Package {DPSynth}

ACS PUMS-like sample

Description

A synthetic microdata sample modeled on American Community Survey PUMS structure, used as an example in DPSynth.

Usage

acs_pums_sample

Format

A data frame with 500 rows and 5 variables:

age

age in years

income

annual income in USD

educ

educational attainment

married

married indicator

weeks_worked

weeks worked last year

Anonymized UCI Adult subsample

Description

A synthetic, anonymized subsample modeled on the UCI Adult census-income dataset, used as an example throughout DPSynth. No real personal records are included.

Usage

adult_sample

Format

A data frame with 500 rows and 8 variables:

age

age in years

workclass

employment sector

education_num

years of education

marital_status

marital status

hours_per_week

working hours per week

sex

sex

capital_gain

capital gain in USD

income

binary income class

Attribute disclosure risk audit

Description

Assumes an attacker knows known_vars for a target record and uses nearest neighbours in the synthetic data to infer sensitive_var. Reports the normalized prediction error and the rate of high-confidence disclosures.

Usage

audit_attribute_disclosure(
  synth,
  orig,
  known_vars = NULL,
  sensitive_var = NULL,
  n_neighbors = 5
)

Arguments

Value

list with mean_disclosure_error and high_risk_rate

Record linkage risk audit

Description

Estimates how many real records can be uniquely re-identified by matching against the synthetic data: a real record is "linked" when its nearest synthetic neighbour is closer than a quantile threshold of the within-original distance distribution.

Usage

audit_linkage_risk(synth, orig, threshold_quantile = 0.05)

Arguments

Value

list with linkage_rate, linked_records, threshold

Membership inference risk audit

Description

Simulates the simplest membership inference attack: for each original record, compute the distance to its nearest synthetic record. Records whose nearest synthetic neighbour is unusually close are considered at risk of a correct "member" verdict.

Usage

audit_membership_risk(synth, orig, method = "nearest_neighbor")

Arguments

Value

list with risk_score, mean_min_distance and high_risk_records

Check remaining privacy budget

Description

Check remaining privacy budget

Usage

check_synth_budget(budget)

Arguments

Value

named vector with remaining_epsilon, spent_epsilon and fraction_used; also prints a summary

Clip data to bounds

Description

Clamp every column of a data frame (or matrix) to user-specified bounds. Clamping bounds sensitivity and is a prerequisite for calibrated DP noise.

Usage

clip_data(data, bounds)

Arguments

Value

data frame with values clipped to the bounds

DP Gaussian copula synthesis

Description

Separates the synthesis problem into (1) DP marginal distributions – per-column histograms released with Laplace noise – and (2) a DP dependence structure – a Gaussian copula correlation matrix estimated on rank-transformed pseudo-observations, perturbed with Gaussian noise and projected onto the positive-definite cone. Generation inverts the DP marginals through the DP copula (post-processing).

Usage

dp_copula_synth(
  data,
  n_synth = nrow(data),
  epsilon = 1,
  delta = 1e-06,
  n_bins = 20,
  copula_type = "gaussian"
)

Arguments

Value

object of class dp_copula / dp_synthetic

DP Gaussian Mixture Model synthesis

Description

Fits a Gaussian mixture model with diagonal covariances using the perturb-and-postprocess paradigm: sufficient statistics (component counts, means, variances) are perturbed with calibrated Laplace / Gaussian noise so that the released model satisfies (epsilon, delta)-DP. Sampling from the fitted model is post-processing and consumes no additional privacy budget.

Usage

dp_gmm_synth(
  data,
  n_synth = nrow(data),
  K = 3,
  epsilon = 1,
  delta = 1e-06,
  bounds = NULL,
  max_iter = 20
)

Arguments

Value

object of class dp_gmm / dp_synthetic with the synthetic data, model parameters and privacy parameters

DP k-means initialization via the exponential mechanism

Description

Selects K cluster centers from candidate rows using the exponential mechanism with a k-means++ style D2 quality score, so the initialization itself satisfies epsilon-DP.

Usage

dp_kmeans_init(data, K, epsilon)

Arguments

Value

numeric matrix with K rows (the selected centers)

DP marginal synthesis

Description

Synthesizes each column independently from a differentially private histogram (numeric columns) or DP multinomial (categorical columns). No dependence structure is preserved; useful as a fast baseline and for consistent-marginal releases in official statistics.

