Trusted information exchange in peer-to-peer mobile social

CONCURRENCY AND COMPUTATION: PRACTICE AND EXPERIENCEConcurrency Computat.: Pract. Exper. 2012; 24:2055–2068Published online 15 August 2011 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/cpe.1837

SPECIAL ISSUE PAPER

Trusted information exchange in peer-to-peer mobilesocial networks

Basit Qureshi*,†, Geyong Min and Demetres Kouvatsos

School of Computing, Informatics and Media, University of Bradford, Bradford BD7 1DP, UK

SUMMARY

Social networks have recently been gaining popularity. In a peer-to-peer mobile social network (MSN), userswith similar interests establish groups or communities and share information without relying on a centralizedinfrastructure. Users socially interact with each other using handheld mobile devices and membership in agroup/community of MSNs is granted by a pre-existing group member. However, it is possible that a groupof malicious users can collude to promote another untrustworthy user in becoming a group member. More-over, revoking membership without the existence of a central authority in a group is also a grant challenge.To address these problems in peer-to-peer MSNs, we propose a decentralized framework and the relatedalgorithms for trusted information exchange and social interaction among users based on the dynamicityaware graph relabeling system. In contrast to the existing implementations of social networks based on aclient/server paradigm, the proposed framework utilizes a lightweight trust model for identifying trustwor-thy users and aims at creating communities of trusted users while isolating and reducing interactions withuntrustworthy users. Simulation results demonstrated the effectiveness of the proposed framework comparedwith the traditional dynamicity aware graph relabeling system algorithm. Copyright © 2011 John Wiley &Sons, Ltd.

Received 2 January 2011; Revised 20 June 2011; Accepted 11 July 2011

KEY WORDS: delay-tolerant networks; mobile social networks; peer-to-peer networking; trust models

1. INTRODUCTION

With the rapid increase in the number of mobile users, the access to various mobile applicationsand services on the Internet has recently been growing at an enormous rate. Popular mobile webbrowsers such as the Opera Mini running on mobile devices has experienced an exponential growthin terms of the number of downloads [1]. Great interests are also being shown to mobile socialnetwork (MSN) services [2] that are being offered by online social networking sites such as Face-book, MySpace, Twitter, etc. Smart phones and personal digital assistants have become a popularchoice for social networking with the help of email, short messaging or by subscribing to a socialnetworking service provider. Typical users of a social network would have a personal public pro-file advertised on the network including information such as personal interests, photos, videos, etc.Users with common interests would subscribe to share in the social environment.

The MSN poses various challenges at two levels. At the networking communications level, thereare many limitations of providing social networking services to users connected to a mobile network.Frequent disconnections because of power exhaustion, poor signal quality and mobility hinder the

*Correspondence to: Basit Qureshi, School of Computing, Informatics and Media, University of Bradford, Bradford BD71DP, UK.

†E-mail: [email protected]

Copyright © 2011 John Wiley & Sons, Ltd.

2056 B. QURESHI, G. MIN AND D. KOUVATSOS

QoS for mobile applications. Knowing the network features ahead of transmission, such as through-put and delay, can help MSN to classify users based on the best performance routes. This leads to theso-called wireless-aware social networks. Many studies have been conducted in providing QoS andperformance evaluation of routing protocols for mobile ad hoc networks (MANETs) [3–6]. At thesecond level, there are many social-aware or social-inspired wireless networks where the knowledgeof social network users is exploited for the benefit of network design. For example, a social networkanalysis for routing in disconnected delay-tolerant MANETs was analyzed in [7]. The methods fordetecting community behavior in delay-tolerant networks, exploiting the benefit of store and for-warding data in socially interactive users, were presented in [8, 9]. A novel technique determiningthe impact of human mobility on the design of opportunistic forwarding algorithms in delay-tolerantnetworks was reported in [10].

Traditionally, social networks are implemented in a client/server environment. In MSN, userssocially interact with each other using handheld mobile devices while on the move; the membershipin a group/community of MSNs is granted by a pre-existing group member. However, it is possiblethat a group of malicious users can collude to promote another untrustworthy user in becoming agroup member. Moreover, revoking membership without the existence of a central authority in agroup is also a grant challenge. Recent advances in semi-decentralized peer-to-peer (P2P) socialnetworks have been presented in [11, 12]. These techniques rely heavily on encryption protocolsin client-to-server communication but provide no security between P2P interactions. Trust manage-ment in a decentralized P2P network is a challenging task in the absence of global knowledge for allusers; any trust or reputation parameters for a user has to be computed locally [13, 14]. Therefore,there is a great need for a framework for trust management in MSNs. The goal of this frameworkis to identify trustworthy users and allow secure transmissions while isolating untrustworthy usersfrom the community.

