Towards a Machine Learning-driven Trust Evaluation Model for Social Internet of Things: A Time-aware Approach
Subhash Sagar, Adnan Mahmood, Quan Z. Sheng, Munazza Zaib, Wei Emma Zhang
TTowards a Machine Learning-driven Trust Evaluation Model forSocial Internet of Things: A Time-aware Approach
Subhash Sagar, Adnan Mahmood, Quan Z.Sheng, and Munazza Zaib
Department of Computing, Macquarie UniversitySydney, NSW 2109, Australia
Wei Emma Zhang
School of Computer Science, The University of AdelaideAdelaide, SA 5005, Australia
ABSTRACT
The emerging paradigm of the Social Internet of Things (SIoT) hastransformed the traditional notion of the Internet of Things (IoT)into a social network of billions of interconnected smart objectsby integrating social networking facets into the same. In SIoT, ob-jects can establish social relationships in an autonomous mannerand interact with the other objects in the network based on theirsocial behaviour. A fundamental problem that needs attention isestablishing of these relationships in a reliable and trusted way, i.e.,establishing trustworthy relationships and building trust amongstobjects. In addition, it is also indispensable to ascertain and predictan object’s behaviour in the SIoT network over a period of time. Ac-cordingly, in this paper, we have proposed an efficient time-awaremachine learning-driven trust evaluation model to address this par-ticular issue. The envisaged model deliberates social relationshipsin terms of friendship and community-interest, and further takesinto consideration the working relationships and cooperativeness(object-object interactions) as trust parameters to quantify the trust-worthiness of an object. Subsequently, in contrast to the traditionalweighted sum heuristics, a machine learning-driven aggregationscheme is delineated to synthesize these trust parameters to ascer-tain a single trust score. The experimental results demonstrate thatthe proposed model can efficiently segregates the trustworthy anduntrustworthy objects within a network, and further provides theinsight on how the trust of an object varies with time along withdepicting the effect of each trust parameter on a trust score.
CCS CONCEPTS • Security, Privacy, and Trust → Trust and Network Security ; •
Se-curity and Privacy → Intrusion/Anomaly Detection and MalwareMitigation . KEYWORDS
Machine Learning, Social Internet of Things, Trustworthiness Man-agement, Social Similarity, Cooperativeness, Friendship, Community-of-Interest
Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for components of this work owned by others than theauthor(s) must be honored. Abstracting with credit is permitted. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from [email protected].
Conference’17, July 2017, Washington, DC, USA © 2020 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM Reference Format:
Subhash Sagar, Adnan Mahmood, Quan Z. Sheng, and Munazza Zaib and WeiEmma Zhang. 2020. Towards a Machine Learning-driven Trust EvaluationModel for Social Internet of Things: A Time-aware Approach. In . ACM,New York, NY, USA, 8 pages.
Rapid advancements in communication and computing technolo-gies have let to the evolution of billions of smart objects (e.g., smartmachines, smartwatch, and smart cars) equipped with sensing,processing, communication capabilities that not only enables anobject-object communication across the internet but also formsthe well-known paradigm of the Internet of Things (IoT) [1] [2].This paradigm has undeniably produced enormous business, andopened doors to many applications and services in various valuablesectors [3], and led to a tremendous increase in connected smartobjects that are expected to surpass 75.44 billion by 2025 [4]. Fur-thermore, over the past decade, numerous research endeavors haveanalyzed the possibilities of integrating the notion of social net-working into the IoT ecosystem. This integration has led to a newparadigm of the Social Internet of Things (SIoT), wherein, objectsand humans are maintained at social and physical level (the leftpart of Figure 1) in order to facilitate the owners of the objects toset the rules for protecting their privacy. Moreover, each object inSIoT is capable of establishing relationships (as illustrated in Figure2) with other objects in terms of their ownership (relation withthe objects belonging to the same owner), social relationship (rela-tion with the objects belonging to the friends in social network),parental relationship (relation with