Trust Computational Heuristic for Social Internet of Things: A Machine Learning-based Approach
aa r X i v : . [ c s . CR ] F e b Trust Computational Heuristic for Social Internet ofThings: A Machine Learning-based Approach
Subhash Sagar , Adnan Mahmood , Quan Z. Sheng , and Wei Emma Zhang Department of Computing, Macquarie University, Sydney, NSW 2109, Australia School of Computer Science, The University of Adelaide, Adelaide, SA 5005, Australia
Abstract —The Internet of Things (IoT) is an evolving networkof billions of interconnected physical objects, such as, numeroussensors, smartphones, wearables, and embedded devices. Thesephysical objects, generally referred to as the smart objects , whendeployed in real-world aggregates useful information from theirsurrounding environment. As-of-late, this notion of IoT has beenextended to incorporate the social networking facets which haveled to the promising paradigm of the ‘Social Internet of Things’(SIoT) . In SIoT, the devices operate as an autonomous agent andprovide an exchange of information and services discovery inan intelligent manner by establishing social relationships amongthem with respect to their owners. Trust plays an important rolein establishing trustworthy relationships among the physical ob-jects and reduces probable risks in the decision making process.In this paper, a trust computational model is proposed to extractindividual trust features in a SIoT environment. Furthermore, amachine learning-based heuristic is used to aggregate all the trustfeatures in order to ascertain an aggregate trust score. Simulationresults illustrate that the proposed trust-based model isolates thetrustworthy and untrustworthy nodes within the network in anefficient manner.
Index Terms —Social Internet of Things, Trust Management,Machine Learning.
I. I
NTRODUCTION
Over the past decade, the notion of Internet of Things (IoT)has evolved as a new generation of a network of billions of de-vices seamlessly connected over the Internet. These devices areregularly outfitted with sensors which monitor different aspectsof human life for supporting numerous beneficial applicationsand services [1]. A big value of the IoT resides on its ability tocreate a network of resources, i.e., by making resources social,where social relationships facilitate the discovery of resourcesthat have the capabilities required to solve a particular task. Toaccomplish this goal, IoT should be endowed with the abilityto define and manage social relationships between resources inevery aspect [2]. The promising paradigm of Social Internet ofThings (SIoT) has transpired recently which can be seen as thecombination of social networks and IoT, wherein every objectis capable of establishing social relationship autonomouslywith the other objects depending on the rules set by their re-spective owners [3]. Nevertheless, the convergence of physicalobjects, humans, and cyber components in SIoT presents newconcerns for risk, privacy, and security. As the intent of SIoTservices is to make decisions autonomously without the needfor any human intervention, the notion of trust is recognized as *Corresponding Authors: (subhash.sagar, adnan.mahmood)@mq.edu.au a prospective solution for supporting both humans and servicesin order to overcome the perception of insecurity and minimizethe risks when making a decision.The paradigm of trust has been used in several disciplines,including but not limited to, psychology, sociology, and com-puter science [4], [5]. In terms of SIoT, trust can be referred toas a notion of ‘belief’ or ‘confidence’ of a trustor in a trusteeto perform a specific task for satisfying a trustor’s expectationsin a specific context within a particular period of time [6]. Asdepicted in Fig. 1, if a node A (trustor) needs to compute thetrust value for node B (trustee), it computes the direct trust onits own and requests the mutual friends (e.g., nodes C, D, andE) to provide their recommendations for the same. Finally, thecombination of direct and indirect trust provides an aggregatetrust score.Fig. 1: A Schematic Representation of Trust Computation ina SIoT EnvironmentThe main reason for providing the trust management systemfor SIoT is apparent since there are misbehaving nodes whichmay perform different types of attacks (such as ballot-stuffingattack, bad-mouthing attack, self-promoting and whitewashingattack) based on their social relationships with other nodes fortheir malicious advantages at the expense of other IoT deviceswhich provide similar SIoT services. number of trust management schemes have been proposedin recent years. The authors in [7] delineated a subjective andobjective trustworthiness scheme in order to ascertain trust ina SIoT environment, wherein trust of any particular node wasevaluated by aggregating three salient features, i.e., centrality,opinions from its common friends, and the direct experiences.Nevertheless, the influence of each of these features on thetrust aggregation process was ascertained by certain weightingfactors that are themselves extremely difficult to figure primar-ily owing to the reason that trust depends on several complexparameters, i.e., context, time, resources, and the environment.Similarly, the authors in [8] proposed