A Review of Artificial Immune Systems
Zafer Ataser
Kuzey Kıbrıs Turkcell, Bedreddin Demirel Avenue Salih Mecit Street Lefkosa, TRNC, Mersin 10, Turkey
Keywords:
Artificial Immune Systems, Immune Network Theory, Clonal Selection, Danger Theory, Negative Selection
Algorithm.
Abstract:
Artificial Immune Systems (AIS) are class of computational intelligent methods developed based on the prin-
ciples and processes of the biological immune system. AIS methods are categorized mainly into four types
according to the inspired principles and processes of immune system. These categories are clonal selection,
negative selection, immune network and danger theory. This paper reviews the models of AIS and the progress
of them. The fundamental characteristics of AIS models are identified and some major studies of each model
are given. In addition to that, some application areas of AIS models are explained.
1 INTRODUCTION
Inspiration for computational methods has often come
from analogy with biological phenomena. The bio-
logical immune system is one of them, and researches
inspired by the biological immunity caused the emer-
gence of a new area, Artificial Immune System (AIS),
for computational intelligence.
The biological immune system is still an active
research area for the biology discipline, and as a re-
sult of the researches, many theories have been pro-
posed about the mechanism of the biological immune
system. Hence, various models of AIS were devel-
oped inspired by the different biological immune the-
ories. Garret (Garrett, 2005) present these models as
immune network, clonal selection, negative selection
and danger theory.
The immune system constitutes of two main
mechanisms, innate immunity and adaptive (ac-
quired) immunity. The innate immunity is the first
layer defense mechanism, and the second one is the
adaptive immunity. The innate immunity consists of
basic elements the organism is born with such as skin
and physical barriers. The adaptive immunity pro-
vides the ability to adapt over time to recognize spe-
cific pathogens, and it creates immunological mem-
ory after an initial response to a specific pathogen.
The clonal selection theory was developed to describe
the principles of adaptive immunity. The main fea-
tures of the clonal selection theory are (de Castro and
Zuben, 2002):
• Clone activated mature cells and generates ran-
dom changes on clone cells with high rates (so-
matic mutation);
• Elimination of newly differentiated cells which
match self cells;
• proliferation and differentiation on activation of
cells by antigens.
The immune network theory was developed to ex-
plain the adaptive immune system mechanism, and it
had been introduced by Jerne (Jerne, 1974). The the-
ory states that the immune system maintains an id-
iotypic network of interconnected B cells for antigen
recognition. These cells interact with each other, and
they interconnect with each other in definite rules to
stabilize the network. Two interacting cells are con-
nected if their affinities exceed a certain threshold.
The strength of the connection is directly proportional
to their affinity (Al-Enezi et al., 2010).
The biological negative selection describes T cells
maturation process in thymus called T cell tolerance.
T cell’s gene segments are randomly rearranged to-
gether by somatic gene rearrangement, and bases are
inserted to create T cell against antigens. The gen-
erated cells are eliminated when they recognize self
cells as antigens. In the end of this elimination pro-
cess, the remaining cells, mature cells, are released
from the thymus. In this manner, these mature cells
increase the ability of the immune system to detect
unknown antigens. Inspired by the biological nega-
tive selection, the process of negative selection gener-
ates a set of T-cell detectors that can detect any form
of non-self in AIS (Garrett, 2005).
The main idea of danger theory states that the im-
128
Ataser Z..
A Review of Artificial Immune Systems.
DOI: 10.5220/0004553101280135
In Proceedings of the 5th International Joint Conference on Computational Intelligence (ECTA-2013), pages 128-135
ISBN: 978-989-8565-77-8
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
mune system responds to danger instead of non-self
(Matzinger, 2002). Danger theory fundamentally ac-
cepts the need for discrimination as other theories.
On the other, the difference between danger theory
and others is the answer to what should be responded
to. Danger theory responds danger instead of foreign-
ness, and danger is evaluated by damage to cells spec-
ified by distress signals. These signals are sent out in
case of an unnatural death of cells.
2 MAJOR AIS WORKS
Artificial Immune System (AIS) attracted the atten-
tion of researchers after the first studies, and many re-
searches have been done on it. AIS emerged as a new
branch of Artificial Intelligence (AI) as other disi-
plines inspired from biological mechanisms. There
are many studies on AIS, and some of them focused
on the categorization of the proposed AIS methods
based on their porperties (Garrett, 2005),(Al-Enezi
et al., 2010). This section reviews the studies on ex-
isting models.
