Opposition Intensity-Based Cuckoo Search Algorithm for Data Privacy Preservation

2.1 Related Works

In 2018, Afzali and Mohammadi [1] have suggested a data anonymization scheme for fitting big DM. The distinctive characteristics of big data make it essential to consider every rule as an ARM with an appropriate degree. In addition, the proposed methods were compared with the conventional methods, and the quickness of the DM process was offered.

In 2016, Li et al. [13] have focused on privacy-preserving mining on vertically partitioned databases. To ensure privacy they have designed an efficient homomorphic encryption scheme and a secure comparison scheme. They have also proposed a cloud-aided frequent itemset mining solution, which was used to build an association rule mining solution. These solutions are designed for outsourced databases that allow multiple data owners to efficiently share their data securely without compromising on data privacy. These solutions leak less information about the raw data than most existing solutions. Finally, the proposed scheme was compared with other conventional schemes, and the resource utilization at the data owner end was found to be very low.

In 2016, Lin et al. [15] introduced two novel techniques based on the Maximum Sensitive Utility (MSU), which were established to reduce the effects of the sanitization procedure for hiding sensitive high-utility itemsets (SHUIs). Accordingly, the introduced schemes were modeled to remove SHUIs proficiently or lessen their utilities by means of the conceptions of minimum and maximum utility.

In 2018, Chamikara et al. [6] presented an effectual data stream perturbation technique, known as P2RoCAl, that provides improved data utility when distinguished with its contenders. Moreover, the classification accurateness of P2RoCAl data streams was found adjacent to those of the data streams, which were original. Finally, from simulation results, the P2RoCAl was found to offer better resilience in contrast to data reconstruction attacks.

In 2016, Tripathi [24] adopted an antidiscrimination approach like discrimination discovery and prevention in the DM. There were primarily two kinds of adopted models, namely, direct and indirect discrimination. The former existed in the conditions when decisions were attained depending on the sensitive attributes, while the latter exists in the circumstances when decisions were attained depending on the non-sensitive parameters.

In 2016, Upadhyay et al. [25] implemented a geometric data perturbation (GDP) technique by means of data partitioning and 3D rotations. Accordingly, attributes were alienated into three groups, and every group of attributes revolved around varied pairs of axes. Finally, the investigational assessment shows that the implemented technique delivers worthy privacy preservation outcomes and data utility when distinguished from the traditional techniques.

In 2016, Komishani et al. [12] adopted a scheme known as preserving privacy in trajectory data (PPTD), a new methodology for data preservation depending on the perception of privacy. In addition, it intends to strike stability among the contradictory objectives of data privacy and data utility in agreement with the privacy necessities. From the simulation results, PPTD was found to be proficient for preserving personalized privacy.

In 2015, Yun and Kim [30] proposed a fast perturbation process depending on a tree structure that more rapidly carries out database perturbation procedures for avoiding sensitive information from being revealed. In addition, wide-ranging experimental results were performed for the offered method and traditional schemes by means of both real and synthetic datasets, and the results have offered improved and faster runtime when compared with other conventional schemes.

In 2017, Mehta and Rao [18] proposed techniques such as k-anonymity, l-diversity, and t-closeness, which were utilized to de-identify the data; however, the chances of reidentification are always possible because data were collected from multiple sources. Moreover, it is tough to handle large data for anonymization; MapReduce technique was introduced to handle large volume of data. It distributes large data into smaller chunks across the multiple nodes. Therefore, scalability of privacy-preserving techniques becomes a challenging area of research. The authors explored this area, and they introduced an algorithm named scalable k-anonymization using MapReduce for privacy-preserving big data publishing. Finally, they compared their technique with existing techniques in terms of running time that results into a remarkable improvement.

