian-mechanism standard deviation for (epsilon, delta).

gumbel_sample

One standard Gumbel sample.

laplace_sample

One Laplace sample with scale b.

mechanism_name

Human-readable mechanism name.

noisy_summary_statistics

Differentially private summary statistics of the observed objectives.

report_noisy_max_gaussian

Argmax after adding Gaussian noise calibrated for (epsilon, delta).

report_noisy_max_gumbel

Report-noisy-max with Gumbel noise.

report_noisy_max_laplace

Report-noisy-max with Laplace noise at scale 2 Delta / epsilon.

summary_mean_noise_scale

The Laplace scale noisy_summary_statistics uses for the mean release.