=Paper= {{Paper |id=Vol-2404/paper11 |storemode=property |title=Supporting Agent CoT Groups Formation by Trust |pdfUrl=https://ceur-ws.org/Vol-2404/paper11.pdf |volume=Vol-2404 |authors=Giancarlo Fortino,Lidia Fotia,Fabrizio Messina,Domenico Rosaci,Giuseppe M.L. Sarné |dblpUrl=https://dblp.org/rec/conf/woa/FortinoFMRS19 }} ==Supporting Agent CoT Groups Formation by Trust== https://ceur-ws.org/Vol-2404/paper11.pdf

Workshop "From Objects to Agents" (WOA 2019)

Supporting Agent CoT Groups Formation by Trust
Gianfranco Fortino∗ , Lidia Fotia§ , Fabrizio Messina† , Domenico Rosaci‡ , Giuseppe M. L. Sarné §
∗ Department DIMES, University of Calabria, Via P. Bucci, cubo 41c, 87036 Rende (CS) giancarlo.fortino@unical.it
§ Department DICEAM, University of Reggio Calabria, Loc. Feo di Vito, 89122 Reggio Cal., Italy, {lidia.fotia, sarne}@unirc.it
† Department of Mathematics and Informatics, University of Catania, Viale Andrea Doria Catania, Italy, messina@dmi.unict.it
‡ Department DIIES, University of Reggio Calabria, Loc. Feo di Vito, 89122 Reggio Cal., Italy, domenico.rosaci@unirc.it

Abstract—IoT devices dealing with complex tasks require The basic idea is that, the generic consumer agent, when
powerful hardware capabilities or to get resources on the cloud. using some data services (s) from a provider agent, should
When an IoT device is “virtualized” on the Cloud, it can rely on consider its past experiences. When no data about paste
one or more software agent that can exploit its social attitude to
interact and cooperate. In this context, the choice of a partner experience does exist, the agent will exploit the recommen-
to cooperate is a sensitive question but when an agent cannot dation given by the community [15], [16]. In particular, the
perform a reliable choice then, like real communities, it can agents belonging to the same group of the agent who has
ask information to other agents it considers as trustworthy. This requested the opinion/recommendation of the provider agent,
process can be improved by partitioning the agents in groups by will provide the information for free, otherwise a fee has to
using trust relationships to allow agents to interact with the most
reliable partners. To this aim, we designed an algorithm to form be paid for the recommendation/opinion. This approach leads
agent groups based on reliability and reputation information to a competitive scenario on which groups/agents are inter-
and the results of some simulations confirmed its potential ested in accepting/belonging to those agents/groups having
advantages. a high reliability and helpfulness. Moreover, to evaluate the
Index Terms—Cloud Computing; Cloud of Things; Internet of helpfulness of an agent we consider the effectiveness of its
Things; Multiagent system; Reputation; Trust; Voting
recommendations, while for a group it is the average of the
helpfulness of its members.
I. I NTRODUCTION
In order to maximize the benefits of an agent to join
Recently, the “Internet of Things” (IoT) and Cloud Com- with a group (and vice versa), we exploit trust measures to
puting (CC) converged into the so called Cloud-of-Things [1], model a distributed group formation process. In particular,
[2] (CoT) for supporting computational and storing require- we designed a distributed algorithm matching devices and
ments [3] of ubiquitous and heterogeneous IoT devices, also groups to improve individual and global satisfaction [9], [17]
in nomadic scenarios [4]. Moreover, to promote cooperation into the CoT on the basis of trust measures considering the
among IoT devices, they can be also associated with software agent helpfulness in providing useful recommendations. In
agents [5]–[8] for taking benefit from their social attitudes. this respect, as it happens in real user communities, in place
In this context, the choice of a reliable partner needs of of the global reputation, we adopt a local reputation [18]
suitable information that can be also required as recommen- approach where the reputation value is based on the opinions
dations to trustworthy agents. To this aim, we propose of coming from the friends (or friends of friends and so on)
supporting this process by encouraging agents to form groups of an agent. This local approach gives important benefits in
of reliable recommenders exploiting some type of social a CoT context, among which i) heavy computational tasks
relationships existing among the group members [9], [10]. and communication overloads can be avoided when collecting
For instance, an important property within a community is a opinions and evaluating the trustworthiness of their sources
high level of mutual trustworthiness among its members [11], and ii) the system reactivity is increased.
[12]. Therefore, we consider the trust-based processes to form Moreover, likely processes having place in human soci-
agent groups of reliable recommenders over a CoT context as eties [19], groups are formed by using a voting mechanism,
potentially capable to significantly improve the IoT devices where each vote combines reliability and local reputation
activities. measures. Finally, to form groups with a high level of mutual
To this aim, we consider a CoT environment where trust among its members we designed a distributed algorithm
heterogeneous devices consume/produce services and/or ex- for group formation (see Section III) that we verified, in terms
tract/exchange knowledge assisted by personal software agents of efficiency and effectiveness, by means of some experiments,
working over the CC. We take into account a specific scenario on a simulated agent CoT scenario, which confirmed our
where each IoT device and its associated agent are considered expectations.
a single entity; moreover, we also take into account the The rest of the paper is organized as follows. Section II
dynamicity of agents in the CoT environment, i.e. their ability describes the adopted local trust model and voting mechanism,
to change groups based on their own convenience [13], [14]. while Section III presents an algorithm to form groups. The

