INTRODUCTION

With the development of technology, ultra large databases appear. At the same time, people show much concerns about the informational privacy. The problem of privacy-preserving is first proposed by Agrawal and Srikant (2000). Data mining is an efficient tool to discover valuable knowledge from a great deal data. But the general data mining is based on the assumption that complete access to data is available, either in centralized or federated form. In fact, privacy and security concerns often prevent sharing of data which may not be possible due to either legal or commercial reasons. In legal terms, medical data cannot be released for any purpose without appropriate anonymization. In commercial terms, data is often a valuable business asset. So, the research direction in data mining incorporating privacy concerns becomes fruitful. People also show much interest in the problem of privacy-preserving data mining. The environment of data mining can be classified into two cases: centralized and distributed data. In the centralized environment, the focus is on the query restriction. In the distributed environment, the data is distributed in different sites. Now, in the distributed environment, the data is classified into horizontally partitioned and vertically partitioned case. For horizontally partitioned data (Yu et al., 2006a; Kantarcioglu and Clifton, 2004), the data matrix A`s rows including all input features are divided into groups belonging to different entities. While for vertically partitioned data, the data matrix A`s columns are divided into groups belonging to different entities (Yu et al., 2006b; Vaidya and Clifton, 2002). The reason is that feature values for each individual are stored as rows of a data matrix, while a specific feature values for all individuals are represented by columns of a data matrix.

Data mining has many applications in the real world. One of the most important and widely found problems is that of classification. The goal of classification is to build a model that can predict the value of one variable, based on the other variables. Support Vector Machine is one of the most important classification methods in machine learning. While SVM is one of the most actively developed classification methodology, so there has been wide interest in privacy-preserving support vector machine (PPSVM) classifier. Recently, people show much interest in the area of privacy-preserving support vector machines, PPSVM for short. We briefly review some of the relevant work. PPSVM classifiers were gained on vertically partitioned data by adding random perturbations to the data by Yu et al. (2006b). PPSVM classifiers using nonlinear kernels on horizontally partitioned data were obtained by Yu et al. (2006a). But it can only deal with binary feature data. Other privacy preserving classifying techniques include wavelet-based distortion and rotation perturbation (Chen and Liu, 2005).

The kernel matrix computed in the above study is equal or close to that computed with the original data. They refer to compute inner product securely (Ioannidis et al., 2002). In this study, we propose another approach to the classification and testing. The kernel matrix computed in our method is not equal to that computed with original data. But the accuracy is comparable with the original data. We propose a unified high efficient novel model of PPSVM classifiers on horizontally and vertically partitioned data that is different from the existing PPSVM classifiers for such partitioned data. During the classification and testing, we use the perturbed data. The study is based on the two ideas. The first idea is that for a given data matrix A, we add the same random real number to the same column of the matrix A. Though the rows of the matrix A are modified, the relative location of each row of matrix A is not changed. We prove that this model has comparable accuracy with the original data. Besides, this experiment gives the evidence. The second idea is that each entity generates a random row vector that has the same dimension with that of its input feature space. The entity holds it privately. In the training and testing process, it does not change. The data for training and testing is perturbed with the same data. By employing the two ideas, we shall describe algorithms that protect the privacy of each partitioned data either horizontally partitioned or vertically partitioned. The generated PPSVM classifier has comparable accuracy comparing to that of an ordinary SVM classifier.

We first describe the notation. An mxn matrix A represents m data points in an n-dimensional input space. An mxn diagonal matrix D contains the corresponding labels (i.e., +1 or -1) of the data points in A. (A class label D_ii, or d_i for short, corresponds to the i-th data points x_i in A). All vectors are column vectors unless transposed to a row vector by a prime superscript^T. The scalar (inner) product of two vectors x and y in the n-dimensional space Rⁿ is denoted by x^Ty.

We assume that we can public the class labels D corresponding to data matrix A. Each entity does not collude and does follow the proposed protocol correctly.

