Abstract
Over the years, t-closeness has been dealt with in great detail in Privacy Preserving Data Publishing and Mining. Other methods like k-anonymity fail in terms of attribute disclosure and background knowledge attack as demonstrated by many papers in this field. l-diversity also fails in case of skewness attack. t-closeness takes care of all these shortcomings and is the most robust privacy model known till date. However, till now t-closeness was only applied upon a single sensitive attribute. Here, a novel way in determining t and applying t-closeness for multiple sensitive attributes is presented. The only information required beforehand is the partitioning classes of Sensitive Attribute(s). Since, t-closeness is generally applied on anonymized datasets, it is imperative to know the t values beforehand so as to unnecessarily anonymize data beyond requirement. The rationale of using the measure of determining t is discussed with conclusive proof and speedup achieved is also shown.
Introduction
In the field of Privacy Preserving Data Publishing (PPDP), researchers must navigate the delicate balance between data utility and individual privacy. While foundational models like $k$-anonymity and $l$-diversity provide basic protections, they remain vulnerable to sophisticated attacks, such as skewness or background knowledge exploitation. $t$-closeness has emerged as the gold standard for robustness because it ensures that the distribution of a sensitive attribute in any group is "close" to the distribution in the overall dataset.
Despite its strengths, $t$-closeness faces two significant practical hurdles: the "Multiple Attribute" problem, where real-world datasets often contain several sensitive attributes (e.g., both Disease and Salary), and the "Magic Number" problem, where there is no standardized method for selecting the threshold $t$. This paper introduces a systematic approach to pre-calculate $t$ for Multiple Sensitive Attributes (MSA), significantly reducing the computational overhead of trial-and-error anonymization.
Methodology
The proposed method utilizes the concept of "Niche Overlap" to determine the optimal $t$ value mathematically. The workflow involves four primary stages:
- Attribute Partitioning: Sensitive attributes are divided into logical classes to simplify analysis. For example, the "Education" attribute can be grouped into "School" and "College" hierarchies.
- Numerical Coding and Normalization: Categorical data is mapped to numerical values, and continuous data is binned. Attributes are then normalized to a Standard Normal Distribution (SND) to ensure that data skewness does not distort the $t$ calculation.
- Niche Overlap Calculation: The researchers employ the "Niche Overlap" metric ($x$) to measure the intersection area between the distributions of various sensitive attributes.
- The Threshold Formula: The final threshold is determined by the relationship $t = 1 - x$, where $x$ is the minimum overlap area between the attributes.
Experimental Evaluation
The methodology was evaluated using two benchmark datasets: CUPS and ADULT. In the CUPS dataset, the researchers analyzed the overlap between "Disease" (categorical) and "Salary" (numerical) attributes.
For the CUPS dataset, the method identified an optimal $t \approx 0.20$. In the ADULT dataset, a minimum overlap of 79% ($x = 0.79$) was found among attribute groups including hours-per-week, education, and workclass, leading to a calculated $t$ of 0.21. Validation using the ARX-Flash Anonymization Toolbox confirmed that these calculated $t$ values aligned with the points of least information loss across Discernibility, Precision, and Non-Uniform Entropy metrics.
Conclusion
By treating attribute diversity as a geometric overlap problem, this research provides a mathematical bridge to find the "sweet spot" between privacy and utility. Knowing the bound for $t$ before running anonymization algorithms—which are often NP-hard—saves significant execution time. While the method is specifically designed for multiple sensitive attributes and depends on accurate class partitioning, it offers a more efficient path for robust data anonymization. Future work will explore extending this model to more complex attribute structures and refining the $t$ approximation via tolerance values.
Summary updated