Values Of Sensitive Attributes Research Articles

Adversarial learning for counterfactual fairness

In recent years, fairness has become an important topic in the machine learning research community. In particular, counterfactual fairness aims at building prediction models which ensure fairness at the most individual level. Rather than globally considering equity over the entire population, the idea is to imagine what any individual would look like with a variation of a given attribute of interest, such as a different gender or race for instance. Existing approaches rely on Variational Auto-encoding of individuals, using Maximum Mean Discrepancy (MMD) penalization to limit the statistical dependence of inferred representations with their corresponding sensitive attributes. This enables the simulation of counterfactual samples used for training the target fair model, the goal being to produce similar outcomes for every alternate version of any individual. In this work, we propose to rely on an adversarial neural learning approach, that enables more powerful inference than with MMD penalties, and is particularly better fitted for the continuous setting, where values of sensitive attributes cannot be exhaustively enumerated. Experiments show significant improvements in term of counterfactual fairness for both the discrete and the continuous settings.

Associated Attribute-Aware Differentially Private Data Publishing via Microaggregation

Releasing raw data sets with sensitive personal information will leak privacy. Therefore, various differential privacy methods have been proposed for efficient data sharing while preserving privacy. However, they focus on noise processing of all quasi-identifier attributes, which results in high time-space complexity and low data utility. In this paper, we propose a Differential Privacy Protection model considering the Correlations between Attributes, denoted DPPCA. DPPCA first computes the degree of correlations between the quasi-identifier attributes and the sensitive attributes, and determines the pair of attributes with maximal degree of correlation. Then based the attributes with the maximal degree of correlations, it uses microaggregation to partition the data set into clusters of size $k$ ( $k\geq 2$ ) according to three types of attributes, i.e., numerical, non-numerical, and hybrid attributes, such that there are $l$ ( $l ) values of sensitive attributes in a cluster. Finally, noise is added to each cluster separately such that each cluster satisfies $\varepsilon $ -differential privacy. While keeping the same degree of preserving privacy, our experimental results demonstrate that DPPCA substantially reduces the amount of added noise to 11% for the Census data set and the Adult data set. Therefore, DPPCA greatly improve the data utility while reaching the same degree of differential privacy.

Towards the design of a secure data outsourcing using fragmentation and secret sharing scheme

ABSTRACTDatabase outsourcing is a new model that offers reliable storage and efficient query execution to clients. Maintaining the confidentiality of data from the service provider is the challenge presented by the growth of database outsourcing. The current approaches for secure database outsourcing are based on encryption. Such approaches impose high computational load on the server. In this paper, we propose a new model based on fragmentation and a secret sharing scheme which partition data among multiple cloud service providers. The privacy requirements are formulated as confidentiality constraints expressing the sensitivity of attributes and their associations. The fragmentation is used to break the sensitive associations among attributes and secret sharing scheme is used to split the values of sensitive attributes among fragments using interpolation. Experimental results show that our proposed model offers better privacy protection from service providers and executes queries efficiently by reducing the computation and communication overhead.

Duplication with Trapdoor Sensitive Attribute Values: A New Approach for Privacy Preserving Data Publishing

Privacy Preserving Data Publishing addresses the problem of publishing the data collected from data owners by the data holder or publisher such that personal sensitive information of the individual is preserved and the published data is highly useful. The anonymization techniques such as Generalization, Suppression, Swapping, Bucketization and Randomization suffers from either individual identity disclosure or the significant loss in information which reduces the usefulness of the data. In this paper, we present a novel Privacy Preserving Data Publishing scheme based on tuple duplication. We introduce the notion of Semantically Equivalent Attribute Values for sensitive attributes and Reputation Loss by Disclosure to hide the sensitive information of an individual in the published data. The Trapdoor Attribute Values for sensitive attributes are defined which helps in recovering the original dataset from the published dataset. We evaluate the proposed scheme with the existing attack models and show that our scheme counters those attacks. We define our own privacy criterion and show that the published data achieves the same. We assess the utility of the published data by using the existing utility metrics and our own defined utility metric. We show that the utility of the published data using the proposed sanitization mechanism is high.

Cascading SOFM and RBF Networks for Categorization and Indexing of Fly Ashes

The objective of this work is to categorize the available fly ashes in different parts of the world into distinct groups based on its compositional attributes. Kohonen's self-organizing feature map and radial basis function networks are applied in a cascading fashion for the classification of fly ashes in terms of its chemical parameters. The basic procedure of the methodology consists of three stages: (1) apply self-organizing neural net to ascertain possible number of groups, delineate them and identify the group sensitive attributes; (2) find mean values of sensitive attributes of the elicited groups and augment them as start-up prototypes in k -means algorithm and find the refined centroids of these groups; (3) incorporate the centroids in a two layer radial basis function network and fine-tune the delineated groups and develop an indexing equation using the weights of the stabilized network. Further, to demonstrate the utility of this classification scheme, the so formed groups were correlated with their performance in High Volume Fly Ash Concrete System [HVFAC]. The categorization was found to be excellent and compares well with Canadian Standard Association's [CSA A 3000] classification scheme.

Extending ℓ-diversity to generalize sensitive data

Generalization is an important technique for protecting privacy in data dissemination. In the framework of generalization, ℓ-diversity is a strong notion of privacy. However, since existing ℓ-diversity measures are defined in terms of the most specific (rather than general) sensitive attribute (SA) values, algorithms based on these measures can have narrow eligible ranges for data that has a heavily skewed distribution of SA values and produce anonymous data that has a low utility. In this paper, we propose a new ℓ-diversity measure called the functional ( τ, ℓ)-diversity, which extends ℓ-diversity by using a simple function to constrain frequencies of base SA values that are induced by general SA values. As a result, algorithms based on ( τ, ℓ)-diversity may generalize SA values, thus are much less constrained by skew SA distributions. We show that ( τ, ℓ)-diversity is more flexible and elaborate than existing ℓ-diversity measures. We present an efficient heuristic algorithm that uses a novel order of quasi-identifier (QI) values to achieve ( τ, ℓ)-diversity. We compare our algorithm with two state-of-the-art algorithms that are based on existing ℓ-diversity measures. Our preliminary experimental results indicate that our algorithm not only provides a stronger privacy protection but also results in better utility of anonymous data.

