Data analysis with performance and privacy enhanced classification

1.1 Overview

Data mining and analysis have influenced everyone’s life. You start searching any term on google and end by selecting various auto-search word combinations and getting suggestions based on your search query. You go for online shopping and many recommendations like “users who purchased this, also viewed these items.” With this, one can experience personalized shopping. Many websites ask to register and log in before giving access to the data needed by the user. Here, everyone wants to keep their information secure. Due to this, fake data are shared by users on the Internet affecting data analytics. Extracting valuable knowledge from data while preserving the privacy of data has become essential now [1]. In Industry 4.0, data privacy is very important while sharing data for analytics [2]. When the data collection is huge and needs to be mined, it requires high-capacity servers and storage, which gives rise to the issue of real-time mining. Vast data storage as well as sharing of information over the Internet brings out different issues of data security. It also suffers from communication costs. So, some fast and privacy-preserving data processing techniques are needed for analysis and classification as hackers are also there to hack the sensitive information of users despite using security protocols and encryption. In such situations, privacy is defined as keeping the personal as well as social information of individuals in a secure manner when there is a presence of a third party who performs the computation based on their details [3]. Consequently, the next subsection explains the procedure of privacy-preserving process and motivation of this research work.

1.2 Privacy-preserving data mining (PPDM) and motivation

PPDM methods have been established to permit the abstraction of information from datasets while avoiding the revelation of delicate data or data subject’s identities. Two or more researchers can collaborate on a single dataset with the PPDM technique. In a multilateral environment, sensitive and confidential information about dataset can be obtained by performing a data mining process without disclosing the data of each party to other parties. In PPDM, it is necessary to protect sensitive information by either masking data or changing the data while maintaining classification accuracy. The PPDM algorithms are judged based on the rate of unique data recovery from modified data, loss of information, and an effect on accuracy [4]. Domadiya and Rao proposed privacy-preserving association rule mining for handling the issue related to the healthcare system in IoT applications [5]. In this COVID-19 pandemic, the healthcare data of COVID patients are collected for ministry decisions, to know the requirement of oxygen beds and the requirement of medicine supply in coming days, etc. Here, patients are reluctant to share their data. The real challenge in data collection is to project on the privacy of patients. It is again necessary to provide privacy to the data while protecting the classification accuracy of sensitive data, which is a very cumbersome task [6]. Usually, data are transformed to preserve its privacy. It may reduce the usefulness of data. The effect of the privacy preservation algorithm is calculated by means of the risk of data disclosure from deployed data. The degree of privacy preservation is expressed as the variance of recovered and original deployed data. The characteristics of PPDM are like preserving the privacy of sensible information and the accuracy of data mining, and it should not compromise the access to sensitive data [7].

1.3 Research contribution

In this research, a data security method is developed and it eases data classification by reducing the computational time and storage space requirements while retaining the accuracy of classification. Since original data get transformed into another form, they become secure. This also proves very useful as no one can hack or make misuse of individual’s data. The data remain unrecognizable and still give accuracy in mining. This article contributes a privacy-preserving method, which works very effectively on the level of privacy preservation, data hiding, and classification accuracy preservation while reducing the complexity of data processing. The highlights of this research are as follows:

• The proposed technique uses an overlap technique that adds the transformation to get output in a reduced format. Since input gets processed in an overlapped fashion, technique uses FFT algorithms which need less time.

• Data utility and mining accuracy are maintained despite privacy preservation, and data are transformed into reduced form leading to less storage and computation time.

• Data transformation is irreversible so one can share transformed data without security concerns.

• With the help of the proposed C-PPA method, the classification algorithm gives more accurate results in the classification process.

Additionally, the novelty of this research is explained here. Generally, in machine learning algorithms, adding many more features at first will give a more accurate result. However, the output of model will reduce after a certain level, with the rising number of elements. This is due to the high dimensionality of features presented in the dataset. This problem arises due to the exponential decrease of samples with increasing dimensionalities. The dimensionality of the feature space expands more and more when the features must be added without increasing the number of training samples. And, this leads to the problem of overfitting in machine learning algorithms. This is reduced with the help of the proposed C-PPA technique which reduces the features and obtains the collection of principle features. In this research, the proposed feature reduction algorithm gives novelty, and by this algorithm, the classifier achieves better classification accuracy.

1.4 Organization of the study

This study is categorized as follows. The next section elaborates on existing research works in the domain under consideration. The third section focuses on the proposed method of privacy preservation and privacy-preserving against member inference attacks. The fourth section describes the simulation outcomes of proposed privacy-preserved data, and the last section concludes the paper.

3.1 Data pre-processing

Generally, pre-processing process is considered as a primary process that directly affects the success rate of proposed methodologies. Performing other processes without pre-processing step reduces the data complexity, since the real-world data are unclean. Likewise, the repeated data or duplicate data also decrease the training process of the proposed methodology. In order to overcome these issues, this research performs two different kinds of a process named data imputation and Min–Max normalization, and they are performed before applying the proposed feature reduction technique.

