BLUE: A REPUTATION-BASED MULTI-AGENT SYSTEM TO
SUPPORT C2C IN P2P BLUETOOTH NETWORKS
Gianluca Lax
University Mediterranea of Reggio Calabria
via Graziella, loc. Feo di Vito, 89060 Reggio Calabria - Italy
Giuseppe M. L. Sarn
`
e
University Mediterranea of Reggio Calabria
via Graziella, loc. Feo di Vito, 89060 Reggio Calabria - Italy
Keywords:
Mobile Commerce, Reputation-based Systems, Multi Agent Systems, Bluetooth Networks.
Abstract:
In this paper we propose a multi-agent system, called Blue, providing a reputation mechanism to promote
mobile C2C commerce in large and highly dynamic environments based on the Bluetooth wireless technol-
ogy. In such a scenarios, detecting malicious users assumes a great importance but traditional approaches to
obtaining reputation of a user are unfeasible or meaningless. For this purpose, we propose a centralized repu-
tation mechanism, used in a asynchronous way, to manage users’ reputation based on trade feedbacks (given
by users) suitably weighted. Agents employ such a reputation mechanism to choose their trade counterparts
trying to avoid malicious users. A number of experiments show the effectiveness of our proposal.
1 INTRODUCTION
The emerging wireless technologies and devices, as
mobile phones and personal digital assistants, will al-
low mobile-commerce to become the dominant form
of trading (Dholakia, 2004). Currently, a user can ex-
ploit his mobile phone (or another wireless device)
to connect in Internet using centralized services, of-
fered mainly by mobile communication companies,
Wi-Fi and Wi-Max providers, or to make an hotel
reservation from a train or to purchase stocks from
a bank while he walks or downloads music onto an
MP3 player and so on. Vice versa, wireless technolo-
gies not relying on a centralized server, as Bluetooth
(see Section 2), allow different users to realize infor-
mation exchanges in a peer-to-peer fashion.
According to the wireless technology improvement
in terms of operating range, hardware costs and power
consumption, more and more highly dynamic
1
wire-
less networks, where users can exchange a large
amount of data, will be realized. In these networks
there are three main issues, well-known from other re-
search contexts (Ramchurn et al., 2004; Resnick et al.,
2000), that have to be dealt with. Such issues regard
how a user may: (i) found interesting information; (ii)
1
From the point of view of a user, each time another
user U enters in (resp., exits from) the operative range of
his device, it is viewed as the U s joining to (resp., leaving
from) the network.
protect himself against malicious users; (iii) guaran-
tee his privacy. Solutions of such problems can be
provided both by centralized services and by cooper-
ative and distribute techniques.
In particular, centralized services, as Napster
(http://www.napster.com, 2006), require that all users
inform a given server about the resources (files) they
offer. A user needing a resource, has to query the
server that provides him with a list of users able to
satisfy his request, along with their reputation judg-
ment provided by other users (a famous example of
this approach is eBay (http://www.ebay.com, 2006)).
A centralized solution, that works well in a wired en-
vironment, fails in our context, mainly for the follow-
ing reasons: i) service availability is not assured, usu-
ally for connection problems; ii) the access time to the
services can be significant in a highly dynamic con-
text; iii) the server can become a bottleneck because
of the high number of updates it has to manage due to
the dynamism of the network; iv) the users reported
by the server can be unavailable since they are too far
w.r.t. the operative range of the mobile device.
A different philosophy is employed in a distribute
approach where resources and reputation information
must be propagated by the users during their interac-
tions (Mui et al., 2002). Also this approach can be
unsuitable in our context for the following reasons: i)
it presumes that the networks are stable enough both
in composition and in time living; ii) in wide and dy-
97
Lax G. and M. L. Sar G. (2006).
BLUE: A REPUTATION-BASED MULTI-AGENT SYSTEM TO SUPPORT C2C IN P2P BLUETOOTH NETWORKS.
In Proceedings of the International Conference on e-Business, pages 97-104
DOI: 10.5220/0001426300970104
Copyright
c
SciTePress
namic communities the acquired knowledge of infor-
mation and reputation could be broadly partial and
consequently ineffective.
In this paper we propose a framework to pro-
mote Bluetooth-based trading activities (file for file
or money for file, though they could occur also for
free) in a profitable and safe way, independently of
telephonic or networking providers. It exploits the
opportunities offered by Bluetooth networks, smart-
phones, agent and e-payment technologies. More in
detail, Bluetooth is used to realize P2P communica-
tions in highly dynamic networks. To realize fast and
sure transactions, a Multi-Agent System, called Blue,
is implemented. Moreover, a centralized agency sup-
ports agents by managing a reputation system.
