A multi agent based decision mechanism for incident reaction in telecommunication network

A Multi-Agent based Decision Mechanism for
Incident Reaction in Telecommunication Network
Cédric Bonhomme, Christophe Feltus and Djamel Khadraoui
Centre for IT Innovation
Public Research Centre Henri Tudor
29, Avenue John F. Kennedy, L-1855 Luxembourg
Email: cedric.bonhomme@tudor.lu
Abstract—The main objective of this paper is to provide a
global architectural and decision support solution built on the
requirements for a reaction after alert detection mechanisms in
the frame of information systems security and more particularly
applied to telecom infrastructures security. These infrastructures
are distributed in nature, therefore the targeted architecture is
developed in a distributed perspective and is composed of three
basic layers: low level, intermediate level and high level. The low
level constitutes the interface between the main architecture and
the targeted infrastructure. The intermediate level is responsible
of correlating the alerts coming from different domains of the
infrastructure and to smartly deploy the reaction actions. The
architecture is elaborated using the multi-agents system that
provides the advantages of autonomous and interaction facilities,
and has been associated to the OntoBayes model for decision
support mechanism. This model helps agents to make decisions
according to preference values and is built upon ontology
based knowledge sharing, Bayesian networks based uncertainty
management and influence diagram based decision support. The
major novelty of this Decision Support System is the layered
view of the infrastructure thanks to MAS architecture, which
enables the decision making with different levels of knowledge.
The proposed approach has been successfully experimented for
data access control mechanism.
Index Terms—security; decision system; reaction; distributed
network; bayesian network; multi-agents system.
I. INTRODUCTION
Today information systems and mobile computing networks
are more widely spread and mainly heterogeneous. This
basically involves more complexity through their opening,
their interconnection and their ability to make decisions [9].
Consequently, this has a dramatic drawback regarding threats
that could occur on such networks via dangerous attacks (i.e.:
introduction of a malicious code or evil-minded modication
of a the DNS configuration file) [26]. This continuously
growing amount of carry out malicious acts encompasses new
and always more sophisticated attack techniques, which are
actually exposing operators as well as the end user.
State of the art in terms of security reaction is limited
to products that detect attacks and correlate them with a
vulnerability database but none of these products are built to
ensure a proper reaction to attacks in order to avoid their
propagation and/or to help an administrator to deploy the
appropriate reactions [19], [23]. In the same way, [22] says that
at the individual host-level, intrusion response often includes
security policy reconfiguration to reduce the risk of further
penetrations but doesn’t propose another solution in term of
automatic response and reaction. It is the case of CISCO based
IDS material providing mechanisms to select and implement
reaction decision.
Information security management and communication sys-
tems is actually in front of many challenges [12] due to the
fact that it is very often difficult to establish central or local
permanent decision capabilities, have the necessary level of
information, quickly collect the information, which is critical
in case of an attack on a critical system node, or launch
automated counter measures to quickly block a detected attack.
Based on that statements, it appears crucial to elaborate
a strategy of reaction after detection against these attacks.
Our previous work around that topic has provided first issues
regarding that finding and has been somewhat presented in [12]
and [15]. These papers have proposed an architecture to high-
light the concepts aiming at fulfilling the mission of optimizing
security and protection of communication and information
systems which purpose was to achieve the following:
• Reacting quickly and efficiently to any simple attack but
also to any complex and distributed ones;
• Ensuring homogeneous and smart communication system
configuration, that are commonly considered and the main
sources of vulnerabilities.
One of the main aspects in the reaction strategy consists
of automating and adapting policies when an attack occurs.
In scientific literature a large number of definitions for policy
and conceptual model exist. The most famous are Ponder [10],
Policy Description Language [5] and Security Policy Language
[2]. For the purpose of that paper, we prefer the one provided
in [10]: Policies are rules that govern the behavior of a system.
The provided policy adaptation is considered as a regu-
lation process. The main steps of the policy regulation are
described in Figure 1, which shows the process that takes
the business rules as input, and maps them onto technical
policies. These technical policies are deployed and instantiated
on the infrastructure in order to have a new state of temporary
network security stability adapted to the ongoing attack. This
policy regulation is thereafter achieved in modifying/adding
new policy rules to reach a new standing (at least up to the
next network disruption) policy based on the observation of
the systems current situation.

