Abstract
Cloud services are cloud computing resources and applications deployed on the Internet or cloud computing platform, and users can access the required cloud services at any time. However, users face the diversity and complexity of quality of service (QoS) when evaluating and selecting cloud services. Therefore, it is important to study and establish an effective and objective trust model to improve user satisfaction and interaction success rate. In this paper, a model based on combining weights and gray correlation analysis is proposed. Firstly, direct trust, recommendation trust, and reputation together form a comprehensive trust, resulting in a more accurate overall trust. Second, rough set theory and analytic hierarchy process (AHP)-based method are used for the direct trust. Meanwhile, the degree of similarity recommendation trust is calculated by a gray relational analysis method. In order to ensure the accuracy of direct trust, this paper proposes a dynamic trust update mechanism. Finally, the simulation experiment is carried out to verify that the cloud services trust evaluation model (CSTEM) is more robust than the other three methods. It protects against malicious entities; at the same time, it can increase user satisfaction and interaction success rate.
1. Introduction
Google introduced the concept of cloud computing in 2006. The cloud computing center uses virtualization technology to organize a large number of idle resources, forming a huge “virtual resource pool”, and users can request personalized cloud services through the network. As the scale of cloud services continues to expand, more and more service providers provide services with similar functions and different quality of services. Users select the most suitable cloud services from among many cloud services. In a cloud environment, trust is subjective and dynamic and is influenced by many factors. In today’s increasingly competitive environment, the highly scalable technology of cloud services brings vitality to enterprises, and it also challenges users’ trust in cloud services. At present, the trust problem of cloud computing is the most concerned issue of most enterprises, and it is difficult to select cloud services. Therefore, it is important to study and establish an effective and objective trust model to improve user satisfaction and interaction success rate.
At present, home and foreign scholars have conducted research on the trustworthiness evaluation and selection of cloud services. A multidimensional trust evaluation system based on compliance is proposed [1]. Ding et al. [2] analyzed the security issues facing cloud computing and defined trusted cloud services. Tang et al. [3] proposed a trustworthiness cloud service selection framework and developed corresponding evaluation middleware. Wu et al. [4] proposed a custom choice decision based on reputation. Therefore, the reputation evaluation and selection research of cloud services has a good theoretical basis.
It is important to study and establish an effective and objective trust model to improve user satisfaction and interaction success rate. We propose a CSTEM based on combining weights and gray correlation analysis. The contributions of this paper are as follows:
(1) Direct trust, recommendation trust, and reputation together form a comprehensive trust, resulting in a more accurate overall trust. The direct trust also considers transaction time and transaction amount, so the final direct trust is very accurate and effective.
(2) We propose a gray correlation analysis method used to calculate the degree of similarity recommendation trust.
(3) In order to ensure the accuracy of direct trust, a dynamic trust update mechanism is proposed.
After this Introduction, Section 2 introduces some related works; Section 3 describes the system model and its trust evaluation process; Section 4 presents the proposed trust evaluation algorithm and trust dynamic updating mechanism; Section 5 describes simulation experiment results and analysis. Finally, Section 6 presents conclusions and possible future works.
2. Related Work
There are many evaluation methods for cloud service: service transactions [5], cloud storage [6, 7], cloud application [8], and cloud security [9, 10]. A trust mining model for identifying trusted cloud services when negotiating a Service-Level Agreement (SLA) is proposed [11]. Wang et al. [12] proposed a dynamic cloud service selection strategy named DCS. Shaikh et al. [13] identified some of the available checklists to evaluate and select the parameters of the service provider to select the optimal service. Yang et al. [14] proposed a QoS-aware cloud service selection strategy, using AHP to help cloud customers choose the appropriate cloud service. Lartigau et al. [15] proposed a method based on intuitionistic fuzzy sets, which can perform similarity evaluation.
For trust evaluation of cloud services, scholars at home and abroad have relevant research in this field. Li et al. [16] proposed a multiattribute service quantization algorithm that uses information entropy and rough set theory to evaluate service quality. Na S H and Huh E N [17] proposed an SLA-based evaluation model, analyzing security threats based on service types and a cooperative model for achieving universal consensus. Wang et al. [18] proposed the CTDSS method, which divided the trust into different communities and finally selected the appropriate cloud services. A CCIDTM (Cloud Computing Incentive and Detection Trust Model) proposed by paper [19] does not consider the user’s demand preference and direct trust update problem. Li et al. [20] proposed a trust-based selection scheme, and users can select cloud services through trust values. Wang et al. [21] proposed a dynamic cloud service selection strategy called DCS (Dynamic Cloud Service, DCS). Fan et al. [22] proposed a new model based on evidence reasoning. Ding et al. [23] proposed a CSTrust model, which can evaluate cloud service credibility by combining QoS prediction and customer satisfaction.
