Traian Marius Truta

Privacy Protection: p-Sensitive k-Anonymity Property

In this paper, we introduce a new privacy protection property called p-sensitive k-anonymity. The existing kanonymity property protects against identity disclosure, but it fails to protect against attribute disclosure. The new introduced privacy model avoids this shortcoming. Two necessary conditions to achieve p-sensitive kanonymity property are presented, and used in developing algorithms to create masked microdata with p-sensitive k-anonymity property using generalization and suppression.

Data and Structural k-Anonymity in Social Networks

The advent of social network sites in the last years seems to be a trend that will likely continue. What naive technology users may not realize is that the information they provide online is stored and may be used for various purposes. Researchers have pointed out for some time the privacy implications of massive data gathering, and effort has been made to protect the data from unauthorized disclosure. However, the data privacy research has mostly targeted traditional data models such as microdata. Recently, social network data has begun to be analyzed from a specific privacy perspective, one that considers, besides the attribute values that characterize the individual entities in the networks, their relationships with other entities. Our main contributions in this paper are a greedy algorithm for anonymizing a social network and a measure that quantifies the information loss in the anonymization process due to edge generalization.

K -anonymization incremental maintenance and optimization techniques

New privacy regulations together with ever increasing data availability and computational power have created a huge interest in data privacy research. One major research direction is built around k-anonymity property, which is required for the released data. Although many k-anonymization algorithms exist for static data, a complete framework to cope with data evolution (a real world scenario) has not been proposed before. In this paper, we introduce algorithms for the maintenance of k-anonymized versions of large evolving datasets. These algorithms incrementally manage insert/delete/update dataset modifications. Our results showed that incremental maintenance is very efficient compared with existing techniques and preserves data quality. The second main contribution of this paper is an optimization algorithm that is able to improve the quality of the solutions attained by either the non-incremental or incremental algorithms.

Generating microdata with p-sensitive k-anonymity property

Existing privacy regulations together with large amounts of available data have created a huge interest in data privacy research. A main research direction is built around the k-anonymity property. Several shortcomings of the k-anonymity model have been fixed by new privacy models such as p-sensitive k-anonymity, l-diversity, (α, k)-anonymity, and t-closeness. In this paper we introduce the Enhanced PK Clustering algorithm for generating p-sensitive k- anonymous microdata based on frequency distribution of sensitive attribute values. The p-sensitive k-anonymity model and its enhancement, extended p- sensitive k-anonymity, are described, their properties are presented, and two diversity measures are introduced. Our experiments have shown that the proposed algorithm improves several cost measures over existing algorithms.

Disclosure risk measures for microdata

We define several disclosure risk measures for microdata. We analyze disclosure risk based on the disclosure control techniques applied to initial microdata. Disclosure Control is the discipline concerned with the modification of data containing confidential information about individual entities, such as persons, households, businesses, etc. in order to prevent third parties working with these data from recognizing entities in the data and thereby disclosing information about these entities. In very broad terms, disclosure risk is the risk that a given form of disclosure will occur if a masked microdataset is released. Microdata represents a series of records, each record containing information on an individual unit. The disclosure risk measures presented in the paper are validated in our experiments.

(p + , α)-sensitive k-anonymity: A new enhanced privacy protection model

Publishing data for analysis from a microdata table containing sensitive attributes, while maintaining individual privacy, is a problem of increasing significance today. The k-anonymity model was proposed for privacy preserving data publication. While focusing on identity disclosure, k-anonymity model fails to protect attribute disclosure to some extent. Many efforts are made to enhance the k-anonymity model recently. In this paper, we propose a new privacy protection model called (p+, alpha)-sensitive k-anonymity, where sensitive attributes are first partitioned into categories by their sensitivity, and then the categories that sensitive attributes belong to are published. Different from previous enhanced k-anonymity models, this model allows us to release a lot more information without compromising privacy. We also provide testing and heuristic generating algorithms. Experimental results show that our introduced model could significantly reduce the privacy breach.

Assessing global disclosure risk in masked microdata

In this paper, we introduce a general framework for microdata and three disclosure risk measures (minimal, maximal and weighted). We classify the attributes from a given microdata in two different ways: based on their potential identification utility and based on the order relation that exists in their domain of value. We define inversion and change factors that allow data users to quantify the magnitude of masking modification incurred for values of a key attribute. The disclosure risk measures are based on these inversion and change factors, and can be computed for any specific disclosure control method, or any combination of methods applied in succession to a given microdata. Using simulated medical data in our experiments, we show that the proposed disclosure risk measures perform as expected in real-life situations.

On-the-fly generalization hierarchies for numerical attributes revisited

Generalization hierarchies are frequently used in computer science, statistics, biology, bioinformatics, and other areas when less specific values are needed for data analysis. Generalization is also one of the most used disclosure control technique for anonymizing data. For numerical attributes, generalization is performed either by using existing predefined generalization hierarchies or a hierarchy-free model. Because hierarchy-free generalization is not suitable for anonymization in all possible scenarios, generalization hierarchies are of particular interest for data anonymization. Traditionally, these hierarchies were created by the data owner with help from the domain experts. But while it is feasible to construct a hierarchy of small size, the effort increases for hierarchies that have many levels. Therefore, new approaches of creating these numerical hierarchies involve their automatic/on-the-fly generation. In this paper we extend an existing method for creating on-the-fly generalization hierarchies, we present several existing information loss measures used to assess the quality of anonymized data, and we run a series of experiments that show that our new method improves over existing methods to automatically generate on-the-fly numerical generalization hierarchies.

A privacy preserving efficient protocol for semantic similarity join using long string attributes

During the similarity join process, one or more sources may not allow sharing the whole data with other sources. In this case, privacy preserved similarity join is required. We showed in our previous work [4] that using long attributes, such as paper abstracts, movie summaries, product descriptions, and user feedbacks, could improve the similarity join accuracy under supervised learning. However, the existing secure protocols for similarity join methods can not be used to join tables using these long attributes. Moreover, the majority of the existing privacy-preserving protocols did not consider the semantic similarities during the similarity join process. In this paper, we introduce a secure efficient protocol to semantically join tables when the join attributes are long attributes. Furthermore, instead of using machine learning methods, which are not always applicable, we use similarity thresholds to decide matched pairs. Results show that our protocol can efficiently join tables using the long attributes by considering the semantic relationships among the long string values. Therefore, it improves the overall secure similarity join performance.

Disclosure risk measures for the sampling disclosure control method

In this paper, we introduce three microdata disclosure risk measures (minimal, maximal and weighted) for sampling disclosure control method. The minimal disclosure risk measure represents the percentage of records that can be correctly identified by an intruder based on prior knowledge of key attribute values. The maximal disclosure risk measure considers the risk associated with probabilistic record linkage for records that are not unique in the masked microdata. The weighted disclosure risk measure allows the data owner to compute the risk of disclosure based on weights associated with different clusters of records. The weights allow a flexible specification of the relative importance of varying cluster sizes in probabilistic record linkage. We show that weighted disclosure risk measure is always between the values of minimal and maximal disclosure risk measures, and moreover for certain values of the weights, the weighted disclosure risk measure is equal to one of the other two measures. Using simulated medical data in our experiments, we show that the proposed disclosure risk measures perform as expected in real-life situations.