In this paper, we focus our work on policy deployment
and on policy modification decision-reaction challenges as
highlighted in the rounded rectangle of Figure 1. This twofold
challenge has already been addressed by other researches like
in [28]. Torrellas explains that facilitating timely decision-
making may achieve much greater productivity benefits by
engineering network security systems using multi-agents. In
[30], Yu developed the concepts of tele-service and proposed
an implementation of an e-maintenance platform based on a
Multi-Agent System (MAS). Yu explained how a Case-Based
Reasoning [1] method may be used to improve the autonomous
decision-making ability. Others works propose rather similar
solutions like [21], [8] but none are explicitly dedicated to the
management of security alerts reaction in the field of open
networks.
Consequently, the paper propose a system that combine a
reaction mechanism with the decision support. Such a problem
has never been addressed before the recently emerging agent-
based applications for reaction after detection infrastructures
as presented in [16].
Fig. 1: Policy regulation
The next section introduces the MAS architecture. Section 3
exposes the decision support system as well as its combination
with the MAS, and the last section concludes the paper.
II. MULTI-AGENT SYSTEM ARCHITECTURE
MAS is composed of several agents, capable of mutual
interaction. The interaction can be in the form of message
passing or producing changes in their common environment.
Agents are pro-actively, reactively and socially autonomous
entities able to exhibit organized activity, in order to meet
their design objectives, by eventually interacting with users.
An agent is collaborative by being able to commit itself to
society and/or another agent.
An agent encapsulates a state and a behavior and provides
moreover a number of facilities such as: control of its behavior,
the ability to decide even if external events influence its
decision, the possibility to exert its control in various manners
(reactively, directed by goals, socially). Moreover, MAS have
several control flows while a system with objects has a priori
only one control flow.
The agents also have global behavior within the MAS, such
as the cooperation (agents share the same goal), collaboration
(agents share intermittently the same goal) or competition
(incompatible goals between agents).
To manage several different systems, due to their location,
their business domain or their organization type, a distributed
system is appropriate. Furthermore, a distributed solution
brings some autonomy to the managed systems. Robustness,
survivability and availability are also impacted.
Fig. 2: Reaction architecture overview
The distributed architecture introduced in this paper is
composed by several components, called “operators”, which
have different responsibilities. Those operators are organized
in two dimensions, as presented in Figure 2.
The vertical dimension, structured in layers relative to
the managed network organization, allows adding abstraction
in going upward. Indeed, the lowest layer is closed to the
managed system and thus being the interface between the
targeted network and the management system. The higher
layer exposes a global view of the whole system and is able to
take some decisions based on a more complete knowledge of
the system, business, and organization. Intermediate levels (1
to n-1) guarantee flexibility and scalability to the architecture
in order to consider management constraints of the targeted
infrastructure. Those middleware levels are optional but allow
the system to be better adapted to the complexity of a given
organization and the size of the information system.
The horizontal dimension, containing three basic compo-
nents, is presented in Figure 2 and its three main phases are
described below:
1) Alert: Collect, normalize, correlate, analyze the alerts
coming from the managed networks and represent an
intrusion or an attack. If the alert is confirmed and coher-
ent, it is forwarded to the reaction decision component.
(Alert Correlation Engine-ACE).
2) Reaction Decision: Receive confirmed alerts for which
a reaction is expected. Considering the knowledge of:
policy, the systems’ organization and specified behavior,