Compared to other selection methods, the CSTEM in this paper considers subjective weight and objective weight in the calculation of direct trust degree, and the calculated trust degree is more accurate. For the dynamic update of trust degree, the recommended trust degree is calculated by the gray correlation method. However, the methods in [16, 17] are not perfect for the malicious evaluation of users and the feedback of users after using cloud services. The CCIDTM does not consider the user’s demand preference and direct trust update problem [19]. Hu et al. [24] proposed a novel method; it does not consider the level of service and the accuracy of credible services span tree. There is no guarantee that the recommended service is the best service. The EigenRep trust model is a global reputation calculation model. There is no reputation penalty correction for node malicious behaviour. Even if the service with the best reputation is selected, the service is not guaranteed to be true and is the most suitable for users. At the same time, the EigenRep model requires an iterative calculation of the global reputation in the entire network for each transaction, and the system overhead is large.
In this paper, we propose an evaluation model that is based combining weights and gray correlation analysis, it introduces the direct trust dynamic update mechanism. Meanwhile, the model can dynamically evaluate the comprehensive trust of cloud services, and obtain the cloud service with the highest comprehensive trust, effectively helping users to choose the best cloud service.
3. Trust Evaluation Model
3.1. Model Definitions
Definition 1 (entity). This paper defines the ability of an entity to have self-behavior. Define cloud users CSU= , User2, ···. Cloud service provider is defined as CSP= , CSP2, [9].
Definition 2 (scoring matrix). After the user uses the cloud service, the user scores the cloud service, and the evaluation of the cloud service indicates that the article uses to indicate the trust evaluation.
3.2. Model Framework
The CSTEM is composed of five modules: service requestor, cloud service registration center [9], CSP, CSTMC, and trust feedback monitoring, shown in Figure 1. The CSTMC is composed of direct trust, recommended trust, reputation, and trust dynamic update mechanism. It performs the evaluation and selection of cloud services.
3.3. Trust Evaluation Process
The comprehensive trust consisted of the direct trust, recommendation trust, and reputation. The direct trust relationship means that both sides have historical interaction experience. Recommended trust has no historical interaction of both interactive experiences; relevant factors include the degree of similarity recommendation respondents and the level of the respondents. The reputation represents the evaluation of all the cloud service users.
This trust evaluation process is as shown in Figure 2.
The specific steps of the trust evaluation are as follows.
Step 1 (register). Cloud service provider resources are registered by CSAPI.
Step 2. Request.
Step 3 (calculate the comprehensive trust CST). History interaction records are inquired by the cloud trust management center. A gray relational analysis method is used to calculate the similarity recommended trust and get the recommended trust RT, and the reputation of is calculated by all users of the cloud service. Meanwhile, three different types of trust are assigned different weights for evaluation.
Step 4. If CST≥θ, indicating that the service meets the requirements of the user, go to Step 5; otherwise, if the cloud service does not meet the user requirements, please go to Step 2.
Step 5. Users select the cloud service with the highest CST.
Step 6. Update the direct trust according to formula (30).
4. Trust Evaluation Algorithm
4.1. Direct Trust Calculation
In order to guarantee the weight of evaluation is right, evaluation data should be adjusted to a uniform interval, the data preprocessing of this paper with a minimum-maximum normalization method, and it is calculated by paper [25].
4.1.1. Objective Weight
Rough set theory is founded by Polish Scientist Z. Pawlak [26]; it is a treatment imprecise, inconsistent, incomplete information and other effective tools.
Set S = (U, A, V, f) as an information system, the definition of fuzzy relation R, V represents an attribute value, represents an attribute value for the object i in the j-th attribute, and is a knowledge representation function, for each x∈U, q∈A, f (x, q) ∈V.
Definition 3 (set R as the equivalent relation on U). It is defined as follows: . The similarity of the object and the object is defined as , and represents a property value of data preprocessing.