2.2 Review

Table 1 shows the methods, features, and challenges of conventional techniques based on the privacy data preservation techniques. At first, Fuzzy logic was adopted in [1], which minimizes unwanted side effects and offers better implementation on a huge amount of data. However, there was no contemplation on information loss. In addition, ARM was suggested in [13] that provides reduced complexity, and it also achieves high level of security; anyhow, there was less consideration on the utility of the proposed scheme. In [15], MSU-MAU and MSU-MIU were proposed, which speed up the evaluations, and along with that, they offer bias among the generated side effects and preservation. However, they require effective consideration on flexible fitness function. Similarly, k-nearest neighbor (kNN) was adopted in [6], which presents better classification of privacy-preserving data and improved accuracy, but there was no contemplation on effectiveness of sampling methods. In [24], ARM was suggested, which provides reduced complexity, and moreover, it achieves a high level of security. However, there was no description of carrying out clustering. In addition, GDP was adopted in [25], which offered enhanced privacy along with high variance, but the decrease in variance increases the chance of attacks. Accordingly, PPTD was proposed in [12], which eliminated critical moving points together with the minimization of attacks. Anyhow, there was no consideration of PPTD with various sensitive attributes. Finally, fast perturbation algorithm (FPA) was presented in [30], which evades the privacy breaches, and it also provides better runtime and scalability. However, there was no consideration on combination with web mining. There, these limitations have to be considered for improving the PPDM techniques effectively in the current research work.

3.1 Objective Function

The proposed model OI-CSA intends to attain the objective function for preserving data as given by Eq. (1).

In Eq. (1), F1,F2,F3, and F₄ are the objectives, which demonstrates the importance of the corresponding function as defined in the section below.

In Eq. (2), F₁ denotes the normalized HF rate, f₁ denotes the HF rate, where max(f1) is considered as the worst f₁ of all iterations. In Eq. (3), F₂ denotes the normalized IP rate, and f₂ denotes the IP. In Eq. (4), F₃ denotes the normalized IP rate; f₃ denotes the FR rate. In Eq. (5), F₄ denotes the normalized DM rate; f₄ denotes the DM. Here the original database is considered as O, and the sanitized database is indicated as O′.

HF rate denoted by f₁ is defined as the fraction of sensitive rules which is depicted in O′ as revealed by Eq. (6). Here, the count of sensitive rules available in O′ is described as f1=|B′∩SRs|. In Eq. (6), B signifies the association rule produced prior to sanitization, B′ denotes the association rules obtained from O′, and SRs denotes the sensitive rules.

IP rate denoted by f₂ is described as “the rate of non-sensitive rules which are concealed in O′”. It is the reciprocal of information loss as revealed in Eq. (7).

FR denoted by f₃ is described as “the rate of artificial rules produced in O′” which is demonstrated by Eq. (8).

DM denoted by f₄ is portrayed as the count of modifications carried out in O′ from O as demonstrated by Eq. (9), in which dist points out the Euclidean distance found between O and O′.

3.2 Proposed Architecture

The overall framework of the adopted OI-CSA scheme is portrayed by Figure 1. Initially, the data preservation comprises two major processes, such as data sanitization and data restoration. In data sanitization process, a key is generated to preserve the sensitive data in a protective approach. The key has to be generated such that it must hide the sensitive data effectively for which OI-CSA is deployed for optimal key generation. The authorized person at the receiver side could then restore the sanitized data by exploiting the same key. This is because to sanitize and restore the data faster, a symmetric key is used by the sender as well as the receiver.

Figure 1:

Overall Framework of the Proposed Data Preservation Model.

3.3 Sanitization Process

Binarization of O and pruned key matrix A₂ are performed during sanitization. The resultant key matrix in binarized form is consecutively provided onto the rule hiding process, in which an XOR function is performed with binarized form of O with identical matrix dimensions and added up with one that generates the O′, as specified by Eq. (10). Moreover, O′ obtained from sanitization process attains SRs and association rules following sanitization B′. Similarly, O extracts the relative association rules prior to sanitization B for attaining the above mentioned objectives. The structural design of the proposed sanitization process is illustrated by Figure 2.

Figure 2:

Sanitization Process of the Implemented Data Preservation Algorithm.

