Trustworthy Service Discovery for Dynamic Web Service Composition

Abstract

As the number of services available on the Web increases, it is vital to be able to identify which services can be trusted. Since there can be an extremely large number of potential services that offer similar functionality, it is challenging to select the right ones. Service requestors have to decide which services closelysatisfy their needs, and theymust worry about the reliability of the service provider. Although an individual service can be trusted, a composed service is not guaranteed to be trustworthy. In this paper, we present a trust model that supports service discovery and composition based on trustworthiness. We define a method to evaluate trust in order to discover trustworthy services. We also provide a method to perform trust estimation for dynamic service composition, and we present results of two experiments. The proposed model allows for service requestors to obtain the most trustworthy services possible. Our mechanism uses direct and indirect user experience to discover the trustworthiness of the services and service providers. Moreover, composing services based on quantitative trust measurements will allow for consumers to acquire a highly reliable service that meet their quality and functional requirements.

1. Introduction

A service-oriented architecture (SOA) is a very popular architecture paradigm that can be used to design and develop modern distributed systems such as cloud computing. SOA is also increasingly used in enterprise information systems, especially in the form of Web services, because it can help realize business goals that require easy, flexible integration with legacy systems, streamlined business processes, reduced costs, innovative service to customers, and agile adaptation and reaction to opportunities and competitive threats[1].In order to realize the potential of service oriented computing (SOC), an SOA should bedesigned to overcome the many possible challenges encountered in the enterprise, including the need for complex distributedservices, managing business processes, ensuring transaction Quality of Service (QoS), complyingwith agreements, and leveraging different computing devices such as personal computersand cell phones [2].

In order to build a service-oriented application, an application developer, or service requestor, canselect services from a variety of providers on the Web. Since there is an extremely large number of potential services with similar functionality, service requestors need to differentiate between them. The only differentiatingfactor between similar services may be their non-functional properties, which can beconsidered to be part of the criteria for service selection.In this context, it is challenge to properly select the right services. Service requestors have to decide which services satisfy to their needs, and they have to worry about the trustworthiness of the service provider. Trust is a non-functional property that can be usedas a criterion for service selection. A service requestor,or trustor, may select a service from a service provider, the trustees, based on trustworthiness.Thus, trust can help requestors make their service selection decision.

Trust is a vast, subjective field that has been defined in the context of e-commerce, Web services, and peer to peer (P2P) networks. For SOA and Web services, Chang et al. have defined trust as “the belief the trusting agent has in the trustedagent’s willingness and capability to deliver a mutuallyagreed service in a given context and in a given time slot”[3].Trust management is of paramount importance to reducethe risk involved in transactions between services andservice consumers.The challenge in implementing trust mechanisms is to find good models from which trust can be derived directly from use experience, recommendations from third parties, and social relationships. In particular,such models aim to develop mechanisms that can provide robustness against attacks suchas misleading recommendations, reentry and Sybil attacks. Therefore, an effective trust model is important to address such issues.

However, it is difficult to solve this problem using the current service discovery techniques such as Universal Description, Discovery and Integration (UDDI). Information for Web services in the UDDI registry cannot ensure that these would be reliable and usable. According to the Web services report in [4], 48% of the production UDDI registry (tModels tested only) has unstable links. Recent research on trust modeling provides us with a promising starting point to find a solution for service selection[5]. Trust is the key basis of interaction in an opensetting, indicating that there is a relationship between the parties involved. For example, in SOCenvironments, a party A may trust another party B, because Aexpects B will provide a service of a certain desired functionality and quality. Wedefine trust-aware service selection as selecting the desired services based on the trustplaced in their ability to deliver specified values of specified qualities.

Intuitively, thetrustworthiness of a service should be estimated according to both direct and indirectexperience. Direct experience indicates the service had been previously received with a certain quality, whereas indirect experience indicates that derived from referrals by peers. It is not straightforward to estimating trust from direct experience with a servicebecause some services may not directly expose details of their composition to theirconsumers. Although individual services can be trusted, composed servicesare not guaranteed to be trustworthy. A consumer may interact with a composite service without knowingmuch about the qualities of the services that underlie it, and in such a case, evaluatingthe trustworthiness of the service is nontrivial. In order to select a service, servicecompositions should be modeled with respect to how the qualityof a component service can affect the whole composition.In this case, we adopt the concept of a social relationship network for e-commerce. According to the e-commerce social relationship network, the model from the buyer’s point of view calculates the trust value list of every seller. For Web services, every consumer has many services available, and every consumer is willing to refer to the reputation reported by other consumers, which is the theoretical basis of our approach.