This paper presents a trust-based framework for membership management in MSNs. The dynam-icity aware-graph relabeling system (DA-GRS) [15, 16] is adopted to label nodes in the networkwith a trust level indicator. These trust labels are used to compute trust ratings at the individual andgroup levels. A group of users utilize these trust level indicators to communicate with new usersand invite them to become members. The goal is to create a groups of users with high trust rat-ings while identifying untrustworthy users and isolating them from the community, thus revokingtheir membership. This proposed method of community-based trust management is more effectivein reducing the amount of computations required at the local level in a distributed environment. Wepropose algorithms based on the greedy concept using the DA-GRS system. Two cost functions tomeasure the trustability of a group of users in a network are also presented. Simulation results showthat trust-based greedy algorithms create a much better quality of trusted groups compared with thetraditional DA-GRS algorithm.

The rest of this paper is organized as follows: in Section 2 we discuss the related work followedby mobile social networking trust requirements in Section 3. The proposed algorithms for MSNtrust management are presented in Section 4. Section 5 discusses the various simulation parametersand analysis of the performance results. Finally, Section 6 concludes this study.

2. RELATED WORK

A trust-based decentralized MSN can be constructed in a delay-tolerant MANET topology. Theliterature review is presented below.

2.1. Delay-tolerant networks

A MANET is a wireless network set up temporarily without a wired infrastructure (routers, switches,servers, cables, access points, etc.) [17]. The wireless nodes in a MANET may move around andneed to forward packets for other components in the network. Because they can be deployed quickly,MANETs can be used for disaster rescue, battlefield communication, sensor networks, etc. Routingprotocols for MANETs focus on establishing routes to the destination. The route is maintained untilthe destination is accessible or until the route is no longer used. This assumption makes MANET

Copyright © 2011 John Wiley & Sons, Ltd. Concurrency Computat.: Pract. Exper. 2012; 24:2055–2068DOI: 10.1002/cpe

TRUSTED INFORMATION EXCHANGE IN P2P MOBILE SOCIAL NETWORKS 2057

routing protocols unsuitable for environments where disconnections are frequent and potentiallylong term [7].

The delay-tolerant network (DTN) is a type of MANET characterized by long delay paths andfrequent disconnections and network partitions [3, 6]. In a DTN, information may be carried bya mobile node and forwarded opportunistically across partitions, therefore allowing communica-tion between areas of the network that are never connected by an end-to-end path. Recently, thesetypes of opportunistic forwarding scenarios has become popular. Mobile nodes enable indirect dataexchange among disconnected portions of the overall network, typically using a store-and-forwardapproach and opportunistic forwarding [5]. Fall [4] presented the architecture for DTN consisting ofan overlay called bundle, which is defined as a number of messages to be delivered together. DTNnodes implement the bundle layer forming an overlay that employs persistent storage to overcomenetwork interruptions. A routing protocol for DTNs based on mobility patterns of nodes was pre-sented in [18]. More recently, a detailed review of challenges of disconnected DTNs was presentedin [19].

2.2. Mobile social networking

Social networks are personal or professional sets of relationships between individuals. MSNs aretechnology-enabled services that adopt wireless and mobile communications to increase the close-ness of the social relationship of users The applications of social network services on the Internethave grown rapidly and recruited a significant number of members. Twitter, MySpace, Facebook,Friendster and Dodgeball [2] are a few examples of commercialized applications. In addition, Plazes[20] is a location-aware interaction system that helps mobile users hook up with friends or otherlike-minded people anywhere on the globe. Jambo Networks [21] uses WiFi-enabled laptops, cellphones, and PDAs to match people within walking distance who have similar interests and wouldlike to meet face to face. It must be noted that most of these online applications are based on thecentralized client/server architecture.

Because of the popularity of online social networks, most service providers have started to extendthis service to mobile devices. Most approaches simply extend the web interface of the socialnetwork to the mobile device. In other words, a user can view the social network through themobile phone without considering the issues of mobile communications. Problems such as lowbattery, poor communication signal, moving out of communication range, interoperability, etc. can,however, cause disconnections to the service provider. Efforts on implementation of decentralizedmobile social networking have been discussed in [22–24]. Users connect to a centralized server toauthenticate themselves and later can communicate in a P2P fashion [12]. Implementation of mobilenetworking applications based on P2P communication model were discussed in [13, 17].

