A Meta-heuristically Optimized Fuzzy Approach towards Multi-metric
Security Risk Assessment in Heterogeneous System of Systems
I
˜
naki Eguia and Javier Del Ser
TECNALIA, 48170 Zamudio, Bizkaia, Spain
Keywords:
Multi-metric Security Assessment, Fuzzy Logic, Meta-heuristics
Abstract:
Security measurement of complex systems is a challenging task since devices deployed over the so-called
System of Systems (SoS) are extremely heterogeneous and hence imply an interoperability effort in order to
enable a common resilient security measurement language. Moreover, systems demand more features beyond
security concept, require to preserve privacy and claim for dependable structures in order to seek a holistic and
aggregated security and safety view. This paper addresses this need by capitalizing the availability of multiple
security metrics through an hybrid meta-heuristic fuzzy aggregation and composition approach that takes into
account the expertise compiled by the security manager, towards the generation of visual dashboards reflecting
the SPD (Security, Privacy and Dependability) risk status of the system at hand.
1 INTRODUCTION
In the security field one of the most sounding
paradigms resides in the security interoperability
within a heterogeneous device landscape. In this
scenario each of such devices performs with different
protocols and scales. Interoperability is by itself a
challenge for engaging the security paradigm as a
built-in approach. Many efforts have been devoted
to define security patterns (Yoshioka et al., 2008;
Heyman et al., 2007; VanHilst & Fernandez, 2007;
Fernandez et al, 2007) so as to develop and implement
a understandable security backdrop.
Unfortunately, this background is usually quite
complex. In this context, the setup elaborated by
the nSHIELD initiative (NSHIELD, 2013) proposes
a System of Systems representing a set of assets: sub-
systems, software components, protocols and devices
and boards. These elements are configured along
node, network and middleware-overlay layers. On the
other hand, risks are denoted as the probability that a
threat can become real impacting in a vulnerability
of one or more components from those listed above.
Threats are numerous and can be predictable or
unpredictable. Those which can be predicted may
be guarded by metrics, whereas those which are
unpredictable could be challenged by composable,
heuristics techniques. Therefore, metrics will be
mapped to threats and problems to be measured.
For example, if a problem of a given system is
the network latency, the network latency must be
measured. This obvious affirmation is the key for
starting the process for selecting the correct metrics
for any heterogeneous system. Consequently, risk
analysis is the first process for creating a set of
security metrics in scenarios with a diverse device
spectrum.
This paper joins this research trend by outlin-
ing and sketching a novel multi-metric approach for
heterogeneous systems capable of measuring risks
impacting on different vulnerabilities in an inter-
operable and normalized fashion. This practical ap-
proach requires understanding and identifying before-
hand security needs of the system at hand, as well as
those of its constituent components. In other words,
current vulnerabilities and threats must be identified
so as to discriminate those vulnerabilities and threats
more likely to take place in the future. To this end, this
paper elaborates in advance of several evident, rele-
vant observations: 1) systems’ complexity yields an
accordingly complex and also non-interoperable se-
curity management; and 2) without loss of generality,
metrics considered in this approach are restricted to
operational activities beyond security organizational
metrics and measurements. Moreover, this paper de-
picts SPD metrics as resiliently built-in artifacts that
are aggregatedly delivered from the engineering pro-
cess to the operation phase, thus rendering a holistic
SoS SPD view. Finally, a self-optimized hybrid meta-
heuristic fuzzy system will be conceptualized as a de-
231
Eguia I. and Del Ser J..
A Meta-heuristically Optimized Fuzzy Approach towards Multi-metric Security Risk Assessment in Heterogeneous System of Systems.
DOI: 10.5220/0004876802310236
In Proceedings of the 4th International Conference on Pervasive and Embedded Computing and Communication Systems (MeSeCCS-2014), pages
231-236
ISBN: 978-989-758-000-0
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
cision support system capable of providing the sys-
tem security manager with an easy and understand-
able mechanism to tune and trigger security-related
actions and rules based on fine-grained input metrics
and measurements.
2 SPD METRICS IN
HETEROGENEOUS DEVICE
ECOSYSTEMS
The meaning of metric can be understood from a
naive business standpoint in the sense that metrics
should be made similar to human understanding
because they need to be recognized and understood
by both business and technical engineers operating on
the systems. Informational SPD metrics are deemed
an important factor when making reliable, well-
grounded decisions about several aspects of security
and dependability, ranging from the design of SPD
architectures and controls to the effectiveness and
efficiency of SPD operations. SPD metrics strive to
offer a quantitative and objective basis for security
assurance.