SABRE: a Sensitive Attribute Bucketization and REdistribution framework for t-closeness

Today, the publication of microdata poses a privacy threat: anonymous personal records can be re-identified using third data sources. Past research has tried to develop a concept of privacy guarantee that an anonymized data set should satisfy before publication, culminating in the notion of t-closeness. To satisfy t-closeness, the records in a data set need to be grouped into Equivalence Classes (ECs), such that each EC contains records of indistinguishable quasi-identifier values, and its local distribution of sensitive attribute (SA) values conforms to the global table distribution of SA values. However, despite this progress, previous research has not offered an anonymization algorithm tailored for t-closeness. In this paper, we cover this gap with SABRE, a SA Bucketization and REdistribution framework for t-closeness. SABRE first greedily partitions a table into buckets of similar SA values and then redistributes the tuples of each bucket into dynamically determined ECs. This approach is facilitated by a property of the Earth Mover's Distance (EMD) that we employ as a measure of distribution closeness: If the tuples in an EC are picked proportionally to the sizes of the buckets they hail from, then the EMD of that EC is tightly upper-bounded using localized upper bounds derived for each bucket. We prove that if the t-closeness constraint is properly obeyed during partitioning, then it is obeyed by the derived ECs too. We develop two instantiations of SABRE and extend it to a streaming environment. Our extensive experimental evaluation demonstrates that SABRE achieves information quality superior to schemes that merely applied algorithms tailored for other models to t-closeness, and can be much faster as well.

Extended k-anonymity models against sensitive attribute disclosure

p-Sensitive k-anonymity model has been recently defined as a sophistication of k-anonymity. This new property requires that there be at least p distinct values for each sensitive attribute within the records sharing a set of quasi-identifier attributes. In this paper, we identify the situations when the p-sensitive k-anonymity property is not enough for the sensitive attributes protection. To overcome the shortcoming of the p-sensitive k-anonymity principle, we propose two new enhanced privacy requirements, namely p + -sensitive k-anonymity and ( p , α ) -sensitive k-anonymity properties. These two new introduced models target at different perspectives. Instead of focusing on the specific values of sensitive attributes, p + -sensitive k-anonymity model concerns more about the categories that the values belong to. Although ( p , α ) -sensitive k-anonymity model still put the point on the specific values, it includes an ordinal metric system to measure how much the specific sensitive attribute values contribute to each QI-group. We make a thorough theoretical analysis of hardness in computing the data set that satisfies either p + -sensitive k-anonymity or ( p , α ) -sensitive k-anonymity. We devise a set of algorithms using the idea of top-down specification, which is clearly illustrated in the paper. We implement our algorithms on two real-world data sets and show in the comprehensive experimental evaluations that the two new introduced models are superior to the previous method in terms of effectiveness and efficiency.

L -diversity

Publishing data about individuals without revealing sensitive information about them is an important problem. In recent years, a new definition of privacy called k -anonymity has gained popularity. In a k -anonymized dataset, each record is indistinguishable from at least k − 1 other records with respect to certain identifying attributes. In this article, we show using two simple attacks that a k -anonymized dataset has some subtle but severe privacy problems. First, an attacker can discover the values of sensitive attributes when there is little diversity in those sensitive attributes. This is a known problem. Second, attackers often have background knowledge, and we show that k -anonymity does not guarantee privacy against attackers using background knowledge. We give a detailed analysis of these two attacks, and we propose a novel and powerful privacy criterion called ℓ-diversity that can defend against such attacks. In addition to building a formal foundation for ℓ-diversity, we show in an experimental evaluation that ℓ-diversity is practical and can be implemented efficiently.

Cardinality-based inference control in data cubes*

This paper addresses the inference problem in on-line analytical processing (OLAP) systems. The inference problem occurs when the exact values of sensitive attributes can be determined through answers to OLAP queries. Most existing inference control methods are computationally expensive for OLAP systems, because they ignore the special structures of OLAP queries. By exploiting such structures, we derive cardinality-based sufficient conditions for safe OLAP data cubes, Specifically, data cubes are safe, from inferences if their core cuboids are dense enough, in the sense that the number of known values is under a tight bound. We then apply the sufficient conditions on the basis of a three-tier inference control model. The model introduces an aggregation tier between data and queries. The aggregation tier represents a collection of safe data cubes that are pre-computed over a partition of the data using the proposed sufficient conditions. The aggregation tier is then used to provide users with inference-free queries. Our approach mitigates the performance penalty of inference control, because partitioning the data yields smaller input to inference control algorithms, pre-computing the aggregation tier reduces on-line delay, and using cardinality-based conditions guarantees linear-time complexity.

Protection of health information in data mining

The paper studies issues related to privacy protection of health information in data mining. We introduce the implications of the Health Insurance Portability and Accountability Act (HIPAA) and the privacy of protected health information (PHI). We present an attribute analysis framework of health information and a technological approach for protecting PHI in data mining. Specifically, we develop such effective privacy-enhancing techniques as data filter, discretisation and randomisation and give an example of inducing the decision-trees from training data in which the values of sensitive attributes have been either removed or modified by using these techniques. The results show that we can achieve comparative predictive accuracies without accessing the original values of the sensitive attributes.