In this paper, we first construct a model from the direct experience of a service consumer. The model represents attributes of quality that are observable from the consumer's point of view. Then we present an approach that models service compositions via graph theory in terms of partially observable factors. The approach captures the dependency for providing a good quality of service between elementary and composite services. It shows how a composite services quality depends upon the quality of its elements. Those two models are combined to provide a quality distribution for the services, including information on how much each elementary service contributes to the composition. The rest of the paper is organized as follows: Section 2 briefly describes some previous work with respect to the trust computing model, and Section 3 defines the concept of trust and formalizes the problem from point of view of the service consumer. Section 4 specifically introduces a method to evaluate trust derived from each service, and it also describes our approach to estimate the trust value for a composite service. Section 5 gives experimental results. Finally, Section 6 summarizes and proposes the work for the next phase.

 

2. Related Work

Fig. 1 presents different trust-based service selection approaches[6]. In terms of the direct experienceapproach, requestors build a trust profile for services after utilizing them. However, since there is aneed to trust services before they are executed, a Trusted Third Party (TTP) approach enablesconsumers to consult a trusted third party to determine the trustworthiness of the services. For the indirect experience approach, reputation is apublic opinion of the characteristics of an entity, and it represents the collective evaluation ofthat entity. Recommendation systems aggregate recommendations and matchrecommenders with others searching for recommendations.Finally, a referral approach is a decentralized approach based on software agents and communities, where agents can assisteach other in finding services by giving referrals to services that were useful for them. Therefore,each agent builds the trust of other agents according to the perceived quality of the services and thereferrals that guide them to the services. In a matchmaking TTP approach, a servicedescription is matched with a requestor’s request and with the trust preferences. The hybrid approachinvolves a combination of different approaches that aim to improve the weaknessesof some of the approaches by combining them with other approaches. The automated trust negotiationapproachbuilds mutual trust between service requestors and service providers. In thisapproach, trust is assessed in two directions: a requestor trusts a service and the service truststhe requestor. The trust negotiation approach depends on the disclosure of digital credentialsbetween the two parties.

Fig. 1.Approaches for Trust-Based Service Selection

Researchers investigating trust have also attracted great attention to SOC. However, the literature on SOA trust is still immature. There are some the Web service security standards that ensure hard security mechanisms for SOA applications that have been addressed by IBM and Microsoft. These organizations cooperate to define a unified approach for managing message security exchangein a Web Services environment[2].Although Web services security technologies, such as WS-Security andWS-Trust, increase security in Web Services,they cannotdetermine the entities that have hidden motives or other trust issues, such as detecting unfair ratings,determining the level of trustworthiness of entities, rewarding positive behavior, punishingmalicious behavior, and reputation bias [7].

In [8], Wu et al. model a consumer’s assessment of the quality of a service using a naiveBayesian network. They apply a fuzzy representation to express the levels of the capabilities of the service. Theirapproach enables consumers to estimate the overall quality. Lin et al. select services according to a consensus of the group preferences in order of various qualities[9]. They use fuzzy logic to resolve the conflicts betweenthe subjective interpretations of the service qualities from each consumer. In [10], Yue et al. apply Bayesian networks to model the relationships between elementaryservices, and their approach constructs web service Bayesian networks(WSBN) based on the invocations between the services. In [11], Paradesi et al.build a trust framework for web service compositions. Theyadopt the trust representation from[12]and introduce operators to combine trust in different types of service compositions, including sequence,concurrent, conditional, and loops. Maximilien and Singh develop a trust-aware approach that selects servicesbased on a well-defined ontology which provides a basis for describing consumers’requirements and providers’ advertisements[13]. The ontology enables consumers todefine nonfunctional properties, but this approachdoes not take into account service composition. Wang and Singh developed a trust model for multi-agent systems that formalizeshow agents map evidence to trust and vice versa[12]. In [14], Artz and Gil compare several definitions of trust in different research areas in the field of computer science. In particular, the authors discuss the relevance of trust and the semantic Web and point outsome unique trust management challenges in that area. Fernandez-Gagoet al.perform a trust management survey focusing onwireless sensor networks[15], and their survey is an overview of existing trust management solutionsfor ad-hoc and the peer-to-peer (P2P) wireless sensor networks.A few surveys focus on reputation-based trust management systems. For instance, in [16], Marti and Garcia-Molina exploit a taxonomy technique to classify different reputation-based trust management systems. In [17], Jøsang et al. discuss general ideas of trust (e.g.,trust classes and trust purpose), explain the overlapping notions between trustand reputation terms, and compare a few trust models. In [18], Silaghi etal. investigate whether existing trust management approachescan be applied to Grid environments, and a few guidelines are given in the surveythat may be useful to future research and the development of trust managementsystems in Grids. In [19], Wang and Vassilevapresent a systematic review of several trust and reputation systems in which they classify these systemsinto three categories including centralized versus decentralized, persons/agents versus resources, and global versus personalized.

