Module private_hyperparameter_optimization 

Differentially private hyperparameter optimization.

Module layout

module	responsibility
types	configuration and the search-component data types
budget_manager	privacy budget allocation and accounting
optimizer	the PrivateHyperparameterOptimizer driver
results	private aggregation and the selection report
selection	the private selection mechanisms and noisy statistics
gaussian_process	the GP surrogate and acquisition function
random_search	NoisyOptimizer for random search
bayesian_optimization	NoisyOptimizer for Bayesian optimization
functions	the NoisyOptimizer trait and function aliases
trait_impls	the Default impls

0.3.2 notes

Hyperparameter selection is now differentially private. Previously HyperparameterNoiseMechanism was stored and never matched on: the choice was the exact argmax over utilities computed from the private data, which is precisely what private HPO exists to avoid.

Behavioural and API changes:

• optimizer::PrivateHyperparameterOptimizer::new rejects a configuration with private_model_selection: true that declares no objective sensitivity.
• types::PrivateHPOResults gained selection, which records whether the returned configuration was chosen privately, by which mechanism, and at what cost.
• types::SelectionParameters gained delta (needed to calibrate Gaussian selection).
• PrivateResultsAggregator::aggregate_results takes &mut self, because selecting now spends budget.
• The 16 *_traits.rs shells were collapsed into trait_impls; the two that held real NoisyOptimizer implementations were renamed to random_search and bayesian_optimization.

Re-exports

pub use budget_manager::AdaptiveBudgetController;

pub use budget_manager::HPOBudgetManager;

pub use budget_manager::DEFAULT_SELECTION_BUDGET_FRACTION;

pub use gaussian_process::encode_configuration;

pub use gaussian_process::ConfigurationEncoding;

pub use gaussian_process::ExpectedImprovement;

pub use gaussian_process::GaussianProcessFit;

pub use optimizer::PrivateHyperparameterOptimizer;

pub use results::PrivateResultsAggregator;

pub use results::SelectionReport;

pub use results::PRIVATE_TOP_K;

pub use selection::exponential_mechanism_index;

pub use selection::exponential_mechanism_probabilities;

pub use selection::gaussian_sigma;

pub use selection::laplace_sample;

pub use selection::mechanism_name;

pub use selection::noisy_summary_statistics;

pub use selection::report_noisy_max_gaussian;

pub use selection::report_noisy_max_gumbel;

pub use selection::report_noisy_max_laplace;

pub use selection::summary_mean_noise_scale;

pub use selection::NoisySummary;

pub use selection::SelectionOutcome;

pub use selection::OBJECTIVE_SENSITIVITY_KEY;

pub use selection::SUMMARY_QUANTILES;

pub use functions::*;

pub use types::*;

Modules

bayesian_optimization

NoisyOptimizer for private Bayesian optimization.

budget_manager

Privacy budget accounting for hyperparameter optimization.

functions

Auto-generated module

gaussian_process

Gaussian-process surrogate and acquisition function for private Bayesian optimization.

optimizer

The private hyperparameter optimizer driver.

random_search

NoisyOptimizer for private random search.

results

Private aggregation and reporting of hyperparameter search results.

selection

Differentially private selection of hyperparameter configurations.

trait_impls

Default implementations for the private-HPO components.

types

Auto-generated module