The Risk-Utility Tradeoff for Data Privacy Models

Abstract

Nowadays with growth of information technologies, organizations are constantly collecting information about individuals. Public availability of these datasets can considerably benefit the society. To ensure data privacy of a released dataset, various privacy models have been introduced. While many privacy models and techniques have been proposed for data sanitization, the area of sanitized data evaluation has received less attention. This paper investigates the four most well-known data privacy models: $k$-anonymity, $l$-diversity, $t$-closeness, and $\epsilon$-differential privacy. We evaluate the data utility (usefulness of sanitized data) and the disclosure risk (re-identification risk of an individual) of the sanitized data for each model. We use a combination of several data utility and risk metrics to measure the impact of a privacy parameter (e.g., $k$, $\epsilon$) on a particular privacy model. This enables us to compare the risk-utility tradeoff of semantic privacy models such as $\epsilon$-differential privacy to the early syntactic models such as $k$-anonymity on the same scale. We used the Adult dataset from the UCI machine learning repository to conduct our experiments. Experimental results show that $\epsilon$-differential privacy outperforms other privacy models in terms of both data utility and disclosure risk.