71
Workshop "From Objects to Agents" (WOA 2019)

b
ωi,r = 2−( l(i,r) −1)
where bl(i,r) is the shortest path between ai and the recom-
mender agent ar . Now, by assuming that ai , in its ego-network,
is able to exploit a number p of recommenders to receive
recommendations about aj , then σi,j can be calculated as:
p
1 X
σi,j = · ǫi,r · ωi,r · recr,j .
p r=1
The trust measure that an agent ai has about an agent aj
Fig. 1. The ego-networks of the agent a including all the nodes of the virtual can be computed by combining reliability and local reputation
community (nodes from a to f ) for which a direct link (red colored) to a (which also takes into account the helpfulness) as:
there exists and some other agents indirectly connected to a by a path of
length 2 (e.g., all the agents connected to a by red and blue links). τi,j = αi · ρi,j + (1 − αi ) · βi · σi,j
where α and β are two parameters ranging in [0, 1] ∈ R. The
experimental results are dealt in Section IV and in Section V parameter α simply weights reliability and local reputation for
the related literature is presented. Finally, in Section VI some giving more or less relevance to one or other. The parameter
conclusions are drawn. β is computed as βi = p/kEi (x)k and takes into account the
dependability of σi,j on the number of p nodes belonging to
II. T HE L OCAL T RUST M ODEL Ei that contributed to compute σi,j (indeed, if the number of
these nodes is small then the local reputation measure loses
For convenience, we represent the agents trust relationships
of relevance because ai will not have a sufficient information
as a graph G, in which a direct edge linking two nodes (i.e.,
from its Ei about aj ). Note that for a newcomer agent, suitable
agents) is associated with the trust level (ranging [0, 1] ∈ R,
“cold start” values of reliability, reputation and helpfulness are
where 0/1 means the minimum/maximum value) an agent has
adopted.
in another agent, and the ego-network Ei of an agent ai ∈ A
The “trustworthiness” of a group g, as perceived by ai (i.e.,
as a sub-graph Ei ⊆ G including those nodes (i.e., agents)
τi,g ), is determined by simply averaging all the trust measures
connected to ai in a fixed depth (see Figure 1).
computed by ai for all the agents belonging to g. Similarly,
For the generic nodes i, j ∈ G (i.e., the associated agents ai
the “trustworthiness” of an agent ai , as perceived by a group
and aj ), the measure of the local trust τi,j that i has about j
g (i.e., τg,i ), is obtained by averaging all the trust measures
combines the reliability ρi,j (i.e., a measure of the confidence
about ai computed by all the agents belonging to g.
that ai has about the capability of aj of providing good
suggestions) and the local reputation σi,j (i.e., a measures of Finally, when a decision about a new membership with a
how much, on average, the agents of Ei estimate the capability group g has to be taken, all the agents belonging to g give a
of aj of having good interactions). preference (i.e., a vote) v ∈ {0, 1} to accept or not this agent
into g (e.g., 0/1 means “not accept” “accept”) [20]. The vote
Usually, ρi,j 6= ρj,i is an asymmetric measure computed as:
depends from i) the local trust measure that the voter computed
q
1 X k about the candidate, also exploiting the recommendations com-
ρi,j = · fi,j ing from its ego-network and ii) a suitable threshold Γg ∈ [0, 1]
q
k=1
that worth 0 (i.e., 1) if τ < Γg (i.e., τ ≥ Γg ). In the following,
by means of all the feedback fi,j ∈ [0, 1] ∈ R assigned by ai we represent the voting process referred to a group g for a
to aj for each of the q interactions carried out with it. To this potential new member y by adopting the voting criterion v
aim, let recr,j ∈ [0, 1] be the suggestion given by ar about proposed above, as the output of a function V (g, v, y). For
aj and let ǫi,r ∈ [0, 1] be the (average) helpfulness perceived instance, a reasonable strategy may be of adopting a majority
by ai about the capability of ar to provide suggestions1 . In criterion for accepting a requester into a group.
detail, the helpfulness ǫi,r of ar perceived by ai is computed,
with respect to the feedback released by ai for each of the m III. T HE D ISTRIBUTED AGENT G ROUPING A LGORITHM
accepted suggestions provided by ar to ai about other agents,
as: This section presents the distributed agent grouping algo-
m rithm formed by two procedures respectively executed by each
1 X agent: i) belonging to the CoT for finding the “best” groups to
ǫi,r = · |fs − recs |
m s=1 join with, in terms of average value of τi,g (where g identifies
To give relevance to the recommender agents in Ei which a generic group); ii) acting as group administrator to evaluate
are closer to ai , it is used a parameter ω computed as: if affiliating a new member with the group itself based on the
mutual trust among the group members and the potential new
1 If any recommendation was provided by a to a , then the helpfulness of
r i
member. The symbols used in the description of the algorithm
ar perceived by ai will be ǫi,r = 0. are listed in Table I.