A novel service composition algorithm is proposed in [20]. The algorithm models the service composition as multi-domainscheduling problem with minimal service resources and timeconstraint. They define each service as an exclusive resourceduring its execution period. By computing the inter-domaincommunications and available services in each domain, thedomain with optimal utilization rate is obtained to arrangeservices. In [21], a framework for a TrustbasedDynamic Web service Composition is presented. The framework not only usesfunctional and non-functional attributes provided by the Webservice description document but also filters and ranks solutionsbased on their trust rating. They define a method to calculatea trust rating per service using Centrality measure of SocialNetworks. In [22], Wang et al. formulate the problem of service composition andservice binding as a multi-objective optimization (MOO)problem, minimizing the service cost, while maximizing thequality of service (QoS) and quality of information (QoI). They develop a trust-basedservice composition and binding protocol, and demonstrate thatthe trust-based scheme outperforms the counterpart non-trustbasedscheme. Wu et al. propose a trust based service composition and optimization method in [23]. They define the trust of service composition in terms of the trust of component service selection processes, composition processes and optimal binding plans.They use filtration, interface based service clustering, and trust evaluation to guarantee the trustworthiness of service selection and composition.

Trust management is one of the most important issues in the area of informationsecurity, and several surveys have been conducted in this regard.Our work presents howto evaluate the trustworthiness of services that are elementary components of a composite service.We provide a systematic way to represent qualities of service compositions viaoperators that correspond to typical ways of composing services. Our contributionis to provide a method that can handle service selection according to trustworthiness in the context of service compositions.

 

3. Problem Formulation

3.1 The Concept of Trust

Trust is widely accepted as a major component of human social relationships. In general, trust is a measure of confidence that an entity will behave in an expected manner, despite the lack of ability to monitor or control the environment in which it operates. OASIS (Organization for the Advancement of Structured Information Standards) defines trust as “the characteristic that one entity is willing to rely upon asecond entity to execute a set of actions and/or to make assertions about a set of subjects and/or scopes”[24].

However, it is difficult to objectively judge whether or not a particular entity is trustworthy. Therefore, the concept of trust is incorporated into a trust management system that establishes monitors and adjusts trust relationships among the participants. The core concept of trust management is represented by a trust model that defines primary factors of trust relationships and describes how to calculate the resulting trust values.

To establish a trust model for SOA, we consider trust as the belief that one agent has in another agent’s capability to deliver a quality of service in a given context and in a given time slot.We assume that an agent can be an abstract entity in a distributed network and can represent an ordinary service user, a service, or a service provider. Trust is formalized by the concept of a trust relationship between agent aiand aj from the set of agents A={a1, a2, .., ai, .., aj, .., an}, and it is expressed as a binary relation R⊆A×A. The trust relation aiRaj represents a directed link between agent aiand aj in a directed graph which is called a trust graph (TG) denoted by an ordered pair TG =(A, R).

We formulate the knowledge as trustworthiness in a specific time and context. That is, the knowledge is quantified according to past experience during aspecific time and context. To establish a trust model, we use the time and context dimensions. The time dimension T is a set of time values ti when the interaction between the agents took place, and the context dimension C contains a set of propositions that can be formulated as a sequence of possible events.