Blue adopts temporary agent reputation credentials
that are required by each agent to the agency and is-
sued on the basis of suitable criteria. Before each
trade activity, agents exchange their credentials to
evaluate the counterpart reputation and, then, they
decide on the opportunity of continuing (or not) the
transaction. Note that such credentials can not be ver-
ified at the same time of the transaction by means of
the issuer or third parties. A more detailed description
of this aspect will be provided in Section 4.
Furthermore, the proposed solution solves many
questions arisen from centralized and distribute ap-
proaches; more in detail: i) centralized services and
business transactions happen at different time over
different communication channels bypassing prob-
lems of services availability; ii) the Blue mechanism
is quick and independent of composition and life of
the network; iii) users maintain privacy about trans-
action contents w.r.t. third parties; iv) a reciprocal
reputation knowledge between two agents is realized;
v) all agent communications are inexpensive, support-
ing also micro-trading activities (in fact, the connec-
tion costs are usually incompatible with the effective
transaction value or constitute themselves a signifi-
cant part; in any case agent-server communications
must happen over an Internet communication chan-
nel); the usage of cryptography and digital certificates
improves the security in Bluetooth that currently is an
unsafe environment (Shaked and Wool, 2005). On the
contrary, the main open question is that asynchronism
does not allow users to exploit certificates really up-
dated. For this reason Blue is riskier than other P2P
systems. However, uncorrect agents are detected and
isolated in a short time.
The rest of the paper is organized as follows: in the
next section some preliminary notions are introduced.
Sections 3 presents the architecture of Blue, consist-
ing of a centralized agency and a number of agents.
The adopted reputation model is described in Section
4. Some experiments and results obtained by simu-
lating our framework are reported in Section 5. The
comparison with other approaches and techniques are
presented in Section 6, and finally, in Section 7 we
draw our conclusions.
2 PRELIMINARY NOTIONS
A general overview is presented now about two issues
widely exploited in this paper, which are the Blue-
tooth technology and the cryptography techniques.
Bluetooth
2
(http://www.bluetooth.com, 2006;
Dursch and Yen, 2004) realizes radio-frequency
communications for short-range connectivity among
devices (as personal digital assistants, mobile phones,
laptops, printers and so on) in an inexpensive and
low power way; it provides fast transmissions of
voice and data, also in noisy radio environments
3
,
and implements data error correction, cryptography
and authentication methods. Bluetooth allows us to
provide both point-to-point and point-to-multipoint
connectivity. In the first modality, the communication
channel is shared only between two Bluetooth peers,
while in the other modality a small group of Bluetooth
units, called piconets, can exploit the communication
channel in a given time. In a piconet a device acts as
master and the others are slaves synchronized to the
master’s clock. More piconets over a same physical
area can form a larger network, called scatternet,
where various piconet-to-piconet communications at
a time can be realized by employing the respective
master units. The current Bluetooth standard (ver. 2)
permits a transfer rate of 2.1 Mbps, but the Bluetooth
SIG members are examining some technologies, as
Ultrawide band (http://www.uwbforum.org, 2006),
for an integration with Bluetooth to further improve
both the transfer rate and the operative range.
The second issue presented in this section regards
the “digital sign” that guarantees authenticity and in-
tegrity of a message. It employs a Public Key Cryp-
tosystem (PKC) based on a cryptographic technique
as RSA (Rivest et al., 1978). A PKC requires two
complementary cryptographic keys, usually named
“secret” and “public”. What it is encrypted with one
of the two keys can be decrypted only with the other
one and vice versa. A trusted authority assures the
public key validity and the identity of its owner by
means of a certificate. The public key (K
P
) should
2
The original Bluetooth Promoter Group is constituted
by the five companies (Ericsson, IBM, Intel, Nokia and
Toshiba) that in 1998 had formed a Special Interest Group
(SIG) on this technology and by Microsoft, 3Com, Lucent
and Motorola.
3
The adopted unlicensed working frequency band (2.4
GHz) assures an interaction among Bluetooth units also if
this band is shared with other devices signals, as garage
door openers. Besides, Bluetooth is compliant with global
emissions rules and airline regulations.
ICE-B 2006 - INTERNATIONAL CONFERENCE ON E-BUSINESS
98
be accessible to each user while the secret key (K
S
)
must be not shared with anyone. Moreover, a hash
function, as MD5 (Rivest, 1992) or the Secure Hash
Algorithm (SHA) (NIST/NSA, 2002), computing a
one way message digest is exploited. In order to dig-
itally sign a message M, the user U signs the digest
of M (H(M)) with his K
U
S
obtaining the digital sig-
nature DS = K
U
S
(H(M)). Then the message and
the signature are sent to V , who can check both the
integrity and the authenticity of M by verifying that
H(M) equals K
U
P
(DS).