3.4 Key Generation

The key generation includes solution transformation process, where A, a key representation, is converted with the aid of the Khatri-rao product. Primarily, A is restructured into A₁ with matrix dimensions [M′′O×Omax] in which O_max indicates the highest transaction length and M_O denotes the number of transactions; the nearest highest perfect square of M_O is denoted by M′′O. For example, the restructured process of A={1,2,1} performs row-wise duplication and constructs the reconstructed key matrix, A₁ with dimension [M′′O×Omax] as revealed by Eq. (11).

Thus, the key matrix A₂ with dimensions [MO×Omax] is attained by the Khatri-rao product of two identically restructured A₁ matrixes represented as A1⊗A1 in which the kronecker product is indicated by ⊗ and its dimensions are further reduced in terms of the dimension size of the original database. Based on the Khatri-rao product, the key generation process is carried out and generates a matrix with dimensions identical to O, which produces A2[MO×Omax]. Finally, the process of rule hiding is done to obtain O′ by hiding the sensitive rules. The optimal key generation is made by means of improved Cuckoo Search (CS) algorithm called OI-CSA.

3.5 Restoration Process

During the restoration process, the O′ achieved from sanitization and A₂ from key generation technique could be binarized. The binarized S_d from the binarization block is minimized from the unit step. In the meantime, the database and key matrix, which is binarized, takes on an XOR function following the subtraction, and subsequently the restored database is extracted. The sanitizing key, A₂, is reconstructed by exploiting Eqs. (11), (10), and (1) and the proposed OI-CSA update. It is deployed to generate O′ by which the lossless restoring could be carried out by Eq. (12), in which Ô indicates the restored data. The design of restoration process is given by Figure 3.

Figure 3:

Restoration Process of the Proposed Data Preservation Algorithm.

4.1 Key Encoding

The keys (chromosome) A used for the sanitization process are subjected to OI-CSA for encoding. The number of keys ranging from key A¹ to key A^M is optimized using OI-CSA, and the optimal key is identified. The solution-encoding process is illustrated by Figure 4. Here, the key length (chromosome) is assigned as M′′O.

Figure 4:

Keys for Encoding.

4.2 Cuckoo Search Algorithm

CS algorithm is a method that is established depending on the reproduction of cuckoos [17]. Usually, cuckoos lay their eggs in the nests of erstwhile cuckoos with the expectation of their babies grown up by alternative parents. Certain periods exists, when the cuckoos find out that the eggs in their nests do not possess to them, in such circumstances, the unfamiliar eggs are push out from the nests and are deserted. This approach is dependent on the subsequent three conditions:

1. Every cuckoo chooses a nest arbitrarily and lays an egg in it.

2. The desired nests with more eggs would be considered for subsequent generation.

3. For a predetermined quantity of nests, a host cuckoo can find out another egg with a probability Pε[0, 1]. Under such conditions, the host cuckoo can either push the egg or construct a nest in another place.

The final condition can be estimated by substituting a fraction of the n host nests with novel ones. The fitness or quality F_i of a solution can merely be equivalent to the value of the objective function. From the execution point, the demonstration which is followed is that every egg in a nest indicates a solution, and every cuckoo can lay only one egg, i.e. one solution. Moreover, no reverence can be performed among an egg, a cuckoo, or a nest. The objective is to exploit the novel and capable cuckoo egg (i.e. solution) to substitute a worst solution in the nest. CS approach is very effectual for global optimization issues as it sustains a balance among global random walk and the local random walk. The balance among global and local random walks is adjusted by a switching constraint Pε[0, 1]. The global and local random walks are demonstrated by Eqs. (13) and (14) correspondingly. Accordingly, in Eq. (13), Xit and Xkt denotes the present positions chosen by arbitrary permutation, β indicates the positive step size scaling factor, Xit+1 points out the subsequent position, s denotes the step size, ⊗ indicates the element-wise product of two vectors, F signifies the heavy side function, P symbolizes a variable that is exploited to switch among global and random walks, and ε denotes an arbitrary variable from a uniform distribution. Accordingly, from Eq. (14), N(s,τ) indicates levy distribution exploited to describe the step size of an arbitrary walk.