2.1 Immune Network Theory
Timmis et al. (Timmis et al., 2000), (Timmis and
Neal, 2001) introduced Artificial Immune Network
(AINE) method which uses artificial recognition ball
(ARB) to represent a number of identical B cells de-
scribed in immune network. The stimulated B cells
are subjected to clone and somatic hypermutation en-
sures that these clones are differentiated a relatively
large proportion of the parent cell. The cells stimu-
lated by particular antigen are kept in the immuno-
logical memory for that antigen. Two B cells are con-
nected based on the affinity between them. The affin-
ity is measured using the Euclidean distance between
the two B cells. The two B cells are connected when
the affinity between them exceeds the network affinity
threshold. The connected B cells are called as ARB.
Timmis et al. (Timmis and Neal, 2001) were tested
AINE using Fisher Iris dataset, and AINE generates
the disconnected clusters which provide the variety
which allows the immune system to generalize varia-
tions in the data encountered.
Castro and Zuben (de Castro and Zuben, 2001)
proposed ”Artificial Immune Network Model for
Data Analysis” (aiNet) learning algorithm. This algo-
rithm uses nodes simulating antibodies, while AINE
uses nodes inspired by B cells. Antibodies are recep-
tor molecules that are secreted B cells with the pri-
mary role recognizing and binding with an antigen.
One of the important aims of this algorithm is to in-
crease the generalization of antibodies and the net-
work. In order to do that, aiNet suppresses antibod-
ies with low antigenic and high affinities according
to the suppression threshold. There are two suppres-
sive steps in this algorithm, clonal suppression and
network suppression. Hence, the suppression thresh-
old controls the specificity level of the antibodies, the
clustering accuracy and network plasticity. aiNet pro-
vides the reconstruction of the metric and topological
relationships. Reproducing the topological relation-
ships causes that similar information are mapped onto
closer antibodies, eventually the same one and clus-
tering of the input space.
Liu and Xu (Liu and Xu, 2008) introduced a co-
operative artificial immune network called CoAIN to
improve search ability and search speed. The CoAIN
uses cooperative strategy inspired by particle swarm
behavior. This means that each network cell has the
ability to cooperate with other individuals, and this
cooperation adjusts position according to its own ex-
perience and the experience of the best cell. In this
way, this cooperation ability finds the best position
encountered by itself and its neighbor. In the other
feature of CoAIN, antibodies with fitness dominate
clonal selection, and smaller step size of mutation is
used in clonal selection to find global optimization.
Step size of mutation is decreased smoothly with the
increase of generation to fit for finer search. This pa-
per shows that some basic immune principles together
with simple cooperation behavior makes possible to
solve complex optimization tasks.
Coelho and Zuben (Coelho and Zuben, 2010) pro-
posed Concentration-based Artificial Immune Net-
work (cob-aiNet) to solve single optimization prob-
lems, and they (Coelho and Zuben, 2011) introduce
the extension of cob-aiNet to solve multi-objective
optimization problems. The cob-aiNet exploits the
features of a concentration-based immune model.
Thus, it controls the dynamics of the population, and
uses new mechanisms to stimulate and maintain the
diversity of the individuals in the population.
Zhong and Zhang (Zhong and Zhang, 2012)
present a novel supervised algorithm based on the im-
mune network theory, the artificial antibody network
( ABNet). In this method, every antibody consists of
two important attributes, its center vector and recog-
nizing radius. The antibody can recognize all antigens
within the range of its recognizing radius. ABNet was
designed for classifications of multi-/hyperspectral re-
mote sensing images.
AReviewofArtificialImmuneSystems
129
2.2 Clonal Selection
Castro and Zuben (de Castro and Zuben, 1999),
(de Castro and Zuben, 2000) popularized the arti-
ficial form of clonal selection developing an algo-
rithm called clonal selection algorithm (CSA). They
applied CSA to different problems and compared it
with the standard genetic algorithm (GA) (de Cas-
tro and Zuben, 2000). Then, they (de Castro and
Zuben, 2002) modified the algorithm and renamed it
to CLONALG, which is the most known clonal selec-
tion algorithm. Two forms of CLONALG were intro-
duced, one for optimization tasks and one for pattern
matching. CLONALG takes into account the follow-
ing main immune features;
• maintenance of a specific memory set;
• selection and cloning of the most stimulated anti-
bodies;
• death of nonstimulated antibodies;
• affinity maturation
• reselection of the clones according to their anti-
genic affinity, generation, and maintenance of di-
versity.