3 THE BLUE FRAMEWORK
In this section we describe the Blue framework, that
consists of a centralized agency and a number of
agents providing same features. Each agent, associ-
ated to both a specific user and a SIM, stores and han-
dles both system and user information on one hand
and monitors and supports user’s activities on the
other hand. All agents are affiliated to the agency.
This latter provides them with some tools and infor-
mation and takes care to promote a trust atmosphere.
For this purpose, the agency exploits the reputation
model described in Section 4. Agents exploit Blue-
tooth features to realize P2P connections among them
in such a way agents can support their users during
searching, buying or selling activities.
3.1 The Blue Agency
The Blue Agency, denoted by Ag, is a centralized ser-
vice provider which supports agent activities in or-
der to guarantee the correct behaviour of all affiliated
agents and their trading activities. To this goal the
agency keeps the following information: i) its own
secret and public cryptographic keys (resp., K
Ag
S
and
K
Ag
P
); ii) an ontology that consists of one or more
XML-schema employed by the agents to describe the
resources to sell or to buy; iii) the list of affiliated
agents; iv) for each agent, the value of its reputation
(whose computing is described in Section 4), as well
as a specific digital certificate, called (reputation) cre-
dential, used during transactions are stored. The cre-
dential C of an agent is a tuple hID, R, K
P
, Exi,
where ID is an agent identifer
4
, R is the agent rep-
utation rating, K
P
is the public cryptographic keys of
4
To avoid malicious identity change of a compromised
reputation with a spotless one, easy in virtual communities
but potentially easier in Bluetooth networks, in Blue the
SIM of smart-phone is employed as agent ID (i.e., a mo-
bile phone number used to identify the agent) (Pfitzmann
et al., 1997) and agent data are stored in a persistent way by
the agency. As a consequence each identity change, even
though possible, requires necessarily another SIM.
the agent and Ex is the expiration dates of C, respec-
tively. Observe that the temporal validity of C, (de-
pending on Ex) is tightly correlated to R, so creden-
tials of trustworthy agents have longer validity. More-
over, the agency stores also user’s financial account
information. The agency provides three main activi-
ties, more specifically:
affiliate managing - When Ag receives a new agent
affiliation, it provides the agent with K
Ag
P
, an ini-
tial reputation value (see Section 4) and the current
ontology. Then the agency adds the new agent to
the list of affiliated agents.
credential providing - On demand, the agency sup-
plies an agent with a valid and updated credential,
signed by the agency with K
Ag
S
.
reputation managing - Ag manages and updates
the reputation rate of agents (see Section 4).
3.2 The Blue Agent
Each agent is associated to a user and supports his
activities by managing (in terms of insertion, dele-
tion and updating) the resources he wants to sell or
to buy. The agent is automatically activated (resp.,
deactivated) when the user’s device is on (resp., off
”per se” or for an explicit user’s choice). In order to
support user activities, an agent keeps the following
information: i) the user’s secret and public keys; ii)
the ID of the agency (e.g., the Internet address) and
its public key; iii) user’s financial account informa-
tion; iv) the Blue ontology; v) the resources to sell
(resp., to buy), described using such an ontology and
including also the price of selling (resp., buying).
Observe that each agent has to be affiliated to the
agency. The first time an agent is activated, it receives
the public key K
Ag
P
of the agency, the current ontol-
ogy and an initial credential C (signed by Ag). More-
over, the user (associated to the agent) has to provide
some information exploited by the agency during the
user identification task, as well as some initial para-
meters, like an individual hazard threshold (Falcone
and Castelfranchi, 2001a; Tan and Thoen, 2000), used
to select the counterpart agent for a transaction, and a
list of resources to sell and to buy.
Now we describe the activities performed by the
agents during trading. Without lost of generality, we
consider two agents, named a
b
and a
s
, the former in-
terested in buying a resource, the latter in selling. The
trading activities are:
1. research - Each agent searches another agent close
enough to communicate by a Bluetooth network.
2. presentation - Then each agent verifies the identity
of the other agent, as follows:
BLUE: A REPUTATION-BASED MULTI-AGENT SYSTEM TO SUPPORT C2C IN P2P BLUETOOTH NETWORKS
99
(a) a
b
sends to a
s
its credential C
b
along with its
current time now
b
;
(b) a
s
verifies the integrity of C
b
(that is signed by
the agency) from which extracts ID
b
and the
public key of a
b
, that is K
b
P
. Then a
s
sends to
a
b
its credential C
s
, its current time now
s
and
the signature (by a
s
) of a message M
s
contain-
ing ID
b
and now
b
.