an adaptive trust protocolfor a SIoT system. It employs direct monitoring and indirectobservations from users with similar social activity metrices,i.e., social contacts, honesty, and communities-of-interest, toascertain overall trust. Furthermore, in order to aggregate thesame, an adaptive filtering method was designed to integratedirect and indirect observations with weighting parameters foreach observation. However, the authors did not validated theirenvisaged protocol on a wide range of dynamic environmentalscenarios, wherein assigning the weighting parameters is itselfa complicated chore. In [9], the authors employed community-of-interest as a social trust metric to envisage a trust manage-ment mechanism for purposes of IoT, wherein a Kalman filterwas used as a tool to estimate the trust value of a node beforeinteraction. However, the aggregate trust value was calculatedusing a linear equation with weighting factors for both directand indirect trust.In [10], the authors presented a subjective trust model em-ploying the social features of similarity in respect of commoninterests, honesty, and cooperativeness for evaluating the trustscore of a particular node. To obtain the direct trust, a weightedsum metric was used to aggregate both the current as well aspast experiences. Nevertheless, the model did not account forthe indirect observations or recommendations from other nodesin the network which is essential for the IoT services delivery.Moreover, in [11], a context-based social trust model has beendeveloped for IoT purposes by considering social relationshipsamongst nodes. For computing the trust score of a node, bothdirect and indirect observations with a transaction context wasused. However, to obtain a single trust score, a weighted summetric was used for aggregation. Recently, the authors in [12]suggested a trust framework model based on the social profileof a node, wherein different social features were accumulatedto ascertain the trust score of a node. In addition, a machinelearning-based algorithm was exploited to aggregate only thedirect trust metric and which is not sufficient enough to decidewhether a node is trustworthy or not.In this research work, we have computed trust based on boththe direct observation as well as the indirect recommendations.Our primary contributions in this paper are threefold: • A comprehensive trust model for the SIoT environmenthas been envisaged which specifies the formation of trustvia both direct and indirect observation of the nodes; • A Machine Learning (ML)-based aggregation scheme hasbeen envisaged in contrast to the conventional trust-based heuristics to aggregate the trust attributes for obtaining asingle trust score; and • Performance evaluation of the proposed model has beencomprehensively carried out in a simulation environment.II. T
RUST C OMPUTATION M ODEL
This section presents the details of the computational modelfor an efficient and flexible trust management scheme in a SIoTenvironment. Our trust model in this paper comprises of twometrics,
Direct Trust Metric (DTM) and
Indirect Trust Metric(ITM) , as depicted in Fig. 2, where DTM gives the notion of di-rect observation while ITM provides the reputation of nodes inthe network. The trust assessment for node i (trustor) towardsnode j (trustee) is denoted by T X ( i, j ) where X denotes the so-cial attributes like F riendship Similarity, Community − of − Interest, Reward, and
Cooperativeness . The rangeof T X ( i, j ) varies from [0 , where values nearer to indicateuntrustworthiness while values around indicate trustworthi-ness. After aggregating all the features T X ( i, j ) from the directinteraction via ML-based algorithm, the result is stored in therepository and is used as a direct trust score. For indirect trust(recommendations), the node (trustor) requests the direct trustfrom the other nodes. Subsequently, Algorithm 1 is utilized tocombine both the results (trusts) to obtain the final trust score.Fig. 2: The Proposed Trust Computation Model A. Direct Trust Metric (DTM)
DTM is used to provide direct observation of a trustee priorto interaction. Although a trustee can be assessed via numerousdifferent attributes, in this paper, we have employed four mainattributes for the assessment of any trustee with respect to thetrustor and which are elaborated as follows:
1) Friendship Similarity (FS):
Friendship similarity repre-sents social relationship in terms of interaction among partici-pating objects. It measures the importance of an object amongother objects with reference to a specific task and context. Thisproperty of an object is ascertained as: T F S ( i, j ) = | F i ∩ F j || F i | − (1)where, F i and F j refers to a set of friends of node i and node j respectively and | . | shows the cardinality of a set. ) Community-of-Interest (CoI): This kind of attribute rep-resents the similarity of nodes with respect to the social interestcommunities or groups. Therefore, nodes with high CoI havemore chances of interacting with each other in order to developa trustworthy relationship. CoI-based trust between two nodesis computed as: T CoI ( i, j ) = | C i ∩ C j || C i | (2)where, C i and C j represents set of communities of node i andnode j respectively.