Brownlee (Brownlee, 2005) introduced Clonal Se-
lection Classification Algorithm (CSCA) which is
mainly based on CLONALG. Concern of CSCA in-
creases classification accuracy. CSCA is considered
as a function optimization procedure that maximizes
the number of correctly classified patterns and mini-
mizes the number of misclassified patterns. The algo-
rithm is constituted of four main steps:
1. Initialization - Initialize the antibody population.
2. Training looping - Involve selection and pruning
step, cloning and mutation step and the insertion
the generated clones.
3. Final Pruning - Prepare fitness scores and perform
pruning
4. Classification - Classify using the antibody popu-
lation.
Oliveira et al. (L. O. V. B. Oliveira, 2012) pro-
posed Clonal Selection Classifier with Data Reduc-
tion (CSCDR) which is mainly based on the CSCA.
CSCA is modified to increase the performance and
to decrease the number of memory cells. The muta-
tion process is changed to get better results in search
space process, and a control parameter is inserted to
decrease the number of memory cells produced.
Li et al. (Li et al., 2012) introduced Reconfig-
urable Space Clone Selection Algorithm (RSCSA),
which is focuses on the antibody population size and
antibody search space. RSCSA reduces the search
space and antibody population size. Due to this re-
duction, the algorithm has strong robustness and fast
convergent speed.
2.3 Danger Theory
Aickelin and Cayzer (Aickelin and Cayzer, 2002) pre-
sented the one of the first major studies about the dan-
ger theory from AIS perspective. This study explains
Matzinger’s Danger Theory in the first part. Then,
the danger theory is evaluated from the perspective of
AIS practitioners. In this evaluation, the danger con-
cept is discussed, and the danger theory is compared
with the other AIS models, i.e. negative selection.
Beside this, they debate about how to implement the
danger theory based on the AIS perspective. In the
last section, the AIS applications are evaluated based
on the danger theory.
Greensmith et al. (Greensmith et al., 2004) de-
scribed the danger theory and AIS applications. The
current state of intrusion detection systems (IDS) was
presented. They discussed the application of the dan-
ger theory on IDS, and claimed that significant im-
provements will be provided. Kim et al.(Kim et al.,
2005) and Roper (Roper, 2009) also proposed that the
danger theory is more suitable than other AIS models
to apply on IDS.
Aickelin and Greensmith (Aickelin and Green-
smith, 2007) introduced two algorithms, the Dendritic
Cell Algorithm (DCA) and the Toll-like Receptor al-
gorithm (TLR), developed based on the danger the-
ory. These algorithms were developed inspired by
different aspects of the danger theory. DCA and TLR
proved that it is possible to build feasible AIS algo-
rithms based on the principles of the danger theory.
Zhu and Tan (Zhu and Tan, 2011) proposed a dan-
ger theory based learning (DTL) model, which mimic
the mechanism of the danger theory. The algorithm
was tested with spam filtering problem and compared
with classical machine learning approaches, Support
Vector Machine (SVM) and Naive Bayes (NB). In ex-
periments, the DTL model outperformed SVM, NB.
2.4 Negative Selection Algorithm (NSA)
Artificial Immune System (AIS) covers many models
inspired by the biological immune system. The first
model, negative selection algorithm (NSA), among
AIS models was introduced by Forrest et al. (Forrest
et al., 1994). Many researches have been performed
after the introduction of NSA. These researches pro-
posed various NSA, and they are differentiated in data
representation, detector representation, self definition
and matching rule.
IJCCI2013-InternationalJointConferenceonComputationalIntelligence
130
Figure 1: Generation of valid detector set (Forrest et al.,
1994).