2.3. Trust management

In human society, trust has become the basis of almost all activities. People gradually form mutualtrust and refer to opinions of the third party in assessing the trust [9]. Trust can be regarded as acriterion for making a judgment under complex social conditions and can be used to guide furtheractions. The early stages of trust and security on MANETs relied on authentication, cryptographicencryption and decryption techniques. These schemes for security were shown to be effective; how-ever, they are based on centralized certification authorities. Significant communication overheadsfrom both preprocessing and during processing periods and energy consumption were major chal-lenges, thus rendering these approaches to be poor for delay-tolerant networks. It has recently beenshown that reputation-based techniques are more effective in decentralized mobile networks [25]. Atrust model based on fuzzy recommendation for MANETs was presented in [26]. Moreover, a novelapproach for trust and recommendations in MANETs was proposed in [27].

2.4. Dynamicity aware-graph relabeling system

The DA-GRS is an adaptation of the graph relabeling systems to the paradigm of dynamic and self-organizing networks. The main characteristics of the DA-GRS model — locality and dynamicity —



make it a suitable tool to represent the core mechanism that an application has to deal with in orderto handle an unpredictable changing context.

Figure 1 illustrates the four rules for the DA-GRS algorithm. The DA-GRS algorithm guaranteesto maintain anytime a spanning forest that strives for a spanning tree, using only one-hop con-text information (i.e., it is a purely localized algorithm). The algorithm is composed of four rules,that is,R D fr1, r2, r3, r4g and is based on three operations on a token: circulation, merging andregeneration. Initially, each node has a token J, meaning that each node is a spanning tree in itself,containing exactly one node (itself), and being its own root. When two nodes meet each other, apply-ing rule r3, the two spanning trees merge. Labels 1 and 2 on an edge in the graph mean that it is apart of the spanning tree. The use of two different labels allows a node to know the local route tothe token. When rule r3 applies, one of the two tokens is deleted and one of the nodes is relabeledN, guaranteeing that there is at most one token per tree. Rule r4 codes the circulation of the token ina tree of the forest. When a communication link is broken (i.e., when an edge is deleted), the nodeon the token side has nothing to do with the token maintenance, and simply applies rule r2. Thenode that had the deleted edge label to 1 has lost the route to the token, and is the only one of itsremaining piece of tree to know that. It then regenerates a new token using rule r1.

In the context of MSN in a P2P topology, DA-GRS is an excellent way of constructing and main-taining a decentralized spanning forest of numerous trees of nodes by virtue of a rule-based tokenmanagement. The network is considered essentially a directed graph where edges connect nodes.Each node has a token that is used to make a cluster of nodes by merging with other nodes. Theprocess of creating clusters is further explained in Section 4.

3. MEMBERSHIP IN MOBILE SOCIAL NETWORK

Most of the online social networking services rely on authentication based on centralized certi-fication authorities. Membership in a P2P mobile social network must rely on a decentralizedreputationbased configuration where nodes participate in labeling other nodes with a trust level.Trust management within a partition of a DTN is very difficult because of its dynamicity, decentral-ized nature and nonpermanent connection that can break up into two or more partitions. Any trustmanagement algorithm has to work at the local level because the global knowledge of the networkis scarcely available and cannot be acquired.

3.1. Trust requirements

Each node in the network is assigned with a unique identification, a token for labeling and a trustlevel indicator. The token is an essential part of the DA-GRS labeling system and is primarily usedto randomly merge a node into a group. We consider trust requirements to be a combination ofhuman social trust factors and the QoS in a delay-tolerant disconnected MANET.

Figure 1. Four rules for the DA-GRS algorithm [15].



3.1.1. Social trust and reputation. Trust is one of the most crucial concepts for decisions in makingrelationships in human societies. Trust is indispensible when considering interaction among users inonline societies such as e-commerce, e-government, etc. Many trust-based schemes have been pre-sented in the literature. However, for decentralized applications or networks, trust is defined as beingbased on a history of a user’s encounters with other users [14, 28]. On the other hand, reputation-based systems compute trust based on recommendations from other users of the system [29]. In thispaper we utilize the concept of computing trust for an individual user and a group of users basedon reputation. Section 3.2 discusses the detailed method for computing the trust values for bothindividual users and the user as a part of a group.

3.1.2. Trust as a QoS metric in MANETs. We also define the trust level for a particular node tobe a measure of its QoS. It is based on criteria such as low battery, node being out of range, poorcommunication signal, etc. The trust level of a user is decreased if the user’s device encounters oneof the above problems. Users with the higher trust level have the luxury to stay connected for longerperiods of time and communicate with a large number of users. Such users are able to store andforward data from adjacent nodes while serving as an intermediate router. Nodes with the lowertrust level should not be permitted to store and forward data from other users because of a higherprobability of a failed delivery, and therefore must be isolated from the group.