Given the application scope in Industrial Systems
of Systems (SoS) environment tackled in this paper,
we will analyze and combine traditional operational
indicators with SPD metrics. Metrics in industrial
operations (i.e. metrology) have always been regu-
lated and certified by a higher authority. In the IT
area this methodology has been applied in a less rigor-
ously manner, since safety has always prevailed over
security. However, SoS environments require both se-
curity and safety (part of dependability) requisites, as
threats could proceed from both virtual and real (un-
predicted/failure threats) worlds.
The spectrum of SPD metrics derived in
nSHIELD constitute the first attempt in the related lit-
erature to correlate operational and SPD metrics so
as to develop a business continuity approach for in-
dustrial sectors. It is important to highlight that con-
trol systems such as SCADAS or ICS (Industrial Con-
trol Systems) not only depend on the operational pro-
cess (which is linked directly to business), but also
is becoming progressively more dependent on robust-
ness, resilience and security factors that preserve op-
eration from malicious attacks and large failures. In-
deed, the dependability concept guarantees this fact:
dependability mechanisms deal with availability (e.g.
threats against DDoS), which is the most important
feature for industrial operations. However, security
may also be threatened in terms of integrity; for in-
stance, a man-in-the-middle attack for value modi-
fication in the communications link from smart me-
ters to concentrators could cause less profit to elec-
tric vendors and an unequal operation for distribu-
tion system operators when balancing energy offer-
ing & demand. On the other hand, the notion of pri-
vacy implies confidentiality and anonymity, and is be-
coming essential as Big Data gets involved within in-
dustrial and large organizational settings. Therefore,
nSHIELD metrics are set towards business continuity:
1) heterogeneous but measurable; 2) understandable
by the human being with comprehensible, possibly
fuzzy glossary; and 3) composable since inputs are
aggregated through an expert system. This envisaged
portfolio of requirements makes the nSHIELD view
fulfill with the so-called Security by Design (SbD)
principles (Cavoukian & Dixon, 2013): nSHIELD
SPD metrics apply to security, privacy and depend-
ability built-in concepts and functionalities within the
whole engineering process. Furthermore, such func-
tionalities are not seen as the final patch but as an in-
trinsic conceptualization of the whole and holistic en-
gineering and operation procedure. SPD metrics are
extracted from a set of SPD requirements and risks
statements within SoS scenarios.
Bearing this rationale in mind, the nSHIELD
multi-metric approach follows a quantitative focus.
This approach provides a metric template for metrics
identification and gathering process. In the first spec-
ification approach more than 60 metrics were iden-
tified and structured in layers {SPD, L
x
}. Metrics
identified as heterogeneous sometimes overlap in dif-
ferent layers. The result of measurements according
to these metrics has to be described quantitatively.
The following list oversees some of the most relevant
metrics concluded after this study
1
:
1. Code execution (Node Layer): verification that
only authorized code (booting, kernel, applica-
tion) runs on the system.
2. Network delay (Network Layer): this is a perfor-
mance metric used for measuring the delay in-
duced by a node in retransmitting incoming data.
3. Network Capacity (Network Layer): this is a per-
formance metric used for measuring the networks
capacity, which shall be large enough to allow
the necessary traffic to go through. As a rule of
thumb, at normal operation, the traffic should be
about 60-70% of the network capacity, so as to
avoid bottlenecks when there will be traffic peaks.
4. Discovery frequency (Middleware Layer):
amount of discovery events per protocol and unit
1
Upon its acceptance a more detailed list of metrics will
be presented and discussed, along with their corresponding
formulae.
PECCS2014-InternationalConferenceonPervasiveandEmbeddedComputingandCommunicationSystems
232
of time.
5. Metric Composition statistics (Middleware
Layer): statistics on the availability of the compo-
sition service: queue length (momentarily count
of outstanding composition requests); Thread
count (count of threads serving composition re-
quests); used memory (amount of heap memory
reserved by the service); wait time (time from
arrival to sending, averaged for all requests per
interval); and response time (time from start of
sending to receiving response, averaged for all re-
quests per interval).
6. Profit per agent type (Middleware Layer): it
stands for the profit that each agent type (namely,
evil or malicious, disturbing, selfish and honest)
can obtain according to its behaviour. The profit
is higher if an agent cheats on another agent.
7. Uptime (Overlay Layer): this metric denotes the
uninterrupted system availability.
8. Attack surface (Overlay Layer): used to count the
number of data inputs to an overlay node.
9. Failed authentication (Overlay/Middleware
Layer): used to count the number of data inputs
to an overlay node.
10. Detection accuracy (Overlay Layer): Measure
of the detection accuracy as the ratio of the
number of undetected attacks to the total amount
of detected attacks.