(c) a
b
in its turn, verifies the integrity of C
s
, and
extracts K
s
P
to verify M
s
authenticity. Then a
b
sends to a
s
a signed (by a
s
) message M
b
con-
taining ID
b
and now
b
. M
b
(resp. M
s
) is the
proof for a
s
(resp. a
b
), of the encounter with a
b
(resp. a
s
), and will be used if the transaction will
be performed;
(d) finally, a
s
completes the presentation task by
verifying the authenticity of M
b
.
3. evaluation - At this point, each agent extracts the
reputation rating of the other agent from the cre-
dential previously received and will continue the
trading only if the counterparts agent is considered
trustworthy (i.e., such a reputation rating must be
higher than the hazard threshold), otherwise it ter-
minates the communication.
4. buy offer - a
b
sends to a
s
a list of resources it wants
to buy. These resources are described using the
common agent ontology and in order to preserve
a
b
privacy, only the hash values
5
of each item are
reported (Han et al., 2004). Then, a
b
receives the
list of resources of a
s
matching the request of a
b
,
along with their prices.
5. sell offer - When a
s
receives the buy offer from
a
b
it compares the hash values of the resources (to
sell) it owns with the hash values received from a
s
in order to find a matching. Finally, the list con-
taining all the matching resources, along with their
price, is sent to a
b
. In the case of an empty result, a
message for closing the protocol is sent to a
b
.
6. transaction - In this step, the trade can be carried
out if there is the matching between offer and de-
mand. At the end of the transaction, a
b
signs again
the proof of the encounter (M
b
) generating
M
b
that
is sent to a
s
and is the proof that the transaction has
occurred. In its turn, a
s
performs the same opera-
tion.
Observe that in our system the transaction is com-
pletely performed locally. This is one of the main
advantages given by our proposal. On the other hand,
for the same reason some fraud can happen since no
check can be performed at the same time of the trans-
action, but it is guaranteed that fraud is detectable,
traceable and unprofitable.
5
For an efficient implementation we may assume that
hash values are pre-computed and stored.
4 THE REPUTATION MODEL
In this section we describe the reputation model used
in our framework. In the following, according to
(Abdul-Rahman and Hailes, 2000), we consider that
the reputation is an expectation about an agent’s be-
havior based on information about or observations of
its past behavior”.
When economical interests are involved, the ac-
tors should be trustworthy. To this aim, trade agent
activities need of reputation information that can be
be based on a direct or, more usually, indirect agent
knowledge exploiting in this case some reliable prop-
agation mechanisms where the number of (indepen-
dent) information sources (Falcone and Castelfranchi,
2001b) and their credibility will be relevant.
In the previous sections we have seen that the
agency manages the reputation rating R of each agent
exploiting the feedbacks provided confidentially by
the past agent’s trade counterparts. To prevent ma-
licious behaviours, before a trade activity, each agent
presents, as a visit card, its temporary Reputation Cre-
dential described in Section 3.2.
The Blue reputation mechanism should satisfy
some properties (Xiong and Liu, 2004; Ramchurn
et al., 2004) and more specifically: i) taking into ac-
count the trade history of each agent; ii) differentiat-
ing dishonest from honest feedbacks, to avoid mali-
cious reputation manipulations; iii) recognizing dif-
ferent transaction contexts, to avoid reputation gain
in small transaction value for cheating in high one;
iv) identifying agents provided of dynamic person-
ality that alternate their behaviours between honesty
and cheat, independently of the transaction context.
Observe that, assuming a large agent population
and a dynamic environment, the probability of reen-
counters among agents is very low. As a consequence,
agents’ reputation based on subjective agents’ im-
pressions (direct dimension of reputation) is not sig-
nificative; in fact, in Blue agent reputation is based
only on indirect information aggregated in a central-
ized way (indirect dimension of reputation). Another
consequence is that the presence of collusive agent
coalitions, that could maliciously influence reputation
evaluations, can be neglected, since collusive agents
should encounter effectively their victims (see the ob-
servation above).
Thus, the proposed reputation metrics is based
only on the following factors: i) The value of feed-
backs obtained; ii) The number of feedbacks ob-
tained; iii) The credibility of the feedbacks sources;
iv) The transaction contexts. To describe the rep-
utation model proposed, we consider a framework
consisting of the centralized agency Ag and a set
of agents {a
1
, . . . , a
n
} that, suitably supported by
Ag, interact to carry out trading activities focused on
the research/sell/buy/exchange of resources in a prof-
ICE-B 2006 - INTERNATIONAL CONFERENCE ON E-BUSINESS
100
itable way over Bluetooth networks.