3) Cooperativeness (CoP):
CoP manifests whether a trusteeis socially cooperative with a trustor or not. Since CoP refers toa measure of balance in the interaction between the nodes, wecan employ the entropy function delineated in [13] to calculateCoP-based trust as: T CoP ( i, j ) = − T p log ( T p ) − (1 − T p ) log (1 − T p ) (3)where, T p represents fraction of messages during the interac-tion.
4) Reward/Punishment:
In order to maintain both trustwor-thy relationships and punish misbehaving nodes, we utilize anexponential downgrading formula to provide the incentive tohonest nodes and penalties to misbehaving nodes as: T Reward ( i, j ) = | Int − Int U || Int | e − ( | IntU || Int | ) (4)here, Int highlights the total number of interactions and
Int U refers to the count for the number of unsuccessful interactionsbetween node i and node j .Traditionally, to aggregate the overall trust, a linear equationwith the weighting factor is used as shown in Eq. (5), however,this approach has numerous disadvantages and challenges (asdiscussed in Section I) while determining the appropriate valueof weights. T Direct ( i, j ) = w T F S ( i, j ) + w T CoI ( i, j ) + w T CoP ( i, j ) + w T Reward ( i, j ) (5)Therefore, to overcome this drawback, we hereby propose anew machine learning-based approach which combines directand indirect trust to ascertain an overall trust value. Addition-ally, this approach also identifies the impact of each of thesefeatures on the aggregated trust value. B. Indirect Trust Metric (ITM)
A reputation metric (ITM) is employed in order to ascertainthe trustee based on the opinion of other nodes in the network.Nevertheless, the reputation of objects vary from node to node,and therefore, it is not optimal to take account of all the nodesin the network for computing the reputation of a trustee. Thus,in this paper, reputation value is requested from nodes havingat least a single friend in common between trustor and trustee.Finally, in order to ascertain a single trust value, we developan algorithm (i.e., Algorithm 1) so as to aggregate both direct
Algorithm 1
Trust Score Estimation → U ntrustworthy, → T rustworthy, → N eutral Input : Direct Trust {
0, 1 or 2 } , Recommendations { | T | , | U | , and | N | } , Output : Single Trust Score { } , | T | → N o : of T rustworthy Recommendations , | U | → N o : of U ntrustworthy Recommendations , | N | → N o : of N eutral Recommendations , θ → T hreshold (%) , P U → U ntrustworthy Recommendations (%) , P T → T rustworthy Recommendations (%) if There are no Recommendations then
F inal T rust = Direct T rust end if if Direct T rust == 0 then if ( | U | > = | T | || ( | N | > = | T | && | N | > = | U | ) then Node is Untrustworthy else P T = | T | T otal Recommendations +1 if P T > = θ then N ode is
Trustworthy else
N ode is
Untrustworthy end if end if else if
Direct T rust == 1 then if ( | T | > = | U | ) || ( | N | > = | T | && | N | > = | U | ) then Node is Trustworthy else P U = | U | T otal Recommendations +1 if P U > = θ then N ode is
Untrustworthy else
N ode is
Trustworthy end if end if else if ( | T | > | U | ) then N ode is
Trustworthy else
N ode is
Untrustworthy end if end if and indirect trust. The said algorithm takes into considerationboth of the direct trust and recommendations as an input, andaccordingly, provides a single trust of a node, i.e., trustworthyor untrustworthy. Our trust score estimation algorithm dependsmore on direct trust as can be noticed from lines 13-16. If thedirect trust is or untrustworthy and more recommendationsare untrustworthy ( | U | ) or neutral ( | N | ), the node is marked asuntrustworthy. Here, neutral manifests that the node is neithertrustworthy nor untrustworthy. Furthermore, if the direct trust a) FS and CoI (b) FS and Reward (c) FS and CoP(d) CoI and Reward (e) CoI and CoP (f) Reward and CoP Fig. 3: Clustering on Different Pairs of Featuresis or untrustworthy and the number of trustworthy recom-mendations are greater than untrustworthy recommendations,i.e., ( | T | > | U | ), then our algorithm does not mark the