Usage

dp_marginals_synth(data, n_synth = nrow(data), epsilon = 1, n_bins = 20)

Arguments

Value

object of class dp_marginals / dp_synthetic

PATE synthetic data for discrete / mixed-type data

Description

Partitions the private data into disjoint teacher datasets, fits a teacher model on each partition (no privacy cost), and releases each per-record query through a noisy aggregation of teacher votes. The privacy cost scales with the number of released records.

Usage

dp_pate_synth(data, n_synth = nrow(data), epsilon = 1, n_teachers = 10)

Arguments

Value

object of class dp_pate / dp_synthetic

Differentially private synthetic data generation

Description

Top-level orchestrator that dispatches to a DP synthesizer, optionally evaluates utility and audits disclosure risk, and attaches a privacy report.

Usage

dp_synthesize(
  data,
  method = "copula",
  n_synth = nrow(data),
  epsilon = 1,
  delta = 1e-06,
  utility_eval = TRUE,
  risk_audit = TRUE,
  ...
)

Arguments

Value

object of class dp_synthesis_result

See Also

dp_copula_synth, dp_gmm_synth, dp_marginals_synth, dp_pate_synth

Examples

set.seed(1)
dat <- data.frame(x = rnorm(200), y = rnorm(200) + 0.5 * rnorm(200))
res <- dp_synthesize(dat, method = "copula", epsilon = 2)
print(res)

Downstream task performance (TSTR)

Description

Trains the same linear model on the synthetic and the original data and compares out-of-sample RMSE on held-out real data. A utility_ratio close to 1 indicates good analytic utility.

Usage

evaluate_downstream(
  synth,
  orig,
  target_var,
  predictors = NULL,
  test_prop = 0.2
)

Arguments

Value

list with RMSEs and the utility ratio

Multivariate fidelity evaluation

Description

Compares the correlation structure (numeric columns) of original and synthetic data via the relative Frobenius-norm distance between correlation matrices and Gaussian-approximated mutual information.

Usage

evaluate_multivariate(synth, orig)

Arguments

Value

list with correlation distance and mutual-information summaries

Propensity score utility (pMSE)

Description

Fits a logistic model to distinguish synthetic from original records. Under perfect synthesis the expected pMSE is c(1-c) with c = n_synth / (n_orig + n_synth); the reported ratio is 0 for indistinguishable data and grows as fidelity degrades.

Usage

evaluate_propensity(synth, orig)

Arguments

Value

list with pMSE, ratio (standardised) and c

Univariate fidelity evaluation

Description

Compares original and synthetic columns with Kolmogorov-Smirnov statistics, Hellinger distance (histogram-based for numeric columns, Jensen-Shannon divergence for categorical columns) and relative moment errors.

Usage

evaluate_univariate(synth, orig)

Arguments

Value

data frame of per-variable fidelity metrics

Standardized utility evaluation

Description

Runs any combination of univariate, multivariate, propensity (pMSE) and downstream (TSTR) fidelity metrics and returns a dp_utility_report.

Usage

evaluate_utility(
  synthetic_data,
  original_data,
  metrics = c("univariate", "multivariate", "propensity"),
  target_var = NULL
)

Arguments

Value

object of class dp_utility_report

Gaussian mechanism

Description

Release a statistic under (epsilon, delta)-DP by adding Gaussian noise calibrated to the L2 sensitivity via the classical analytic bound \sigma = \Delta \sqrt{2 \log(1.25 / \delta)} / \epsilon.

Usage

gaussian_mech(value, sensitivity, epsilon, delta = 1e-06)

Arguments

Value

numeric vector: noisy release of value

Create a privacy budget tracker

Description

Create a privacy budget tracker

Usage

new_synth_budget(
  epsilon,
  delta = 0,
  accounting = c("basic", "advanced", "rdp")
)

Arguments

Value

object of class synth_privacy_budget

Print a synthesis result

Description

Print a synthesis result

Usage

## S3 method for class 'dp_synthesis_result'
print(x, ...)

Arguments

Value

No return value, called for side effects. The input x is returned invisibly.

Laplace noise

Description

Draw n independent Laplace(location, scale) variates using inverse-CDF sampling. The Laplace mechanism adds Lap(sensitivity / epsilon) noise to achieve pure epsilon-DP.

Usage

rlaplace(n, location = 0, scale = 1)

Arguments

Value

numeric vector of length n

Spend privacy budget on a step

Description

Records a DP release consuming epsilon (and optionally delta) and updates the accumulated spend under the configured composition rule. Issues a warning when the total budget is exceeded.

Usage

spend_synth(budget, epsilon, delta = 0, description = "")

Arguments

Value

updated budget object