The noise a budget implies, and the sample size at which that noise
stops mattering.
Usage
morie_private_accuracy_tradeoff(
sensitivity = 1,
epsilon = 1,
n = 100,
confidence = 0.95
)
Arguments
- sensitivity
query sensitivity.
- epsilon
privacy budget.
- n
sample size the query averages over.
- confidence
coverage for the half-width.
Value
list with noise_scale, noise_sd,
half_width, relative_error, noise_to_signal_n.
Details
The number worth reporting is noise_to_signal_n, which is
\(2/\epsilon^2\): past that sample size the privacy noise is
smaller than the sampling error already in the estimate, so privacy
costs nothing anyone can measure. Quoting an epsilon alone tells a
reader nothing about whether that is the case here.
References
Dwork, C. and Roth, A. (2014). FnTTCS, 9(3-4).
Examples
morie_private_accuracy_tradeoff(1, 0.5, n = 1000)$noise_to_signal_n
#> [1] 8