72
Workshop "From Objects to Agents" (WOA 2019)

TABLE I Algorithm 1 The procedure executed by a CoT agent.
TABLE OF THE MAIN SYMBOLS Input: Hi ⊂ Gr, W, πi , θi ; Y = {g ∈ G} a set of groups randomly
T S
selected : kY k = M ≤ W , Hi Y = { }, Z = (Hi Y )
Symbol Description
A set of agents associated to the IoT devices 1: for g ∈ Z : t̂g > πi do
G graph representing the set of agents and their relationships G = hN, Li 2: Compute τi,g by exploiting the agents belonging to g.
Ei set of agents belonging to the ego-network of ai , with E ⊆ G 3: end for
Gr set of all the groups S 4: m ← 0
Hi set of the groups which ai is affiliated, with Hi = gi ⊆ A
Kg set of agents affiliated with a group g 5: Let be Sc = {g ∈ Z : τi,g ≥ θi }, with kSc k = W
M maximum number of new groups the single agent is able to analyze 6: for all g ∈ Sc : g 6∈ Hi do
W maximum number of groups that an agent can join with 7: send a join request to the agent administrator of g
R maximum number of agents belonging to a group 8: if g accepts the request then m ← m + 1
Sc set of candidate groups
V (·) voting function 9: end if
Y set of groups randomly chosen, with kY k ≤ M 10: end for
a agent 11: for all g ∈ Hi : g 6∈ Sc do
ag agent administrator of the group g 12: Sends a leave message to g
g generic group
t̃ time elapsed from the last execution of the procedure for an agent
13: m←m−1
t̂ time elapsed from the last execution of the procedure for a group 14: if (m==0) then break
θ threshold on the level of trust between an agent and a generic group 15: end if
φ time threshold fixed by the agent administrator of a group
16: end for
π time threshold fixed by an agent
τ trust
Algorithm 2 The procedure executed by a group administrator.
S
Input: Kg , R, ai , φ, X = Kg {ai };