these components decide if a reaction is needed or not
and define he reaction, if there is any. The reaction will
be modification(s), addition(s) or removal(s) of current
policy rules. (Police Instantiation Engine-PIE).
3) Reaction: Instantiation and deployment of the new poli-
cies, on the targeted networks. The deployment (Policy
Deployment Point PDP) and enforcement (Policy En-
forcement Point PEP) of these new policies, lead to a
new security state of the network.
Fig. 3: The three basic components
An issue is raised considering which layer is allowed to take
a decision reaction: only one layer, two, several, or all? If more
than one layer can trigger a reaction on the same object(s),
there will be a conflict issue. Thus, the system should be
able to provide mechanisms to solve conflicts between several
selected reactions. Another issue concerns the agreement: at
which level should it be asked? A solution could be to ask at
the same level (or at an upper one) that the reaction decision
is made; this should be specified by the user. A possible
solution is a distributed, vertically layered and hierarchical
architecture. The layer’s number could be adapted according to
the managed systems organization. In our case, three layers are
sufficient (local, intermediate and global). The reaction system
is composed of three main parts: the alert management part,
the reaction part and the police definition-deployment part.
Three trees (alert, reaction and policy) could be placed side
by side, as presented in Figure 2. These trees are alike but
their operators have different functions. The alert tree collects
the alerts with the local operators and correlate them in several
steps, one step by layer. A certain response time is used by the
system from intrusion detection to reaction application. This
time is increased if the reaction process is propagated to upper
layers, as presented in Figure 4. The global goal is of course
to shorten it.
The next step of our research development is firstly the
definition of a reaction engine that encompasses both, archi-
tecture components and the communication engine between
these components. This engine is based on a message format
and on a message exchange protocol based on standards such
as [11]. Secondly, real cases are studied in order to experiment
with the architecture and its associated protocol.
The message format is defined in XML format and is struc-
tured around a number of attributes that specify the message
source, the message destination and the message type (alert,
reaction, policy request, policy modification, policy modifica-
tion validation, decision and synchronization). The protocol
defines the exchange format and the workflow of messages
between the architecture components. It encompasses a set a
rules governing the syntax, semantics, and synchronization of
communication. The technical requirements request the opera-
Fig. 4: Response time
tor structure must be flexible in order to be able to reorganize
itself, if an operator fails or disappears. Each operator also
has to be autonomous in order to permit reorganization. Given
these requirements, the use of a MAS appears as a solution
to provide autonomy, flexibility and decision mechanisms to
each operator that are consequently represented by agents.
As studied in the state of the art presented in [14], a
set of agents could be managed and controlled through an
organization. An organization is a set of agents playing roles,
gathered in a normative structure and expecting to achieve
some global and local objectives. Several models like the roles
model, the tasks model, the interaction model or the norms
models specify an organization.
In our context we need an interaction definition in order to
specify communication protocols between agents representing
operators. We also need roles in order to specify which agent
will have to communicate or act in order to detect intrusions
and then react. Based on this needs, the use of an electronic
institution based on agents is one of the possibilities that we
will investigate.
The main goal of the reaction policy enforcement engine
is to apply policies in terms of specific concrete rules on
“technical” devices (firewall, fileserver, and other systems
named PEP). For that, we need means to make ACE, PIE,
PDP and PEP interact and collaborate.
The multi-agents systems concept already defines archi-
tectures and models for autonomous agents organization and
interaction. Existing platforms like JADE (Java Agent DEvel-
opment framework) [4], [3] implement agents’ concepts as
well as their ability to communicate by exchanging messages
and the reaction components integration could be simplified.
This is a solution, which will be detailed hereafter. The Foun-
dation for Intelligent Physical Agents (FIPA) [13] promotes
the success of emerging agent-based applications, services
and equipment. It makes available internationally agreed spec-
ifications that maximize interoperability across agent based
applications, services and equipment pursue this goal. This is
realized through open international collaboration of member

