Data Mining Based On Perturbation Technique

Published Date: 02 Nov 2017

Abstract

To the success of today’s data mining techniques preserving personal and sensitive information is critical. Privacy Preserving Data Mining (PPDM) directs such issues by balancing both the preservation of privacy and the utilization of data.

Traditionally, the widely used methods for preserving clustering are Geometrical Data Transformation Methods (GDTMs). The drawback of these methods is that geometric transformation functions are invertible, that results in a lower level of privacy protection. In this paper, the technique that preserves the privacy of delicate information in a multiparty clustering scenario called the Principal Component Analysis based technique is proposed. The performance of the technique is further judged by using a classical K-means clustering algorithm and machine learning-based clustering method on synthetic and real world datasets. The PCA based transformation method surpasses the traditional GDTMs by providing superior privacy protection and better performance.

Introduction

Ameliorations in data compression, the advent of low storage costs and the increasing usage of the World Wide Web have played a great role in acquire and storing huge volumes of data, a relevant goal for these data is their use to gain insight into invisible patterns by the use of data mining tools and techniques. With these advancements, there are amplifying concerns about the privacy of personal and confidential information. People don’t take interest in sharing their personal data, which can result in skewing the result of the data mining because the data gathered may then have wrong or partial information. An overview of the risks that are linked with data mining, along with the strategies that have been proposed over the years to lighten these risks, is given in. Most of the traditional PPDM algorithms preserve the privacy of data by changing the real data in a way that the applicability of data is unchanged. The ability to analyze private data without breaching the confidentiality of the individuals has afforded to the reputation of PPDM.

Defending the delicate attribute values of objects that are exposed to clustering analysis is the concept behind privacy preserving clustering (PPC). Some investigations recommend that synthetic data production using the ISPO family of methods is a mode to sustain data privacy in clustering and discovered different sorts of situations in Privacy Pre-serving Clustering. The scenarios are

1. PPC over horizontally partitioned data

Suppose that Alice A and Bob B have two samples that are unlabeled, DA and DB. Suppose that both sample in DA and DB has all of the characteristics or that the data sets are horizontally partitioned amongst A and B. Alice and Bob desire to cluster the joint data set DA U DB without disclosing the separate items in their data sets.

2. PPC over centralized data

The scenario involves two parties A and B, party A having a dataset D and party B requiring to mine it for clustering. The dataset is supposed to be a data matrix Dmxn, where each of the m rows denotes a unit or object, and each unit has values for each of then attributes. The matrix Dmxn might include binary, categorical, or numerical attributes. Before distributing the dataset D with party B, party A has to convert D to retain the privacy of the independent data records. Though, the transformation that is applied to D must not endanger the resemblance among the objects.

3. PPC over vertically partitioned data

Consider a situation where k parties, k>= 2, have dissimilar attributes for a common set of objects. The target here is to accomplish a join over the k parties and then to cluster the common objects. After implementing a join over the k parties, the challenge of PPC over vertically partitioned data turn out to a problem of PPC of centralized data.

Literature Review

The prototype for clustering horizontally partitioned or centralized data sets using a simple PCA based transformation approach addresses the problem of handling centralized data and horizontally partitioned data with the accuracy preserved but the model falls behind in addressing vertically partitioned data [1]. Aims to associate the same property of IPSO (The Information Preserving Statistical Obfuscation) to the FCRM (fuzzy c-regression models) synthetic data generator, also at the identical period of time, to judge the association between the information losses produced when generating synthetic data with FCRM and the clustering similarity between the original and synthetic data. The generation of synthetic data by FCRM is as appropriate as it is when generating synthetic data with the IPSO family of methods and this protected data is planned to be studied or modeled with either crisp or fuzzy clustering algorithms and the results show that the FRI clustering similarity measure is inversely proportional to the probabilistic information loss incurred when replacing the original data set with the synthetic data generated by FCRM [2].

Problem definition

The data outburst is a result of data processing techniques and growth of the internet. The data that is available may contain sensitive information that can endanger privacy of individuals if misused.

Solution Methodology

A technique namely Principal Component Analysis based technique is proposed that upholds the confidentiality of sensitive data in a multiparty clustering scenario. The functioning of the technique is additionally assessed by means of a classical K-means clustering algorithm and machine learning-based clustering method on artificial and factual world datasets. Various approaches like Clustering, Multiparty clustering scenario and Data transformation are of greater significance.

Data Transformation

Several algorithms for privacy preserving transformations are present in the literature. Liu et al. have suggested a random projection-based transformation technique for data protection in distributed mining scenario. Principal Component Analysis [PCA] is availed for significantly minimizing the dimension of the data. PCA developed by Karl Pearson is a statistical technique that makes use of orthogonal transformations for replenishing the initial variables of a set of data with a lower number of irrelevant variables usually called the principal components. Its main intensity resides in its proficiency to distinguish data patterns with greater dimension. PCA can then sort out this data in a style that causes minimal information loss. Thus, it efficiently lowers the difficulty and the dimension of the data in a way that the major sections have the utmost possible variance. PCA offers a linear estimation that most of the deviation in original data is shown. The existence of only scarce modules makes it simpler to label every dimension with an perceptive sense.