A trust value is formulated when an element r∈R has a degree of trust from a domain set of possible trust values D. When ri,j =aiRaj for aiand ajinA, the trust value for ri,j is defined as the trust degree of an agent ai toward agent aj in time t and in context c, and it is represented by τi,j(t, c)∈D where t∈T and c∈C. We will then describe τi,j(t, c) using direct and indirect measurements in the next section. Fig. 2 shows an example of a trust graph in some time and context. On the trustgraph, the directed links between agents have weight valuesτi,jfrom the domain set D.

Fig. 2.Trust graph during some time and in some context

Trust has several main characteristics that remain trueregardless how trust is defined, as discussed below:

3.2Web ServicesModel

In SOC environments, services from different providers might have to be combined to establish a complete service. For example, consider the composition of three services (S1op1S2) op2S3 as shown in Fig. 3, where s are services and s are arbitrary composition operators. Many business domains require composing existing services in order to deliver new functionality. Service composition therefore extends the notion of service discovery by enabling automatic composition of services to meet the requirements of a given a high level task description.

Fig. 3.An example of a composition

In general, service compositionis defined by an open, standards-based approach for connecting services together to create higher-level business processes. Standards are designed to reduce the complexity required to compose services, hence reducing time and costs, and increasing overall efficiency in businesses. Standards for Web services, such as BPEL4WS (Business Process Executable Language for Web Services), enable creating of compositions of Web services as well as defining business processes as coordinated sets of Web service interactions.

When the trust level of each service is known (e.g. untrusted, partially trusted, or fully trusted), what is the trust level of combined service? Suppose that services S1 and S2 are only partially trusted, but that service S3 is fully trusted. Evidently, the service consumer cannot fully trust the composed service. Although the individual service can be trusted, the composed service is not guaranteed to be trustworthy. Thus we need to find a way to determine whether the composed service can be trusted [27]. To select QoS guaranteed services or credible service providers for composite one as well as individual one, we define a trust model to evaluate the trust level of individual service and deduce the trust level of composed service.

BPEL4WS describes the control logic for Web services coordination in a business process. For example, sequential control between activities is provided by , , , , and the serial variant of . In BPEL4WS, there are four different structures to describe composite services: sequential flow, conditional branching, looping, and parallel processing [28]. To construct a composite service, we consider the execution flow of four structures.

3.3Trust Model

Traditional approaches model trust qualitatively, based on an intuition of hard security. If one cannot definitely determine that a particular party has the stated identity, then thatis sufficient reason not to deal with it at all. Yet, in many cases, requiring an all-or-nothing decision about trust can be too much to ask for, especially when we think not of identitybut more broadly of whether a given party would support one’s plans.

This paper develops a well-formulated mathematical system with which to approach trust. Intuitively, thetrustworthiness of a service should be estimated based on both direct and indirectexperience.Direct experience means that there is a previous quality of service that has been received fromthe service, whereas indirect experience comes from referrals by peers. Thus we model trust using Equation (1). The trust value τi,j(t, c)represents the trust degree of an agent ai toward agent aj in time t and context c.

In Equation (1), Ei,j(t0, tt) represents the experience factor that indicates the knowledge of agent ai toward agent ajthat was acquired through direct interaction between ai and aj within the time interval [t0, tt]. We assume that as the number of interactions increases, the knowledge of agent ai toward agent aj increases. However, as time lapses, the knowledge level converges to some point. This paper thereforedefines the experience factor as a logarithmic growth curve:

In Equation (2),Ni,j is the total number of interactions between agents ai and aj, and e is a constant approximately equal to 2.7183 as a base of the natural logarithm. Exponential decay models of this form can model learning curves. In the exponential decay model shown in Fig. 4, the experience factor grows rapidly in the beginning, and then levels off to become asymptotic to an upper limit.

Fig. 4A learning curve representing the experience factor

In Equation (1), QoSi,j(t, c) and RFi,j(t, c) are as follows:

 

4.Trust Evaluation

4.1 Trust measurement

To evaluate the level of trust, we first assume that the users of the service can assess trustworthiness of the service providers for every transaction. That is, before agent ai makes interaction with agent aj, s trustworthiness towards aj has to be assessed according to the feedback on the history from the trust system. In order to evaluate and assign the trust level to each service and service provider to select quality-guaranteed services or credible service providers, we consider the concept of social experience formed by direct and indirect participation of users within their own societies. Trust level is considered with direct and indirect experiences. Thus we defineconfidence from direct experience and the reputation from indirect experience for the service and the service provider.