organizations, which are companies and universities active in
the agent field. FIPA’s specifications are publicly available.
They are not technologies for specific application, but generic
technologies for different application areas, and not just inde-
pendent technologies but a set of basic technologies that can
be integrated by developers to make complex systems with a
high degree of interoperability.
The used multi-agent framework is JADE. We base our-
selves on a survey made in [6] to argue that this agent platform
responds to the expectations in terms of agents’ functionalities,
security, performance, standardization, and secure communi-
cation between agents.
Figure 5 introduces the developed architecture. The flow
is supposed to begin with an alert detected by the IDS,
positioned on a network component. This alert is send to
the ACE agent (or LAN ACE if it concerns a precise Local
Area Network): This ACE agent confirms or not the alert to
the PIE. This decision to confirm the alert is explained in
section 3. Afterwards, the PIE decides to apply new policies
or to forward the alert to an ACE from a higher layer (upper
ACE). Its PIE agent sends the policies to the PDP agent, which
decides which PEP is able to implement it in terms of rules or
script on devices (firewall, fileserver, etc.) Then, the PDP agent
sends the new policy to the concerned PEP agent that knows
how to transform a policy into a rule or script understandable
by the associated device (for example fileserver).
On Figure 5, dash dot lines stand for flow of messages
encompassing alert or alert confirmation. Full lines stand for
flow of messages containing policies information, and dot lines
are reserved for decision support mechanisms.
Fig. 5: Multi-Agent System reaction architecture
A focused analysis of the PDP shows that it is composed by
several modules. For the multi-agent system point of view, the
Component Configuration Mapper results from the interaction
between the PDP agent and the Facilitator Agent while the
Policy Analysis module is realized by the PDP agent. The
Facilitator manages the network topology by retrieving PEP
agents according to their localization (devices registered with
IP address or MAC address) or according to actions they could
apply and their type (firewall, file server, etc.). For that the
Facilitator uses white pages and yellow pages services. The
JADE platform already provides implemented facilitator and
searching services. Besides, the use of a multi-agent system
as the framework provides flexibility, and openness. Actually,
when we decide to add a new PEP, we just have to provide its
PEP Agent with the ability to concretely apply the policies that
will register itself through the Facilitator, which will update
the databases.
III. DECISION SUPPORT ARCHITECTURE
Section 2 explains the developed MAS architecture that
guarantees a telecommunication security incident reaction.
This section explains the implementation of the decision
mechanism for incident reaction, the main objective of this
paper. For that reason the MAS architecture has voluntarily
been explained before the Decision Support System (DSS)
part because components of this architecture are used for the
illustration of the DSS.
One important challenge of the DSS is the management of
uncertainty. In our context uncertainty is defined as situation
“caused by a lack of knowledge about the environment when
an agent need to decide the truth of statement.”
Decision is a process [20] and consequently, it may be
represented using its input and its output. For the security
incident reaction, inputs of the decision mechanism are for
instance: the severity, duration and frequency of the alerts,
the impact on the system, or the network criticality for the
business whereas outputs are for instance: the escalation of
the alert to upper ACE or its confirmation to the PIE.
As explained by Yang [29], the decision-making mechanism
is composed of four pillars: Ontology, Bayesian Networks
(BN), Influence Diagram (ID) and Virtual Knowledge Com-
munity (VKC). In the framework of that paper, the VKC
will not be treated because the use of the 3 first pillars is
enough to understand the decision mechanism. The approach
preferred to design the decision mechanism is adapted from the
research performed by Yangs thesis for the incident reaction
through a MAS architecture. As a consequence our solution
differs from and completes the Yang research since our DSS
is illustrated by a real architecture for incident reaction that is
really deployed in our research labs.
A. Ontology
Ontology is the first pillar and is defined by a formal,
explicit specification of a shared conceptualization [24]. Ontol-
ogy may be categorized as domain ontology when it concerns
concepts and their relations from a same and well-defined
domain or top-level ontology when it concerns very general
domain-independent concepts. Ontology is the most important
pillar in that, it will be adapted to support the second pillar
concerning the Bayesian Network and the third pillar concern-
ing the Influence Diagram.
For the incident reaction system, ontology is defined using
the Web Ontology Language (OWL). Resource Development