a) Algorithm 1: It is executed by the agent ai to improve 1: for all k ∈ Kg do
2: if t̃i ≥ φ then ask to k for updating local trust values of ai
its configuration of groups in terms of overall mutual trust 3: end if
with the related peers. More in detail, let Hi ⊂ Gr be the 4: end for
set of the groups which ai belongs to and for which ai stores 5: if kXk < R then
the local trust measure τi,g of each group g ∈ Hi ⊂ Gr 6: if V (g, v, ai ) == 1 then Send an accept message to ai
contacted in the past and let t̂g be the time elapsed from its 7: else Send a reject message to ai
8: end if
last updating. Moreover, let W be a parameter specifying the 9: else
maximum number of groups that an agent can join with, let 10: for all k ∈ X do compute τk,ai
M be the maximum number of groups the generic agent is 11: end for S
capable to analyze, let πi be a time threshold fixed by the 12: Let X ′ = {k1 , k2 , . . . , kkKg k+1 } with ki ∈ X {ai },
agent ai and, finally, let θi ∈ [0, 1] be a threshold on the trust ordered by trust with τg,m ≥ τg,n iff m < n
13: if X[kKg k + 1] == ai then Send a reject message to ai
value between the agent ai and the generic group g ∈ Hi . 14: else
Firstly, the values of τi,g are updated if older than πi (lines 15: Send a leave message to the node X[kKg k + 1]
1-3). Then, it is built a set of candidate groups Sc , with kSc k < 16: Send an accept message to ai
W , sorted in decreasing order based on the values τi,g of the 17: end if
18: end if
groups, while Y Sis a set of groups randomly chosen and with
the set Z = Y H. The sets Y , Z and Sc might store the
groups already belonging to Hi , while some others might be 1) ||X|| < R (line 6), then all the agents in g give a vote.
new groups that were selected at random and put into the set The function V (·), see Section II, combines all the votes
Y . Based on the groups in Sc not belonging to Hi , the agent ai to determine if the agent ai is admitted or not in g.
could improve the quality of its choices by joining with those 2) ||X|| = R and the agent ai is admitted into the group in
groups. The two loops in lines 6-16 represents the kernel of place of another agent. To make comparable the agents,
the procedure, after that Hi = Sc . a natural measure is the trust of the group vs the agent
b) Algorithm 2: It is performed by the administrator ag itself, which is computed as explained in Section II (line
of a CoT group g once an agent, denoted as ai , sends a join 16). In particular, τg,n denotes the current value
S of trust
request to ag . Let Kg ⊂ Gr be the set of the agents affiliated between the group g and the agent kn ∈ X {ai }.
to g, where kKk ≤ R (with R the maximum number of agents S The first scenario is dealt with in lines 6 − 11, while the
allowed to be affiliated with g), let the set X be X = Kg ai , second one into lines 12 − 18 of Algorithm 2.
where ai is the agent candidate to be affiliated with g and let
φ a time threshold fixed by the administrator ag . Moreover, IV. E XPERIMENTS
the administrator ag of a group g stores the values of the local Some experiments have been carried out to test the capabil-
trust computed by the members of its group for ai which desire ity of the proposed algorithm to form groups having a higher,
to join with, and the timestamp t̃i of its retrieval. in average, mutual trust among their members of that obtained
Firstly, the administrator ag asks to the members of its group from different compositions. The reader may refer to Table II
the updated local trust values about ai (lines 1 − 5), then if: for the list of experimental parameters.

73
Workshop "From Objects to Agents" (WOA 2019)

4 TABLE II
LP
3.5 HP E XPERIMENTAL S ETTINGS

3 Parameter Value
General
2.5 No. of Agents (kAk) 1000
No. of Feedbacks per step (Poisson distrib.) λ = 50
P.D.F.

2 Agents Performance (Reliability and trust)
Low Performance (Normal Distribution) mean = 0.9; stdDev = 0.1
1.5 High Performance (Normal Distribution) mean = 0.2; stdDev = 0.1
1 Cold start value of trust 0.5
Ratio of reliable/unreliable agents 0.5
0.5 Group formation
K (Max no. of agents per group) 20
0 M (Max no. of groups an agent analyzes) {5, 10, 15, 20}
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 kGrk (No. of groups) 50
Feedback lM ax (Maximum recommender distance) {1,2}
θ (Minimum value of trust for a group to be 0.2
selected as candidate for group formation)
Fig. 2. a) Generated feedback values.