3.1.3. Gaining membership. We utilize the DA-GRS algorithm to discover and merge a node withothers. Assuming users A and B have discovered each other and are willing to communicate, User Ais already a member of a group X, whereas B seeks membership of this group through A, as shownin Figure 2(a). In this case user B can merge with group X if the tokens of A and B, that is, TX andTY can merge. If B is a part of a trusted group Y, then the groups X and Y can merge into a largergroup Z, as shown in Figure 2(b). It is worth noting that the new group Z now possesses only onetoken TZ .

3.1.4. Trust labeling. The trust level of a node can be assigned in a cooperative manner by thetrusted adjacent nodes based on its degree of connectivity (number of connections) and thresholdfor the remaining uptime. This threshold is determined by a set of factors such as running out ofbattery, node being out of range, poor communication signal or the user’s discretion. The purpose

b

a

Figure 2. Nodes A in group X and B in group Y merge in to group Z.



of introducing the threshold is to discourage connections to intermediate nodes that have a lowerremaining uptime, thus reducing the number of connections to nodes with high degrees. Nodes withthe higher trust levels can connect to a larger set of nodes and share information, whereas nodes withthe lower trust levels are isolated. Trust for a group of nodes is computed using two cost functions,group_cost function and isolation_cost function as detailed in Section 3.2.

3.1.5. Membership revocation. If the trust level of a node falls to 0, it is detached from the groupand its membership is effectively revoked. All members of the group remove the untrustworthy userfrom their individual lists of trusted users.

3.2. Trust computation

We define trust level of a node to take values from 0 (lowest) to 3 (highest). Typically, a node with atrust level of 3 can be connected to a large number of nodes (higher degree) and have a low possibil-ity of disconnection (high threshold) and therefore is more likely to complete its task. Alternatively,a node with a low trust level such as 1 is considered to be an isolated node and must therefore bemarginalized. Table I shows a comparison of various trust levels.

3.2.1. Computing trust for a single user. To compute the trust for a user, we utilize the recom-mendations from other users who have recently been in contact with the intended user. Each usermaintains a list of users with which they have a direct interaction. Every user has an opinion aboutanother user and labels it as trustworthy, unknown or untrustworthy, taking the values +1, 0 and–1, respectively. Typically, a user may trust or distrust another user; a new user having no previousencounters with a trusted user is labeled as unknown, that is, 0. The trust of a user is computed bythe following equations

T .x/D†i�t_list .Trust.i/ � opinioni .x//

†i�t_listTrust.i/(1)

where x is the node whose trust is to be computed, i is a node in the list of trusted users (t_list)and the function opinioni .x/ indicates the opinion of user i towards user x. The value for T .x/ isalways in the interval (1, –1). A trustworthy user will obtain a positive value, whereas a negativevalue indicates an untrustworthy node. Trust.x/ labels the node x with a trust value based on thevalue of T .x/ given by

Trust.x/D

8̂<:̂

3 1> T .x/> 0.52 0.5 > T .x/ > 01 0 > T .x/> �0.50 �0.5 < T .x/ < �1

(2)

The trust values for all users in contact are stored in the t_list and are updated frequently. If a trustrating is requested for a particular user, the latest value stored in t_list is forwarded to the requestinguser.

3.2.2. Computing trust for a group of users. If user A requests to join group X and gain its mem-bership, its token TA is to be merged into group X’s token TX . In this case, the group trust levelis computed by the user possessing the token TX . The trust level for a group is computed by two

Table I. Definition of trust levels in nodes of the network.

Trust level Degree Threshold Example

3 High High Trustable store and forward intermediate node2 Low High Trustable intermediate node1 High Low Isolated node0 Low Low Node with membership to be revoked



cost functions group_cost() and isolation_cost(). The trust level for the whole group indicates thequality of the trusted group therefore a higher value indicates a desirable trusted group. We com-pute the values of these cost functions to compare the trust values of groups in various environmentsettings.

� Group_cost() function: This function computes the cost of trust for the group. The cost of groupG is determined by two factors, degree of trusted connections and trust level for each node inG. It is given by

Group�cost.G/D†.trust.x/ � t�conn.x// (3)

where t_conn() for a node x is the number of trustable connections to other nodes and trust(x)indicates the trust level of node x. As an example, the Group_cost() for the group shown inFigure 3(a) is 16. Similarly, for the group in Figure 3(b), the group cost computed is 21. Thisshows that the group of users in Figure 3(b) has a higher trustability compared with the groupin Figure 3(a). Having connections with nodes that have a higher trust level is desirable forlong-term communication. Node D in Figure 3(a) has a trust level of 3 and has three activetrustable connections so therefore is more trustable than node A in Figure 3(b), which has atrust level of 1 and three active connections. To have an optimal trust level in a group, nodeswith lower trust levels should be isolated with a minimum number of connections, whereas thehigher trust level nodes should be allowed to establish more connections.