In this practice, the procedure of making a PCA-based transformation matrix for privacy preservation begins with choosing a sample data from the real multi-attribute set of data in a way that of all of the classes of data from the original set of data such that the samples are descriptive. Then, a transformation matrix is formulated from the sample by means of PCA. ‘Shifting factor’ a security improvement which could be a modest scalar value, is used. The transformation matrix might be transferred by multiplying it with a randomly chosen shifting factor, if required. The act will additionally surge the security in contrary to any opposite methods that can be used to estimate the original data. Now, the original records are changed by treating them with the transformation matrix. The dimension of the altered data will be all times lower than that of the original number of dimensions.

Clustering

A technique that consigns data objects into similar clusters in a way that the objects that are in every cluster show like qualities. Cluster analysis handles the challenge of ordering vector sets into many clusters. Clustering has an extensive sort of purposes in multiple fields, such as finance, medicine marketing, bioinformatics and insurance. Clustering can aid researchers in developing theories by attaining important information from clustered files. In recent time, Lee and Olafsson suggested a latest measure of ‘‘cluster quality’’ which is on basis of decreasing the consequence of interruption among objects that would perfectly be clustered together.

The performance of a transformation based on PCA is assessed by two clustering techniques, namely K-means and Self Organizing Map (SOM) based clustering. Clustering was impersonated on the original and change archives and judgments were made, to calculate the precision of the clustering. K-means is one of the easiest and highly famous clustering algorithms that use learning without any supervision. To produce k clusters, it begins with describing k centroids. Each object is allotted to a cluster based on its nearness to the centroid. Then, the centroid is calculated again and objects are allotted to clusters taking new centroid values as basis. This procedure is iterated till new centroids cannot be located any longer or no extra alterations can be made. The purpose of this algorithm lies in reducing the criteria of sum of squares.

Steps in evaluating the framework

An interpretation of the steps concerned in fulfilling and estimating the proposed methodology followed.

1. Prepare a multi-attribute synthetic data set S with dimension d using the Gaussian distribution function. Let N=|S| and C= {C1, C2,.. .,Cc} be the known classes of S.

2. Choose sample s is a subset of S such that S∩Ci ≠ null for all i=1, 2,..., c.

3. Prepare a transformation Matrix Ts corresponding to s using a PCA-based transformation such that dim (Ts) =d1<d.

4. Form a shifted transformation matrix Tsr from Ts by using a shift factor ‘r’ (if necessary).

5. Project S on Ts or Tsr to obtain the new reduced dimensional data set Ss.

6. Cluster the original data set S by using K-means and SOM algorithms and let the new clusters correspondingly be C1 and C2.

7. Cluster the transformed data set Ss by using K-means and SOM algorithms and let the new clusters be Cs1 and Cs2, respectively.

8. Obtain the Rand Index (RI) for the pairs (C,C1), (C,C2), (C,Cs1), and (C,Cs2),and estimate the accuracy of the clustering.

9. Repeat the evaluation from step 1 by varying the values of the parameters d and N.

Multi-Party Clustering Scenario

The fundamental sense of secure multi-party computation is that a computation is protected only if, at the termination of the computation, no party understands anything excluding its self-input and the outputs. Privacy preserving data mining deals the necessity of several parties with personal inputs to execute a data mining algorithm and to study the results over the united data without disclosing any pointless information.

In a privacy preserving multi-party clustering scenario, the target is to permit a group of parties to accomplish a combined cluster analysis deprived of disclosing the values of their delicate attributes. The proposed methodology can be used to preserve the privacy of horizontally partitioned or centralized data in such a multi-party clustering scenario.

Fig.1 illustrates the application of the proposed methodology in a multi-party clustering scenario in which two parties, A and B, are included. Party A forms a PCA-based transformation matrix. For a greater level of security, B changes this transformation matrix by a shifting factor that is unknown to A. Now, the parties A as well as B show their data using this shifted transformation matrix. These changed sets of data are mixed to achieve collaborative clustering without disclosing the real values of attributes of both parties A and B.

Fig.1. Applying the proposed transformation method for a multi-party clustering scenario.

Results

Comparison with GDTM’s

The recommended method is contrasted with GDTMs. Stanley R.M. Oliveira and Zaiane presented a family of GDTMs to deal privacy preservation in clustering analysis. Conventionally, the confidentiality is offered by a perturbation technique, which is calculated as the variance between the original and the perturbed values. This metric, used to calculate the efficiency, is given by Variance(X-Y), where X denotes a sole real attribute and Y is the partial attribute. This evaluation can be & scale-invariant relating to the variance of by representing the security as Variance(X-Y)/Variance(X).

It is likely that GDTMs might be preventive in regard to the privacy based on the results. Apart from the challenge of minimal confidentiality, a geometric alteration function is reversible so that one can be able to guess the original values of the data under clustering. Nevertheless, the advised PCA-based methods give an improved secrecy level matched with all of the conventional GDTMs.

Conclusions and Future work

The PCA-based transformation model was lucratively applied for privacy-preserving clustering of centralized and horizontally partitioned data. Certain results from previous studies have revealed that precision at times suffers as a result of safety. Though, the proposed technique retained precision and, in certain situations, the accuracy was nearly identical to that of the real set of data. The PCA-based transformation technique preserved the important features of the data even in its altered state. The suggested technique can be used to cover delicate data while showing it on a widely handy platform such as the web. Additionally, the demand is that the suggested technique can be used in a multi-party clustering scenario to accomplish collaborative clustering without disclosing the Real attribute values of the two parties.

Upcoming study may deal with other cases, like vertically partitioned data. Privacy preservation in data mining is a legitimately new subject of research, and it will be fascinating to perceive where such study will take us in the future time.