SVM OVERVIEW

The Support Vector Machine (SVM) is a classifier, originally proposed by Vapnik (2000), that finds a maximal margin separating hyperplane between two classes of data. The aim of Support Vector Classification (SVC) is to devise a computationally efficient way of learning good separating hyperplane in a high dimensional feature space. SVM finds the separating hyperplane ((ω.x)+b = 0) that maximizes the margin, denoting the distance between the hyperplane and the closest data points (i.e., support vectors). When we can not find a hyperplane to separate the data perfectly, we introduce the soft margin. To maximize the margin while minimizing the error, the standard SVM solution is formulated into the following primal program:

(1)

ξ_i is the slack variable in the constraint. This shows that SVM allows error or the soft margin. The slack or error is minimized in the objective function. C is the margin parameter which is used to tune the margin size and the error. The weight vector ω and the bias b will be computed by this optimization problem. Then we can determine the class of a new data object x by f(x) = (ω.x)+b, where the class is positive if f(x)>0, or else negative.

To solve the primal problem, we can solve its dual problem by applying the Lagrange multipliers:

(2)

The coefficients α are to be computed from the dual problem. Where

is the kernel function where for linear kernel. We can also apply a nonlinear kernel for K(x_i, x_j) (e.g.,

for RBF kernel, K(x_i · x_j) = ((x_i · x_j))+1^p for polynomial kernel.) The weight vector and

Then the classification function can be computed.

For more information, see the tutorial written by Burges (1998).

A UIFIED MODEL FOR PPSVM

If we can access the data points xi, i = 1,2,.., l freely, we can get the original SVM model:

(3)

Solving the problem, we can use the method earlier. While privacy and security prevent access the data points ξ_i, i = 1,2,.., l freely. One technique for privacy and security is perturbation (Agrawal and Ramakrishnan, 2000), that is perturbs the data points with some data. In this research, we give the same bias to the same column, that is each data point has the same bias. In the privacy-preserving support vector machines, privacy and security prevent to access the original data.

In present model, the original data points x_i εRⁿ, i = 1,2,.., l become x_i+a, where aεRⁿ. We can access the data points x_i+a, i = 1,2,.., l. The purpose of introducing vector a is to realize the requirement of privacy and security. The PPSVM model is as following.

(4)

We give the proof that using the PPSVM model has comparable accuracy comparing with that of the original SVM model.

Lemma 1: In the original SVM model (3) and the PPSVM model (4), the distance of the data points x_i, x_j is the same with that of the corresponding perturbed data points x_i+a,x_j+a.

Lemma 2: The original SVM model Eq. 3 is equivalent to:

(5)

Similarly, the PPSVM model Eq. 4 is equivalent to:

(6)

Lemma 3: Given the same parameter C, solving the original SVM model (5), we get the decision function f(x) = (ω·x)+b and solving the PPSVM model we get the decision function then we can get for linear kernel. For the original SVM model (5), the Lagrange multipliers are α_i, i = 1,2,..., l. For the PPSVM model (6), the Lagrange multipliers are If the slack variable ξ_i is same in the original model (5) and PPSVM model (6) (In Lemma 4, we proof that the assumption is reasonable), then for both linear and nonlinear kernels.

Proof: For the original SVM model (5), the Lagrange function is:

(7)

where α_i are Lagrange multipliers. Its KKT conditions are:

(8)

(9)

(10)

(11)

(12)

(13)

We replace original data points x_i, i = 1,2,...,l with x_i + α_i, i = 1,2,...,l and replace ω, x_i,α_i,b with in the Eq. 8-13. We get the KKT conditions of the PPSVM model (6).

From Eq. 8-9, we get

For linear kernel:

So, ω does not change for linear kernel.

If the slack variable ξ_i is same in the original model (5) and PPSVM model (6), the parameter C is the same in (5) and (6), from the Eq. 10, we can know that the Lagrange multipliers do not change from the Eq. 10. That is:

(14)

From Eq. 14, we know that if x_i is support vector for original SVM model, then x_i+a is support vector for PPSVM model.

Lemma 4: If

(ω,b,ξ) is the optimal solution of the original SVM model Eq. 5, then there is such that is the feasible solution of the PPSVM model Eq. 6.

Proof: Because so there is subjected to So:

(ω,b,ξ) is the optimal solution of the original SVM model Eq. 5, so Then So is the feasible solution of the PPSVM model Eq. 6

Theorem 1: For linear kernel, given the same parameter C, the decision function f (x) = (ω·x)+b got from the original SVM model has the same function value with the decision function

Proof: From Lemma 3, we know

Then for any i, α_i>0:

(15)

(16)

So we get Theorem 1.

Suppose from the original SVM model (5), we get the support vectors are x_si, i = 1,2,...,k. We give the notations. x_jmax, i = max {x_sij, i = 1,2,...,k}, x_jmin, i = max {x_sij, i = 1,2,...,k}, j =1,2,..,n, where x_sij is the j-th element of vector x_si. We define x_max = (x_1max, x_2max,..., x_nmax), x_min = (x_1min, x_2min,..., x_nmin).