� Isolation_cost() function: To create the better quality trusted groups we try to isolate nodeswith low trust levels (trust level 61) and low number of connections. We simply compute thegroup_cost() for low trust nodes in the group and subtract this from the group_cost of thatgroup. As an example, the isolation_cost for group G in Figure 3(a) would be 13, whereas inFigure 3(b) it is 16.

b

a

Figure 3. Examples of a group in a MSN.



4. ALGORITHMS FOR TRUST MANAGEMENT

Because of the decentralized nature of mobile social networking in a delay-tolerant environment,maintaining trust management in groups of nodes at the global level is very difficult. Instead, a trustmanagement algorithm must work at the local level. The proposed algorithms modify the DA-GRSto construct groups of nodes in the MSN and compute trust in the network.

4.1. Dynamicity aware-graph relabeling system

The DA-GRS algorithm is improved for trust management in this study. The trust level of a groupis computed whenever a user/node seeks to communicate to another user in a group. In otherwords, the tokens of the two nodes willing to communicate are compared. If the trust levels andthe group_cost() and isolation_cost() values are acceptable, the merger is completed and a largergroup is formed. Consider Figure 4 as an example, Node A in group X has a trust level of 3; thegroup_cost() value is 27 and the isolation_cost() value is 21. Node B in group Y has a low levelof trust where the group_cost() is 16 and isolation_cost() is 6. Node A has a higher trust level ingroup X that has the higher group trust level compared with node B in group Y. Also in group Y, theratio of group_cost() versus isolation_cost() is 15 to 6, indicating a high percentage of nodes thathave a low level of trust and are isolated in the group. The DA-GRS algorithm in this case wouldallow groups X and Y to merge. It must be noted that this algorithm does not consider the trust ofindividual nodes or the group trust level while merging.

4.2. Greedy labeling

The proposed Greedy DA-GRS algorithm is an improvement of the DA-GRS algorithm by addingthe greedy concept. The idea behind this concept is to select a group from a set of groups that hasthe highest group trust level and merge with it. Algorithm 1 shows the greedy labeling using DA-GRS. The algorithm initializes the values of the token with highest trust rating, referred to as Tbest ,and can be determined in O.n/ time. If no Tbest can be found, that is, the current trust rating isthe highest, then the token should be forwarded as explained in the DA-GRS algorithm previously.Otherwise the token will allow the node to merge with the node with Tbest . As an example of thegreedy labeling algorithm, in Figure 4, the algorithm would merge node B with group X containingnode A. Node B with a trust level of 1 would prefer to merge with node A which has a higher trustlevel of 3 instead of node C which has a trust level of 2. The greedy labeling algorithm improves theoverall group trust level.

Figure 4. Merging of Groups. Each node in a group has a token, a trust level, the group trust cost g(x) andisolation trust cost i(x).



Algorithm 1 Greedy DA-GRS algorithm.1: initialize Tbest //Token2: initialize X as list of neighbors, where node x � X3: Tbest best�trust(X)4: if Tbest¤ null then5: merge�with�group(Tbest, Tx)6: else7: move�token(Tx)8: end if

4.3. High group trust labeling

The high group trust (HGT) labeling algorithm focuses on group level trust rather than merging anode’s trust level. A group with a higher level of group_cost() value can be considered a robusttrusted group with a long duration of time to live. In other words, in terms of performance, thegroup has the longest available connection time and therefore is more reliable. Algorithm 2 showsthe process of labeling using HGT. The process for determining Tbest is similar to the greedy algo-rithm. Computing the group_cost() adds complexity to this algorithm. If the computed group_cost isless than the determined Gbest value, the node or group would merge with the group having Gbest .Gbest is determined in a similar manner to Tbest and takes O.n/ time. As shown in Figure 4, nodeB prefers to merge with group X with a group cost of 27 rather than group Z with a group cost of10. Larger groups with higher group trust cost can be considered most reliable. This algorithm isessentially a greedy algorithm based on DA-GRS where the group_cost of a group is considered forcomparison instead of individual node trust level.

4.4. Optimal group trust labeling

The optimal group trust (OGT) labeling algorithm focuses on the quality of group trust. A groupwith the lowest percentage of isolated nodes is preferable to larger groups with a high percentageof isolated nodes. Algorithm 3 shows the process for OGT-labeling algorithm. Similar to HGT, thevalues of Tbest , Gbest , and Ibest are determined in O.n/ time, where Ibest represents the best iso-lation_cost. If the difference in group_cost and isolation_cost is less than the difference in values ofGbest and Ibest then the node or group merges with the group having Ibest . To explain OGT further,Figure 4 can be taken as an example where group X has a ratio of 21 to 27, group Y has a ratio of6 to 15 and group Z has a ratio of 8 to 10. This indicates that group Z has the highest optimal trustvalue, that is, the least number of isolated nodes. The OGT algorithm is also a greedy algorithmbased on DA-GRS. It focuses on quality of trusted groups in terms of group trust coherence.