node astrustworthy immediately. Instead, a percentage of trustworthyrecommendations ( P T ) is computed, and if P T is greater thanthreshold ( θ ), which in our case is or . , then the nodeis marked as trustworthy (lines 17-22). The value of θ totallydepends on an individual application and the reason for sucha high value of θ in our case is to accord higher authority tothe trustor node rather than the recommendations from othernodes in the network to cope with the issue of good mouthingand ballot-stuffing attack. Similarly, this algorithm follows thelines 24-34 provided that the direct trust is or trustworthy .At the end, if a direct trust score is or neutral , it impliesthat the trustor node does not possess its own observation forthe trustee and the trustworthiness of a node is decided on thebasis of recommendations. If | T | > | U | , then node is markedas trustworthy, otherwise, it is untrustworthy.III. S IMULATION S ETUP
To ascertain the aforementioned trust features, i.e., T F = { F S, CoI, Reward, CoP } , to be employed for our ML-basedmodel, we have used sigcomm-2009 dataset which comprisesthe traces that could be mapped in the form of the promisingparadigm of SIoT. These traces contain the social informationof device/user (i.e., friendships and interested groups informa-tion, activities, and messages logs), which have been utilized https://crawdad.org/thlab/sigcomm2009/20120715/ for computing the trust features mentioned in Section II. Thisdataset comprises nodes and , interactions for overa period of four days. The trust features are ascertained foreach pair of nodes with at least one interaction between themand there are , pair of nodes in total.Subsequent to obtaining of the trust features T F , an unsuper-vised learning algorithm (i.e., k-means clustering) is employedso as to label the features in different classes [14]. The primaryreason to use the k-means clustering is due to the unavailabilityof the labeled training set. The k-means algorithm requires twoinputs: initial centroid points ( C ) and the number of clusters( C k ). Initially, we have allocated random centroid positions toall the clusters ( C k , k = 1 to ) and executed the algorithmuntil the convergence. For the proposed model, we needed twoclusters, namely trustworthy or untrustworthy, nevertheless, weused the elbow method to acquire the optimized cluster ( C opt )size with lower k-means cost function, which in our case is, C opt = 3 . This implies that the k-means algorithm segregatesthe data in three clusters – trustworthy, untrustworthy, and neu-tral, wherein neutral means that the node is neither trustworthynor untrustworthy and nodes with values near the origin (0 , are marked as untrustworthy.In order to train the model subsequent to clustering process,we employed the multi-class random forest classification algo-rithms [15] to identify the best decision boundary which segre-gates the trustworthy and untrustworthy interactions. Randomforest is fairly suitable for feature engineering, i.e, to identifythe most important features amongst all the available features a) FS and CoI (b) FS and Reward (c) FS and CoP(d) CoI and Reward (e) CoI and CoP (f) Reward and CoP Fig. 4: Decision Boundaries for Different Pairs of Featuresin the dataset, and in our case, it is extremely indispensable toknow the more imperious features as well as the weightage ofeach feature during aggregation of the overall trust. In addition,random forest avoids overfitting problem by constructing mul-tiple decision trees on the same dataset with random selection.For training and testing purposes, , samples were utilizedwhich represents the interaction between each pair of nodes.Out of the same, of them were used for training purposes,whereas, of them were used to evaluate the accuracy ofthe proposed model.IV. R ESULTS AND D ISCUSSION
