Fix gaussian tau

CHANGELOG.md

@@ -6,6 +6,10 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
+## [0.9.27] - 2024-12-11
+### Fixed
+- Fix gaussian tau.
+
 
 ## [0.9.26] - 2024-12-05
 ### Fixed

Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 authors = ["Nicolas Grislain <ng@sarus.tech>"]
 name = "qrlew"
-version = "0.9.26"
+version = "0.9.27"
 edition = "2021"
 description = "Sarus Qrlew Engine"
 documentation = "https://docs.rs/qrlew"
@@ -34,7 +34,7 @@ r2d2_postgres = "0.18"
 rust_decimal = { version = "1.35", features = [ "tokio-pg" ] }
 statrs = "0.16.0"
 sqlx = { version = "0.6.3", features = ["mssql", "runtime-tokio-native-tls", "offline", "any"], optional = true }
-tokio = { version = "1", features = ["full"], optional = true }
+tokio = { version = "1.37.0", features = ["full"], optional = true }
 
 # bigquery dependencies
 gcp-bigquery-client = { version = "=0.20.0", optional = true }

src/differential_privacy/dp_event.rs

@@ -1,8 +1,5 @@
 use itertools::Itertools;
-use statrs::{
-    distribution::{ContinuousCDF, Normal},
-    prec::F64_PREC,
-};
+use statrs::distribution::{ContinuousCDF, Normal};
 use std::fmt;
 
 /// An object inspired by Google's [DPEvent](https://github.com/google/differential-privacy/blob/main/python/dp_accounting/dp_event.py)
@@ -169,9 +166,9 @@ pub fn gaussian_noise_multiplier(epsilon: f64, delta: f64) -> f64 {
     ((2. * (1.25_f64 / delta).ln()).sqrt() / epsilon).clamp(0.0, f64::MAX)
 }
 
-pub fn gaussian_tau(epsilon: f64, delta: f64, sensitivity: f64) -> f64 {
+pub fn gaussian_tau(epsilon: f64, delta: f64, max_privacy_unit_groups: f64) -> f64 {
     let dist = Normal::new(0.0, 1.0).unwrap();
-    let scale = gaussian_noise(epsilon, delta, sensitivity);
+    let scale = gaussian_noise(epsilon, delta, max_privacy_unit_groups.sqrt());
     // TODO: we want to overestimate tau
-    1. + scale * dist.inverse_cdf((1. - delta / 2.).powf(1. / sensitivity)) + F64_PREC
+    1. + scale * dist.inverse_cdf((1. - delta).powf(1. / max_privacy_unit_groups))
 }

src/differential_privacy/dp_parameters.rs

@@ -39,7 +39,7 @@ impl DpParameters {
 
     pub fn from_epsilon_delta(epsilon: f64, delta: f64) -> DpParameters {
         // These default values are underestimating the bounds
-        DpParameters::new(epsilon, delta, 0.5, 100.0, 0.1, 1)
+        DpParameters::new(epsilon, delta, 0.5, 100.0, 0.1, 5)
     }
 
     pub fn with_tau_thresholding_share(self, tau_thresholding_share: f64) -> DpParameters {

src/differential_privacy/group_by.rs

@@ -107,7 +107,7 @@ impl PupRelation {
         // Apply noise
         let name_sigmas = vec![(
             COUNT_DISTINCT_PID,
-            dp_event::gaussian_noise(epsilon, delta, max_privacy_unit_groups as f64),
+            dp_event::gaussian_noise(epsilon, delta, (max_privacy_unit_groups as f64).sqrt()),
         )];
         let rel = rel.add_gaussian_noise(&name_sigmas);