jax_privacy.experimental.monte_carlo.delta_calculation.get_base_delta#
- jax_privacy.experimental.monte_carlo.delta_calculation.get_base_delta(num_samples, target_delta, support_bound=1.0)[source]#
The base_delta for Monte Carlo verification to achieve target_delta.
In more detail, if we use num_samples samples to verify that a mechanism satisfies (epsilon, base_delta)-DP, then we can formally report that the mechanism satisfies (epsilon, target_delta)-DP.
This method assumes we are doing many Monte Carlo verifications of different ordered mechanisms, and then picking the highest-utility mechanism that meets our target privacy guarantee. If we are only verifying one mechanism (which requires some chance of an empty output which is usually undesirable in practice), a tighter bound is possible.
- Parameters:
num_samples (
Union[_Buffer,_SupportsArray[dtype[Any]],_NestedSequence[_SupportsArray[dtype[Any]]],complex,bytes,str,_NestedSequence[complex|bytes|str]]) – The number of samples used in each Monte Carlo verification.target_delta (
float) – The value such that we want to report the overall mechanism of Monte Carlo verification and then running the best verified mechanism is (epsilon, target_delta)-DP.support_bound (
Union[_Buffer,_SupportsArray[dtype[Any]],_NestedSequence[_SupportsArray[dtype[Any]]],complex,bytes,str,_NestedSequence[complex|bytes|str]]) – An upper bound on each (weighted) sample of the hockey-stick divergence, for each verification. Seeget_overall_delta().
- Return type:
float- Returns:
The base_delta such that we use Monte Carlo verification to check if each mechanism satisfies (epsilon, base_delta)-DP.