11. Total attack impact (Overlay Layer): quantifies
the impact of the attack as the ratio of the number
of nodes attacked to the total amount of nodes.
2.1 Design of the proposed Multi-metric
Decision Making Approach
Once the overall set of metrics under consideration
has been defined, a multi-metric decision making
approach is applied to infer and trigger security
actions rules based on the values of the metrics in a
understandable fashion. Such expert system builds
upon the following design steps:
A. Selection of Metrics: system operators will se-
lect specific security metrics according to the require-
ments and risk factors of the scenario at hand. In
particular, the nSHIELD approach addresses more
than 60 types of SPD metrics structured in 4 layers:
node, network, middleware and overlay. System op-
erators must decide which metrics refer better to their
business operation: some might prioritize integrity to
availability, whereas other could trade reliability for
privacy, etc.
B. Normalization and Regression: Each of the
metrics identified previously may feature different
units and value range. This means that such values
must be normalized in order to have a common
value range domain and not to bias subsequent
processing stages of the decision making engine.
This is of utmost importance in order to define a
common value range for their mutual comparison,
and is indeed one of the purposes of the SPD metric
triplet: to homogenize the representation of metrics
in a threefold risk-valued domain. In other words,
metrics quantify phenomena of very diverse nature,
and therefore samples can be represented in different
units (e.g. seconds, KW or Celsius) and within
distinct numerical ranges (i.e. from discrete sets such
as key par length – 64, 128, 256, 512 – to continuous-
valued quantities such as time lapse in nanoseconds).
For instance, if one assumes that the network latency
takes on valid values from a positive region between
4 and 10 milliseconds, correct SPD values for this
metric will be declared if all fall within this range.
This axiom shall only be valid in a particular domain
concerning a certain business process. If is hence
between 4 and 10 milliseconds, the valid range for
the network latency could be set to [0, 10].
At this stage it must be pointed out that al-
though time is an objective and measurable concept,
the S level (Security Threat Level) is a subjective,
experience-earned concept that the system operator in
the nSHIELD approach must set up. The same con-
clusion holds for the P and D levels as they corre-
spondingly relate to privacy and dependability. The
S level represents the security threat level that will
be bounded by its valid value range and its normal
performance region. All these subjective SPD indi-
cators must be determined by expert engineers and
system operators who have the operational experience
and who are in the position to shape universal secu-
rity principles to the specific scenario under opera-
tion. Therefore, the decision making approach pro-
posed in this work not only takes into account the ex-
perience of the operator, but also learns from past de-
cisions in order to set up the best indicators for the
considered metrics.
Table 1: Example of S level values for a temporal variable
mapped by an expert.
Time 0 1 2 3 4 5 6 7
S level 0 2 4 6 8 11 14 17
Time 8 9 10 11 15 25
S level 20 23 26 32 44 58
Returning to the aforementioned example, the S
level associated to the first time values [0, 3] could
be labeled as unlikely to happen under the criterion
AMeta-heuristicallyOptimizedFuzzyApproachtowardsMulti-metricSecurityRiskAssessmentinHeterogeneousSystem
ofSystems
233
Input scaling
Fuzzification
Rule
Defuzzification
Output scaling
application
SPD
regression
SPD
regression
SPD
regression
[S,P,D]
metricA
.
.
.
[S,P,D]
metricZ
metricZ
metricB
metricA
S P D
Node
Network
Middleware
Overlay
Security operator
metricD
metricB
metricA
metricZ
.
.
.
Meta-heuristic optimizer
. . .
KB1 KB2 KB3 KB4 KBN
KB under evaluation
Comparison between output
by the system and output
by the operator
Iterative
S
metricZ
metricZminZ
maxZ
100
0
Figure 1: Overview of the proposed meta-heuristic fuzzy multi-metric aggregation scheme. The term KB (Knowledge Base)
stands for any set of ruleset, scaling and membership functions within the population tuned by the meta-heuristic solver.
and experience of the system administrator. Correct
values fall between [4, 10], within where the threat
level is kept linear to reflect that systems are able to
handle this security context in an scalable fashion.
However, the mapped level of security context is
set to grow exponentially beyond 10 milliseconds
and logarithmically above 15 miliseconds in order
to represent a sharp increase of the severity of the
security context. A similar methodology should
be adopted for privacy and dependability concepts
referring to each of the considered metrics.