Theorem 2: Given the same parameter C in the PPSVM model and the original SVM model. We get the classification function f(x) = sgn((ω·x)+b) solving the original SVM model. We get the classification function The original data distribution function is P(x), its mean value is μ and standard dubitation is σ.

Under the assumption that the slack variable ξ_i is same in the original model (5) and PPSVM model (6), for 0-1 loss function c(x,d,f(x)), the difference of the expect risk of classification function f(x) and is not more than P(x_max)-P(x_min). Then we say that the PPSVM model has comparable accuracy comparing with the original SVM model.

Proof: From Lemma 3, we know that under the assumption the corresponding support vectors do not change, that is if x_i is a support vector then x_i+a is a support vector, too. From lemma 3, we know that the difference of classification of f(x) and only may be the data points between the support vectors. Then difference of expect risk of decision functions is:

(17)

We get that the PPSVM model has comparable accuracy comparing with the original SVM model.

PRIVACY-PRESERVING METHODS AND QUANTIFYING PRIVACY

Present basic approach to preserving privacy is to let users provide a modified value of its original data. We consider the technique for modifying values:

Table 1:	Quantifying privacy

Value perturbation: Return a value x_i+r instead of x_iεR where r is a random value drawn from certain distribution. We consider mainly two random distributions. In fact, we can draw r from more random distributions which can improve the privacy:

Table 1 shows the privacy preserved by the methods of Uniform and Gaussian. We fix the perturbation of an entity. So, it has the same bias for the same attribute of all data points.

We use the same quantifying methods with the study by Agrawal and Srikant (2000). For quantifying privacy provided by a method, we use a measure based on how closely the original values of a modified attribute can be estimated. If it can be estimated with c% confidence that a value x lies in the interval [x₁, x₂], then the interval width x₁, x₂ defines the amount of privacy at c% confidence level.

Table 1 shows the privacy offered by the different methods using this metric.

ALGORITHMS OF PRIVACY-PRESERVING SUPPORT VECTOR MACHINES

Before give the algorithms, we define a notation firstly.

Definition 1 : If data and vector then Aa = ((x₁+a))^T, (x₂+a)^T,..., (x_n+a)^T.

Lemma 5: For two vectors a,bεRⁿ and AεR^mxn, then Aab = Aba.

Algorithm of PPSVM on horizontally partitioned data: The dataset using to obtain a classifier consists of m points in Rⁿ represented by the m rows of the matrix AεR^mxn. Each row contains values for n features associated with a specific individual, while each column contains m values of a specific feature associated with m different individuals. The data matrix A is divided into q blocks of m₁, m₂, ..., m_q rows with m₁+m₂+...+m_q = m, The blocks of matrix A are held by q entities P₁, P₂, ..., P_q, respectively. That is the entity P_i owns the data A_i. Each entity is unwilling to make public or share its data with others. Furthermore, they do not reveal its data for various reasons such as commercial reason. Data is often a valuable business asset for a factory. Each row of AεR^mxn is labeled belonging to the classes by a corresponding diagonal matrix DεR^mxn. Class label D_ii, or d_i for short, corresponds to the i-th data points in A. We assume that entities make public the matrix D.

In the horizontally partitioned case, we need an entity P₀ that is independent from the participating entities P₁,P₂...., P_q. It actually owns no data and its task is only computing the global classification function. It is the protocol initiator. Each entity P_i, except entity P₀, generates a random vector r_i, i = 1,2,..., q having the same dimension with the row vector dimension of data matrix A_i. The random vector r_i is privately held by the entity and is never made public and keeps the same during the procedure of training and testing. The elements of these random vectors are drawn from the Uniform distribution or Gaussian distribution or other distributions. Each row of the data owned by all entities is added to the sum of random vectors r_i, i = 1,2,..., q. Then the data matrix A becomes Ar₁r₂...r_q. After that, we can use the data set and the PPSVM model to compute the classification function f(x) = sgn((ω·x)+b). When one entity has a new data point x whose class label the entity wants to know. The initiator P₀ can predict its label using the data x+r₁+r₂+...+r_q.

Now we completely describe algorithms for horizontally partitioned data using PPSVM model. The whole algorithm is described as Algorithm 1 and Algorithm 2.