As deliberated in Section III, k-means clustering has beenemployed in order to classify the trustworthy and untrustwor-thy relationships. Nevertheless, it is the elbow method whichcategorizes the data into three clusters instead of the two, i.e.,trustworthy, untrustworthy, and neutral, as portrayed in Fig. 3.For demonstration purposes, we only took into considerationthe pairs of trust features for clustering instead of employingall the features at once since it is not feasible to visualize all ofthem together. However, we can use the Principle ComponentAnalysis (PCA) technique for dimension reduction, e.g., fromfour to two in our case, to ascertain the results for visualizingall the features at once [16]. As can be clearly observed fromFig. 3a and 3b, the distribution of trust values by comparing
F S with
CoI and
F S with
Reward respectively demonstratethat the region where
CoI > = 0 . and Reward > = 0 . arethe trustworthy regions, whereas, region where CoI < = 0 . and Reward < = 0 . are the untrustworthy regions. However,these figures signify a clear dominance of CoI and
Reward on F S as trust value primarily depends on
CoI and
Reward . Furthermore, all other subfigures reflect a mutual contributionto the overall trust score as depicted in Fig. 3c-3f.With successful investigation of the labels, the next step is totrain our model to identify whether the futuristic interactions ofSIoT nodes are trustworthy or not. After applying a supervisedlearning algorithm (random forest), Fig. 4 portrays the decisionboundary to successfully classify the nodes with minimal error.It is quite evident from Fig. 4 that our model is fully capableof classifying the futuristic interactions as trustworthy, neutral,or untrustworthy. Fig. 5 manifests the accuracy of the classifi-cation algorithm ( . ) and the significance (i.e., weightage)of each individual feature ascertained after applying our modelon the dataset. A cc u r a c y ( % ) Fig. 5: Feature Weightage and Model AccuracyIt can be easily observed that
CoI ( . ) and Reward ( . ) have more impact on the overall trust score followedy CoP ( . ) and F S ( . ). The key reason for such ahigher weightage of CoI is owing to the fact that the objectsbelonging to the same group tends to be more trustworthy andinteract more frequently. Similarly, the
Reward feature showsthat more unsuccessful interactions lead to untrustworthinessmaking it an important factor in calculating the trust score.As we have identified three categories of nodes, i.e., trust-worthy, untrustworthy, and neutral, nevertheless, by and large,we only need to identify a node as either a trustworthy one oran untrustworthy one. To cope with the issue of neutral nodes,we applied our trust score estimation algorithm, i.e., Algorithm1, on the results obtained from the k-means clustering so as toaggregate the direct trust with the recommendations (indirecttrust) and accordingly reduced the size of the clusters to twoto be useful for real-world applications. We used a percentagethreshold ( θ ) to combine the trust score, if and when, there isa conflict of interest between the trustor and the other nodes inthe network. For instance, if direct trust = trustworthy andthe number of untrustworthy recommendations ( | U | ) are morethan trustworthy recommendations ( | T | ), then our algorithminvestigates the percentage of untrustworthy recommendations( P U ). If P U > θ , then the node is marked as untrustworthy. A cc u r a c y ( % ) Fig. 6: Trust Estimation AccuracyTo ascertain the optimum value of θ , we compared differentthresholds, and when θ = 70% , our algorithm manifests themaximum accuracy of . as depicted in Fig. 6. However,the accuracy of our algorithm after combining both direct andindirect trust is bit lower than the accuracy of the model whenonly the direct trust has been taken into contemplation. Thisis primarily owing to the fact that there had been no previoushistory of interactions between the nodes which led to highertrust values at the start. Thus, to avoid such circumstances, rec-ommendations are always employed to ascertain the concreteresults which lowers the overall accuracy.V. C ONCLUSION AND F UTURE W ORK