execution of Algorithm 1, which are then mixed with groups
A network of 1000 different CoT agents (each one as- already present in the set Hi , in the new set Sc . Therefore,
sociated with a IoT device), 1000 initial trust relationships the higher the parameter M , the higher the number of new
and |Gr| groups, randomly formed, were generated. Trust groups analyzed in the algorithm 1, the higher the probability
values were set by adopting a normal distribution and with to join with a new group containing distrusted agents and
the ratio between trusted/distrusted agents set to 0.5. During replacing that showing the worst value of trust (by increasing
the simulation the initial sparsity of the trust network will the MAMT value because, sooner or later, distrusted agents
decrease for the availability of new reliability information. will leave groups). Moreover, the presence into the all groups
At each simulation step some interactions among a subset of distrusted agents at different simulation steps per different
of the agents was simulated and their “quality” evaluated values of M is shown in Figure 4. Results confirm that almost
by simulated feedback. For unreliable and reliable agents, distrusted agents are replaced by trusted agents into the groups.
the values of feedback were generated based on a normal Therefore, the execution of the distributed algorithm for
distribution; these and the other simulation parameters are group formation leads to a configuration of groups with a
shown into Table II. More in detail, for each simulation step: high level of (average) mutual trust among their members. In
1) a number of interactions is simulated among agents; particular, in a simulated environment, the convergence of the
2) 100 execution of the algorithm are simulated by trigger- algorithm towards a group configuration with trusted agents
ing the algorithm 1 on 100 different agents randomly can be very fast, when the algorithm parameters are properly
selected. For each agent request to join with a group, the set (e.g. parameter M ), leading to ruling out the unreliable
administrator executes the algorithm 2 to decide whether agents from the groups very quickly.
or not to accept the requiring agent;
3) some statistics are computed. V. R ELATED W ORK
To evaluate the simulation results, the measure Average
In open, competitive and distributed contexts a large number
Mutual Trust among the components of a group g as:
of potential threats exist and, to this aim, trust systems can
kgk avoid to be engaged with unreliable partners [21]–[26].
1 X
AM T g = (τi,j + τj,i )
2kgk i,j=1
i6=j 0.85
and the Mean Average Mutual Trust, for a certain configuration 0.8
at a certain time-step, as: 0.75
kGrk
1 X 0.7 M=5
M=6
M AM T (Gr) = AM T gi
MAMT

M=7
kGrk i=1 0.65
M=8
M=9
0.6 M=10
were defined.
The first set of results is shown in Figure 3 and reports the 0.55

median value of MAMT measured after each single step of 0.5
the simulation for the different values of M = [5 ÷ 10] for 0.45
the first 30 steps of the simulation. For M = 5 is shown a 0 5 10 15 20 25 30
slow convergence of the MAMT values, while for M ≥ 6 there Step

exist a radical change. Indeed, the parameter M represents the
number of new groups analyzed by the single agent ai in the Fig. 3. MAMT - results until 30 Steps

74
Workshop "From Objects to Agents" (WOA 2019)

500 sensors monitoring traffic flows on the roads is described. Each
450 agent-based sensor of the grid is associated with a road and
400 gathers, analyzes and aggregates acoustical signals generated
350
Untrusted agents