the objects from the same man-ufacturer), co-location and co-work relationships (relation with theobjects from the same location and work environment), and enablesautonomous inter-object interaction based on these relationships[5][6].Trust plays an important role in establishing and maintainingsocial relationships since social objects only involve themselves in arelationship when the participating objects are trustworthy enoughto reduce the probable risk in decision-making [5]. Trustworthyrelationships among the objects make it easy for them to onlyrespond to the service requests from the familiar objects in thenetwork, thereby, reducing the exposure to malicious objects [7].The notion of trust has been studied in many disciplines, i.e.,sociology, psychology, and computer science [8][9], nevertheless,the perception of trust in each of these disciplines is different, andtherefore, there is no generally acknowledged definition of trust.As a result, it is essential to look at trust from the SIoT point ofview. In SIoT, trust is characterized as the “confidence” of a trustorin a trustee to achieve an objective under a particular setting inside a r X i v : . [ c s . CR ] F e b onference’17, July 2017, Washington, DC, USA Subhash Sagar, Adnan Mahmood, Quan Z. Sheng, and Munazza Zaib and Wei Emma Zhang Figure 1: Illustration of SIoT components and process of trust evaluation a particular timespan. Trust is a dynamic process that involvestrustor, trustee, and the underlying context. The trust computationprocess in this paper follows three steps: 1) social relationshipsrequest, 2) interactions in terms of the number of successful pack-ets transmission, and 3) trust evaluation by quantifying the trustmetrics from step 1 and step 2 and trust update. The right part ofFigure 1 illustrates the trust process by utilizing the services of alocal authority as a server, wherein, if an object 𝐴 needs to computethe trustworthiness of an object 𝐵 , it will first request the socialprofile (i.e. friends, interest groups, and working relationship) ofthe object 𝐵 from the server. Subsequently, after the interaction,object 𝐴 evaluate and update the trustworthiness of object 𝐵 in thenetwork through a local authority. Figure 2: SIoT relationships
As of recently, research into SIoT primarily focused on the defi-nition and construction of these SIoT relationships, nevertheless,the SIoT paradigm still lacks in some fundamental aspects such asunderstanding how these relationships can be used and quantified to build a reliable trustworthy system based on the behaviour of theobjects [10]. Based on these observations, in this work, we herebypropose a trust evaluation model, wherein, the trustworthiness ofan object is computed dynamically by employing social relation-ships as well as the interactions among the objects. Overall, themain contributions of this paper are summarized as follows:1) A time-aware trust evaluation model has been proposed thatascertains the trust of an object by exploiting social relation-ships (i.e., friendship and community-of-interest), co-workrelationship, and cooperativeness in terms of interactionsamong the objects. Moreover, both the direct trust and rep-utation (indirect trust) are employed to identify between atrustworthy and, an untrustworthy node.2) In contrast to the traditional aggregation techniques, a ma-chine learning-driven aggregation scheme has been envis-aged to aggregate all the trust metric for a single trust score;and,3) Finally, the experimental evaluation of the proposed modelgives insight on how the trustworthiness of an object varieswith respect to time along with depicting the effect of eachtrust metric while computing the trustworthiness of an ob-ject.The remainder of this paper is organized as follows. The recentwork in the literature is discussed in Section 2. Section 3 formallydefines the problem and introduces the proposed trust evaluationmodel. Section 4 provides the detailed simulation setup and the ma-chine learning-driven aggregation scheme. Results and discussionare described in Section 5. Finally, Section 6 concludes the paperand discusses the future work.