Algorithm 2 Greedy high group trust algorithm.1: initialize Tbest, Gbest //Token, Group best value2: initialize X as list of neighbors, where node x � X3: initialize group�cost compute�g�cost(x)4: Tbest best�trust(X)5: if Tbest¤ null and group�cost < Gbest then6: merge�with�group(Tbest,Tx)7: else8: move�token(Tx)9: end if



Algorithm 3 Greedy optimal group trust algorithm.1: initialize Tbest, Gbest, Ibest //Token, Group best value, Isolation best value2: initialize X as list of neighbors, where node x � X3: initialize group�cost compute�g�cost(x)4: initialize isolation�cost compute�i�cost(x)5: initialize diff group�cost - isolation�cost6: Tbest best�trust(X)7: if Tbest¤ null and diff < ( Gbest - Ibest )then8: merge�with�group(Tbest,Tx)9: else

10: move�token(Tx)11: end if

5. SIMULATION EXPERIMENTS AND RESULTS

In the simulation experiments, each user in a MSN was equipped with a mobile device that had anOmni directional transmission range. Users were mobile and could communicate and stay connectedwhile on the move. Simulations in this work considered three real-world environment categoriesthat were selected in terms of mobility and concentration of users. Users in a university campusand office building floor were considered to be less mobile. Users in a shopping mall and on citystreet networks were considered to have the higher mobility. To ensure the validity of simulation,we generated three different networks for each category of environment (12 networks in total).Table II shows the properties of each of these networks. We considered that each network consistedof 100 users. The total duration for each simulation was 20 s with 40 simulation steps taken at 0.5 sintervals.

The simulation duration was selected carefully to reflect changes in networks that have the highermobility (street network). Changes in city street networks are more frequent than in campus, shop-ping mall or office building floor networks. Figure 5 shows an example of each of the four types ofnetworks.

As stated before, determining an optimal trusted group for a decentralized dynamic network isextremely difficult. However, because networks used in this study were generated using the simu-lator, the configuration of a network can be predetermined. Therefore, the robustness of suggestedalgorithms can be evaluated by calculating the group_cost function and the isolation_cost functionof each of these networks. For each of these networks we ran the simulation 400 times. The averagepercentage of trusted nodes (T .x/ >1) was 30, in this simulation. We also compared the number ofuntrustworthy nodes (trust(x/ 6 1) and investigated whether these were successfully isolated from

Table II. Properties of four sets of each category of networks (campus, shopping mall, city street, andbuilding floor). Total number of users in each network is 100.

Campus1 Campus2 Campus3 Mall 1 Mall 2 Mall 3

Max no. of connections 20 40 60 20 40 60Min no. of connections 0 0 0 1 1 1Avg. no. of connections 5.8 19.1 33.2 4.2 17.3 28.6Total no. of connections 708 1045 1389 688 943 1073

Street 1 Street 2 Street 3 Building1 Building2 Building3

Max no. of connections 50 70 90 20 40 60Min no. of connections 2 2 2 0 0 0Avg. no. of connections 9.2 11.6 12.8 22.5 32.1 38.2Total no. of connections 322 379 437 813 1450 1883



c d

a b

Figure 5. Examples of Network used (a) Campus Network, (b) Shopping Mall (c) City Street and (d) Officebuilding floor.

groups. It is worth noting that a node with a trust(x/ value less than 1 is considered untrustworthyand its membership is revoked from the group.

5.1. Results for campus networks

Table III shows the results for the average values of group and isolation cost functions for thesuggested algorithms. The campus network was chosen because of its low mobility and high con-nectivity feature. The results reveal that greedy labeling algorithm yields the highest group cost,thus resulting in the most number of trustable groups. The percentage of successfully isolated nodeswith trust(x/ less than 1 is 83%. The isolation cost for the HGT algorithm is higher than the greedyalgorithm therefore resulting in a better quality group. It must be noted that the group cost for theOGT algorithm is lower than both greedy and HGT algorithms but it provides the best isolation costthus creating the best quality trust groups. The OGT provides the highest percentage of successfullyisolated untrustworthy nodes.

Table III. Averages of group and isolation cost functions for campus networks.