Frameworks (RDF) syntax is the most commonly used method
to model information or meta-concepts in OWL. It may
be implemented in web resources and is structured based
on the triple (object, subject, predicate). Figure 6 illustrates
RDF graph. Both, object and subject are resources whereas
predicate is an attribute or a relation used to describe a
resource.
Fig. 6: RDF graph
Fig. 7: Dependency graph
In parallel to the MAS architecture developed in section 3,
we need a DSS to decide the transfer of an alert from the
IDS to the in-LAN ACE1
, for the forward of that alert to an
upper ACE, and for the confirmation of the alert to the PIE.
This is formalized using OWL as explained in Figure 8. On
that figure, ovals stand for OWL class, solid arrow lines stand
for RDF predicate, dash arrows for influence relations and
rounded rectangles for set of domain value.
Fig. 8: Decision system for alert transfer using OWL
B. OntoBayes
Ontology developed in the previous section permits to
formalize the concept encompassed in the MAS architecture as
well as their relations. However, at that the ontological level of
formalization, uncertainty challenge remains unaddressed and
decision mechanism remained needed for the agents to take
the decision.
1ACE agent located in the Local Area Network where alert is sent.
OntoBayes is an extension of OWL with two features:
Bayesian Networks (BNs) and Influence Diagram. BN address
the uncertainty and ID support the decision mechanism pro-
cess.
1) Bayesian networks extension: In probabilistic, Bayes
Theorem is a simple mathematical formula used for calculating
conditional probabilities [27]. It means that the calculations
of probability depend on prior knowledge that could be
considered as uncertain. I.e.: the probability of having a high
impact on the system if we have an alert of medium severity.
This probability is written P(alert.severity|system.impact).
The BNs extension of OWL introduces the parameters
of that formula by specifies the following two perspectives:
a qualitative perspective and a quantitative perspective. The
qualitative perspective specifies the random variables explic-
itly as well as their dependencies and the later associates’
quantitative information to those variables.
The specification of random variable and their dependency
is performed by introducing the new OWL property ele-
ment <owl:ObjectProperty rdf.ID="dependsOn"
/> and could be graphical represented as illustrated on Fig-
ure 7.
Accordingly, the qualitative extension may be represented
by 2 Bayesian graph models (Figure 9) extracted from the
OWL graph model from Figure 7.
Fig. 9: Bayesian graph models
The ovals represent Bayesian variables and the arrows
specify their relations. The graph is to be read i.e. 1.: The
alert that is forwarded from the out-LAN ACE2
to the network
upper ACE has influence on the confirmation of the alert that
is send from the upper ACE to the PIE. I.e. 2.: The severity
of the alert has influence on the action to send an alert to the
in-LAN ACE. The last examples may be translated using the
new OWL dependsOn element as in Figure 10.
The quantitative extension is performed with the association
of probability table to the Bayesian variables. In the case
of the above example, the Table I provides the quantitative
probability P(alert.severity|system.impact) and is represented
on Figure 10 by the Bayesian variables database.
For example according to Bayes’ formula for conditional
probability the first line means that if the severity is low, the
2In opposition with in-LAN ACEs, out-LAN ACEs are agents located in
others Local Area Networks than the LAN where alert is sent.