From Equation (1), for the sake of convenience, the two variables t and c are omitted, as in Equation (3). The trust mediator determines the trust level by using the Equation (3) according to historical data of past interactions. To represent the reputation, we use feedback indicating the level of the customer’s satisfaction after use. We assume that the ratings provided by the service consumers are available and are valid.

(1) Direct measure

In Equation (3), QoSi,j represents the satisfaction degree of an agent ai towards an agent aj according to QoS attributes. This is to directly acquire a satisfaction degree of the service by evaluating the non-functional aspect of each service using several QoS properties.

We choose the quality factors of the service level quality measurements defined by the OASIS standard as QoS attributes: response time, maximum throughput, and availability. Since a service could be provided by third parties and could be invoked dynamically via a network, the service performance might vary depending on the network speed or the number of users that are connected at a given time. Service level quality measurements are a set of quantitative attributes that describe the runtime service responsiveness from the point of view of consumers. This quality factor represents how quickly and soundly Web services can respond and can be measured numerically on a system[29].

Moreover, we also consider the price as another factor. The price of a Web service can be determined by a service provider. A service consumer considers the price of a Web service according to the functions, contents, and the quality of the web service in order to make decisions on whether he uses the web service. Thus the QoS value on each quality attribute defined by [29] is as follows:

LetQP,QR,QT and QA stand for price, response time, throughput, and availability, respectively. Only three attributes are considered in this paper, but more attributes can be added depending on the context. For each QoS attribute qx in Q = {QP, QR, QT, QA},a weight is assigned according to its importance in such a way that the following conditionholds: for eachqx in Q (1≤x≤4), qx has the weight wx where 0 ≤wx≤ 1 and∑wx =1.

In order to have a uniform measure independent of units, we normalize the QoS values using the method defined in [30]. Liu et al. define a two-pass normalization method for which we adopt only the first normalization step to distribute the computed QoS value over the closed interval [0, 1]. Using this normalization, the QoS value provides a uniform index to represent the servicequalities for each provider. The provider can increase and decreasehis/her quality index by entering a few parameters. Moreover, they can set a threshold regarding those qualities. To compute the QoS value using the above fourcriteriaQ = {QP, QR, QT, QA}, we can obtain the following matrix QM.Assuming thatS = {s1, s2, …, sn} is a set of Web services that have the same functional properties. Each row inQMrepresents a Web service sy, while each column representsone of the QoS criteriaqx in Q.

To normalize the matrix, we need to define two vectors. The first vector is N = {n1, n2,n3, n4}where the value of nj(1≤ j≤ 4) can be 0 or 1. nj= 1is used in the case where the increase inqibenefits theservice requester while nj= 0 is used for the case wherethe decrease of qibenefits the service requester. Thesecond vector is C = {c1, c2, c3, c4}. Here cj(1≤ j≤ 4)is a constantwhich sets the maximum normalized value. Eachelement in vector Vwill be normalized using Equation (5) and (6).

Applying these twoequations to QM, we get matrix QM´which is shown below:

Then we can finally compute QoSi,j using Equation (8). Through the interaction between agent ai and aj, each QoS attribute is measured according to historical data of past interactions, such as logging data, actual execution time, and the number of response times to the user’s requests.

, where vi,k is the element of QM’ and wk is a weight for the quality value vi,k.

(2) Indirect measure

In real situations, people evaluate others based on local views of the world. To form an opinion about an unknown person, we turn to people we know and ask about the unknown person’s reputation. In most cases, we do not rely on some centralized authority to determine the reputation each person deserves. One of the fundamental characteristics of indirect measurements is that it is a purely local analysis. To indirectly measure the trustworthiness of a service, we modify the reputation inference algorithm defined in [25]. In this paper, as shown in Fig. 6, our method begins with A’s adjacent nodes and expands out to infer a reputation rating of node B or B’ by starting at node A.

Fig. 3.Indirect measure between two nodes with no interactions

Table 1 showsthe algorithm used to compute RFi,j in Equation (3). In the following algorithm, the sink is the node for which a rating is desired and the source is the node for which the rating will be made. That is, the source is the agent ai and the sink is the agent aj for RFi,j. In this metric, the source polls each of the neighbors to which it has given a trust rating. Each of the source’s adjacent nodes will return their rating for the sink. The source will then average these ratings. The mean value is the inferred trust rating from source to sink, and the function Rating(a, b) gets a trust value QoSa,b from direct measures between agents a and b.