ALGORITHM 1

•	Step 0 (initialize): P0 is the protocol initiator. The entity Pi (i = 1,2,..,q) generates its random vector ri and holds it privately. The vector has the same dimension with the row of its dataset owned and keeps the same during the procedure of training and test.
•	Step 1: The first step begins by the initiator P0 sending an empty dataset to its neighbor P1. Note that, to prevent P0 from contributing to training set, P1 must reject P0`s start request if P0 sends a non-empty set in the first round.
•	Step 2: For i = 1,2,...q-1 entity Pi sends the dataset to entity Pi+1.
•	Step 3: Pq send the dataset

to P₀.

•	Step 4: P0 removes and send the remaining dataset to entity P1.
•	Step 5: For i = 1,2,...,q-2, Pi sends the dataset to P0 and sends

to P_i+1.

•	Step 6: Pq-1 sends to P0.
•	Step 7: P0 gets dataset

(Lemma 5) . Go to Algorithm 2.

ALGORITHM 2

All entities make public the class matrix D_ll = ±1, l = 1,2,..,q for the data matrices A_i, i = 1,2,..q.

(18)

Get the optimal α*.

Get the classification function is f(x) = sgn ((ω·x)+b).

Assuming that the participating entities do not collude with the initiator and do follow the given protocol correctly, the initiator successfully acquires the global SVM model without disclosing any information on the private data of any entity.

One may ask why not one entity adds its random row vector to each row of its private dataset and sends to the initiator entity P₀ directly. The reason is that random row vectors generated by entities may be different. Then the same column of the data horizontally partitioned has different bias values comparing to the original data. Thus it effects the classification accuracy. Another problem is that the initiator entity P₀ must send the empty dataset to its neighbor P₁ in the step 1 of the Algorithm 1. The key is that if P₀ sends data, in the end, it will know the bias value of its data. Because each row of the dataset has the same bias value (r₁+..+r_q), the initiator will know the data privately owned by the entities. It violates the privacy of the datasets.

From the algorithms described above, we know that each row of all datasets has the same bias value. The initiator P₀ constructs the PPSVM model using the perturbed dataset. Then we discuss how to perform testing/classification using the PPSVM model constructed by our algorithms. Suppose the entity P_i has the dataset New_i to know its classification. The entity P_i could simply send the original dataset New_i to initiator P₀ to classify. But this would violate this constraint that the initiator should learn nothing about the private data of any entity. Furthermore, the training datasets have the bias value. Using the original dataset to classify has a low accuracy. To be corresponding with the training dataset, the testing/classifying datasets should carry out the algorithms described above. Then the initiator would have the perturbed datasets for testing/classifying using the constructed PPSVM model.

As discussed earlier, the initiator is prohibited from contributing data to the testing set. That is to prevent it from obtaining any information. Thus P₁ must reject P₀`s start request if P₀ sends a nonempty set in the Step 1 of Algorithm 1. No data is disclosed to any entity in this process. The details of security can be found in the paper by Agrawal and Srikant (2000).

Algorithm of PPSVM on vertically partitioned data: Suppose there are P₁,P₂,...,P_q participating entities, having the datasets A₁,A₂,...,A_q, correspondingly. All of these entities hold some feature values of the same group. The group is constituted with m individuals. The entity P_i has f_i features, i = 1,2,..,q, where f₁+f₂+...+f_q = n. The whole data is A = A[A₁ A₂ ...A_q]εR^mxn. Suppose there is a volunteer entity P₁ that is willing to compute the global PPSVM model. We call it initiator and other entities are called passive entities. Similar to the horizontally partitioned case, our goal is to obtain the perturbed dataset that has the same bias value for each column. Entity P₁(i = 2,3,...,q) generates a random vector r_i whose dimension is the same with the row dimension of it owned dataset. Each element of the vector r_i is drawn from Uniform distribution, Gaussian distribution or other distributions. Entity P_i holds the vector r_i privately. We now describe the algorithm (Algorithm 3) for PPSVM model on vertically partitioned data.