Forrest et al. (Forrest et al., 1994) proposed NSA
inspired by discrimination between self and non-self
in immune system. Therefore, NSA imitates the T
cell maturation process, which gives the ability to T
cells to make discrimination between self and non-
self. To implement NSA, there are some critical de-
velopment considerations; data representation, self
definition, detector coverage, matching rule. This
study converts the given data to binary representation,
so detectors are also represented in binary form. Self
was defined as the string to be protected, and other
(non-self) to be any other string. The self string is log-
ically split into equal-size segments to generate valid
detectors. This produces the collection S of self sub-
strings. In the second step, detectors are randomly
generated, and generated strings match strings in S
are eliminated. Strings that do not match any string
in S are added to the detector set. Two strings match,
if they match at least r contiguous locations. Figure 1
shows all these the detector generation phase.
Freitas and Timmis (Freitas and Timmis, 2007)
discussed the application of NSA for data mining, so
features of NSA were explored. The basic framework
of the negative selection process to generate detectors
were given in Algorithm 1.
Hofmeyr and Forrest (Hofmeyr and Forrest,
1999),(Hofmeyr and Forrest, 2000) proposed the ar-
tificial immune system (ARTIS) method, and they ap-
plied it to intrusion detection. This method repre-
sents detectors as bit strings, and uses the r contigu-
ous matching rule. Beside these, the study defines the
lifecycle of a detector, so it provides dynamic detector
populations and adaptation ability in a continuously
changing environment. In this lifecycle, a detector
can be in one of the five states: immature, mature, ac-
tivated, memory or death. Figure 2 presents the lifecy-
cle of a detector. Randomly generated detector is con-
sidered as immature detectors, and if it does not match
Algorithm 1: Pseudocode of the Negative Selec-
tion Process to generate detectors (Freitas and Tim-
mis, 2007).
Data: a set of normal (self) data instances (S)
Result: a set of mature detectors that do not
match any instance in S
repeat
Randomly generate an immature detector
Measure the affinity (similarity) between
this detector and each instance in S)
if the affinity between the detector and at
least one instance in S is greater than a
user-defined threshold then
discard this detector
else
output this detector as a mature immune
detector
until stopping criterion;
Figure 2: The lifecycle of a detector (Hofmeyr and Forrest,
2000).
self data during the tolerization period, it becomes a
mature detector. A mature detector becomes an acti-
vated, when it exceeds the activation threshold (match
threshold). After that, a human security officer’s con-
firmation (costimulation) is needed for an activated
detector to make it a memory detector. Immature,
mature and activated detectors can die, but memory
detectors go to activated state, when they match non-
self. An immature detector dies, if it matches self. A
mature detector death is occurred, when it does not
exceed activation threshold during lifetime (life ex-
pectancy). An activated detector dies, if it does not
receive confirmation in a time period (costimulation
delay).
Gonzalez et al. (Gonzalez et al., 2003a) pre-
sented the effects of the low-level representation and
its matching rules on the performance of NSA in
covering the non-self space. They explored and
compared the different binary matching rules: r-
contiguous matching, r-chunk matching, Hamming
AReviewofArtificialImmuneSystems
131
distance matching, and Rogers and Tanimoto match-
ing. This study indicates that the matching rule for
NSA needs to be chosen when it represents data ac-
curately in problem space.
Gonzalez et al. (Gonzalez et al., 2002) pro-
posed a Real-Valued Negative Selection (RNS) algo-
rithm. RNS algorithm uses real numbers to represent
self/non-self space RNS algorithm and binary NSA
were compared for anomaly detection problem. Then,
advantages and disadvantages of the real-valued rep-
resentation were presented based on the binary repre-
sentation. Real-valued representation advantages are:
closer to original problem space, allowing the use of
methods from computational geometry to speed-up
the algorithms, facilitating the use of other machine
learning methods to find useful high level knowledge
i.e. (Gonzalez and Dasgupta, 2003). Disadvantages
of real-valued representation are: making analysis of
the problem space harder, not suitable for the repre-
sentation of categorical attributes.
Dasgupta and Gonzalez (Dasgupta and Gonzalez,
2002) explored positive selection and negative selec-
tion, and they were compared using real-valued rep-
resentation. Detectors are represented as rectangle
with real numbers. Based on this comparison, advan-
tages and disadvantages of these approaches were de-
scribed. This comparison showed that positive selec-
tion is more precise, but it needs more time and space
resources. The negative selection is less precise, but
it needs fewer time and space resources.