Table 1.The computation algorithm forRFi, j

Each interaction between agents has to accumulate information with respect to the sequence of eventsthat has occurred. In addition, every agent should have its own trust policy to check for compliance with trust-related requirements. Declarative policy languages, such as WS-Policy, are used to express the capabilities and requirements of the entities in a Web-service-based system.

4.2 Trust estimation for composing services

Web service composition is a compilation of several services aggregated to execute in sequence. A complex composition problem with a parallel implementation, branching, and loops is usually an NP-hard problem. To select a service, we have a set of candidate services. Depending on the user need, non-functional aspects can be selected.

In [5], Hang and Singh give some example on how some quality metrics are composed in the type of interactions. They propose quality metrics such as SWITCH, MAX, MIN, SUM and PRODUCT. For example SUMyields the composite quality value as the sum of the quality values obtained from all constituent services. However, the highest total sum of the quality value cannot guarantee that the composite service is the most trustworthy because the summation depends on the number of constituent services.

The trustworthiness of a composite service is computed via propagation of the trust values of atomic services. Here three strategies can be utilized according to Kuter and Golbeck [31]: Overly-Cautious, Overly-Optimistic and Average. All these strategies aim to find a composition with the highest trust value. The goal of the first strategy is to maximize the minimum expected trust value that the composer user has in the atomic services of the composite service. In other words, it assumes that if something bad could happen, it would definitely happen, and thus it avoids incorporating low-trustservices. In contrast to the first strategy, an Overly Optimistic strategy promotes the influence of highly trusted atomic services into trust of the composite service because it believes that nothing bad happens if low-trust services are taken into account for composition. The last strategy is an intermediate approach that looks for compositions with a maximum average trust.

We adopt anOverly-Cautious strategy for defining the trust in a service S that aims to maximize the minimum expected trust value that the composer user has inthe sub-processes of S. Consider composition operators that integrate services to build a business process. As described in Section 3.3, there are four structures that can be used to compose services: sequential flow, looping, conditional branching, and parallel processing.Let OP be the set of composition operators for describing four structures. Let Sbe a composite service and {s1, s2, …, sn} be a set of available atomic services to construct the composite service S. We denote S = {s1, s2, …, sn}. Each atomic service si in S has the trust value t(si) of si computed by the average of the trust values on the incoming edges into node si.

In Equation (9), τk,i is the trust value of agent ak toward agent si by Equation (3), where ak is an adjacent node of si.Then S is defined by business process models and atomic services are executed on the business process model. Let BP(S) = {bp1, bp2, …, bpm} be a set of business process models used to construct the composite service S. Each bpj is defined by a tuple bpj = (OP’, S’) where OP’⊆OP and S’⊆S. In formal, a trust value t(S) for S is denoted as in the following Equation (12):

For example, consider the process model BP, as shown in Fig. 6, to build a composite service S = {s1, s2, …, s7} based on business processes BP = {bp1, bp2, bp3}. For each service si in S, t(si) is the trust value of si. Assume that (t(s1), t(s2), …, t(s7)) = (0.78, 0.25, 0.34, 0.36, 0.52, 0.69, 0.83). Then Q(bp1) = min(0.78, 0.25, 0.36, 0.52, 0.69) = 0.25, Q(bp2) = min(0.34, 0.36, 0.52, 0.69, 0.83) = 0.34, and Q(bp3) = min(0.78, 0.36, 0.52, 0.69, 0.83) = 0.36. The final trust value of the composite service S is 0.36 of max(bp1, bp2, bp3).

Fig. 4.An example of the service composition

 

5. Experimental evaluation

We performed experiments to evaluate the models presented in this paper. First of all, Fig. 7 shows the framework used to implement the experiments. Ourmodel consists of the trust manager, composition engine, and QoS analyzer.Once a service provider enrolls service descriptions in the UDDI service registry, the trust manager constructs trust information of the service and the service providers. Each time the serviceis evaluated through service monitoring, the trust manager updatestrust information of the service. After assigning the service’s trust information, the trust manager modifies thepast service profile with respect to the trust level. The QoS analyzer should have event structures that maintaintransactions during specific time-intervals. Changes to the trust level of the services in agroup have no effect on other service groups. This manages trust information efficiently.To implement the trust model, we need away to describe the QoS attributesof a service. The QoS of the service can be attached to the ServiceProfile in OWL-S, a new class as a subclass of the ServiceParameter class already defined in OWL-S.Table 2 is an example of QoS requirements described in OWL-S.