4.3 OI-CSA

The conventional CS algorithm is an uncomplicated and efficient global optimization algorithm; anyhow, it could not be exploited directly to resolve multimodal optimization issues. Hence, the traditional CS approach is improved by modifying the opposition intensity denoted by γ as given in Eq. (15). In Eq. (14), Xi(w) indicates the worst solution, Xit denotes the current solution, and γ varies from 0 to 1.

The fundamental phases of the OI-CSA algorithm depending on their conditions is given by Algorithm 1.

5.1 Simulation Procedure

The proposed OI-CSA method for privacy preservation of sensitive data was simulated in JAVA, and the results were obtained. The analysis was carried out using four datasets, namely, T10, Chess, Retail, and T40. Moreover, the results were compared with conventional models such as PSO [31], GA [27], DE [32], and CSA [3] algorithms. In addition, the results that were obtained by varying the γ value were also described for the four adopted datasets.

5.2 Performance Analysis

The performance analysis of the proposed OI-CSA model for four datasets is given by Figure 5. From Figure 5A, it can be noted that the presented model for chess dataset in terms of F₁ is 17.36%, 14%, 12.24%, and 6.99% better than the PSO, GA, DE, and CSA designs. Also, for F, the suggested scheme is 0.76%, 0.49%, 0.28%, and 0.23% superior to the PSO, GA, DE, and CSA algorithms. Also, from Figure 5B, a retail dataset was implemented, where the implemented scheme for F₂ is 1.89%, 3.02%, 2.64%, and 2.64% better than the PSO, GA, DE, and CSA algorithms. Also, for F, the proposed OI-CSA model is 1.89%, 3.02%, 2.64%, and 2.64% superior to the PSO, GA, DE, and CSA algorithms. Similarly, from Figure 5C, the performance analysis for the T40 dataset can be obtained, where the suggested design for F₃ is 86.08%, 83.7%, 83.2%, and 82.29% better than the PSO, GA, DE, and CSA methods. For F, the presented method is 2.36%, 1.47%, 0.89%, and 0.29% superior to the PSO, GA, DE, and CSA algorithms. Moreover, from Figure 5D, the T10 dataset can be attained, where the F₄ analysis for OI-CSA model is 52.38%, 38.98%, and 10.94% better than the GA, DE, and CSA approaches. Also, from Figure 5D, the performance analysis for the T10 dataset for F can be attained, where the implemented scheme is 0.32%, 0.31%, 0.29%, and 0.1% superior to the PSO, GA, DE, and CSA schemes. Therefore, the enhancement of the proposed OI-CSA model has been substantiated successfully.

Figure 5:

Performance Analysis of the Proposed and Conventional Approaches for Data Preservation Algorithm Using Satabase (A) Chess, (B) Retail, (C) T40 and (D) T10.

5.3 Effect on Varying γ

The effect of varying γ using four datasets, namely, chess, retail, T40, and T10, is demonstrated by Tables 2–5, respectively. The values of γ were varied from 0.0, 0.2, 0.5, 0.7, to 1. Accordingly, from Table 2, the normalized HF rate, F₁ has attained the values of 0.24875, 0.01890625, 0.24575, 0.24775, and 0.020227273 for γ = 0.0, γ = 0.2, γ = 0.5, γ = 0.7, and γ = 1, respectively, using the chess dataset. From Table 3, the normalized rate, F₂, has attained the values of 0.490418, 0.487253, 0.490492, 0.490418, and 0.487179 for γ = 0.0, γ = 0.2, γ = 0.5, γ = 0.7, and γ = 1, respectively, using the retail dataset. In addition, from Table 5, the normalized IP rate, F₃, has attained the values of 0.025027958, 0.019158457, 0.04072148, 0.015477353, and 0.025347075 for γ = 0.0, γ = 0.2, γ = 0.5, γ = 0.7, and γ = 1, respectively, using the T40 dataset. Also, from Table 5, using the T10 dataset, the normalized FR rate, F₄, has acquired the values of 0.072494, 0.030037, 0.052792, 0.025071, and 0.09723 for γ = 0.0, γ = 0.2, γ = 0.5, γ = 0.7, and γ = 1, respectively. Therefore, the superiority of the presented OI-CSA approach has been validated effectively.