New Open-Source Framework trasgoDP Enables Privacy-Preserving Release of Noised Tabular Microdata

The publication of open data stands as a cornerstone of the open science movement, yet the release of sensitive microdata—raw records containing individual information such as hospital patient details or municipal census data—presents significant privacy challenges. Standard anonymization techniques often fail when data is updated, as later records can potentially compromise earlier privacy guarantees. To address this critical gap, researchers have developed trasgoDP, an open-source Python framework designed to release noised tabular microdata while maintaining formal, record-level privacy guarantees through the application of Local Differential Privacy (LDP) and metric-privacy notions.

Differential privacy (DP) offers a rigorous mathematical framework for protecting information during data analysis. While global DP (GDP) typically applies noise when computing statistics from a trusted curator, local DP (LDP) perturbs each individual record *before* it is shared. This record-level protection is vital for microdata publication, but it introduces a fundamental trade-off: LDP mechanisms generally require substantially more noise than global models, which can severely impact data utility. trasgoDP is engineered to navigate this balance, implementing sophisticated mechanisms across numerical, categorical, and location-based attributes.

For numerical attributes, trasgoDP utilizes mechanisms adapted from classical DP theory. The Laplace mechanism, traditionally used for $\epsilon$-global-DP, can be adapted for the local case. For a function $f$ mapping a domain $D$ to $\mathbb{R}^k$, the Laplace mechanism is defined as $ML(x, f(\cdot), \epsilon) = f(x) + (Y_1, \dots, Y_k)$, where $Y_i$ are independent and identically distributed (i.i.d.) random variables drawn from a Laplace distribution centered at zero with a scale of $\Delta_1(f)/\epsilon$. Similarly, for $(\epsilon, \delta)$-global-DP guarantees, the Gaussian mechanism is employed, defined as $MG(x, f(\cdot), \epsilon, \delta) = f(x) + (Y_1, \dots, Y_k)$, where $Y_i$ are i.i.d. random variables from a Gaussian distribution $N(0, \sigma^2)$. However, the core distinction lies in sensitivity. In global DP, sensitivity scales with the range of the data divided by the sample size $n$, meaning noise decreases as $n$ increases. In contrast, LDP sensitivity is fixed at the full range of the data, independent of $n$, leading to inherently higher variance and noisier outputs.

Categorical data requires specialized approaches. trasgoDP implements three distinct strategies: the Exponential mechanism, which outputs a value $r$ with probability $P[r] = \exp(\epsilon \cdot g(D, r) / 2\Delta(g)) / \sum_{r’ \in R} \exp(\epsilon \cdot g(D, r’) / 2\Delta(g))$, where $g$ is a utility function and $\Delta(g)$ is the sensitivity; the Randomized Response (RR) mechanism for binary attributes, which returns the true value $y$ with probability $p(\epsilon) = \frac{e^{\epsilon}}{e^{\epsilon}+1}$ and the opposite value $\neg y$ with probability $1-p(\epsilon)$; and the k-ary Randomized Response for non-binary categorical data. These mechanisms allow users to transform categorical attributes into synthetic or noisy representations while preserving formal $\epsilon$-DP guarantees at the individual record level.

A particularly complex area addressed by trasgoDP is location-based data, which is highly sensitive to re-identification. The framework introduces the concept of metric-privacy, also known as d-privacy, which incorporates geographical distance into privacy solutions. This is formally defined such that for any inputs $y, y’ \in D$ and any output $r
otin R$, the probability of observing $r$ satisfies $P[M(y)=r] \le e^{\epsilon} \cdot d(y, y’) P[M(y’)=r]$, where $d$ is a metric on the domain $D$. This leads to the concept of geo-indistinguishability, which requires that the mechanism provides $\epsilon$-metric-privacy within a specified radius $x$. In practice, trasgoDP achieves this by perturbing true latitude and longitude coordinates using a mechanism that draws a random angle $\theta \sim Unif[0, 2\pi)$ and a random distance $r \sim Gamma(2, 1/\epsilon)$, outputting perturbed coordinates $(lat^*, lon^*)$ based on the Earth’s radius $R_{\oplus}$. This ensures that location points are indistinguishable within a defined proximity.