Definition 4 (S = (U, A, V, f)). represents the fuzzy similarity class of the ,, they are expressed as
Definition 5 (S = (U, A, V, f)). The equivalence class of x is represented by I(x), x∈U, x represents the fuzzy relationship of U, , represents the upper approximation set, and represents the lower approximation set.
For a given threshold , the definition upper approximation set of variable precision rough set is
The under approximation set of is
Definition 6. Set R∈A, X as a division of property, , and the approximate classified quality is defined as
Definition 7. Set S = (U, A, V, f), and represents the importance of properties:
Definition 8. Set S = (U, A, V, f), A = , and represents the weight of attribute in A:
4.1.2. Subjective Weight
Analytic Hierarchy Process (AHP) is put forward a qualitative and quantitative combination of decision analysis method by A.L.Saaty in the 1970s [27]. AHP model for cloud service selection is as shown in Figure 3.
The process of calculating subjective weights using AHP is as follows.
Step 1. It clears a problem, and then establish hierarchical structure.
According to the important attribute index summarized by the cloud service, the relationship between the influencing factors of the cloud service selection process is analyzed, and the hierarchical model is established.
(1) Target Layer G. It chooses a most suitable cloud service from multiple cloud services.
(2) Criterion Layer B. It analyzes the factors that affect the choice of cloud service and the three basic elements of cloud service: cost, performance, and reputation.
(3) Attribute Layer C (Including Subattribute Layer).
(4) Object Layer.
The evaluation indicators B1, B2··· have an effect on the target G. We use a pairwise comparison method to determine their impact in the proportion of G. The results of all the comparisons are expressed by the matrix B = () n n.
Step 2. Set up the comparison matrix of the two pairs.
Step 3. Hierarchy Single Sort.
As shown in Figure 3, we use the judgment matrix B and then get a single ordering vector of the attribute index in the criterion layer B with respect to target G. It is a feature vector that satisfies , it represents , and is represented by the corresponding position element.
(1) We normalize each column element of the judgment matrix, and the general term of the element is
(2) Each column is normalized after the judgment matrix, and the line added is
(3) , is the approximate solution of the required eigenvector.
(4) .
Step 4 (consistency check). Since matrix B is subjectively determined, it is necessary to check satisfaction consistency.
(1) Judgment matrix consistency index CI (consistency index):
(2) Random consistency index RI.
(3) Consistency ratio CR (it is used to determine the permissible range of B’s inconsistency).
When n<3, the judgment matrix is always completely consistent. The random consistency ratio CR (consistency ratio) is as shown in the following equation.
Step 5 (hierarchical sorting). The hierarchical total sort vector is a relative weight vector that calculates the importance of the target layer for all factors at a certain level; the process is carried out from high level to low level. As shown in Figure 3, if the level B has m influencing factors B1, B2,, , the total ranking weights for the target layer G are . The next level C of the influencing factor Bj contains the n attribute indices C1, C2,..., , and their hierarchical single order weights are , and the C level attribute is given the total ordering weight of the target layer G:
4.1.3. Combining Weights
In this paper, the rough set theory is used to calculate the user’s objective weight, the AHP is used to calculate the subjective weight of the user, and finally their combined weights are calculated. The objective weight calculation formula is as shown in (9), and the subjective weight calculation formula is as shown in (15). represents the combined weight, m is the number of attributes, and the combined weights values are shown in the following equation:
4.1.4. Attenuation Factor
(1) Transaction Time. The user’s preference will decay with time. According to the Ebbinghaus forgetting curve, the time decay factor is established. The time decay factor is shown in the following equation.
When the user requests the service and interacts with the cloud service , represents the evaluation time after the user uses the cloud service , and the latest evaluation time after the user uses the cloud service is represented by . The earliest evaluation time after the user uses the cloud service is represented by .
(2) Transaction Amount. The transaction amount is a more important factor. indicates the transaction amount between the user and the cloud service at the i-th transaction. As shown in (18), it indicates attenuation factor .
is a constant greater than 1, which is used to adjust the intensity of the change in the amount of different transactions. It can produce discrete effects and better distinguish between the effects of different transactions, when can achieve a good discrete effect, so we set in this model.
represents the final attenuation factor, and .
4.1.5. Final Direct Trust
represents a trusted evaluation matrix of the user for the cloud service, combined with (16) and (19), and the final direct trust values are shown in the following equation:
4.2. Recommended Trust
Professor Deng Julong proposed a grey system theory [28]; it is a new method for less data and poor information uncertainty. In this paper, it is used to calculate the degree of similarity of the requester with other user’s recommendation. By the geometry of the column reference data and comparison data columns similarity, it is used to determine their similarity tightness.