Campus network

Group�cost Isolation�cost Percentage of isolated nodes

DA-GRS 559.2 455.3 30%Greedy labeling 683.3 581.4 83%HGT 635.6 588.1 92%OGT 621.0 603.9 97%



5.2. Results for shopping mall networks

Results for the averages of group and isolation cost functions for shopping mall networks can beseen in Table IV. The shopping mall networks have slightly higher degrees of mobility comparedwith campus networks. Because of higher mobility, the average numbers of connections are lower.It can be seen from the results that greedy labeling algorithm performs better compared with HGTand OGT algorithms in creating trustable groups. Nevertheless, the ratio of group cost and isolationcost function is lower for greedy algorithm, which indicates that the trust quality of groups is lower.The comparison of the ratio of group cost and isolation cost for all algorithms indicates that OGTperforms better in terms of creating high quality trusted groups. It can also be seen that the groupcost function for HGT yields almost similar values for OGT.

5.3. Results for city street networks

The results for the averages of group and isolation cost functions can be found in Table V. Usersmoving on a city street are considered to be highly mobile compared with the earlier defined net-works. Results show that the dynamicity of the network yields fewer trusted connections thereforethe average cost function values are lower compared with campus and mall networks. An interestingfact observed in the simulation indicates that because of the higher mobility the group cost for OGTis not similar to HGT. A possible reason could be the decrease in performance owing to the costof computing the ratios. Apart from this issue, OGT still performs better in terms of creating betterquality trusted groups.

5.4. Results for office building floor network

Table VI shows the results for the office building floor network. Users in an office building floor areconsidered to have low mobility. Because of the lower mobility rates, the impact of mobility on thenumber of connections and disconnections is lower; therefore overall, a higher number of connec-tions are made. Greedy algorithm performs better compared with both HGT and OGT in creatingtrusted groups. Also, the isolation rate for greedy algorithm is comparable yet slightly lower thanHGT. OGT has the lower value for group cost function but the ratio between group cost and isolationis better than the others. The results show that the OGT algorithm performs better compared withother algorithms in building better trusted groups.

Table IV. Averages of group and isolations cost functions for shopping mall networks.

Shopping mall network



Table V. Averages of group and isolation cost functions for city street networks.

City street





Table VI. Averages of group and isolation cost functions for office building floor networks.

Office building floor



6. CONCLUSIONS

Mobile social network services implemented by popular online social networking organizationsextend the user interface to mobile devices without addressing the inherent problems caused bymobile communications. To date very few decentralized MSNs have been implemented because ofthe enormous challenges posed by the dynamic nature of networks. Trust management in dynamicdecentralized mobile networks is receiving attention because of its immense application. This paperpresents a decentralized framework and the related algorithms for decentralized trust managementin P2P MSNs based on a dynamicity aware graph relabeling system. The proposed algorithms arebased on the greedy concept and the performance results affirm their benefits.

Although simulating human behavior for trust and reputation assignment is unpredictable, wepresented a method to compute trust of users based on a reputation model where users give theiropinion about other users. Two cost functions to measure the trustability of a group of users werepresented. The simulation results showed that the trust-based greedy algorithms create the betterquality of trusted groups compared with the traditional DA-GRS algorithm. The results of extensivesimulation experiments also revealed the performance of the proposed algorithms when tested inscenarios such as a campus, a shopping mall, an office building floor and a city street. The pro-posed solution helps identify users with low trust ratings, isolate untrustworthy users and effectivelyrevoke their membership.

REFERENCES

1. Opera Mini website http://www.opera.com/mini/ [5 June 2010].2. Ziv ND, Mulloth B. An Exploration on Mobile Social Networking: Dodgeball as a Case in Point. Proceedings of the

International Conference on Mobile Business, ICMB ’06, 2006; 12–21.3. Jain S, Fall K, Patra R. Routing in a delay tolerant network. In Proceedings of the 2004 conference on Applications,

technologies, architectures, and protocols for computer communications (SIGCOMM ’04). ACM: New York, NY,USA, 2004; 145–158.

4. Fall K. A delay-tolerant network architecture for challenged internets. In Proceedings of the conference on Applica-tions, technologies, architectures, and protocols for computer communications (SIGCOMM ’03). ACM: New York,NY, USA, 2003; 27–34.

5. Zhang Z, Zhang Q. Delay/disruption tolerant mobile ad hoc networks: latest developments. Wireless Communicationsand Mobile Computing 2007; 7(10):1219–1232.

6. Barbosa L, Siqueira I, Loureiro AA. Evaluation of ad hoc routing protocols under a peer to peer application.Proceedings of IEEE Wireless Communications and Networking Conference, WCNC, 2003; 1143–1148.

7. Daly E, Haahr M. Social network analysis for routing in disconnected delay-tolerant MANETs. Proceedings of the8th ACM international symposium on Mobile ad hoc networking and computing (MobiHoc), 2007; 32–40.

8. Hui P, Crowcroft J, Yoneki E. BUBBLE rap: social-based forwarding in delay tolerant networks. Proceedings of the9th ACM international symposium on Mobile ad hoc networking and computing, Hong Kong, China, 2008; 241–250.