Let us consider the original value to be N and the normalized value tends to N ′ . The range of N can be given as [ min , max ] , and the new range is represented as [ mi n ′ , ma x ′ ] . For the mapping of the original value from one range to a new range, the normalized value is given by,

Using equation (1), Min–Max normalization has been calculated for all values in the dataset. Before going to feature extraction process, we have generated one example for the Min–Max normalization process. After converting words into numerical values, Min–Max normalization is applied. The strength of the Min–Max normalization procedure is that the correlation among the values of original data is maintained. The data redundancy can be diminished, and the uprightness is improved. After the pre-processing, the feature reduction technique is performed on the pre-processed data to diminish the dimensionality of the features. Moreover, the feature reduction method is detailed in the following subsection.

3.2 Proposed C-PPA technique

After performing the pre-processing steps, C-PPA is applied to the pre-processed data to obtain reduced features for classification purposes. Convolution is one of the operations in digital signal processing (DSP) applications which converts every single value in the input sequence by delta function or scaling and shifting of unit impulse. From this, it is observed that the dominant dimensions in the input result in output with the same dominance. The first step of C-PPA is to divide data tuples into groups, with each group containing elements nearly the same as the closest, 2 n . These groups act as input to C-PPA. After the application of C-PPA, the output contains a reduced set of elements. Basically, C-PPA is based on the circular convolution of the signals expressed as,

The significant property of Discrete Fourier Transform (DFT) is the circular convolution which is expressed in equation (2). The multiplication of DFTs of two sequences shows the same results as the circular convolution of the same in the time domain [41,42,43,44,45,46]. The proposed C-PPA procedure is given below.

From the above algorithm, one can reach Y(K) by taking the FFT of the output of C-PPA. But the hacker will get stuck at the point-by-point multiplication stage as numerous possible combinations can be there. So, Step 5 in the algorithm is irreversible. One cannot recover the original data from the extracted data. Hence, convolution-based privacy-preserving algorithm is the loss method. Likewise, the observed reduction in features is more effective when applied to data tuples containing a large number of features. The calculation of DFT is performed using Fast Fourier Transform (FFT) algorithms which proves beneficial in terms of calculations and hence computation time due to the following reasons:

1. Computation with FFT algorithm – for N Number of values, DFT calculation needs N 2 complex multiplications whereas FFT calculation needs ( N ∣ 2 log 2 N ) . Similarly, DFT calculation needs ( N 2 − N ) complex additions whereas FFT calculation needs ( N log 2 N ) . Hence, the speed improvement is observed. Hence, FFT is a fast algorithm to calculate the DFT and with FFT, and the convolution operation is carried with high speed. Convolution performed using FFT is also called as high-speed convolution.

2. Overlapped Computations – here, computations are performed in overlapped fashion and then, the results are added to get the final data sequence

3. The data tuple features present in input undergo convolution and are transformed into a new set of features containing a smaller number of features. The new features are nothing but a combination of input features. The dominance of features is preserved in output; hence, we get better classification accuracy of output.

With the proposed system architecture, the data will be collected only at the central server and processed there. Now, the main server can share the transformed data ‘ n ’ number of times with other servers for data analytics purposes, and it can be shared through any network. The data will remain safe and trustworthy for analysis. Also, the central web server can store only the transformed representation of data, i.e., output of C-PPA, as it is the reduced form and gives better accuracy than the original data. The proposed algorithm gives output in reduced dimensions, so it proves better in terms of computational time and space requirements. Likewise, it also has some advantages when compared to the perturbation-based methods which are given below:

• reduction in data;

• extraction of data in a different form so that original data can’t be predicted;

• data privacy is achieved;

• data utility is maintained;

• data mining accuracy is preserved;

• reduction in computational time;

• reduction in storage space;

• improved accuracy of mining;

• reduced communication costs over the Internet.

In this proposed C-PPA algorithm, the data tuple can be broken into a group of features for parallelism. Every group is processed parallelly in a overlapped manner. The output of these small groups is computed fast and then output features are fitted one over another to get the final output. This is known as overlap-add C-PPA algorithm which transforms the input in some anonymized and reduced form. From the algorithm, it is clear that N is nearly similar to a number of features in each group. If we consider an example of Parkinson’s disease, the total features in one tuple are 754. We divide features into two groups with 377 in each group. So, we decide N = 512 (nearest 2 r ). Hence, the features in output after the application of C-PPA with N = 512 will be 512. Here, the output has reduced features. Also, the output is a transformed version of the input which is not reversible. Now, the output of C-PPA can be shared with different servers for analysis without data leakage risk. After performing the feature reduction process, the classification is done by using different machine learning algorithms like like SVM, DT, CNN and NB classifiers. These classifiers achieve better classification accuracy by using the proposed feature reduction technique. Subsequently, the next subsection explains the privacy preserving against the member inference attacks.