ALGORITHM 3

•	Step 0 (Initialize): Choose the initiator Pi. Entity Pi (i = 2,3,...,q) generates its random vector ri. The random vector does not change during the training and testing. All q entities agree on the same labels (matrix D) for each data point. If an agreement on D is not possible, we can use semi-supervised learning (Fung and Mangasarian, 2001b) to handle such data points. It is the future work.
•	Step 1: For i = 2,3,...,q, entity Pi (i = 2,3,...,q) executes the operation :Airi. Then the entity Pi sends Airi to the initiator P1. Thus the initiator Pi now owns the data Airi, i = 2,3,..q. The initiator P1 owns the data where 0 is the vector whose elements are zero. Its data point is represented with
•	Step 2: Using all data, initiator P1 solve the PPSVM dual problem:

(19)

Get the optimal α*.

The classification function is f(x) = sgn ((ω·x)+b).

Using the classification generated in step 3, we can get the label of x+r_i that is also the label of x.

In the Algorithm 3, the initiator P₁ does not send its privately owned data to any entity, so it does not disclose any information of its owned data. For the entity P_i (i = 2,3,,..,q), it sends dataset A_ir_i to initiator P₁. Because each element of the random vector r_i is generated from certain distribution, which is known to P_i, it is hard for the initiator P₁ to get any information of A_i. When entity P_i has data New_i to know its class label, P_i sends New_ir_i to P_i as described in Algorithm 3. There is not any information disclosed in the protocol. Our algorithm is secure. The details of security can be found in the stduy by Agrawal and Srikant (2000).

EXPERIMENTAL RESULTS

The present research demonstrate the efficiency of our protocol in three ways. First, our experiments show that the accuracy obtained using this PPSVM is the same as that of SVM when the data is centralized. Second, our experiments show that using our approach, entities can obtain classifiers with lower misclassification error than that obtained using only one entity`s data alone.

We choose three datasets from UCI repository. To simulate a situation in which each entity has only a subset of the feature space for each data point, we randomly distribute the features among the entities such that each entity receives about the same number features. Similarly, to simulate the situation in which each entity has only a subset of the individuals with the same features, we randomly distribute the data points among the entities such that each entity has about the same number individuals. The SVM model is created using the LIBSVM. During the experiments, we choose the same SVM model for one dataset.

Comparison of this approach with classifiers obtained when the data is original centralized: We demonstrate the classification accuracy of our approach is the same with that of the case when the data is centralized. The accuracy is shown in the Table 2 no matter horizontally partitioned or vertically partitioned.

From the Table 2, we can observe that the accuracy of our approach obtained the same with that obtained when the data is original centralized. Besides, the accuracy does not change with the increasing of entities number.

Table 2:	Accuracy comparison of our approach no matter horizontally or vertically partitioned data case with the original centralized data case

Table 3:	Comparison the classification accuracy of our approach with that of using only one entity`s original data in the horizontally partitioned case

Table 4:	Comparison the classification accuracy of our approach with that of using only one entity`s original data in the vertically partitioned case

Comparison of our approach with classifiers obtained when the data is part of original centralized: We compare the classification accuracy of this approach with that of using only one entity`s original data. The accuracy is shown in the Table 3 in the case of horizontally partitioned. Table 4 shown the accuracy of vertically partitioned case.

From Table 3 and 4, we can see that this approach has lower classification error comparing with that using only one entity`s original data in the most cases, especially when the number of the data points is not large.

DISCUSSION

Though our approach is efficient and secure, there are some challenges. The requirement of an entrusted intermediator seems overly restrictive. A key challenge for the future work is to remove the need of the intermediator, while still efficiently constructing the SVM model. In the vertically partitioned case, we can deal with data points using semi-supervised learning methods when an agreement on class label D is not possible. We also can immerge our approach with other SVM model such as RSVM (Lee and Huang, 2007), PSVM (Fung and Mangasarian, 2001a) etc. to improve the accuracy and reduce running time. There is room for future work in PPSVM. Privacy-preserving Support Vector Regression may be a prospective study direction.

CONCLUSION

This study proposes a unified model for privacy-preserving classification on horizontally and vertically partitioned data. We use the distinct character of classification. Giving the same bias value to the same column of the training and testing datasets, it does not affect the accuracy of classification. During the data passing, no data information is disclosed. Our protocol is secure. Besides, our experiment shows that our approach is scalable as the increasing of entities number. It costs lower time. PPSVM proposed can deal with the real number data points comparing with the paper by Yu et al. (2006a). The initiator in our approach can deal with the received dataset as its original dataset. It is convenient for initiator to construct the global SVM model and test the new data.

ACKNOWLEDGMENT

This research is supported by the National Science Foundation of China under Grant (No. 10571109, 60603090 and 90718011).