Fig. 5.Trust evaluation model prototype

Table 2.OWL-S description of service requirements

We have conducted two experiments running on a framework implemented on jUDDIv3 and MySql. We collect 120 services from e-commerce systems. The services are classified into three logical groups providing similar functionalities with different quality levels. Registered serviceshave functional capabilitiesand QoS information.

The goal of the first experiment is to verify the fact that the proposed trust evaluation methodprovides the optimal trusted service that fulfills the user’s requirements. Each service is invoked 50 times with different QoS requirements, and at the end of eachinteraction, the system updates the corresponding trust graph. To show the effectiveness of our approach, we compare the proposed method with the averageQoS values in all invocations.We compare the servicehaving the highest value of trust level using Equation (3) with the selected servicebased on QoS information.

For example of computation of the trust value, Table 3 shows the values of each quality attribute for arbitrary five services satisfying the requirements in Table 2 among the collected 120 services.The experiment executes on four quality attributes: price (10,000 won per service), response time (millisecond), throughput (the number of requests completed in second), and availability (%). The five services are in the candidates. As a result of the trust evaluation using Equation (3), the service having the highest trust level is returned.

Table 3.The value of quality attributes

For each service and its adjacent nodes, the computation of the direct measure is the first. Note that the trust level of each service S1, S2, S3, S4, and S5 is evaluated by the adjacent nodes. If a service S1 is adjacent with S7, S12, and S20, then the experience factor E7,1, E12,1, and E20,1 and the direct measureQoS7,1, QoS12,1, and QoS20,1 are computed. For example, when S1 and S7 are interacted with each other in 5 times, E7,1 = 1-e-2.5. For the convenience of explanation, if S={S1, S2, S3, S4} is a set of adjacent nodes to S7, QM is defined as the following and is normalized in [0, 1] according to Equation (5) and (6) as the following QM’. Then QoSi,j is computed by weighted summation of elements in the j-th row of QM’. That is, QoS7,1 = (1×0.25)+ (1×0.25)+ (1×0.25)+ (1×0.25) = 1 where the weight is even.

For the services S2, S3, S4, and S5, the computation of the direct measure is computed similarly. If two nodes S1 and S7 are not adjacent with each other,we have to compute the indirect measure. The algorithm in Table 1 returns the value of RF7,1 as the indirect measure. Then we get the trust values τ7,1, τ12,1, and τ20,1.Consequently, the trust level of the service S1 averages those trust values. Table 4 shows the trust level for the services in Table 3.

Table 4.P: Price, RT: Response Time, T: Throughput, A: Availability

Assume that the services that have an appropriate quality level for the QoSattributes are selected first. Then some of them are chosen to evaluate the reputationbecause the reputation is the only factor for selecting services without trustworthiness.Uponthe submission of a user’s requests, the service candidates that satisfy functional and qualityattributes for the user’s requirements are selected. Among the service candidates, only theservice with the highest trust level is returned.In Table 4, the service S5 remains the most trustworthy service independent with the weight of quality attributes. Thus the proposed method returns always the service S5 as a trustworthy service fulfilling the user’s requirements. However the experimental results do not come to the conclusion that the attribute weight is independent with the trust level. It is needed to execute further refined experiments with several weight variations.

Fig. 8 shows the results of our first experiment where the ratio of the right selection of two approaches ismuch the same as when the number of times of the invocation is less than 100. Since our trustmodel is based on a direct interaction, the proposed model is helpful whenexperience data is sufficient after a certain amount of time has lapsed. The right selection ration on average of our approach is 83.63%, and that of the average based approach is 71.58%.

Fig. 6.The correct selection ratio

In the second experiment, the composite service “Purchase Order” is chosen as an experimental scenario. It is based on the “Purchase Order Process” example taken from the BPEL 2.0 specification as shown in Fig. 9. The WSDL service descriptions of the example are used to define the composite service “Purchase Order”[32]. To compose the service, we implement forty candidate services with differentiated QoS values. The goal of this experiment is to show whether the proposed trust evaluation method also helps to estimate the trust of the composite service based on the user’s QoS requirements.