Recently, there has been an increasing interest by a number of re-searchers to model the trustworthiness management in SIoT since it owards a ML-Driven Trust Evaluation Model for SIoT: A Time-aware Approach Conference’17, July 2017, Washington, DC, USA is one of the fundamental concerns in SIoT research [11][12][13][14].A wide range of facets can be contemplated for ascertaining thetrustworthiness, including but not limited to, social relationships,reputation, quality-of-service, context (i.e., energy and time), etc.In [11], Boa et al. proposed a trust evaluation mechanism, wherein,three trust parameters are considered to evaluate the trustworthi-ness of an object. The considered parameters are cooperativeness,honesty, and community-of-interest, and to aggregate these pa-rameters, an adaptive weighted sum technique has been suggested.Moreover, their trust model considers direct trust and indirect trust(as recommendations) while quantifying the trust of an object.Anuoluwapo et al. [12] suggested a dynamic trust evaluationmodel CTRUST, wherein, objects decide the functional parametersin terms of collaborative context, to estimate the trustworthiness ofan object in the SIoT network. Besides, CTRUST includes the decayfactor to ensure the trust degradation process more efficiently and ina consistent manner. Nevertheless, weighted trust parameters havebeen equipped to aggregate the final trust. Each node decides theweights of each parameter at run-time to adjust the trustworthinessof an object in various contexts. Moreover, an adaptive trust proto-col for service-oriented architecture is proposed in [13], whereinsocial similarity and trust feedback are considered to measure thetrustworthiness of an object, and the protocol’s effectiveness isdemonstrated through service-oriented applications. However, anobject in the discussed models need to discover the best trust pa-rameters and their respective weights as per the environmentalconditions.There are numerous research works on the trust computationalmodel delineated in the literature targeting different applicationsof IoT such as vehicular networks [15][16], peer-to-peer networks[17][18], recommendation systems [19][20], and mobile crowd-sourcing [21][22]. Alnasser et al. [15] proposed a recommendation-based trust model for vehicle-to-everything (V2X) communicationwhere direct trust and recommendations from credible vehicles arecombined to obtain the trust score of a vehicle. Subsequently, adynamic adaptive weighted sum method is used to aggregate bothdirect trust and recommendations. Yuanyi et al. [19] presented atime-aware smart object recommendation system, wherein, socialrelationships in terms of SIoT knowledge graph and social simi-larity together with collaborative filtering model is amalgamatedto recommend an object in SIoT application. Likewise in [21], anexperience and reputation trust evaluation model is delineated torecruit the mobile nodes for mobile crowdsourcing. The experi-mental evaluation is performed on a real-world dataset and theeffectiveness of the same is validated by comparing the model withthe state-of-the-art schemes.Numerous studies in the literature also suggest the idea of uti-lizing various techniques such as a fuzzy logic-based model [10],machine learning-based schemes [23][24], and regression model[25] to ascertain the single trust score as the traditional weightedsum method has many drawbacks. Wang et al. [25] presented a logitregression-based model to estimate the trust which is based on theconditional probability of a credible service by a service provider.Nevertheless, this technique requires more observation to deal withdifferent recommendation-based attacks. Xin et al. [10] designeda context-aware fuzzy logic-based trust model to build a reliabletrustworthy relationship among the objects, wherein, centrality and community-interest are considered to quantify the trust of anobject. Also, fuzzy logic-based inference rules are formed to syn-thesize the selected trust parameters in order to ascertain a singletrust score. Finally, one of the most recent works related to ourresearch was carried out by Upul et al. [23]. The authors staged amachine learning-based trust framework model based on the socialprofile of a node, wherein different social features are accumulatedby exploiting machine learning-based to ascertain the direct trustmetric of any node in the IoT network. However, their approachlacks the idea of incorporating recommendations as indirect trustalong with the direct trust observation.
This section formally defines the problem focused in this paper, andsubsequently introduces the proposed trust evaluation model.