To quantify the utility loss incurred by these privacy mechanisms, trasgoDP implements several sophisticated metrics. Divergence measures, including Total Variation Distance (TVD), Jensen-Shannon divergence (JS), and Kullback-Leibler divergence (KL), are used to measure information loss between the original and perturbed columns. Furthermore, the framework introduces a novel correlation loss metric, expressed as a percentage. This metric assesses how well the correlations between a feature and a set of other features are preserved after DP application. The process involves calculating the correlation matrix $R$ for the original dataset and $R’$ for the privatized dataset using a chosen method like Pearson, Spearman, or Kendall. The correlation loss is then defined as $L = 100 \cdot \mu_{\Delta} / \mu_R$, where $\mu_{\Delta}$ is the mean of the absolute differences between corresponding correlations ($\|\rho_{ij} – \rho’_{ij}\|$) and $\mu_R$ is the mean of the absolute original correlations. To mitigate numerical instability when $\mu_R$ is close to zero, the function incorporates a reference confidence factor derived from classical statistical results regarding the standard error of correlation coefficients, providing a warning if the original correlation is too weak for reliable comparison.

The framework also provides insights into how the choice of privacy budget ($\\epsilon$) and the size of the dataset ($n$) interact with utility. When testing the mechanisms on the Adult dataset, researchers observed that a smaller $\epsilon$ significantly increases correlation loss, while a larger $\epsilon$ substantially reduces it. For numerical mechanisms, the use of the $\delta$ parameter in the Gaussian mechanism allowed for up to three times more correlation loss compared to the Laplace mechanism for the same $\epsilon$. Crucially, when scaling the analysis to larger sample sizes, such as $n \in \{500, \dots, 30,000\}$, the correlation loss decreased on average for both numerical and categorical variables. However, the authors clarify that this improvement is not due to noise reduction, as LDP sensitivity is constant regardless of $n$. Instead, the improvement stems from the fact that the correlation itself is an aggregate statistic estimated over $n$ rows, and as $n$ increases, this aggregate estimate becomes more stable and less sensitive to the individual noise applied to each record.

Building upon these findings, trasgoDP offers practical heuristics for budget allocation. Instead of splitting the total privacy budget uniformly across all columns, the library recommends allocating a larger share of the budget to attributes whose correlation loss decreases more slowly as $\epsilon$ increases. Practitioners can use the library’s built-in utility metrics, such as correlation_loss and divergence measures, as a guide: starting with an initial budget allocation and iteratively reallocating budget from columns with low loss to those with high loss, while adhering to the constraints of sequential composition. This approach provides a practical, albeit heuristic, method for optimizing the privacy-utility trade-off in real-world microdata release scenarios.

The development of trasgoDP adheres to rigorous open-source software engineering standards. The project is maintained on GitHub, utilizing Conventional Commits for standardized version control and employing Continuous Integration/Continuous Delivery (CI/CD) workflows via GitHub Actions. Unit testing, using the unittest framework, achieves over 97% code coverage for version v2.1.0, ensuring the reliability of the implemented mechanisms. Furthermore, trasgoDP is positioned as the third component of a comprehensive toolkit alongside pyCANON and anjana, forming a complete open-source pipeline that covers anonymization, differential privacy application, and subsequent privacy auditing of tabular datasets. This integrated approach allows researchers to move beyond simply anonymizing data to actively quantifying the statistical information loss associated with a chosen privacy level.

Subject of Research: Development of a local differential privacy framework for releasing noised tabular microdata

Article Title: trasgoDP: An open source framework for releasing noised tabular microdata under local differential privacy

Article References: Sáinz-Pardo Díaz, J., & López García, Á. (2026). trasgoDP: An open source framework for releasing noised tabular microdata under local differential privacy. SoftwareX, 36, Article 103171. https://doi.org/10.1016/j.softx.2026.103171

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103171

Keywords: Local Differential Privacy, Data Privacy, Microdata Release, Differential Privacy, Open Source Software, Metric Privacy, Data Utility, Tabular Data, Machine Learning, Geo-indistinguishability, trasgoDP, open