<owl:Class rdf:ID="alert.severity">
<owl:Restriction>
<owl:onProperty>
<owl:ObjectProperty red:ID="dependsOn" />
</owl:onProperty>
<owl:hasValue rdf:resource="system.impact" />
</owl:Restriction>
</owl:class>
Fig. 10: Dependency encoding
TABLE I: Bayesian variables value probability
ProbCell HasPParameters HasPValue
Cell 1 alert.severity=low|system.impact=low 0.6
Cell 2 alert.severity=medium|system.impact=low 0.3
Cell 3 alert.severity=high|system.impact=low 0.1
Cell 4 alert.severity=low|system.impact=medium 0.2
Cell 5 alert.severity=medium|system.impact=medium 0.5
Cell 6 alert.severity=high|system.impact=medium 0.3
Cell 7 alert.severity=low|system.impact=high 0.1
Cell 8 alert.severity=medium|system.impact=high 0.3
Cell 9 alert.severity=high|system.impact=high 0.6
probability that the impact is low will be relatively high. Of
course if the severity is high the chance to have a low impact
on the system is minimal (generally):
P[alert.severity = high | system.impact = low] = 0.1
In order to follow the law of total probability we must have :
P[alert.severity = low ∩ system.impact = low]
+ P[alert.severity = medium ∩ system.impact = low]
+ P[alert.severity = high ∩ system.impact = low]
= 0.6 + 0.3 + 0.1
= 1
The conditional probability from Table I is encoded as
follows (Figure 11):
<owl:Class rdf:ID="Alert">
<CondProbDist rdf:ID="table_1">
<hasPCell>
<ProbC rdf:ID="Cell_1">
<HasPValue rdf:Iddatatype="#float">0.6</HasPValue>
<HasParameters rdf:datatype="#string">
alert.severity=low|system.impact=low
</HasParameters>
</ProbC>
</HasPCell>
...
</CondProbDist>
</owl:Class>
Fig. 11: Bayesian variables value probability encoding
2) Inﬂuence diagrams extension: IDs extension aims at
representing and analyzing a decisional model to support
the decision-making process. The review of the literature
that treats ID [17], [18] shows that decision mechanisms are
composed by three types of nodes: 1) Chance nodes that rep-
resent variables that are not controled by the decision maker,
2) Decision nodes that represent choices available for the
decision maker, and 3) Utility nodes that represent agent utility
functions. Additionally, [25] explains that three type of arcs
express the relationship between nodes: I) Information arcs
(isKnownBy) that point out the information that is necessary
for the decision maker, II) Conditional arcs (influenceOn)
that point out the probabilistic dependency on the associated
variable, and III) Functional arcs (attributeOf) that point
out variables used by utility nodes as decision criteria.
Fig. 12: IDs graph model of alert transfer
Based on that structure of decisional model, the alert trans-
fer may be represented in Figure 12. Ovals stand for Chance
nodes, rectangles stand for Decision nodes, and diamonds
stand for Utility nodes. The information arc relates to all
information observed to make a decision and the conditional
arc relates to data issued from Chance node and considered
as evidence for the Decision nodes.
Additionally, to make a decision, the agent that takes a
decision needs to have its preferences quantiﬁed according to
a set of attributes. In [7], Butler introduces the theory of multi-
attribute utility (MAUT) to quantify a preference with numer-
ical value. The best preference has the higher value whereas
the worst has the lower one. To achieve that, the Utility node
is associated with a utility table that gathers the preferences of
all decision choices. Table II shows these preferences for the
in-LAN ACE alert sending decision taking mechanism and is
represented by the utility database in Figure 5.
TABLE II: Utility table for in-LAN ACE alert sending
UtilityCell HasUParameters HasUValue
Cell 1 send(alert.InLanACE)=yes|severity.alert=low -80
Cell 2 send(alert.InLanACE)=yes|severity.alert=medium 50
Cell 3 send(alert.InLanACE)=yes|severity.alert=high 100
Cell 4 send(alert.InLanACE)=no|severity.alert=low 80
Cell 5 send(alert.InLanACE)=no|severity.alert=medium 40
Cell 6 send(alert.InLanACE)=no|severity.alert=high -100
The Figure 13 shows the encoding of Table II utility table