Each transaction T
i
performed by a
i
is a tuple
hID
j
, val, M
j
, r
j
i where: ID
j
is the identifier of the
trading agent counterpart, val is the monetary value
of the transaction describing the transaction context
(i.e., its relevance),
M
j
is the proof that an encounter
with a
j
has occurred and it is carried out during the
agent presentation (see Section 3.2); r
j
is the appre-
ciation of a
i
about the transaction. The value of such
an appreciation ranges from 0 (for unsatisfying trans-
actions for example, when a fraud occurs) to 1 (for
full satisfaction). Such a value is explicitly provided
by the user at the end of the transaction.
Since each agent credential has an expiration date,
periodically an agent has to require a new C to the
agency. On this occasion, the agent transfers to the
agency the list of transactions done. The agency on
the basis of the transaction received, updates the rep-
utation of each agent involved in the transaction. In
particular, for each transaction T performed by a
i
, as-
suming a
j
be the counterpart of the transaction, then
the reputation of a
j
(that is R
j
) is updated as follows:
R
j
= (1 α) ·
f
R
j
+ α · r
j
with α = R
i
·
val
V al
max
, where the new value of the
reputation of a
j
is computed weighting in a comple-
mentary way, by means of α, two contributions, (1) its
previous reputation value (denoted by
f
R
j
) and (2) the
appreciation about the trade event done by counter-
part agent a
i
(that is r
j
). α allows us to tune the two
components. In particular, the weight of the second
contribution is considerable whenever the value of the
transaction (val) is high or the counterpart agent is au-
thoritative (it has a high reputation R
i
).
Observe that we assume the maximum value of a
transaction is V al
max
. It is computed by the agency
that stores the value of the transactions performed.
Moreover, the initial reputation rating is fixed to 0.5.
This choice takes into account two opposite needed:
the former is of not penalizing new agents (Ramchurn
et al., 2004) while the latter is of penalizing agents
having a bad reputation that want reenter into the sys-
tem (Zacharia and Maes, 2000).
5 EXPERIMENTS
In this section we describe a number of experiments
performed in order to test the efficacy of our proposal.
For this purpose we implemented a C++ prototype
simulating the interactions among (buyer and seller)
agents during trading. Now we describe the parame-
ters and the evaluation metrics considered in our ex-
periments.
Number of agents. We considered a population of
1,000 agents.
Number of malicious agents. This parameter varies
from 0.1% to 25% of the overall number of agents.
Malicious behaviours. In order to considerate differ-
ent behaviours of malicious agents, we assumed that
a malicious agent behaves incorrectly with a proba-
bility M B. In particular, MB = 1 means that the
agent always misbehaves, whereas MB = 0.5 means
that each 100 transactions, it behaves well during 50
(randomly chosen) transactions and badly in the re-
maining ones. We varied M B between 0.1 and 1.
Number of transactions. The number of transactions
performed by each agent ranges varies from 10 to 100.
Thus, the overall number of transactions varies from
500 to 50000.
Value of transactions. We represented the value of
each transaction by a random integer between 1 and
10, thus assuming that the maximum value is 10.
Initialization. We initially set the agent rate to 0.5 and
the hazard threshold of each agent to a random value
between 0.3 and 0.7.
Evaluation metrics. We have computed several
evaluation measures (Rijsbergen, 1979; Srivihok
and Sukonmanee, 2005; McLaughlin and Herlocker,
2004), that are Precision, Recall and Accuracy. Be-
fore defining such measures, we introduce some no-
tations. Let TP (true positive) be the overall num-
ber of good transactions taken place, where the word
good means that the transaction ends positively. In
words, TP represents the number of time the repu-
tation model forecasts correctly that the counterpart
agent is not malicious. Analogously, let TN (true
negative) be the number of bad transactions taken
place, let FP (false positive) be the number of good
transactions that did not occur (because the counter-
part agent reputation was not enough), and finally, let
FN (false negative) be the number of bad transactions
that did not occur. Observe that TP and FN repre-
sent cases in which the reputation model worked well,
conversely TN and FP represent incorrect predictions
of the reputation model. We are ready to define Preci-
sion, Recall and Accuracy. Precision P re =
T P
T P +F P
is the percent of positive predictions that are correct
and is indicative of the model correctness. Recall
Rec =
T P
T P +T N
is the percent of the positive cases
that are caught by the model and denotes the model
completeness. Accuracy Acc =
T P +F N
T P +T N+F P +F N
is
the percent of predictions that are correct. Obviously,
the higher the value of such measures, the better the
accuracy of the system is. Moreover, as a metrics
we consider also a system property, named Malicious
Rate, that is the average reputation rate of malicious
agents. We expect it to decrease as the number of
transactions increases.
BLUE: A REPUTATION-BASED MULTI-AGENT SYSTEM TO SUPPORT C2C IN P2P BLUETOOTH NETWORKS
101
10
2
10
3
10
4
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
number of transactions
Precision
Recall
Accuracy
Mal. Rate
Figure 1: Results vs. number of transactions.