4.1 Dataset description

Datasets are taken from the UCI repository for the evaluation of the proposed algorithm. UCI is a large dataset repository which contains different kinds of datasets to perform various operations. From this repository, six disease datasets are selected, which are pre-processed by missing value imputation and Min–Max normalization technique. The proposed algorithm is evaluated on selected datasets. Table 1 shows the datasets under consideration with their dimensions before and after the application of C-PPA. The observations show that the C-PPA algorithm also plays a major role in feature reduction too. The reduction in features is possible due to the selection of N and division of total attribute of a tuple in different groups which is an inbuilt methodology of the C-PPA algorithm.

The binary classification datasets listed in Table 1 undergo the C-PPA algorithm. CNN, DT, SVM, and NB classifiers are selected for evaluation of the proposed algorithm as it does not require any domain knowledge for its construction. Also, its execution is fast and easy to understand for humans [33]. To check the usefulness of data for further analysis, the mining accuracy is recorded for each dataset after the application of all the classifiers.

Table 1 clearly displays that the original features of Parkinson’s disease dataset which are reduced from 754 to 512. The total reduction for Parkinson’s disease dataset is 242 features. Here, the C-PPA transforms the original data dimensions into lower dimensions. The data is transformed or extracted in reduced form with the application of C-PPA. In this method, the important or high weighted features maintain their weight in output after convolution, so the accuracy of convolved data does not get reduced [42].

Bertino et al. [43] presented the desired dimensions for the evaluation of the PPDM algorithm. It mentions efficiency, scalability, level of privacy achieved, quality of data after application of PPDM, and data hiding failure [44,45,46]. All these parameters are essential during the assessment of the quality of the algorithm under consideration. The efficiency is determined in terms of performance, i.e., computation cost, and space requirement in general. Tables 1 and 2 play a major role in deciding the efficiency of the C-PPA algorithm. Table 1 speaks with respect to space requirement whereas Table 2 shows effectiveness in execution time. The application of C-PPA also performs feature reduction and hence reducing space requirement and computation cost.

Scalability includes the efficiency of data mining for increased data size. The feature reduction of C-PPA works for it. C-PPA extracts the input dataset into lower dimensions. More % reduction in features is observed for the datasets having half of the total features near any 2^N. A high level of privacy is achieved by C-PPA as the output of C-PPA is not reversible. The original tuple features are transformed into a smaller number of features which do not have any correlation with the original values. The quality of data is important in the sense of its analysis purpose. If the information is to be mined for an application like health care record, then its accuracy of mining is crucial in decision making. In this regard accuracy, data consistency plays a significant role to assess the quality of data. The accuracy of data mining depends majorly on the information loss after the application of PPDM. From Table 2, it is clear that after the application of C-PPA, the accuracy of classification is either retained or improved. Table 2 shows the computation time elapsed with and without the application of C-PPA. Table 2 observations show that computation time is less with the application of C-PPA as the number of attributes is also reduced. Here, the time complexity of circular convolution using the FFT algorithm, which is the heart of C-PPA, is O (N log N) [47,48].

Table 3 summarizes the observations on accuracy of DT, SVM, NB and CNN classifiers with and without using C-PPA algorithm as well as different feature extraction algorithms like Principal Component Analysis (PCA) and Singular Value Decomposition (SVD). It is observed that CNN gives the same or better accuracy as compared to without C-PPA algorithm. The use of C-PPA algorithm shows a 14.15% average gain in accuracy for the CNN classifier. Additionally, the comparison with previous feature reduction techniques shows that the proposed algorithm achieves better accuracy results than existing methods.

From Table 4, it is clear that after the application of C-PPA, the accuracy of classification is either increased or remains close to the original accuracy. For classification, a CNN is applied to the output of C-PPA as CNN achieves higher classification accuracy when compared to other classifier algorithms. Here, C-PPA acts as a feature-reduction technique for classification purposes.

In the C-PPA approach as stated earlier, the features can be grouped into more than two groups, processed parallelly then execution time can be reduced further. It is known as overlap-add approach of convolution. Table 5 compares the accuracy without C-PPA with Overlap-add convolution-based PPA algorithm (OAC-PPA). An average gain in accuracy with Overlap-add C-PPA is 9.11, 9.78, 08.15, 12.79, and 10.58 for DT, SVM, NB, and CNN classifiers. Tables 4 and 5 strongly prove the accuracy gain factor of the proposed algorithm by using different classifiers. It is clear from the results that despite of gain in accuracy the important feature is a reduction of features which in turn causes a reduction in processing time and space as well.

Table 6 lists the performance of different classifiers on different parameters by performing the proposed feature reduction technique on different healthcare datasets. Here, the classifiers like SVM, DT, CNN, and NB are analyzed to show the efficiency of the proposed methodology. These classification algorithms are implemented by ourselves, and the comparison is made for different medical datasets. From the comparison, it is shown that the DT classifier displays better results in terms of different parameters like accuracy, precision, recall, and F-measure.