rebalancedNoise: Perturbative Statistical Disclosure Control for Magnitude Tables

rebalancedNoise implements the EZS method for magnitude tables as described in the Paper Using Perturbative Methods for Magnitude Tables in Statistical Disclosure Control from Sabolová et al. (2025). It provides a high-performance framework for applying record-level noise with a dynamic rebalancing algorithm that preserves data quality in non-sensitive cells while ensuring additive consistency across complex hierarchies.

Key Features

• Hierarchy Integrity: Ensures that all hierarchical aggregates remain perfectly additive after perturbation.
• Dynamic Rebalancing: Minimizes noise in non-sensitive cells by "balancing" noise directions within hierarchical nodes.
• High Performance: Rebalancing algorithm is implemented in C++ with OpenMP support for fast processing of large inputs.
• Built-in Diagnostics: Powerful $summarize() tools to evaluate rebalancing efficiency and data utility via high-density statistical reports.
• Two Interchangeable Interfaces: A stateful R6 engine for interactive table production and a functional, pipe-friendly API for scripts and reproducible pipelines.
• Seamless Integration: Designed to work directly with data.table and sdcHierarchies.

Installation

# Install from Github
devtools::install_github("sdcTools/rebalancedNoise")

Quick Start

The same workflow, first with the R6 engine, then with the functional interface:

library(rebalancedNoise)
library(data.table)

# Generate dummy data
N <- 100
countries <- c("AT", "DE", "NL", "SE", "FR", "IT")
set.seed(1)
dt <- data.table(
  country = sample(countries, N, replace = TRUE),
  turnover = runif(N, 10, 1000),
  direction = sample(c(1, -1), N, replace = TRUE),
  noise_multiplier = 0.05
)

# Define hierarchy
dims <- list(
  country = sdcHierarchies::hier_create("Total", nodes = countries)
)

# Option 1: R6 engine
sdc <- rn_setup(data = dt, sensitive_params = list(n_threshold = 5))
sdc$rebalance(dim_list = dims, num_var = "turnover")
sdc$perturb(dim_list = dims, variables = "turnover", name = "table_a")
sdc$summarize(table = "table_a")
sdc$get_results("table_a", format = "long") # or "wide"

# Option 2: functional interface
res <- dt |>
  rn_init(sensitive_params = list(n_threshold = 5)) |>
  rn_rebalance(dim_list = dims, num_var = "turnover") |>
  rn_perturb(dim_list = dims, variables = "turnover")

rn_summarize(res)
rn_format(res, format = "long") # or "wide"

Documentation

Two vignettes document the package in both styles:

# Overview of the method and both interfaces, with parallel examples
vignette("getting-started", package = "rebalancedNoise")

# Detailed workflows: multiple tables, linear combinations, caching,
# rounding, evaluation, export and import, parallelization
vignette("workflows", package = "rebalancedNoise")