In order to quantify the trustworthiness of an object in the SIoT en-vironment, we use social information (i.e., Friends, Interest Group,Working Relationships, etc) of an object, and 𝑜𝑏 𝑗𝑒𝑐𝑡 − 𝑜𝑏 𝑗𝑒𝑐𝑡 interac-tions as our primary data source. The social information is the collec-tion of object’s friends ( 𝐹 ), object’s social interest communities ( 𝐶 ),and co-work relationship ( 𝐶𝑊 ) information in terms of mutlicast in-teractions, whereas 𝑜𝑏 𝑗𝑒𝑐𝑡 − 𝑜𝑏 𝑗𝑒𝑐𝑡 interactions gives the insight ofcooperativeness ( 𝐶𝑜𝑃 ) in the form of 𝑠𝑢𝑐𝑐𝑒𝑠𝑠 𝑓 𝑢𝑙 and 𝑢𝑛𝑠𝑢𝑐𝑐𝑒𝑠𝑠 𝑓 𝑢𝑙 interactions among the participating objects. The data sources aredefined as the set of triplets { 𝐹, 𝐶, 𝐶𝑊 } , and 𝑜𝑏 𝑗𝑒𝑐𝑡 − 𝑜𝑏 𝑗𝑒𝑐𝑡 inter-actions as cooperativeness ( 𝐶𝑜𝑃 ).Let 𝑂 = { 𝑜 , 𝑜 , ..., 𝑜 𝑛 } represents the set of objects and giventhe set of triplets { 𝐹, 𝐶, 𝐶𝑊 } for each object, and the 𝑜𝑏 𝑗𝑒𝑐𝑡 − 𝑜𝑏 𝑗𝑒𝑐𝑡 interactions, the targeted problem of this paper can be formulatedas quantifying the trust of an object 𝑜 𝑖 towards another object 𝑜 𝑗 at any time interval 𝑡 by using the given set of triplets andcooperativeness, and is expressed as the composition of all the trustparameters as follows: 𝑇𝑟𝑢𝑠𝑡 𝑡 ( 𝑜 𝑖 , 𝑜 𝑗 ) = < 𝐹, 𝐶, 𝐶𝑊 , 𝐶𝑜𝑃 > (1) For the purpose of quantifying trustworthiness of the object fromtheir social information and interactions, we propose a computa-tional model depicted in Figure 3 which considers all of the salientaspects of trustworthiness management. As can be seen in figure,the first step to compute the trustworthiness is to get the socialinformation of trustee, subsequently the next step is the quantifythe selected feature by using the information provided in first step.Feature extraction component addresses the issue of selecting andquantifying the trust metrics suitable for a particular Social IoT ap-plication, and the designated metrics for this paper are
𝐹𝑟𝑖𝑒𝑛𝑑𝑠ℎ𝑖𝑝𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 , 𝐶𝑜𝑚𝑚𝑢𝑛𝑖𝑡𝑦 − 𝑜 𝑓 − 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 , 𝐶𝑜 − 𝑤𝑜𝑟𝑘 𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 , and 𝐶𝑜𝑜𝑝𝑒𝑟𝑎𝑡𝑖𝑣𝑒𝑛𝑒𝑠𝑠 . After that, trustworthiness of an object is com-puted using both the direct trust and the reputation, then the trustaggregation process is employed to accumulate the independentsocial trust metrics to form the final trust score via employing amachine learning-driven scheme. Finally, the trust decision compo-nent provides the information on whether a node is trustworthy or onference’17, July 2017, Washington, DC, USA Subhash Sagar, Adnan Mahmood, Quan Z. Sheng, and Munazza Zaib and Wei Emma Zhang
Figure 3: Proposed SIoT-based trust evaluation model untrustworthy, and is denoted as
𝑇𝑟𝑢𝑠𝑡 𝑡 ( 𝑜 𝑖 , 𝑜 𝑗 ) at time 𝑡 betweenany two objects 𝑜 𝑖 as trustor and 𝑜 𝑗 as trustee. 𝑇 𝑡𝐹𝑆 ). This trust feature signifies theimportance of an object 𝑜 𝑗 vis-à-vis the social relationship of theobject 𝑜 𝑖 locally among its immediate neighbours at any time 𝑡 .Besides, friendship similarity prohibits the malicious objects fromestablishing counterfeit social relationships to get the advantage ofhigher similarity. It is widely accepted that friends are slanted tocooperate with each other and therefore, highly similar objects canbe selected for service discovery and provisioning or a commontask. An object can describe the friendship of a neighbouring objectas follows: 𝑇 𝑡𝐹𝑆 ( 𝑜 𝑖 , 𝑜 𝑗 ) = | 𝐹 𝑜 𝑖 ∩ 𝐹 𝑜 𝑗 || 𝐹 𝑜 𝑖 ∪ 𝐹 𝑜 𝑗 | (2)wherein, 𝐹 𝑜 𝑖 and 𝐹 𝑜 𝑗 refers to a set of friends of object 𝑜 𝑖 and object 𝑜 𝑗 respectively. Furthermore, | . | shows the cardinality of a set whichgives the count on the number of elements in the set. 𝑇 𝑡𝐶𝑜𝐼 ). This property facilitates incomputing the community-based trust feature of a trustee 𝑜 𝑗 vis-à-vis the trustor 𝑜 𝑖 at time 𝑡 , wherein both the objects share commoninterest groups like social groups, e-commerce, and so forth, i.e.,it is an indication of the common interest between them. In SIoTenvironment, object collaborate with at least one interest group,and two objects have a greater chance to build up close contactwith one another if they have a high level of community of interest[26]. In contrast to the example of friendship similarity, the commu-nity of interest introduced in objects do not change instinctively,thus, each object needs to store a list of its owner’s interest group.Mathematically, 𝑇 𝑡𝐶𝑜𝐼 is computed as follows: 𝑇 𝑡𝐶𝑜𝐼 ( 𝑜 𝑖 , 𝑜 𝑗 ) = | 𝐶 𝑜 𝑖 ∩ 𝐶 𝑜 𝑗 || 𝐶 𝑜 𝑖 ∪ 𝐶 𝑜 𝑗 | (3) here, 𝐶 𝑜 𝑖 and 𝐶 𝑜 𝑗 depict the interest group of objects 𝑖 and 𝑗 respec-tively. The higher the degree of common interest (i.e. 𝑇 𝑡𝐶𝑜𝐼 ( 𝑜 𝑖 , 𝑜 𝑗 ) ),the more prominent is the similarity between the objects. 