Real-valued representation is used in many appli-
cations due to the nature of applications’ domains,
i.e. intrusion detection from network traffic. The non-
self coverage gets difficult for the problems with nat-
ural real-valued representation. This is because, the
real-valued space is continuous and the boundary of
self and non-self is ambiguous in this space. There-
fore, the non-self coverage is a major issue for real-
valued NSA (RNSA) (Gonzalez et al., 2002), (Gon-
zalez et al., 2003b), (Ji and Dasgupta, 2009), (Zeng
et al., 2009), (Balachandran et al., 2007), (X. Yuel
and Wangl, 2010). Detector representation and self
definition are the determinant for the non-self cover-
age. A part of research have been focused on the de-
tector representation and distribution in the non-self
space in order to maximize the coverage(Gonzalez
et al., 2003b), (Balachandran et al., 2007),(Ji and Das-
gupta, 2009). On the other hand, the recent research
is focused on adaptive-self that implicates the variable
self radius (Bezerra et al., 2005), (Zeng et al., 2009),
(X. Yuel and Wangl, 2010) . The self radius is an im-
portant value to control the detection rate and false
alarm rate.
In real-valued NSAs, the detectors are usually rep-
resented as circles or rectangles for two dimensional
problems. Nevertheless, some NSAs use mixture of
specific geometrical shapes to represent the detec-
tors. However, some of NSAs generate the detectors
with different sizes. Based on the data and detec-
tor representation, the matching rule is changed, and
Euclidean distance matching is usually used in real-
valued representation.
Dasgupta and Gonzalez (Dasgupta and Gonza-
lez, 2002), (Gonzalez and Dasgupta, 2002) represent
the detectors generated by genetic algorithm as rules.
They present the general form of detector rules (de-
tectors) as follows:
R
j
: If Cond
i
then nonself, j = 1,...,x
Cond
i
= x
1
∈ [low
i
1
, high
i
1
] and ... and x
n
∈ [low
i
n
,
high
i
n
]
where (x
1
,..., x
n
) is a feature vector, and
[low
i
1
,high
i
1
] specifies the lower and upper values in
the condition part. In this definition, m is the num-
ber of detector rules, and n is the number of fea-
ture dimensions. The detectors correspond to hyper-
rectangles in a multidimensional space. In this study,
self region is determined by the level of variability (v)
parameter, which is interpreted as the radius of self
samples.
Gonzalez et al. (Gonzalez et al., 2003b) proposed
a Randomized Real-Valued Negative Selection Algo-
rithm (RRNS). This algorithm takes the detector ra-
dius and the self variability threshold (self sample ra-
dius) as parameters, so each self sample and detector
is represented as circles in two-dimensional problem
space. These circles have a fixed size specified by
the relevant parameter. Based on the self radius pa-
rameter, the algorithm uses Monte Carlo method to
estimate the volume of self region.
Balachandran et al (Balachandran et al., 2007)
present a work focused on developing a framework
for generating multi-shaped detectors in real-valued
NSA. This new extended real-valued NSA uses mul-
tiple shape (sphere, rectangle or ellipse) detectors for
covering two dimensional non-self space. In this
NSA, self space is also specified by the constant self
radius parameter.
Ji and Dasgupta (Ji and Dasgupta, 2009),(Ji and
Dasgupta, 2005),(Ji and Dasgupta, 2004) proposed a
new real-valued NSA, which generates variable size
detectors. In this NSA, the detectors are represented
as circles in two dimensional space and the radii of
these circles are variable. On the other hand, the ra-
dius for all self samples is taken as the constant pa-
rameter and used to check whether a new generated
detector is in any self circle or not. If it is, then dis-
carded, otherwise the distance between the center of
IJCCI2013-InternationalJointConferenceonComputationalIntelligence
132
detector and the nearest self sample is assigned to
this detector radius. This is called boundary-aware
method (Ji and Dasgupta, 2005).
In the work by Bezerra et al (Bezerra et al., 2005),
an adaptive radius immune algorithm (ARIA) was de-
veloped. This is one of the first researches on variable
self radius for each self sample. Although, ARIA is
closer to clonal selection algorithm, this adjusted self
radius has crucial effect in AIS. ARIA considers the
density information to form its representation. ARIA
takes an initial value of self radius and based on the
local density of samples, this initial value is adjusted
for each sample.
Zeng et al (Zeng et al., 2009) introduce a
self-adaptive negative selection algorithm (ANSA).