9. Raento M, Oulasvirta A. Privacy management for social awareness applications. Proceedings of 1st Workshop onContext Awareness for Proactive Systems — CAPS 2005, Helsinki, Finland, 2005; 105–114.

10. Hui P, Yoneki W, Chan S, Crowcroft J. Distributed community detection in delay tolerant networks. Proceedings of2nd ACM/IEEE international workshop on Mobility in the evolving internet architecture, 2007; 55–62.

11. Kellerer W, Despotovic Z, Michel M, Hofstatter Q, Zols S. Towards a Mobile Peer-to-Peer Service Platform.IEEE/IPSJ International Symposium on Applications and the Internet Workshops, pp. (SAINTW’07), Hiroshima,Japan, 2007; 2–10.

12. Tsai FS, Han W, Xu J, Chuan-Chua H. Design and development of a mobile peer-to-peer social networkingapplication. International Journal of Expert Systems with Applications 2009; 36(8):11107–11087.



13. Sabater J, Sierra C. Review on computational trust and reputation models. Artificial Intelligence Review 2005;24(1):33–60.

14. Hang C, Wang Y, Singh M. An adaptive probabilistic trust model and its evaluation. In Proceedings of the 7th inter-national joint conference on Autonomous agents and multiagent systems - Volume 3 (AAMAS ’08), Vol. 3, 2008;1485–1488.

15. Casteigts A, Chaumette S. Dynamicity aware graph relabeling systems (da-grs), a local computation based modelto describe MANET algorithms. International Conference on Parallel and Distributed Computing Systems, 2005;231–236.

16. Casteigts A. Model driven capabilities of the da-grs model. Proceedings of the International Conference onAutonomic and Autonomous Systems ICAS ’06, 2006; 24–30.

17. Wu J. Peer-to-Peer Overlay Abstractions in MANETs, Handbook on Theoretical and Algorithmic Aspects of SensorAd Hoc Wireless, and Peer-to-Peer Networks. CRC Press: Boca Raton, FL, USA, 2005.

18. Leguay J, Friedman T, Conan V. Evaluating Mobility Pattern Space Routing for DTNs. In Proceedings of the IEEEINFOCOM 2006, Vol. 5, April 2006; 2540–2549.

19. Daly E, Haahr M. The challenges of disconnected delay-tolerant MANETs. Ad Hoc Networks 2010; 8(2):241–250.20. Plazes. http://www.plazes.com [Jan 2010].21. Jambo Networks. www.jambo.net/website/Home.html [Jan 2010].22. Lugano G, Saariluoma P. To Share or not to share: Supporting the user decision in Mobile Social Software applica-

tions. In Proceedings of the International User Modelling conference (UM 2007; Corfu, Greece, 25–29 July). LectureNotes in Computer Science, volume 4511. Springer: Berlin, 2007; 440–444.

23. Chang Y, Liu H, Chou L, Chen Y, Shin H. A General Architecture of Mobile Social Network Services. Proceedingsof the International Conference on Convergence Information Technology, Ggyeongju, South Korea, 2007; 151–156.

24. Beach A, Gartrell M, Akkala S, Elston J, Kelley J, Nishimoto K, Ray B, Razgulin S, Sundaresan K, Surendar B,Terada M, Han R. WhozThat? evolving an ecosystem for context-aware mobile social networks. IEEE Network 2008;22(4):50–55.

25. Merwe J, Dawoud D, McDonald S. A survey on peer-to-peer key management for mobile ad hoc networks. ACMComputer Surveys 2007; 39(1):1. DOI: 10.1145/1216370.1216371

26. Luoa J, Liub X, Fana M. A trust model based on fuzzy recommendation for mobile ad-hoc networks. ComputerNetworks 2009; 53(14):2396–2407.

27. Balakrishnan V, Varadharajan V, Tupakula U, Lucs P. Trust and Recommendations in Mobile Ad hoc Networks.Proceedings of the Third International Conference on Networking and Services, 2007; 64–70.

28. Mui L, Mohtashemi M, Halberstadt A. A computational model of trust and reputation for e-businesses. In Proceed-ings of the 35th Annual Hawaii International Conference on System Sciences (HICSS’02), Vol. 7. IEEE ComputerSociety, 2002; 188–196.

29. Yu B, Singh M. A social mechanism of reputation management in electronic communities. In Proceedings of the4th International Workshop on Cooperative Information Agents, The Future of Information Agents in Cyberspace.Springer-Verlag, 2000; 154–165.


Trusted information exchange in peer-to-peer mobile social

Documents

Transcript of Trusted information exchange in peer-to-peer mobile social