for in-Lan ACE alert sending.
<owl:Class rdf:ID="send(alert.InLanACE)">
<owl:Restriction>
<owl:onProperty>
<owl:ObjectProperty rdf:ID="attributeOf" />
</owl:onProperty>
<owl:hasValue rdf:resource=#U />
</owl:Restriction>
...
<rdfs:subClassOf>
<owl:hasValue rdf:ID="DecisionNode" />
</rdfs:subClassOf>
</owl:Class>
<owl:Class rdf:ID="severity.alert">
...
<rdfs:subClassOf>
<owl:hasValue rdf:ID="ChanceNode" />
</rdfs:subClassOf>
</owl:Class>
<owl:Class rdf:ID="U1">
<UtilityTable rdf:ID="table_1">
<hasUCell>
<UtilityCell rdf:ID="cell_1">
<hasUParameter rdf:datatype="#string">
send(alert.InLanACE)=yes,severity.alert=low
</hasUParameter>
<hasUValue rdf:datatype="#float">-80</hasValue>
</UtilityCell>
</hasUCell>
...
<hasUCell>
<UtilityCell rdf:ID="cell_6">
<hasUParameter rdf:datatype="#string">
send(alert.InLanACE)=no,severity.alert=high
</hasUParameter>
<hasUValue rdf:datatype="#float">-100</hasValue>
</UtilityCell>
</hasUCell>
</UtilityTable>
</owl:Class>
Fig. 13: Utility encoding
As seen in Figure 7, a sequential path between all decisions
exists. Indeed, some decision depends on previous decisions
and as a consequence, previous decisions (Decision node)
become Chance nodes for next Chance node. Figure 12 illus-
trates that send(alert.InLanACE) is at the same time a Decision
node and a Chance node that is known be the decision node
alertForward2(InLanACE, UpperACE).
IV. CONCLUSION
In this paper we have presented a global solution developed
for an incident reaction system based on a policy regulation
approach strategy. The solution is composed firstly with a
MAS that offers the advantage to react quickly and efficiently
against an attack while being adapted for distributed networks
and secondly with a decision support system that helps agents
to make decisions based on utility preference values. This is
achieved by taking uncertainty into account through Bayesian
networks and influence diagram.
The decision support system has been explained for the
transfer of an alert from the alert correlation engine to the
policy instantiation engine. Other decision points exist in
the architecture. All of them could be solved using decision
support system but they are not explained in the paper.
An important advantage of this decision support system is
its capability to take decision at different points of the network.
If more knowledge is needed to take a decision, the higher
layer gives a global view of the whole system and is able to
take decisions based on a more precise state of the system.
The future works based on our achievements will be the
specification of a protocol, specification of the messages and
thus the reaction methodology service oriented based. This
protocol and methodology will be dedicated to the architecture
presented in this paper and address the interoperability issues
with regard to the policy representation and modeling.
ACKNOWLEDGMENT
This research was funded by the National Research Fund
of Luxembourg in the context of TITAN (Trust-Assurance for
Critical Infrastructures in Multi-Agents Environments, FNR
CO/08/IS/21) project.
REFERENCES
[1] Agnar Aamodt and Enric Plaza. Case-based reasoning: Foundational
issues, methodological variations, and system approaches. AI Commun.,
7(1):39–59, 1994.
[2] Cataldo Basile, Antonio Lioy, Gregorio Mart´ınez Pérez, Félix J. Garc´ıa
Clemente, and Antonio F. Gómez-Skarmeta. Positif: A policy-based
security management system. In POLICY, page 280. IEEE Computer
Society, 2007.
[3] F. Bellifemine, G. Caire, A. Pogg, and G. Rimassa. Jade - a white paper.
Technical Report 3, Telecom Italia Lab, EXP Online, 2003.
[4] Fabio Bellifemine, Agostino Poggi, and Giovanni Rimassa. Jade - a
fipa-compliant agent framework. Technical report, CSELT, 1999.
[5] Elisa Bertino, Alessandra Mileo, and Alessandro Provetti. Pdl with
preferences. In POLICY, pages 213–222. IEEE Computer Society, 2005.
[6] E. Bulut, D. Khadraoui, and B. Marquet. Multi-agent based secu-
rity assurance monitoring system for telecommunication infrastructures.
Network, and Information Security conference (CNIS 2007), Berkeley,
California, USA, 2007.
[7] John Butler, Douglas J. Morrice, and Peter W. Mullarkey. A multiple
attribute utility theory approach to ranking and selection. Manage. Sci.,
47(6):800–816, 2001.
[8] Carlos Carrascosa, Javier Bajo, Vicente Julián, Juan M. Corchado, and
Vicente J. Botti. Hybrid multi-agent architecture as a real-time problem-
solving model. Expert Syst. Appl., 34(1):2–17, 2008.
[9] A. Cuevas, P. Serrano, J. I. Moreno, C. J. Bernardos, J. Jähnert, and R. L.
Aguiar. Usability and evaluation of a deployed 4g network prototype.
Journal of Communications and Networks, Vol. 7 (2), 2008.
[10] Nicodemos Damianou, Naranker Dulay, Emil Lupu, and Morris Sloman.
The ponder policy specification language. In Morris Sloman, Jorge
Lobo, and Emil Lupu, editors, POLICY, volume 1995 of Lecture Notes
in Computer Science, pages 18–38. Springer, 2001.
[11] H. Debar, France Telecom, D. Curry, Guardian, B. Feinstein, and Inc.
SecureWorks. The intrusion detection message exchange format (idmef).
IDMEF/RFC4765, 2007.
[12] C. Feltus, D. Khadraoui, B. de Rémont, and A.Rifaut. Business
gouvernance based policy regulation for security incident response.
IEEE Global Infrastructure Symposium, 2007.

