Differentially Private Classical Statistical Inference

Implements differentially private (DP) versions of common classical statistical procedures, including descriptive statistics (mean, variance, quantiles, histograms), hypothesis tests (t-test, chi-square, Kolmogorov-Smirnov, one-way ANOVA), and regression (closed-form DP linear regression and DP-SGD for generalized linear models). Provides Laplace and Gaussian mechanisms with analytic calibration, exponential mechanism for medians, privacy-aware confidence intervals that account for both sampling and privacy noise, and privacy budget accounting via basic, advanced, and Renyi differential privacy (RDP) composition. Designed for official statistics and privacy-preserving data analysis research.

DPrivStats — Differentially Private Classical Inference

An R package implementing differentially private (DP) versions of common classical statistical procedures, with privacy budget accounting and privacy-aware confidence intervals.

Features

• Mechanisms — Laplace (pure ε-DP), Gaussian (classic calibration), analytic Gaussian calibration (Balle & Wang 2018), exponential mechanism.
• Descriptive statistics — DP mean, variance, quantiles, median, histogram, all with post-processing to keep outputs valid.
• Hypothesis tests — DP two-sample t-test, chi-square test of independence, Kolmogorov–Smirnov test, one-way ANOVA.
• Regression — closed-form DP linear regression (dp_lm) calibrated to the L2 sensitivity Δ₂ = C·D / λmin(XᵀX), and DP-SGD for GLMs (dp_glm).
• Privacy-aware confidence intervals — analytical intervals combining sampling and privacy noise variance, parametric bootstrap, and a privacy-aware bootstrap that re-privatizes each resample.
• Privacy budget management — an S3 budget tracker supporting basic, advanced, and Rényi (zCDP) composition rules.
• Diagnostics — Monte Carlo coverage validation, utility comparison across ε-grids, composition comparison studies.

Installation

# install.packages("remotes")
remotes::install_github("MukulBijalwan/DPrivStats")

Or from a local clone:

install.packages("path/to/DPrivStats", repos = NULL, type = "source")

Quick start

library(DPrivStats)

data(example_microdata)

# DP mean income under a Laplace mechanism
m <- dp_mean(example_microdata$income, epsilon = 0.5,
             bounds = c(0, 500000))
print(m)

# DP linear regression with privacy-aware confidence intervals
fit <- dp_lm(income ~ education + age + hours, example_microdata,
             epsilon = 2.0, delta = 1e-6,
             bounds = list(y = c(0, 500000)))
dp_confint(fit)

See the vignettes (browseVignettes("DPrivStats")) for full workflows, including official-statistics tabulation under a shared budget.

Author

Mukul Bijalwan — [email protected]

License

MIT © Mukul Bijalwan