The behavior of the SPD metrics with respect to
any given value of the input metric can be analytically
characterized in terms of the expected risk for each
of such values. To this end, a function f
α
λ
:
x
λ
r
α
can be inferred by regressing the values
x
λ
taken by metric λ to a risk value r
α
[0, 100]
for SPD metric α {S, P, D}. The analytical
expression for this function f
α
λ
can be obtained by
performing a piecewise regression (not necessarily
linear) over a set of metric-risk pairs input by the
security manager of the system at hand. As a result of
the overall processing step, one obtains a numerical
triplet [S, P, D] for every security metric input to the
system, which is then processed via a decision making
engine to produce visual indicators of the level of
SPD risk for every layer of the system based on the
expertise of the manager for security operations.
C. Multi-metric Aggregation based on Expert
Systems and Meta-heuristically Optimized Fuzzy
Systems: The last step of the envisaged multi-metric
security approach consists of an aggregation and
decision making engine that processes the numerical
values of the monitored security metrics (part of
which are shown in the enumerated list of Section
2) towards providing a fuzzy representation of the
risk level of the heterogeneous device ecosystem at
hand. The aggregation system should 1) comprise
a set of human understandable linguistic rules (in
the form of IF-THEN-ELSE conditional statements)
to ease their supervision and manual crafting; 2)
automatically infer optimized rule sets based on a
performance index that reflects the error between
the produced decision and that provided by security
experts; and 3) tailor the mapping between the
numerical and linguistic domains corresponding to
the SPD values of the considered metric with their
fuzzy representation and processing through the rule
set. These specifications are deemed of utmost
importance in our attempt at avoiding black-box
decision engines that provide no further benefit than
the blind automation of the decisions to be made.
These specifications call for the adoption of a spe-
cific class of fuzzy systems that incorporates stochas-
tic solvers to optimize the scaling factors and mem-
bership functions that specify the meaning of the nu-
merical SPD inputs in the linguistic domain, as well
as the collection of fuzzy rules that best matches
the decisions taken by security experts (see Figure
1). Traditionally tackled via evolutionary algorithms
(mostly, genetic optimizers (Cordon et al., 2004,
2001)), we plan to analyze and benchmark the perfor-
mance – in the context of optimizing fuzzy rule-based
systems of more recent meta-heuristic solvers. In
particular we will focus on Harmony Search (Geem et
al., 2001), a population-based optimization algorithm
that has been proven to outperform their genetic coun-
terparts in a plethora of application fields such as en-
gineering optimization, routing, resource allocation,
economics and operations research (see (Manjarres et
al., 2013) for a thorough survey). This algorithm re-
sembles the way musicians in an orchestra improvise
with their instruments before playing a piece of music
in order to reach an aesthetically well-sounding har-
mony. This mimicking has implications in the def-
PECCS2014-InternationalConferenceonPervasiveandEmbeddedComputingandCommunicationSystems
234
µ(x) =
0 if S
metricA
< a,
2
λ1
S
metricA
a
ba
λ
if a S
metricA
(a + b)/2,
1 2
λ1
bS
metricA
ba
λ
if (a + b)/2 < S
metricA
< b,
1 2
β1
S
metricA
b
cb
β
if b S
metricA
< (b + c)/2,
2
β1
cS
metricA
cb
β
if (b + c)/2 S
metricA
c,
0 if S
metricA
> c.
(1)
inition of the operators driving the search procedure
of the algorithm, which ultimately leads to a better
adaptability of the explorative and exploitative capa-
bilities with respect to the problem to be optimized.
Back to the pursued aggregation system, the
goal is to produce an intensity, colored, real-valued
indicator of the level of risk for the S, P and D
components at each of the considered layers (node,
network, middleware, overlay), in the form of an
output matrix. To this end, the expert would be first
asked to provide, by means of e.g. questionnaires,
their estimated output to a series of eventual metric
values so as to lay decisional baseline information.
This would permit to compare the output of the
system under differently optimized fuzzy systems
and extract therefrom a performance index (integer
from the range [0-100], in %) quantifying the level
of compliance of the optimized decision maker with
the expected output by the security expert(s). This
performance index would measure the fitness of every
combination of rule set/membership function/scaling
factor iteratively refined by the harmony search
optimizer. In particular, parameterized membership
functions and rules will be jointly optimized by means
of a Pittsburg approach where the population of the
algorithm is formed by separately encoded variable
domains.
Once the fuzzy engine is optimized after a given
number of iterations of the Harmony Search heuris-
tic, the aggregation system is ready to receive out-
puts from the different security metrics and fuzzify
them through the optimized scaling and membership
functions µ(x), which could be parametrized to yield
the generic function formulation in Expression (1) to
be optimized by means of the aforementioned meta-
heuristic solver, where S
metricA
denotes the S com-
ponent of metric A (the same holds for the rest of SPD
components and metrics), and {λ, β, a, b, c} are the
parameters to be optimized. It should be obvious that
λ and β define the shape of right and left slopes of
the membership function at hand, i.e. λ = β = 1
would correspond to the well-known triangular func-
tions with center b and upper and lower extremes a
and c, respectively.