by vehicles in their motion. Based on a distributed trust-
300
M=5 system, each agent improves own performances by interacting
250 M=10
M=15 with other sensors in its neighboring.
200
Finally, some trust systems have been conceived for IoT
150
and CC contexts. For instance, in [41] two interacting IoT
100
devices can mutually trust each other device and propagate
50
their evaluations to the other nodes with a word of mouth
0
0 20 40 60 80 100 approach. In [42] each node evaluates the trustworthiness of
Step its friend nodes and the opinions of the common friends (by
considering reliability and local reputation measures). A Trust
Fig. 4. Sum of untrusted agents vs number of steps of the simulation Management system for a CC marketplace in [43] evaluates a
multidimensional trustworthiness of the CC providers by ex-
ploiting different sources and trust information. CC federation
With respect to the problem of suggesting to a group (i.e., are considered in [44], where a fully decentralized trust-based
a member of a community) if accepting (i.e., joining with) a model for large-scale federations is designed to allow any node
candidate (i.e., a group), several trust-based approaches have to find the most suitable collaborators in an efficient way,
been proposed. For instance, in [10], [27] is proofed that trust- avoiding exploration of the whole node space by including
based groups are more stable over time with respect to groups trustworthiness information about the set of candidate nodes.
formed without to consider trust. Indeed, the expectations
of receiving benefits is higher among the members of trust- VI. C ONCLUSIONS
formed groups. In such a context, the predominance of local In this paper, a CoT scenario supporting the virtualization
trust is particularly true in large communities where each actor of IoT devices over the cloud in a multi-agent context has
usually interacts only with a narrowest share of the community. been presented. The social attitude of software agents has been
Examples of local trust approaches can be find in [28] exploited to form groups for promoting satisfactory agents
and [29]. The first one is TidalTrust which exploits the closer interactions. However, a satisfactory interaction depends on
neighbors to compute its trust predictions, also by ignoring the choice of the partner but in absence of suitable information
part of the neighbors if the trust network is too sparse. The to perform an autonomous choice, some suggestions can be
second techniques, named MoleTrust, performs a backward asked to those agents perceived as the mostly trustworthy.
exploration by fixing a maximum depth in the search-tree of To this aim, we designed a distributed algorithm to guide
the trust network to calculate trust scores by using at depth x the formation of agent groups of reliable recommenders, in
only the trust scores at depth x − 1. a competitive and cooperative scenario, exploiting a voting
Independently from the adopted group formation modalities, procedure focused on the agent capability of providing useful
to reach a decision within a group voting mechanisms [30], recommendation on the basis of reliability, local reputation
[31] optimize the social utility [32] and avoid conflicts [33], and helpfulness measures. In particular, the adoption of local
although any “ideal” voting procedure exists due to the risks reputation measures avoids the heavy computational tasks and
of manipulations [34]. This aspect is very critical for soft- communication overheads required from a global reputation
ware agent communities, where agents can quickly examine mechanism because only a little share of the agent community
the effects of each manipulation strategy [35]–[37]. In this is involved in this process. Some experiments, in a simu-
respect, [38] presents a local trust-based voting, working in lated agent CoT scenario, confirmed the potential advantages
a mobile wireless scenario, where a node is admitted in a deriving by our proposal in improving individual and group
transmission path on the basis of the trustworthiness perceived satisfaction in terms of mutual trust.
by the other nodes. The actual trust of a node is propagated
by mutual acquaintance among neighbors placed at one hop ACKNOWLEDGMENT
of distance on an oriented trust network by combining their This study has been developed at NeCS Laboratory
confidence values considered as trust measures. A node will (DICEAM, University Mediterranea of Reggio Calabria).
be trusted/distrusted by using a local voting scheme. In [39]
faulting sensors are discovered by using a trustworthiness R EFERENCES
measure, named SensorRank, modeled by a Markov chain on
[1] M. Aazam, I. Khan, A. A. Alsaffar, and E.-N. Huh, “Cloud of things:
the sensor network. This value is used in a voting scheme, Integrating internet of things and cloud computing and the issues
named TrustVoting, where each vote implicitly represents the involved,” in Applied Sciences and Technology (IBCAST), 2014 11th
number neighbors referencing the opinions of a node and by International Bhurban Conference on. IEEE, 2014, pp. 414–419.
[2] P. Parwekar, “From internet of things towards cloud of things,” in
weighting each vote proportionally to their proximity to the 2011 2nd International Conference on Computer and Communication
target node on the sensor network. In [40] a grid of agent-based Technology (ICCCT-2011), Sept 2011, pp. 329–333.