In this paper, we have proposed a machine learning-basedtrust aggregation scheme in contrast to the traditional weightedheuristics to ascertain a single trust score for each SIoT node. A trust computational model has been accordingly envisagedto extract key trust features with respect to the SIoT domain.Subsequently, in order to aggregate the trust, the data is labeledby using the k-means clustering for identifying the trustworthyand untrustworthy interactions. A trust prediction scheme hasbeen further proposed for identifying the decision boundariesand to learn the impact of individual features on the aggregatedtrust score. Our simulation results demonstrate higher accuracyin ascertaining the trustworthy interactions.In near future, we intend to incorporate experience as a trustattribute for the computation of direct and indirect trust so as toaccumulate the previous interaction history of the target nodescoupled with few other social features, i.e., social relationshipsin terms of co-location and co-work. This could result in moreprecise determination of trustworthy nodes in a SIoT network.R
EFERENCES[1] L. Atzori, A. Iera, and G. Morabito, “The Internet of Things: A Survey,”
Computer Networks , vol. 54, no. 15, pp. 2787 – 2805, Oct 2010.[2] P. Mendes and P. P. Mendes, “Social-driven Internet of Connected Ob-jects,” in
Proceeding of Interconnected Smart Objects with the InternetWorkshop , Lisbon, Portugal, Mar 2011.[3] L. Atzori, A. Iera, G. Morabito, and M. Nitti, “The Social Internet ofThings (SIoT) – When Social Networks Meet the Internet of Things:Concept, Architecture, and Network Characterization,”
Computer Net-works , vol. 56, no. 16, pp. 3594 – 3608, Nov 2012.[4] M. Nitti, R. Girau, L. Atzori, A. Iera, and G. Morabito, “A SubjectiveModel for Trustworthiness Evaluation in the Social Internet of Things,”in
Proceeding of 23rd IEEE International Symposium on Personal,Indoor, and Mobile Radio Communications - (PIMRC) , Sep. 2012, pp.18–23.[5] G. M¨ollering, “The Nature of Trust: From Georg Simmel to a Theory ofExpectation, Interpretation, and Suspension,”
Sociology , vol. 35, no. 2,pp. 403–420, May 2001.[6] F. Amin, A. Ahmad, and G. Sang Choi, “Towards Trust and FriendlinessApproaches in the Social Internet of Things,”
Applied Sciences , vol. 9,no. 1, p. 166, Jan 2019.[7] M. Nitti, R. Girau, L. Atzori, and S. Member, “Trustworthiness Manage-ment in the Social Internet of Things,”
IEEE Transactions on Knowledgeand Data Engineering , vol. 26, no. 5, pp. 1253–1266, May 2014.[8] I.-R. Chen, F. Bao, and J. Guo, “Trust-based Service Management forSocial Internet of Things Systems,”
IEEE Transactions on Dependableand Secure Computing , vol. 13, no. 6, pp. 684–696, Apr 2016.[9] O. Ben Abderrahim, M. H. Elhdhili, and L. Saidane, “TMCoI-SIOT:A Trust Management System based on Communities of Interest for theSocial Internet of Things,” in Proceeding of 13th International WirelessCommunications and Mobile Computing Conference - (IWCMC) , pp.747–752, July 2017.[10] N. B. Truong, H. Lee, B. Askwith, and G. M. Lee, “Toward a TrustEvaluation Mechanism in the Social Internet of Things,”
Sensors ,vol. 17, no. 6, pp. 1–24, June 2017.[11] S. E. A. Rafey, A. Abdel-Hamid, and M. A. El-Nasr, “CBSTM-IoT: Context-based Social Trust Model for the Internet of Things,” in
Proceeding of International Conference on Selected Topics in MobileWireless Networking (MoWNeT) , Apr 2016, pp. 1–8.[12] U. Jayasinghe, G. M. Lee, T. Um, and Q. Shi, “Machine Learning basedTrust Computational Model for IoT Services,”
IEEE Transactions onSustainable Computing , vol. 4, no. 1, pp. 39–52, Jan 2019.[13] S. Adali, R. Escriva, M. K. Goldberg, M. Hayvanovych, M. Magdon-Ismail, B. K. Szymanski, W. A. Wallace, and G. Williams, “MeasuringBehavioral Trust in Social Networks,” in
Procedding of IEEE Interna-tional Conference on Intelligence and Security Informatics , May 2010,pp. 150–152.[14] L. Breiman, “Random Forests,”
Machine Learning , vol. 45, no. 1, pp.5–32, Oct 2001.[15] X. Amatriain, A. Jaimes, N. Oliver, M. J. Pujol, F. Ricci, L. Rokach,B. Shapira, and P. B. Kantor, “Data Mining Methods for RecommenderSystems,” pp. 39–71, Oct 2011.16] J. Shlens, “A Tutorial on Principal Component Analysis,”