Requester and recommender are for common use a set , , m is the number of recommenders, and trust evaluation for are and , the gray correlation analysis method is used to calculate the similarity of the evaluation of requesters and , and the step is as follows:
(1) represents gray correlation coefficient of the requester and recommender .
means the resolution factor. is smaller, the greater of the resolution, the general value . Set = 0.5, and , , and represent the poles that are maximum difference, minimum difference, and absolute difference for and .
(2) Gray correlation of and is
(3) represents the similarity, and m is the number of recommenders.
represents the trust relationship between recommender and service provider, represents the evaluation similarity, is a global trust of recommender, and recommendation trust value is as shown in the following equation.
4.3. Comprehensive Trust
represents a direct trust, represents a recommendation trust, and represents cloud service itself reputation in time , the initial value is 0.6, and the value with user interaction will be updated. represents the weight of direct trust. The weight of the recommended trust is indicated by , and the weight of the cloud service credibility is represented by . In time , the comprehensive trust degree of the user to the cloud service can be represented by , and the final comprehensive trust degree calculation formula is shown as (26). The pseudocode of trust evaluation process is shown in Algorithm 1.
| ||||||||||||||||||||||||||||||||||||||||||||
4.4. Trust Dynamic Updating Mechanism
Customer Satisfaction. represents each attribute trust preferred weight of cloud service, S represents service vector, maybe a single cloud service or multiple cloud services, represents trust evaluation feedback, and represents the satisfaction on the cloud service i. Customer satisfaction is shown in (28).
Penalty Factor. The user’s preference will decay with time, and trust will drop quickly. Therefore, the direct trust update of this paper introduces a penalty mechanism. Once the transaction fails, that is, the cloud service does not satisfy the user, the service party will be punished. Setting is a penalty factor after the transaction.
If the transaction is successful, it is , the transaction failure is , and the number of failures is indicated by n. In order to prevent the problem of small amount of transaction fraud, the transaction amount is considered. The trust value will be decreased rapidly in the transaction failures by the acceleration factor, so that once the main trading promises to each other, their trust values will decline rapidly.
Dynamic Trust Updates. Trust relationship will be updated by the trust manager center, represents the weight of transaction history , is calculated by (17), and is calculated by (18).
Finally, the direct trust value update formula (30) of the penalty factor is considered.
is a direct trust value, trust needs to be updated, through this service to meet the user set basic satisfaction threshold , and it is used to determine whether the transaction is successful. Rewrite the Cloudsimlet class in Cloudsim, add variables to identify cloudletPrice for the user in the interaction provided by the resources of the cost, and meanwhile, increase the variable cloudletTime, customer satisfaction ST, and penalty factor . Pseudocode of direct trust update is shown in Algorithm 2.
| ||||||||||||||||||||||||||||||||||
5. Simulation Experiment Results
The model is analyzed and implemented on the Myeclipse 2014 platform, its integrated simulation package CloudSim 3.0.3, its hardware environment for the CPU Core i7-3770 3.4GHz, RAM 8.00 GB. CloudSim is a new, universal, and extensible simulation framework.
The number of nodes in the cloud computing network is very large, the relationship between nodes is also very complex, and in this simulation, the experimental process is divided into many rounds: each round has the same nodes, and the trust mechanism is similar. There are 50,100,150,200,250,300 experiments of different sizes, the comprehensive average of the rounds of experiments as the average satisfaction, and the value of the success rate. It is successful to meet the certain threshold of customer satisfaction as a criterion for judging the success of the interaction: when the satisfaction reaches 0.7, the interaction is successful.
In this paper, CSTEM was compared with other CCIDTM, EigenRep, and RSS methods; the first method is a random service selection (RSS), and RSS only randomly selects a service that satisfies the basic functions of the user. It does not perform feedback after the user uses the service. The EigenRep is a global reputation calculation model. There is no reputation penalty correction for node malicious behaviour. The third method is the CCIDTM proposed by paper [19].
5.1. Customer Satisfaction
CSTEM was compared with other CCIDTM, EigenRep, and RSS methods. The average satisfaction of the different methods is shown in Figure 4. The customer satisfaction of different proportions of malicious cloud service entities is shown in Figure 5.