75
Workshop "From Objects to Agents" (WOA 2019)

[3] A. Botta, W. De Donato, V. Persico, and A. Pescapé, “Integration of [23] A. Jøsang, R. Ismail, and C. Boyd, “A survey of trust and reputation
cloud computing and internet of things: a survey,” Future Generation systems for online service provision,” Decision support systems, vol. 43,
Computer Systems, vol. 56, pp. 684–700, 2016. no. 2, pp. 618–644, 2007.
[4] G. Aloi, G. Caliciuri, G. Fortino, R. Gravina, P. Pace, W. Russo, [24] M. Momani and S. Challa, “Survey of trust models in different network
and C. Savaglio, “Enabling iot interoperability through opportunistic domains,” arXiv preprint arXiv:1010.0168, 2010.
smartphone-based mobile gateways,” Journal of Network and Computer [25] P. R. Vamsi and K. Kant, “Systematic design of trust management
Applications, vol. 81, pp. 74–84, 2017. systems for wireless sensor networks: A review,” in Advanced Comput-
[5] D. Uckelmann, M. Harrison, and F. Michahelles, “An architectural ing & Communication Technologies (ACCT), 2014 Fourth International
approach towards the future internet of things,” in Architecting the Conference on. IEEE, 2014, pp. 208–215.
internet of things. Springer, 2011, pp. 1–24. [27] P. De Meo, F. Messina, D. Rosaci, and G. M. L. Sarné, “Forming time-
[6] G. Fortino, R. Gravina, W. Russo, and C. Savaglio, “Modeling and sim- stable homogeneous groups into online social networks,” Information
ulating internet-of-things systems: A hybrid agent-oriented approach,” Sciences, vol. 414, pp. 117–132, 2017.
Computing in Science & Engineering, vol. 19, no. 5, pp. 68–76, 2017. [28] J. Golbeck, “Computing and applying trust in web-based social net-
[7] P. De Meo, F. Messina, M. N. Postorino, D. Rosaci, and G. M. L. Sarné, works,” in PhD Thesis. University of Maryland, Department of
“A reputation framework to share resources into iot-based environ- Computer Science, 2005.
ments,” in 2017 IEEE 14th International Conference on Networking,
[29] P. Massa and P. Avesani, “Trust-aware recommender systems,” in Proc.
Sensing and Control (ICNSC). IEEE, 2017, pp. 513–518.
of the 2007 ACM Conference on Recommender systems. ACM, 2007,
[8] G. Fortino, F. Messina, D. Rosaci, and G. M. L. Sarné, “Using trust
pp. 17–24.
and local reputation for group formation in the cloud of things,” Future
Generation Computer Systems, vol. 89, pp. 804–815, 2018. [30] S. J. Brams and P. C. Fishburn, “Voting procedures,” Handbook of social
[9] C.-M. Chiu, M.-H. Hsu, and E. T. Wang, “Understanding knowledge choice and welfare, vol. 1, pp. 173–236, 2002.
sharing in virtual communities: An integration of social capital and [31] N. R. Council et al., Public participation in environmental assessment
social cognitive theories,” Decision support systems, vol. 42, no. 3, pp. and decision making. National Academies Press, 2008.
1872–1888, 2006. [32] L. Xia, “Computational voting theory: game-theoretic and combinatorial
[10] P. De Meo, E. Ferrara, D. Rosaci, and G. M. L. Sarnè, “Trust and com- aspects,” Ph.D. dissertation, Duke University, 2011.
pactness in social network groups,” ACM Transactions on Cybernetics, [33] T. C. Beierle and J. Cayford, Democracy in practice: Public participa-
vol. 45, no. 2, pp. 205–2016, 2015. tion in environmental decisions. Resources for the Future, 2002.
[11] P. De Meo, F. Messina, D. Rosaci, and G. M. L. Sarné, “Combining [34] V. Conitzer and T. Sandholm, “Universal voting protocol tweaks to make
trust and skills evaluation to form e-learning classes in online social manipulation hard,” arXiv preprint cs/0307018, 2003.
networks,” Information Sciences, vol. 405, pp. 107–122, 2017. [35] J. Pitt, L. Kamara, M. Sergot, and A. Artikis, “Formalization of a voting
[12] C. Castelfranchi and R. Falcone, Trust theory: A socio-cognitive and protocol for virtual organizations,” in Proc. of the 4th Int joint Conf. on
computational model. John Wiley & Sons, 2010, vol. 18. Autonomous Agents and Multiagent Systems. ACM, 2005, pp. 373–380.