5.1 Selected Results of Experiments
In the first experiment the behavior of the system as
the number of transactions increases is analyzed. Fig-
ure 1 reports the value of Precision, Recall, Accuracy
and Malicious Rate obtained by a simulation consist-
ing of 50,000 transactions. We have fixed the number
of malicious agents to 10% of the agent population,
and we have set the probability of malicious action
(MB) to 1. We may observe that the considered met-
rics assume a stable value after about 10,000 transac-
tions, that is when each agent has performed about 5
transactions. Precision is always quite high (at least
0.8) and converges to 1. It means that when the rep-
utation model suggests performing a trading with an
agent, the probability that such a transaction ends pos-
itively is very high. The value of Recall ranges from
0.6 to 0.8 in the first 1,000 transactions, then it re-
mains around 0.8. This result shows that a number
(about 20%) of true negative cases may occur, and
this is the price to pay for obtaining a high Preci-
sion. Indeed, the reputation model suggests perform-
ing a transaction only if it is reputed sure and such a
protective policy increases the number of good trans-
actions that are not suggested (true negative cases)
by the model. Also Accuracy is always quite high
(around 0.8). As observed above, we remark that a lot
of wrong predictions regard true negative cases that
are not too penalizing for users.
In Figure 2 we show the value of the adopted met-
rics after a simulation of 50,000 transactions, con-
ducted varying the number of malicious agents. We
note that Malicious Rate is always very low (less than
0.2) and is independent of the tested parameter. Thus
our reputation model works well in the case of both
small and large number of malicious agents, as shown
in all experiments conducted. Precision decreases
slightly as the number of malicious agents increases,
anyway maintaining a high value (always more than
0.95). Finally we observe that the measures of Recall
and Accuracy are always very similar and decrease
50 100 150 200 250
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
number of malicious agents
Precision
Recall
Accuracy
Mal. Rate
Figure 2: Results vs. number of malicious agents.
as the considered parameter increases. Indeed, as the
number of malicious increases, the average of the rep-
utation rate of all agents decreases. As a consequence,
the number of true negative cases increases and this
reduces Recall and consequently Accuracy.
In the last experiment we studied the performance
of the reputation model as the probability of malicious
action varies. For space limitation, we report only the
results obtained for Malicious Rate. However, sim-
ilar considerations may be done also for Precision,
Recall and Accuracy. In Figure 3 the value of Mali-
cious Rate versus the probability of malicious action
is reported. We recall that Malicious Rate is a sys-
tem property computed as the average reputation rate
of malicious agents, thus it ranges from 0 to 1. We
have considered three different times of the simula-
tion: after 500 transaction (short time), after 5,000
transactions (medium time) and after 50,000 transac-
tions (long time). The first result we may observe
is that for long time, Malicious Rate is always very
low (about 0.2), whereas for short time the reputation
model does not produce good results. This may be ex-
plained considering that for short time each agent has
performed (on the average) 1 transaction that is not
enough for evaluation its behaviour. Moreover, ma-
licious agents are more difficult to detect when their
probability of malicious behaviour is low, as shown
in Figure 3. For medium time, we note that the ca-
pability of the reputation model to detect malicious
agents is almost linearly depending on the probability
of malicious behaviour. However, the more interest-
ing result is that our reputation model is able to detect
(after a adequate number of transactions) malicious
agents even though they alternate between good and
bad transactions.
6 RELATED WORK
The relevance of reputation issues is witness by the
rich literature produced in these latter years. In partic-
ICE-B 2006 - INTERNATIONAL CONFERENCE ON E-BUSINESS
102
0
0,1
0,2
0,3
0,4
0,5
0,6
0,2 0,4 0,6 0,8 1
probability of malicious action
Malicious Rate
500 transactions 5,000 transactions 50,000 transactions
Figure 3: Malicious rate vs. probability of malicious action.
ular, different notions of reputation belonging to dif-
ferent disciplines are presented in (Mui et al., 2002),
while different reputation properties and models are
described in (Ramchurn et al., 2004). An interesting
approach is proposed by (Dellarocas, 2003), where
reputation is dealt with game and economic theories.
A well know reputation system is employed in
eBay (http://www.ebay.com, 2006); it has been
deeply investigate and doubtless the main advantage
is its simplicity. On the contrary, it is sensitive to
many malicious behaviours. In the Blue multi-agent
context, the reputation is certified (as in (Huynh et al.,
2004)) and limits the effects of malicious elements.
Differently from (Sen and Sajja, 2002), Blue does not
suffer the presence of liars even when a boolean rate
is adopted.