𝑇 𝑡𝐶𝑊 𝑆 ). Co-work similarity gives the no-tion of trust when two or more objects collaborate with each otherto accomplish a common goal. In this type of trust, more focus is onthe working relation instead of their physical closeness. Precisely, 𝑇 𝑡𝐶𝑊 𝑆 score is measured as the ratio of common multicast interac-tion among the object to the total number of multicast interactionsat any time 𝑡 , and as given in Eq. 4 . 𝑇 𝑡𝐶𝑊 𝑆 ( 𝑜 𝑖 , 𝑜 𝑗 ) = | 𝑀 𝑜 𝑖 ∩ 𝑀 𝑜 𝑗 || 𝑀 𝑜 𝑖 ∪ 𝑀 𝑜 𝑗 | (4)wherein, 𝑀 𝑜 𝑖 represents the multicast interactions of object 𝑜 𝑖 ,whereas, 𝑀 𝑜 𝑗 symbolizes the multicast interactions of object 𝑜 𝑗 . 𝑇 𝑡𝐶𝑜𝑃 ). Cooperativeness manifests whethera trustee is socially cooperative in terms of interactions with thetrustor or not, i.e., an object can behave maliciously for a specificservice and try to manipulate the authenticity of the informationfor the other service, thus CoP maintains the content consistencyand offers trustworthy service. Since CoP refers to a measure ofbalance in the interaction between the objects, so we can utilizethe concept of entropy function delineated in [27] to get CoP-basedtrust feature as: 𝑇 𝑡𝐶𝑜𝑃 ( 𝑜 𝑖 , 𝑜 𝑗 ) = − 𝑇 𝑝 𝑙𝑜𝑔 ( 𝑇 𝑝 ) − (1 − 𝑇 𝑝 ) 𝑙𝑜𝑔 (1 − 𝑇 𝑝 ) (5)where, 𝑇 𝑝 represents fraction of messages during the interactionbetween the object 𝑜 𝑖 vis-à-vis object 𝑜 𝑗 , and is computed as follows: 𝑇 𝑝 = 𝑆𝑢𝑐𝑐𝑒𝑠𝑠 𝑓 𝑢𝑙 _ 𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛𝑇𝑜𝑡𝑎𝑙 _ 𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛 (6)Decisively, to estimate the single trust score, a notable methodol-ogy to combine all trust features is to use a conventional weightedsum method as shown in Eq. (7) while determining the appropriatevalue of weights. owards a ML-Driven Trust Evaluation Model for SIoT: A Time-aware Approach Conference’17, July 2017, Washington, DC, USA
Figure 4: Clustering on different pairs of trust featuresFigure 5: Classification on different pairs of trust features (7)
𝑇𝑟𝑢𝑠𝑡 𝑡 ( 𝑜 𝑖 , 𝑜 𝑗 ) = 𝑛 ∑︁ 𝑖 =1 𝑤 𝑖 𝑇 𝑡𝑋 𝑜𝑖 ( 𝑜 𝑖 , 𝑜 𝑗 ) where 𝑋 represents the trust features (i.e., 𝐶𝑜𝐼 , 𝐹𝑆 , 𝐶𝑊 𝑆 and
𝐶𝑜𝑃 ), 𝑤 symbolize the weight of each trust feature, and 𝑛 gives the counton the total number of trust features used.Nevertheless, the weighted sum approach has numerous dis-advantages including but not limited to an unending number ofconceivable outcomes with regards to assessing a weighting factorfor each feature and inability to recognize which trust feature makesthe most effect on the trust in a specific context. Therefore, to over-come these drawback, we hereby propose a machine learning-basedscheme which combines all trust features to ascertain an overalltrust value, and to identify the impact of each feature on the finaltrust decision. To get the statistical estimate for the trust features mentioned in Sec-tion 3.2, we have utilized the CRAWDED dataset from SIGCOMM-2009 [28] to map these traces of data in the form, to be exploitedfor SIoT environment. This dataset comprises of nodes with , interaction between them for over a period of four days.These traces contain the social information of nodes (i.e., friendship,interested communities, interactions, and message logs). There are , pairs of interactions for these nodes, and trust features foreach pair is computed with at least one interaction between thesepairs and these features are formulated in the form of feature matrix given in Eq. 8. To assess the proposed model, we have detachedthe data into hours and have checked the trust score of arbitrarilychosen nodes over a time of hours. (8) 𝐹𝑒𝑎𝑡𝑢𝑟𝑒 𝑀𝑎𝑡𝑟𝑖𝑥 = CoI FS CWS CoP ... ... ... ...CoI 𝑚 FS 𝑚 CWS 𝑚 CoP 𝑚 where 𝑚 shows the number of interactions. To overcome the shortcoming of the weighted sum method as dis-cussed earlier, a machine learning-driven (ML-driven) aggregationscheme is proposed to accumulate the extracted features. So asto accomplish this, we first utilize an unsupervised learning algo-rithm (i.e., k-means clustering) to distinguish two distinct classes,in particular trustworthy and untrustworthy [29]. The primary mo-tivation to utilize unsupervised learning over supervised learningis because of the unavailability of the labeled training set. The k-means algorithm compels two initial information sources; initialcentroid points ( 𝐶 ) and the number of clusters ( 𝐶 𝑘 ). In the proposedmodel, we have randomly assigned the initial centroids for twoclusters, namely trustworthy and trustworthy, wherein, the objectsclose to 1 are marked as trustworthy and objects with values nearthe origin (0 , are marked as untrustworthy. onference’17, July 2017, Washington, DC, USA Subhash Sagar, Adnan Mahmood, Quan Z. Sheng, and Munazza Zaib and Wei Emma Zhang (a) Time: 4 Hours (b) Time: 8 Hours (c) Time: 12 Hours(d) Time: 16 Hours (e) Time: 20 Hours (f) Time: 24 Hours Figure 6: Trust score of 20 Nodes over a period of 24 hours