[13] FIPA. http://www.fipa.org/.
[14] B. Gâteau. Modélisation et Supervision d’Institutions Multi-Agents. PhD
thesis, École Supérieure des Mines de Saint-Étienne, 2007.
[15] B. Gâteau, D. Khadraoui, and C. Feltus. Multi-agents system service
based platform in telecommunication security incident reaction. IEEE
Global Information Infrastructure Symposium, 2009.
[16] Benjamin Gâteau, Djamel Khadraoui, Christophe Feltus, and Benoˆıt De
Rémont. Multi-agents based architecture for is security incident reaction.
In RIVF. [PDF], 2008.
[17] R. A. Howard and J.E. Matheson. Influence diagrams. Decision Analysis,
pages 127–143, 2005.
[18] Finn V. Jensen. Bayesian Networks and Decision Graphs (Information
Science and Statistics). Springer, July 2001.
[19] L. J. LaPadula. State of the art in anomaly detection and reaction
technical report mp 99b0000020. Mitre, July 1999.
[20] H. D. Lasswell. The decision process; seven categories of functional
analysis. College of Business and Public Administration, University of
Maryland, 1956.
[21] Kun-Yung Lu and Chun-Chin Sy. A real-time decision-making of
maintenance using fuzzy agent. Expert Syst. Appl., 36(2):2691–2698,
2009.
[22] M. Petkac and Lee Badger. Security agility in response to intrusion
detection. In ACSAC, pages 11–20. IEEE Computer Society, 2000.
[23] G.L.F. Santos, Z. Abdelouahab, R.A. Dias, C.F.L. Lima, E. Nascimento,
and E.M. Cochra. An automated response approach for intrusion
detection security enhancement. Software Engineering and Applications,
2003.
[24] Rudi Studer, V. Richard Benjamins, and Dieter Fensel. Knowledge
engineering: Principles and methods. Data Knowl. Eng., 25(1-2):161–
197, 1998.
[25] J.A. Tatman and R.D. Shachter. Dynamic programming and influence
diagrams. IEEE Transactions on Systems, Man, and Cybernetics,
20(2):365–379, 1990.
[26] Joseph Chee Ming Teo, Chik How Tan, and Jim Mee Ng. Denial-of-
service attack resilience dynamic group key agreement for heterogeneous
networks. Telecommunication Systems, 35(3-4):141–160, 2007.
[27] Bayes theorem. http://plato.stanford.edu/entries/bayes-theorem/.
[28] Gustavo A. Santana Torrellas and Luis A. Villa Vargas. Modelling a
flexible network security systems using multi-agents systems: security
assessment considerations. In ISICT, volume 49 of ACM International
Conference Proceeding Series, pages 365–371. Trinity College Dublin,
2003.
[29] Y. Yang. A framework for decision support systems adapted to uncertain
knowledge. PhD thesis, University of Karlsruhe, 2007.
[30] R. Yu, B. Iung, and H. Panetto. A multi-agents based e-maintenance
system with case-based reasoning decision support. Engineering Appli-
cations of Artificial Intelligence, pages 321–333.

A multi agent based decision mechanism for incident reaction in telecommunication network

More Related Content

What's hot (16)

Viewers also liked (16)

Similar to A multi agent based decision mechanism for incident reaction in telecommunication network (20)

More from Luxembourg Institute of Science and Technology (20)

Recently uploaded (20)

A multi agent based decision mechanism for incident reaction in telecommunication network