[13] P. De Meo, F. Messina, D. Rosaci, and G. M. L. Sarné, “An agent-
[36] A. Gibbard, “Manipulation of voting schemes: a general result,” Econo-
oriented, trust-aware approach to improve the qos in dynamic grid
metrica: J. of the Econometric Society, pp. 587–601, 1973.
federations,” Concurrency and Computation: Practice and Experience,
2015. [37] M. Satterthwaite, “Strategy-proofness and arrow’s conditions: Existence
[14] F. Messina, G. Pappalardo, D. Rosaci, and G. M. L. Sarné, “A trust- and correspondence theorems for voting procedures and social welfare
based, multi-agent architecture supporting inter-cloud vm migration in functions,” J. of Economic Theory, vol. 10, no. 2, pp. 187–217, 1975.
iaas federations,” in International Conference on Internet and Dis- [38] T. Jiang and J. S. Baras, “Trust evaluation in anarchy: A case study on
tributed Computing Systems. Springer, 2014, pp. 74–83. autonomous networks.” in INFOCOM, 2006.
[15] L. Palopoli, D. Rosaci, and G. M. L. Sarné, “A multi-tiered recom- [39] X.-Y. Xiao, W.-C. Peng, C.-C. Hung, and W.-C. Lee, “Using sensorranks
mender system architecture for supporting e-commerce,” in Intelligent for in-network detection of faulty readings in wireless sensor networks,”
Distributed Computing VI. Springer, 2013, pp. 71–81. in Proc. of the 6th ACM Int. Work. on Data Engineering for Wireless
[16] ——, “A distributed and multi-tiered software architecture for assessing and Mobile Access. ACM, 2007, pp. 1–8.
e-commerce recommendations,” Concurrency and Computation: Prac- [40] M. N. Postorino and G. M. L. Sarné, “An agent-based sensor grid
tice and Experience, vol. 28, no. 18, pp. 4507–4531, 2016. to monitor urban traffic,” in Proceedings of the 15th Workshop dagli
[17] A. Blanchard and T. Horan, “Virtual communities and social capital,” Oggetti agli Agenti, WOA 2014, ser. CEUR Workshop Proceedings, vol.
in Knowledge and social capital. Elsevier, 2000, pp. 159–178. 1260. CEUR-WS.org, 2014.
[18] P. De Meo, F. Messina, D. Rosaci, and G. M. L. Sarné, “Recommending [41] F. Bao and I.-R. Chen, “Dynamic trust management for internet of things
users in social networks by integrating local and global reputation,” applications,” in Proceedings of the 2012 international workshop on
in Proc. of the 7th Int. Conf. on Internet and Distributed Information Self-aware internet of things. ACM, 2012, pp. 1–6.
Systems, ser. LNCS, vol. 8729. Springer, 2014, pp. 437–446. [42] M. Nitti, R. Girau, L. Atzori, A. Iera, and G. Morabito, “A subjective
[19] D. M. Kilgour and C. Eden, Handbook of group decision and negotia- model for trustworthiness evaluation in the social internet of things,”
tion. Springer Science & Business Media, 2010, vol. 4. in Personal Indoor and Mobile Radio Communications (PIMRC), 2012
[20] L. S. Lai and E. Turban, “Groups formation and operations in the web IEEE 23rd International Symposium on. IEEE, 2012, pp. 18–23.
2.0 environment and social networks,” Group Decision and negotiation,
[43] S. M. Habib, S. Ries, and M. Muhlhauser, “Towards a trust management
vol. 17, no. 5, pp. 387–402, 2008.
system for cloud computing,” in Trust, Security and Privacy in Com-
[21] T. Grandison and M. Sloman, “A survey of trust in internet applications,”
puting and Communications (TrustCom), 2011 IEEE 10th International
IEEE Communications Surveys & Tutorials, vol. 3, no. 4, pp. 2–16, 2000.
Conference on. IEEE, 2011, pp. 933–939.
[22] G. Lax and G. M. L. Sarné, “CellTrust: a reputation model for C2C
commerce,” Electronic Commerce Research, vol. 8, no. 4, pp. 193–216, [44] F. Messina, G. Pappalardo, D. Rosaci, C. Santoro, and G. M. L. Sarné,
2006. “A trust-aware, self-organizing system for large-scale federations of
[26] G. Fortino and P. Trunfio, Internet of things based on smart objects: utility computing infrastructures,” Future Generation Computer Systems,
Technology, middleware and applications. Springer, 2014. vol. 56, pp. 77–94, 2016.