Custom sensitivities may be passed directly to mechanisms
- protect_sensitivity must be disabled in the privacy definition
Expand documentation for each component as well as Python Analysis class
Python Bindings: Analysis initializer only accepts keyword arguments.

Python Bindings: enable protect_floating_point by default.
- Real-valued queries are less susceptible to floating-point attacks, at the cost of utility
- Use sn.Analysis(protect_floating_point=False) to enable the laplace and (analytic) gaussian mechanisms
Fix noise scaling issues in the Gaussian and Analytic Gaussian mechanism
Fixes for Gaussian and Analytic Gaussian accuracy
Postprocess geometric mechanism noise with clamping
Compute sensitivities as integers whenever possible (counts, histograms, sums)
Added runtime sanity checks to detect violations of static properties in pre-aggregated data
O(n^2) -> O(n) runtime performance in exponential mechanism and categorical imputation
Fixed an incorrect inference of dataset size when transforming a dataset with unknown size against a broadcastable scalar
Unions always permitted on public data
Added inference of nature (categories, bounds) to ToInt
Plug-in mean derives bounds for sum in laplace and geometric mechanism

Renamed package to Smartnoise, version number reset
Added snapping mechanism
Added analytic gaussian mechanism
Added DP Linear Regression through the Theil-Sen transform and gumbel mechanism
Added generalized resize for privacy amplification by subsampling
Tightened c-stability checks to protect against adversarial dataset reordering when unioning data partitions
Modified error messages to contain suggested fixes
Bugfix to retain statistics when generating reports

Provide feedback

Saved searches