In accord with Sporas (Zacharia and Maes, 2000),
Blue disincentives the change of identity, but in Spo-
ras a very low initial reputation rate (differently from
Blue) is used and in this way it is hardest for a new
agent to gain reputation. Moreover, Sporas try to
avoid collusive agent alliances for increasing recipro-
cally their reputation rates, by limiting the number of
times an agent may increase the reputation of another
agent. A study of the dynamics of honesty and dis-
honesty behaviours in a semi-competitive multi-agent
environment where agents can have incentives to be
honest or dishonest is presented in (Lam and Leung,
2005).
Reputation models are more trustworthy if more
agents as possible cooperate to provide their evalu-
ations (Birk, 2000). Two approaches exist, in the first
the agents are free to not provide their feedbacks (pos-
itive reputation system), in the other the agents are
penalized (or likely promoted) if they do not provide
their feedbacks (negative reputation systems). The
second approach is needed in Blue, but in usual C2C
contexts the first approach seems to be more effective-
ness (Yamamoto et al., 2004) where, using an iterate
prisoner’s dilemma approach, have been investigate
different strategies.
In REGRET (Sabater and Sierra, 2001) the agent
reputation is computed aggregating (social dimen-
sion) the impressions that agents (altruistic and co-
operative) obtain by direct interactions (individual di-
mension) weighting the terms of a common semantic
(ontological dimension) in accord with their personal
point of views. By knowing these weights, it is possi-
ble for an agent to uniform the rates provided by other
agents to its point of view. A similar approach could
be applied in a future version.
In the field of P2P many works have explored rep-
utation issues. Usually, a great attention is given to
preserve the correctness of the reputation rates from
the effects of the malicious agents by realizing a mix
of different techniques. All the following cited works
implement frameworks where agents do not have par-
ticular difficulty to contact other agents or centralized
services for knowing reputation information, differ-
ently from Blue.
In particular, a safe approach is reported by (Kam-
var et al., 2003) that uses pretrusted peers to minimize
the influence of malicious agents in collusion activi-
ties. It implements a mechanism taking into account
the entire system history. A well formed reputation
models is PeerTrust (Xiong and Liu, 2004), where it
is implemented a reputation-based trust framework in
an adaptive manner able to solve many issues of a P2P
system. Some aspects of this work, also if they are re-
ferred to a different scenario, are similar to Blue. Fi-
nally, in (Damiani et al., 2002) a reputation-based pro-
tocol is proposed to support and preserve anonymous
and secure services, for choosing reliable resources in
P2P networks.
7 CONCLUSIONS
In this paper we have presented Blue, a framework to
support C2C commerce activities over Bluetooth net-
works. The adopted reputation model allows the users
involved in trading to detect possible malicious users.
It exploits temporary agent reputation credentials that
are managed by a centralized agency. A number of
experiments show the effectiveness of our proposal
in finding malicious users and avoiding frauds. This
work has opened many research directions to explore
the different possibilities of the Bluetooth networks
in a P2P networks. Currently, a real prototype is in an
advanced stage of implementation.
REFERENCES
Abdul-Rahman, A. and Hailes, S. (2000). Supporting trust
in virtual communities. In HICSS ’00: Proc. of the
33rd Hawaii Int. Conf. on System Sciences - Vol.
BLUE: A REPUTATION-BASED MULTI-AGENT SYSTEM TO SUPPORT C2C IN P2P BLUETOOTH NETWORKS
103
6, pages 1769–1777, Washington, DC, USA. IEEE
Comp. Soc.
Birk, A. (2000). Boosting cooperation by evolving trust.
Applied Artificial Intelligence, 14(8):769–784.
Damiani, E., di Vimercati, D. C., Paraboschi, S., Samarati,
P., and Violante, F. (2002). A reputation-based ap-
proach for choosing reliable resources in peer-to-peer
networks. In CCS ’02: Proc. of the 9th ACM Conf. on
Computer and communications security, pages 207–
216, New York, NY, USA. ACM Press.
Dellarocas, C. (2003). The digitization of word of mouth:
Promise and challenges of online feedback mecha-
nisms. Management Science, 49(10):1407–1424.
Dholakia, R. R. (2004). Editorial: electronic markets in the
post-euphoric phase: Relationships, values and behav-
iors. Telematic Information, 21(2):115–121.
Dursch, A. and Yen, D. C. (2004). Bluetooth technology:
An exploratory study of the analysis and implementa-
tion frameworks. Computer Standards and Interfaces,
26(4):263–277.
Falcone, R. and Castelfranchi, C. (2001a). Social trust:
a cognitive approach. Kluwer Academic Publishers,
Norwell, MA, USA.
Falcone, R. and Castelfranchi, C. (2001b). The socio-
cognitive dynamics of trust: Does trust create trust? In
Proc. of the Workshop on Deception, Fraud, and Trust
in Agent Societies held during the Autonomous Agents
Conf., pages 55–72, London, UK. Springer-Verlag.