RSS only randomly selects a service that satisfies the basic functions of the user. It does not perform feedback after the user uses the service, and the satisfaction is the lowest. The EigenRep trust model is a global reputation calculation model. There is no reputation penalty correction for node malicious behaviour. Even if the service with the best reputation is selected, the service is not guaranteed to be true and is the most suitable for users. At the same time, the EigenRep model requires an iterative calculation of the global reputation in the entire network for each transaction, and the system overhead is large. The CCIDTM does not consider the user’s demand preference and direct trust update problem, so the user satisfaction is low. The CSTEM in this paper considers subjective weight and objective weight in the calculation of direct trust degree, and the calculated trust degree is more accurate. The dynamic update of trust degree, the recommended trust degree, is calculated by the gray correlation method, so the user satisfaction is the highest, CSTEM> CCIDTM> EigenRep> RSS.
5.2. Success Rate of Cloud Services Selections
RSS only randomly selects a service that satisfies the basic functions of the user. It does not perform feedback after the user uses the service, and the satisfaction is the lowest. The EigenRep trust model is a global reputation calculation model. There is no reputation penalty correction for node malicious behaviour. Even if the service with the best reputation is selected, the service is not guaranteed to be true and is the most suitable for users. At the same time, the EigenRep model requires an iterative calculation of the global reputation in the entire network for each transaction, and the system overhead is large. As shown in Figure 6, the success rate of CSTEM and CCIDTM is higher than EigenRep model. The CCIDTM does not consider the user’s demand preference and direct trust update problem, so the user satisfaction is low. The EigenRep trust model is a global reputation calculation model. There is no reputation penalty correction for node malicious behaviour. Even if the service with the best reputation is selected, the service is not guaranteed to be true and is the most suitable for users. At the same time, the EigenRep model requires an iterative calculation of the global reputation in the entire network for each transaction, and the system overhead is large. The CSTEM considers the time decay of direct trust, using feedback density factor to reduce the impact of malicious trust feedback, and dynamic trust updating mechanism, so the interactive success rate is highest. The interactive success ranking is CSTEM>CCIDTM>EigenRep>RSS.
In order to validate the model, there are malicious cloud service entities ranging from 10% to 40% evaluated. Different proportions and success rates of malicious cloud service entities are shown in Figure 7, and CSTEM is more better than CCIDTM, EignRep, and RSS. The success rates of CSTEM and CCIDTM is higher than that of EigenRep model. RSS only randomly selects a service that satisfies the basic functions of the user. It does not perform feedback after the user uses the service, and the satisfaction is the lowest. The CSTEM considers the time decay of direct trust, using feedback density factor to reduce the impact of malicious trust feedback, and dynamic trust updating mechanism, so the success rate of interaction is higher than the other three methods, CSTEM > CCIDTM >EigenRep> RSS.
(a) Malicious cloud services entity proportion (10%)
(b) Malicious cloud services entity proportion (20%)
(c) Malicious cloud services entity proportion (30%)
(d) Malicious cloud services entity proportion (40%)
6. Conclusion
The CSTEM is based on combining weights and gray correlation analysis. In order to improve user satisfaction and interaction success rate, direct trust, recommendation trust, and reputation together form a comprehensive trust, resulting in a more accurate overall trust. The direct trust also considers transaction time and transaction amount, the objective weight is calculated by rough set theory, and the subjective weight is calculated using AHP, so the final direct trust is very accurate and effective. A gray relational analysis method is used to calculate the degree of similarity recommendation trust, make the recommendation trust become more reasonable, combine with the service provider’s own reputation, and then get the comprehensive trust. This paper proposes a dynamic direct trust update mechanism. Finally, simulation experiments show that the CSTEM performs better than CCIDTM, EignRep, and RSS.
This paper is subject to the following limitations:
(1) The CSTEM could be further improved by considering more trust dynamic update factors of the cloud services evaluation. Considering the trust dynamic update factor of cloud service evaluation, cloud service providers can classify and evaluate and select cloud service trustworthiness in specific environments such as mobile environment and heterogeneous environment.
(2) In order to improve interaction success rate, the future cloud service selection performance, and security, we can further expand the scale of the experiment, build the actual cloud prototype system, and deploy this strategy to verify and improve the model which is the main work of the future.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work was supported by the National Science Foundation of China under Grant No. 61672117 and No. 61602070 and the Medical Research Project of Chongqing Health and Family Planning Commission of China (No. 20142124).