Next the engine would apply the simultaneously
optimized rule set and combine their outputs into a
linguistically encoded output SPD matrix (by means
of conventional methods for accumulating fuzzy
outputs of individual rules, such as normalized and
bounded sums), which is finally defuzzified (via
e.g. the center of gravity method (Van Leekwijck &
Kerre, 1999)) into colored intensity indicators so as to
yield a more intuitive security assessing information
of increased visual understandability. This 4 ×
9 output matrix of integer entries between 0 (low
intensity) and 10 (high intensity) facilitates a global
perspective of the Security, Privacy and Dependability
risk status at node, network, middleware and overlay
layers. All in all, this visual information provides an
holistic understanding of aggregated metrics in a SoS
scenario.
3 CONCLUDING REMARKS
This manuscript has outlined the main design prin-
ciples of a metric aggregation engine technically en-
visaged as a fuzzy expert system hybridized with a
novel meta-heuristically optimized learning mecha-
nism. The paper has gravitated, from a theoreti-
cal perspective, on the feasibility of measuring met-
rics in an heterogeneous complex systems environ-
ment. Mainly motivated by the lack of universal secu-
rity guidelines for highly heterogeneous systems, the
proposed aggregation approach regards system expert
knowledge as an essential basis for security decision
making. However, we foresee that this dependency
could also be circumvented by a learn-as-it-goes ap-
proach, i.e. by supervising the SPD levels of succes-
sively held security incidences and feeding this infor-
mation back to the decision making tool for its au-
tonomous adjustment. Next steps will be oriented to-
wards implementing and validating this procedure in
diverse application scenarios.
AMeta-heuristicallyOptimizedFuzzyApproachtowardsMulti-metricSecurityRiskAssessmentinHeterogeneousSystem
ofSystems
235
REFERENCES
Yoshioka, N., Washizaki, H., Maruyama, K., 2008, A
Survey on Security Metrics, Progress Informatics, N. 5,
pp. 35-47.
Heyman, T., Yskout, K., Scandariato, R., Joosen, W.,
2007, Analysis of the Security Patterns Landscape,
International Workshop on Software Engineering for
Secure Systems. Washington, DC, USA, p. 3.
VanHilst, M., Fernandez, E. B., 2007, Reverse Engineering
to Detect Security Patterns in Code. Proceedings of
the International Workshop on Software Patterns and
Quality. Information Processing Society of Japan, pp. 25-
30.
Fernandez, E. B., Yoshioka, N., Washizaki, H., 2007, Using
Security Patterns to Build Secure Systems, Proceedings
of the International Workshop on Software Patterns and
Quality. Information Processing Society of Japan, pp. 47-
48.
NSHIELD Artemis project, 2013,
http://www.newshield.eu/.
Cavoukian, A., Dixon, M., 2013, Privacy and Security by
Design: An Enterprise Architecture Approach, retrieved
from http://www.ipc.on.ca.
Cordon, O., Gomide, F., Herrera, F., Hoffmann, F.,
Magdalena, L., 2004, Genetic Fuzzy Systems: New
Developments, Fuzzy Sets and Systems, Vol. 141 (1), pp.
1-3.
Cordon, O., Herrera, F., Gomide, F., Hoffmann, F., Mag-
dalena, L., 2001, Ten Years of Genetic-Fuzzy Systems:
A Current Framework and New Trends, Proceedings
of Joint 9th IFSA World Congress and 20th NAFIPS
International Conference, pp. 1241-1246, Vancouver,
Canada.
Geem, Z. W., Kim, J.-H., Loganathan, G. V., 2001, A New
Heuristic Optimization Algorithm: Harmony Search,
Simulation, Vol 76 (2), pp. 60-68 (2001)
Manjarres, D., Landa-Torres, I., Gil-Lopez, S., Del Ser, J.,
Bilbao, M. N., Salcedo-Sanz, S., Geem Z. W., 2013,
A Survey on Applications of the Harmony Search Al-
gorithm, Engineering Applications of Artificial Intelli-
gence, Vol. 26 (8), pp. 1818-1831.
Van Leekwijck, W., Kerre, E. E., 1999, Defuzzification:
Criteria and Classification, Fuzzy Sets and Systems, Vol.
108 (1999), pp. 159–178.
PECCS2014-InternationalConferenceonPervasiveandEmbeddedComputingandCommunicationSystems
236