Han, P., Xie, B., Yang, F., and Shen, R. (2004). A scal-
able p2p recommender system based on distributed
collaborative filtering. Expert Systems with Applica-
tion, 27(2):203–210.
http://www.bluetooth.com (2006).
http://www.ebay.com (2006).
http://www.napster.com (2006).
http://www.uwbforum.org (2006).
Huynh, T. D., Jennings, N. R., and Shadbolt, N. R. (2004).
Fire: An integrated trust and reputation model for
open multi-agent systems. In ECAI 2004: Proc. of
the 16th Europ. Conf. on Artificial intelligence, pages
18–22. IOS Press.
Kamvar, S. D., Schlosser, M. T., and Garcia-Molina, H.
(2003). The eigentrust algorithm for reputation man-
agement in p2p networks. In WWW ’03: Proc. of the
12th Int. Conf. on World Wide Web, pages 640–651,
New York, NY, USA. ACM Press.
Lam, K. M. and Leung, H. F. (2005). A trust/honesty
model in multiagent semi-competitive environments.
In PRIMA 2004: Proc. of the 17th Pacific rim Int.
Workshop, volume Lecture Notes in Computer Sci-
ence Vol. 3371, pages 128–147. Springer.
McLaughlin, M. R. and Herlocker, J. L. (2004). A collab-
orative filtering algorithm and evaluation metric that
accurately model the user experience. In SIGIR ’04:
Proc. of the 27th Int. Conf. on Research and develop-
ment in information retrieval, pages 329–336. ACM
Press.
Mui, L., Mohtashemi, M., and Halberstadt, A. (2002). No-
tions of reputation in multi-agents systems: a review.
In AAMAS ’02: Proc. of the first Int. Joint Conf. on Au-
tonomous agents and multiagent systems, pages 280–
287, New York, NY, USA. ACM Press.
NIST/NSA (2002). Fips 180-2. secure hash standard (shs).
NIST/NSA.
Pfitzmann, A., Pfitzmann, B., Schunter, M., and Waidner,
M. (1997). Trusting mobile user devices and security
modules. Computer, 30(2):61–68.
Ramchurn, S. D., Huynh, D., and Jennings, N. R. (2004).
Trust in multi-agent systems. Knowl. Eng. Rev.,
19(1):1–25.
Resnick, P., Kuwabara, K., Zeckhauser, R., and Friedman,
E. (2000). Reputation systems. Commun. ACM,
43(12):45–48.
Rijsbergen, C. J. V. (1979). Information Retrieval. Butter-
worth.
Rivest, R. L. (1992). The md5 message-digest algorithm.
Network Working Group and RSA Data Security, Inc.
Rivest, R. L., Shamir, A., and Adleman, L. (1978). A
method for obtaining digital signatures and public-key
cryptosystems. Commun. ACM, 21(2):120–126.
Sabater, J. and Sierra, C. (2001). Regret: reputation in
gregarious societies. In AGENTS ’01: Proc. of the
5th Int. Conf. on Autonomous agents, pages 194–195,
New York, NY, USA. ACM Press.
Sen, S. and Sajja, N. (2002). Robustness of reputation-
based trust: boolean case. In AAMAS ’02: Proc. of the
1st Int. Joint Conf. on Autonomous agents and multi-
agent systems, pages 288–293, New York, NY, USA.
ACM Press.
Shaked, Y. and Wool, A. (2005). Cracking the bluetooth
pin. In MobiSys ’05: Proc. of the 3rd Int. Conf. on
Mobile systems, applications, and services, pages 39–
50, New York, NY, USA. ACM Press.
Srivihok, A. and Sukonmanee, P. (2005). E-commerce in-
telligent agent: personalization travel support agent
using q learning. In ICEC ’05: Proc. of the 7th Int.
Conf. on Electronic commerce, pages 287–292. ACM
Press.
Tan, Y.-H. and Thoen, W. (2000). An outline of a trust
model for electronic commerce. Applied Artificial In-
telligence, 14(8):849–862.
Xiong, L. and Liu, L. (2004). Peertrust: Supporting
reputation-based trust for peer-to-peer electronic com-
munities. IEEE Transactions on Knowledge and Data
Engineering, 16(7):843–857.
Yamamoto, H., Ishida, K., and Ohta, T. (2004). Modeling
reputation management system on online c2c market.
Comput. Math. Organ. Theory, 10(2):165–178.
Zacharia, G. and Maes, P. (2000). Trust management
through reputation mechanisms. Applied Artificial In-
telligence, 14(9):881–907.
ICE-B 2006 - INTERNATIONAL CONFERENCE ON E-BUSINESS
104