Subsequently, a multi-class classification algorithm like randomforest [30] is utilized to train the model, and to identify the bestdecision boundary that distinguish trustworthy and untrustworthyinteractions. There are a few motives to employ random forest forclassification. Random forest is appropriate to recognize the mostsignificant features in the dataset as well as the weights of eachtrust feature to acquire the single trust score. Furthermore, thisclassification algorithm makes use of multiple random decisiontrees on the same dataset to avoid overfitting. Finally, for demon-strations purpose, we just consider the pairs of trust features forboth the clustering and the classification as opposed to utilizing allthe features at once.
This section describes the performance evaluation of the proposedmodel deliberated in Section 3. As discussed, we merely considerpairs of trust features for both visualization and distribution oftrust via comparing
𝐶𝑜𝐼 with 𝐹𝑆 , 𝐶𝑊 𝑆 , and
𝐶𝑜𝑃 respectively. Asportrayed in Figure 4, it can be clearly observed how the clusteringalgorithm successfully classifies the interaction as trustworthy oruntrustworthy. Subsequently, as depicted in Figure 5, the classifi-cation algorithm (Random Forest) clearly identifies the pair-wisedecision boundary between the interactions and is capable of clas-sifying the futuristic interactions as trustworthy or untrustworthy.Additionally, the classification algorithm also manifests the weigh-tage (i.e., feature importance) of each individual feature so as toascertain a single trust score, wherein the weightage of each feature{
𝐶𝑜𝐼 , 𝐹𝑆 , 𝐶𝑊 𝑆 , and
𝐶𝑜𝑃 } is { . , . , . , and . }.Besides, we have evaluated the trust model in terms of accuracyand considered the following three evaluation metrics: 1. Precision:
Precision is defined as the accuracy of a model toclassify trustworthy objects as trustworthy and untrustwor-thy objects as untrustworthy. It is measured as the ratio ofcorrectly predicted untrustworthy observations to the totalnumber of untrustworthy observations (as per untrustwor-thy class). The value of precision is class dependent.2.
Recall:
Recall is also class dependent and is referred to theproportion of trustworthy or untrustworthy objects thathave been correctly identified. It is measured as the ratioof correctly predicted positive (i.e., trustworthy or untrust-worthy) observation to all the observations in the actualclass.3.
F1-Score:
It is used to measure the accuracy of a model byconsidering weighted average of precision and recall.Overall, Table 1 gives the score of each of the above evaluationmetrics for both the classes, trustworthy (T) and untrustworthy (U) ,and it can observed that our model gives the high precision andrecall values with an accuracy of . .Furthermore, Figure 6 depicts the trust score of randomly se-lected nodes over a period of hours. As can be seen from thefigure, the trust result of nearly all the nodes remains the same.However, trust of two nodes ( 𝑁𝑜𝑑𝑒 𝑎𝑛𝑑 𝑁𝑜𝑑𝑒 ) varies with time,i.e., trust score of 𝑁𝑜𝑑𝑒 increases with time and turns out to bestable after time 𝑡 = 12 ℎ𝑜𝑢𝑟𝑠 . Besides, the trust score of 𝑁𝑜𝑑𝑒 decreases with time, i.e., starting from the trust score of almost . at time 𝑡 = 4 ℎ𝑜𝑢𝑟𝑠 it drops down to . at the end. Subse-quently, Figure 7 justifies the variation in the trust score of twonodes ( 𝑁𝑜𝑑𝑒 𝑎𝑛𝑑 𝑁𝑜𝑑𝑒 ) and illustrates the effect of each trustfeature on trust of each node. It can be observed from Figure 7(a) owards a ML-Driven Trust Evaluation Model for SIoT: A Time-aware Approach Conference’17, July 2017, Washington, DC, USA (a) CoI, Trust vs Time (b) FS, Trust vs Time(c) CWS, Trust vs Time (d) CoP, Trust vs Time Figure 7: Effect of CoI, FS, CWS, and CoP score on nodes trust and Figure 7(c) that the CoI and CWS score for Node: 8 remains thesame but the change in score of FS and CoP prompt the increasein the trust result of Node:8. Similarly, the decrease in the CWSscore as shown in Figure 7(c) decreases the trust score of Node:19as all other trust features remain the same. Overall, the computedscore for each of the trust parameters for 20 nodes over the periodof time is depicted in Table 2.