ANSA can adapt the varieties of self/nonself space
by adjusting self radii and detectors’ radii. Yuel et al
(X. Yuel and Wangl, 2010) worked on optimization of
self set for real-valued NSA. In order to do that, self
samples are processed in three steps. In the first step,
wrong samples are discarded according to ”3 σ” cri-
terion. In the next step, the self radius is adjusted by
the self’s probability density. In the last step, unnec-
essary self samples, whose covered region is already
overlapped by others, are discarded.
The major characteristics of a negative selection
algorithm can be identified as follows:
1. Negative representation (Ji and Dasgupta, 2007):
NSA identifies and represents the complementary
space of the given samples in training phase. Neg-
ative representation and positive representation al-
gorithms have been compared in many researches
to extract the strength and applicability of neg-
ative representation (Dasgupta and Nino, 2000),
(Dasgupta and Gonzalez, 2002), (Stibor et al.,
2005a), (Stibor et al., 2005b).
2. Usage of detector set as the classification mecha-
nism (Ji and Dasgupta, 2007): Detector set usage
provides the opportunity to NSA to distribute its
processes, i.e. detectors generation.
3. One-class classification (Ji and Dasgupta,
2007)(Freitas and Timmis, 2007): NSA was
developed inspired by the self/non-self discrimi-
nation mechanism of the biological mechanism.
Therefore, NSA is trained with samples from
the one class (self) and then classifies the given
instance into one of two classes (self/non-self).
Although some researches tried to extend NSA
for multiclass classification problems(Dasgupta
and Gonzalez, 2002; Gonzalez and Dasgupta,
2002), large majority of the researches have been
applied to one-class classification problems.
4. Adaptation capability(Hofmeyr and Forrest,
1999), (Hofmeyr and Forrest, 2000) (Chen et al.,
2005): There are many researches to develop
adaptive NSAs inspired by the adaptation ability
of biological immune system. These adaptive
NSAs use some mechanisms, i.e. memory, and
processes, to obtain dynamic change of the
detectors population.
Particularly, NSA was developed for intrusion de-
tection research (Forrest et al., 1994). Negative se-
lection algorithm (NSA) can be used in many do-
mains today, but the most natural application domain
of NSA is intrusion detection (Dasgupta and Gonza-
lez, 2002), (Powers and He, 2006), (Kim and Bentley,
2001), (Hofmeyr and Forrest, 1999).
2.5 Applications
AIS have many application areas today after first pro-
posal. Hart and Timmis (Hart and Timmis, 2008) sur-
veyed AIS studies and classified application areas of
AIS into 12 headings. These categories are presented
in table, and according to the number of researches on
these categories, categories were divided into major
and minor. Based on these categorizations of applica-
tion areas, Hart and Timmis summarized application
areas of AIS as (1) Learning (2) Anomaly Detection
and (3) Optimisation. Application areas were mapped
to these groups: Learning contains clustering, classi-
fication and pattern recognition, robotic and control
applications; Anomaly Detection includes fault de-
tection and computer and network security applica-
tions; Optimisation consists of real-world problems
which essentially include combinatoric and also nu-
meric function optimisation.
Table 1: Application Areas of AIS (Hart and Timmis,
2008).
Major Minor
Clustering/Classification Bio-informatics
Anomaly Detection Image Processing
Computer Security Control
Numeric Function Optimisation Robotics
Combinatoric Optimisation Virus Detection
Learning Web Mining
Garret (Garrett, 2005) surveyed AIS models and
gives the application areas for each of them. NSA ap-
plication areas are change detection, fault detection
and diagnosis, network intrusion detection; Clonal
selection application areas are pattern recognition,
automated scheduling, document classification, uni-
modal, combinatorial and multi-modal optimization;
Immune network application areas are detecting gene
promoter sequences, diagnosis data mining and clus-
AReviewofArtificialImmuneSystems
133
ter analysis; Danger theory application area is intru-
sion detection.
Freitas and Timmis (Freitas and Timmis, 2007)
discussed the application of AIS for data mining.
They evaluated all AIS models and found limitations
in existing AIS for data mining. Limitations they dis-
covered and suggestions for future researches were
mentioned in order to mitigate corresponding limita-
tion.
3 CONCLUSIONS
In machine learning, there are many learning meth-
ods that are inspired by the biological mechanisms.
Genetic algorithm and neural network are the well-
known biologically inspired computational models.
