Randomization has emerged as a useful technique for data disguising in privacy-preserving data mining. Its privacy properties have been studied in a number of papers. Kargupta et al. challenged the randomization schemes, and they pointed out that randomization might not be able to preserve privacy. However, it is still unclear what factors cause such a security breach, how they affect the privacy preserving property of the randomization, and what kinds of data have higher risk of disclosing their private contents even though they are randomized. We believe that the key factor is the correlations among attributes. We propose two data reconstruction methods that are based on data correlations. One method uses the Principal Component Analysis (PCA) technique, and the other method uses the Bayes Estimate (BE) technique. We have conducted theoretical and experimental analysis on the relationship between data correlations and the amount of private information that can be disclosed based our proposed data reconstructions schemes. Our studies have shown that when the correlations are high, the original data can be reconstructed more accurately, i.e., more private information can be disclosed. To improve privacy, we propose a modified randomization scheme, in which we let the correlation of random noises “similar ” to the original data. Our results have shown that the reconstruction accuracy of both PCA-based and BEbased schemes become worse as the similarity increases.

This paper explores the possibility of using multiplicative random projection matrices for privacy preserving distributed data mining. It specifically considers the problem of computing statistical aggregates like the inner product matrix, correlation coefficient matrix, and Euclidean distance matrix from distributed privacy sensitive data possibly owned by multiple parties. This class of problems is directly related to many other data-mining problems such as clustering, principal component analysis, and classification. This paper makes primary contributions on two different grounds. First, it explores Independent Component Analysis as a possible tool for breaching privacy in deterministic multiplicative perturbation-based models such as random orthogonal transformation and random rotation. Then, it proposes an approximate random projection-based technique to improve the level of privacy protection while still preserving certain statistical characteristics of the data. The paper presents extensive theoretical analysis and experimental results. Experiments demonstrate that the proposed technique is effective and can be successfully used for different types of privacypreserving data mining applications.

We present several methods for performing linear regression on the union of distributed databases that preserve, to varying degrees, confidentiality of those databases. Such methods can be used by federal or state statistical agencies to share information from their individual databases, or to make such information available to others. Secure data integration, which provides the lowest level of protection, actually integrates the databases, but in a manner that no database owner can determine the origin of any records other than its own. Regression, associated diagnostics or any other analysis then can be performed on the integrated data.

www.niss.org

Abstract. We consider scenarios in which two parties, each in possession of a graph, wish to compute some algorithm on their joint graph in a privacy-preserving manner, that is, without leaking any information about their inputs except that revealed by the algorithm’s output. Working in the standard secure multi-party computation paradigm, we present new algorithms for privacy-preserving computation of APSD (all pairs shortest distance) and SSSD (single source shortest distance), as well as two new algorithms for privacy-preserving set union. Our algorithms are significantly more efficient than generic constructions. As in previous work on privacy-preserving data mining, we prove that our algorithms are secure provided the participants are “honest, but curious.”

Abstract. In recent years there has been massive progress in the development of technologies for storing and processing of data. If statistical analysis could be applied to such data when it is distributed between several organisations, there could be huge benefits. Unfortunately, in many cases, for legal or commercial reasons, this is not possible. The idea of using the theory of multi-party computation to analyse efficient algorithms for privacy preserving data-mining was proposed by Pinkas and Lindell. The point is that algorithms developed in this way can be used to overcome the apparent impasse described above: the owners of data can, in effect, pool their data while ensuring that privacy is maintained. Motivated by this, we describe how to securely compute the mean of an attribute value in a database that is shared between two parties. We also demonstrate that existing solutions in the literature that could be used to do this leak information, therefore underlining the importance of applying rigorous theoretical analysis rather than settling for ad hoc techniques. 1

Reluctance of statistical agencies and other data owners to share their possibly confidential or proprietary data with others who own related databases is a serious impediment to conducting mutually beneficial analyses. In this paper, we propose a protocol for securely computing matrix products in vertically partitioned data, i.e., the data sets have the same subjects but disjoint attributes. This protocol allows data owners to estimate coefficients and standard errors of linear regressions, and to examine regression model diagnostics, without disclosing the values of their attributes to each other or to third parties. The protocol can be used to perform other procedures for which sample means and covariances are sufficient statistics. 1

We study the problem of private classification using kernel methods.

This paper considers the problem of learning the parameters of a Bayesian Network, assuming the structure of the network is given, from a privacy-sensitive dataset that is distributed between multiple parties. For a binary-valued dataset, we show that the count information required to estimate the conditional probabilities in a Bayesian network can be obtained as a solution to a set of linear equations involving some inner product between the relevant different feature vectors. We consider a random projection-based method that was proposed elsewhere to securely compute the inner product (with a modified implementation of that method). 1.

Many participatory sensing applications use data collected by participants to construct a public model of a system or phenomenon. For example, a health application might compute a model relating exercise and diet to amount of weight loss. While the ultimately computed model could be public, the individual input and output data traces used to construct it may be private data of participants (e.g., their individual food intake, lifestyle choices, and resulting weight). This paper proposes and experimentally studies a technique that attempts to keep such input and output data traces private, while allowing accurate model construction. This is significantly different from perturbation-based techniques in that no noise is added. The main contribution of the paper is to show a certain data transformation at the client side that helps keeping the client data private while not introducing any additional error to model construction. We particularly focus on linear regression models which are widely used in participatory sensing applications. We use the data set from a map-based participatory sensing service to evaluate our scheme. The service in question is a green navigation service that constructs regression models from participant data to predict the fuel consumption of vehicles on road segments. We evaluate our proposed mechanism by providing empirical evidence that: i) an individual data trace is generally hard to reconstruct with any reasonable accuracy, and ii) the regression model constructed using the transformed traces has a much smaller error than one based on additive data-