Table 1: Performance EvaluationAccuracy = . Classes Precision Recall F-Score
Untrustworthy 1.0 0.97 0.99Trustworthy 0.98 1.00 0.99
In this paper, we have envisaged an efficient time-aware trust evalu-ation model to identify untrustworthy objects in the SIoT network.Precisely, the proposed trust model consider SIoT relationshipsin terms of friendship, community-of-interest, co-work similarity,and cooperativeness as the trust parameters. Furthermore, a ma-chine learning-driven aggregation scheme is introduced so as tosynthesize these trust parameters to ascertain a single trust score.Simulation results demonstrate the effectiveness of the model viasegregating trustworthy and untrustworthy objects and furtherprovides the variation in the trust score of each node over a periodof time. In order to further develop our idea, we intend to verify the con-vergence and resilience property of the proposed model by incorpo-rating context-awareness (i.e., environment conditions, energy, andtime) in a dynamically changing SIoT environment. Additionally,quantifying and employing social strength [31] among the objectshaving social connection as a trust parameter would result in theprecise determination of trust.
REFERENCES [1] Giancarlo Fortino and P. Trunfio.
Internet of things based on smart objects: Tech-nology, middleware and applications . Springer, 01 2014.[2] Luigi Atzori, Antonio Iera, and Giacomo Morabito. The Internet of Things: ASurvey.
Computer Networks , 54(15):2787 – 2805, Oct 2010.[3] Sonia Khetarpaul, S. Gupta, and L.V. Subramaniam. Spatiotemporal social (sts)data model: correlating social networks and spatiotemporal data.
Social NetworkAnalysis and Mining
Computer Networks , 56(16):3594 – 3608, Nov 2012.[6] Roopa M.S., Santosh Pattar, Rajkumar Buyya, Venugopal K.R., S.S. Iyengar, andL.M. Patnaik. Social internet of things (siot): Foundations, thrust areas, systematicreview and future directions.
Computer Communications , 139:32 – 57, 2019.[7] Jian An, Xiaolin Gui, Wendong Zhang, Jinhua Jiang, and Jianwei Yang. Researchon social relations cognitive model of mobile nodes in internet of things.
Journalof Network and Computer Applications , 36(2):799 – 810, 2013.[8] M. Nitti, R. Girau, L. Atzori, A. Iera, and G. Morabito. A Subjective Model forTrustworthiness Evaluation in the Social Internet of Things. In
Proceeding of23rd IEEE International Symposium on Personal, Indoor, and Mobile Radio Commu-nications - (PIMRC) , pages 18–23, Sep. 2012.[9] Guido Möllering. The Nature of Trust: From Georg Simmel to a Theory ofExpectation, Interpretation, and Suspension.
Sociology , 35(2):403–420, May 2001.[10] H. Xia, F. Xiao, S. Zhang, C. Hu, and X. Cheng. Trustworthiness InferenceFramework in the Social Internet of Things: A Context-Aware Approach. In proceeding of IEEE Conference on Computer Communications (INFOCOM) , pages onference’17, July 2017, Washington, DC, USA Subhash Sagar, Adnan Mahmood, Quan Z. Sheng, and Munazza Zaib and Wei Emma Zhang
Table 2: Trust Parameters Data for 20 Nodes
Time: 1 Time:2 Time: 3 Time: 4 Time: 5 Time: 6Id CoI FS CWS CoP CoI FS CWS CoP CoI FS CWS CoP CoI FS CWS CoP CoI FS CWS CoP CoI FS CWS CoP1
19 0.50 0.46 0.60 0.47 0.50 0.45 0.55 0.47 0.50 0.47 0.52 0.47 0.50 0.47 0.50 0.47 0.50 0.47 0.47 0.47 0.60 0.47 0.47 0.4720
IEEE Transactions on Dependable and Secure Computing , 13(6):684–696, 2016.[12] A. A. Adewuyi, H. Cheng, Q. Shi, J. Cao, Á. MacDermott, and X. Wang. CTRUST:A Dynamic Trust Model for Collaborative Applications in the Internet of Things.