Fig. 7.The Purchase Order Process Outline [32]

To show the efficacy of our approach, we compare the proposed methodagainst arandom walk in all compositions. To construct a composite service, our approach selects constituentservices using a trust based min-max strategy described in the Section 4.2 while the random walk selects constituentservices arbitrarily from candidate services. We measure the success rate, which is the ratio ofthe operation sequences that are successfully executedby the composite service. In our experiments, to measure the success rates, we generated 10 business process models as defined by different candidate services. For each business process model, all possible operation sequences are generated by composition operators.

Fig. 10 shows the results of our second experiment, indicating the success rates of the execution of composite services. In Fig. 10, the horizontal axis represents 10 business process models with three trials. For thirty total operation sequences, the success rate on average of our approach is 80.1%, and that of the random walk is 59.1%. Our approach can provide trustworthy composite services stably by estimating the trustworthiness of the service as compared against a random approach. Since our trust evaluation of a composite service is based on the average of the trust values acquired from direct and indirect experiences, it may guarantee that we select more trustworthy elementary services to construct a composite service. Consequently, the composite service consisting of trustworthy elementary services can be trusted.

Fig. 8.The success rate of the composition

 

6. Conclusion

Recently, SOA has emerged as a very popular architecture paradigm that can be used to design and develop modern distributed systems, such as cloud computing systems. In distributed, heterogeneous Web services environments, the potential number of services that provide similar functionalities in enterprise softwarecan be extremely large. To provide consumers with useful services,the environment in which services exist should be steady, and all the services in it shouldbe trustworthy. Otherwise, consumers may find that, although the descriptions of servicesmatch their requirements perfectly, the services they select may not be the most trusted. In critical fields, such as medical or financial systems, poorly-informed decisions can have unacceptable negative consequences.

To create more complex, value-added, cross-organizational business processes, single services are combined into composite services. In this case, although an individual service can be trusted, the composed service is not guaranteed to be trustworthy. A consumer may interact with a composite service without knowingmuch about the qualities of the underlying services. In such a case, evaluatingthe trustworthiness of a service is nontrivial. Thus, the service selection should consider service compositions to model how the qualityof a component service can affect the whole composition.Thus a way to efficiently select and compose trusted services is needed.By facilitating QoS-driven discovery and composition, users can easily and efficientlyselect trustworthy services that are most suitable to their needs.

In this paper, we presenta trust model that supports discovery and composition of trustworthy servicesbased on QoS properties. The model includes an evaluation method based on quality attributes observable from direct and indirect experiences of a service consumer and the estimation method to provide the trust value of the composite service. The model prototype is implemented on a QoS-brokerbasedarchitecture for Web services, and we have conducted two experiments running on the framework. The goal of the first experiment is to verify the fact that the proposed trust evaluation methodprovides the optimal trusted service that can fulfill the user’s requirements.The result for the ratio of the right selection of two approaches ismuch the samethe number of times of invocation is less than 100. The experiment shows that the proposed model is helpful when there is enough data that has been collected as time lapsed because our trustmodel is based on direct interactions. The goal of the second experiment is to show whether the proposed trust evaluation method helps to also estimate the trust of the composite service based on the user’s QoS requirements. The result is that the success rate of our approach is 80.1%, and that of a random walk is 59.1%, on the average, for thirty operation sequences in total. Our approach can provide a trustworthy composite service stably by estimating the trustworthiness of the service as compared against the random approach.

Existing research on trust leaves open significant research andtechnological issues which, from a software service perspective, are related to: (1) theassessment of trust in a dynamic deployment and the composition of software services, (2) theevaluation of the accuracy and the risk of trust assessments, (3) acquisition of trust informationdue to deviations of services from normal behaviorin a wide range of contexts, (4) theevaluation of the context relevance for trust assessments, (5) the bootstrapping of trustassessment environments, and (6) theprovision of an interoperable, robust and secureruntime platform that can realize independent trust certification services. Our researchwill therefore be further refined to resolve these issues.Lastly, QoS based service composition will be adapted in the real world safe and sound providedthat QoS integrity is preserved at all service providers, and the OWL-S execution platform that supports the dynamic service grounding is deserved to bedeveloped. These problems remain as future research topics.