IEEE Internet of Things Journal , 6(3):5432–5445, 2019.[13] I. Chen, J. Guo, and F. Bao. Trust management for soa-based iot and its applicationto service composition.
IEEE Transactions on Services Computing , 9(3):482–495,2016.[14] Subhash Sagar, Adnan Mahmood, Jitander Kumar, and Quan Z. Sheng. A time-aware similarity-based trust computational model for social internet of things.In
Globecom 2020 - 2020 IEEE International Global Communications Conference(Globecom) , pages 1–6, 2020.[15] A. Alnasser, H. Sun, and J. Jiang. Recommendation-Based Trust Model forVehicle-to-Everything (V2X).
IEEE Internet of Things Journal , 7(1):440–450, 2020.[16] A. Mahmood, B. Butler, W. E. Zhang, Q. Z. Sheng, and S. A. Siddiqui. A hybridtrust management heuristic for vanets. In , pages748–752, 2019.[17] Li Xiong and Ling Liu. Peertrust: supporting reputation-based trust for peer-to-peer electronic communities.
IEEE Transactions on Knowledge and Data Engineer-ing , 16(7):843–857, 2004.[18] Yusuo Hu, Danqi Wang, and Hui Zhong. Socialtrust: Enabling long-term socialcooperation in peer-to-peer services.
Peer-to-Peer Networking and Applications , 7,12 2014.[19] Y. Chen, M. Zhou, Z. Zheng, and D. Chen. Time-aware smart object recommen-dation in social internet of things.
IEEE Internet of Things Journal , 7(3):2014–2027,2020.[20] Y. Saleem, N. Crespi, M. H. Rehmani, R. Copeland, D. Hussein, and E. Bertin.Exploitation of social iot for recommendation services. In , pages 359–364, 2016.[21] N. B. Truong, G. M. Lee, T. Um, and M. Mackay. Trust Evaluation Mechanismfor User Recruitment in Mobile Crowd-Sensing in the Internet of Things.
IEEE Transactions on Information Forensics and Security , 14(10):2705–2719, 2019.[22] B. Ye, Y. Wang, and L. Liu. Crowd trust: A context-aware trust model for workerselection in crowdsourcing environments. In , pages 121–128, 2015.[23] U. Jayasinghe, G. M. Lee, T. Um, and Q. Shi. Machine Learning based Trust Com-putational Model for IoT Services.
IEEE Transactions on Sustainable Computing , 4(1):39–52, 2019.[24] S. Sagar, A. Mahmood, Q. Z. Sheng, and W. E. Zhang. Trust computationalheuristic for social internet of things: A machine learning-based approach. In
ICC 2020 - 2020 IEEE International Conference on Communications (ICC) , pages1–6, 2020.[25] Yating Wang, Yen-Cheng Lu, Ing-Ray Chen, Jin-Hee Cho, Ananthram Swami,and Chang-Tien Lu. Logittrust : A logit regression-based trust model for mobilead hoc networks. 2014.[26] F. Bao, I. Chen, and J. Guo. Scalable, adaptive and survivable trust managementfor community of interest based internet of things systems. In , pages1–7, 2013.[27] S. Adali, R. Escriva, M. K. Goldberg, M. Hayvanovych, M. Magdon-Ismail, B. K.Szymanski, W. A. Wallace, and G. Williams. Measuring Behavioral Trust in SocialNetworks. In
Procedding of IEEE International Conference on Intelligence andSecurity Informatics , pages 150–152, May 2010.[28] Anna-Kaisa Pietilainen and Christophe Diot. CRAWDAD datasetthlab/sigcomm2009 (v. 2012-07-15). Downloaded from https://crawdad.org/thlab/sigcomm2009/20120715, 2012.[29] Leo Breiman. Random Forests.
Machine Learning , 45(1):5–32, Oct 2001.[30] Xavier Amatriain, Alejandro Jaimes, Nuria Oliver, M. Josep Pujol, FrancescoRicci, Lior Rokach, Bracha Shapira, and Paul B. Kantor. Data Mining Methodsfor Recommender Systems. pages 39–71, Oct 2011.[31] J. Jung, S. Chun, X. Jin, and K. Lee. Quantitative computation of social strengthin social internet of things.