Genetic algorithm mimics the principles and pro-
cesses of natural evaluation. On other hand, neu-
ral network is inspired by the network or circuit of
biological neurons and mimics the properties of bi-
ological neurons. Biological immune system is the
other biological system that has various computa-
tional mechanisms: pattern recognition, memory, dis-
tributed processing, self organizing, etc. Inspired by
the principles and processes of the biological immune
system, many computational intelligent models were
developed, and this type of models is called Artificial
Immune Systems (AIS). These AIS models are cat-
egorized mainly into immune network model, clonal
selection, negative selection and danger theory.
REFERENCES
Aickelin, U. and Cayzer, S. (2002). The danger theory and
its application to artificial immune systems. In In Pro-
ceedings of the 1st International Conference on Arti-
ficial Immune Systems. Springer-Verlag.
Aickelin, U. and Greensmith, J. (2007). Sensing danger:
Innate immunology for intrusion detection. In Infor-
mation Security Technical Report. Elsevier.
Al-Enezi, Abbod, J., and Alsharhan, M. (2010). Artificial
immune systems - models, algorithms and application.
In International Journal of Research and Reviews in
Applied Sciences. ARPA Press.
Balachandran, S., Dasgupta, D., Nino, F., and Garrett, D.
(2007). A framework for evolving multi-shaped de-
tectors in negative selection. In In Proceedings of
the 2007 IEEE symposium on foundations of compu-
tational intelligence. IEEE Xplore.
Bezerra, G. B., Barra, T. V., de Castro, L. N., and Zuben,
F. J. V. (2005). Adaptive radius immune algorithm
for data clustering. In Artificial Immune Systems: 4th
International Conference. Springer-Verlag.
Brownlee, J. (2005). Clonal selection theory and clonalg -
the clonal selection classification algorithm (csca). In
Centre for Intelligent Systems and Complex Processes
(CISCP), Tech. Rep. 2-02. Faculty of Information and
Communication Technologies (ICT), Swinburne Uni-
versity of Technology.
Chen, J., Liang, F., and Yang, D. (2005). Dynamic nega-
tive selection algorithm based on match range model.
In Proceedings of the 18th Australian Joint confer-
ence on Advances in Artificial Intelligence. Springer-
Verlag.
Coelho, G. P. and Zuben, F. J. V. (2010). A concentration-
based artificial immune network for continuous opti-
mization. In IEEE Congress on Evolutionary Compu-
tation. IEEE.
Coelho, G. P. and Zuben, F. J. V. (2011). A concentration-
based artificial immune network for multi-objective
optimization. In International Conference on Evo-
lutionary Multi-Criterion Optimization. Springer-
Verlag.
Dasgupta, D. and Gonzalez, F. (2002). An immunity-based
technique to characterize intrusions in computer net-
works. In IEEE Transactions on Evolutionary Com-
putation. IEEE Press.
Dasgupta, D. and Nino, F. (2000). A comparison of nega-
tive and positive selection algorithms in novel pattern
detection. In IEEE International Conference on Sys-
tems, Man, and Cybernetics. IEEE Xplore.
de Castro, L. N. and Zuben, F. J. V. (1999). Artificial im-
mune systems: part i - basic theory and applications.
In Technical Report DCA-RT 01/99. School of Com-
puting and Electrical Engineering, State University of
Campinas.
de Castro, L. N. and Zuben, F. J. V. (2000). The clonal
selection algorithm with engineering applications. In
Proceedings of the Genetic and Evolutionary Compu-
tation Conference. Morgan Kaufmann.
de Castro, L. N. and Zuben, F. J. V. (2001). ainet: An ar-
tificial immune network for data analysis. In In Data
Mining: A Heuristic Approach. Idea Group.
de Castro, L. N. and Zuben, F. J. V. (2002). Learning
and optimization using the clonal selection principle.
In IEEE Transactions on Evolutionary Computation.
IEEE Press.
Forrest, S., Perelson, A., Allen, L., and Cherukuri, R.
(1994). Self-nonself discrimination in a computer. In
In Proceedings of the 1994 IEEE Symposium on Re-
search in Security and Privacy. IEEE Computer Soci-
ety.
Freitas, A. A. and Timmis, J. (2007). Revisiting the foun-
dations of artificial immune systems for data mining.
In IEEE Transactions on Evolutionary Computation.
IEEE Press.
Garrett, S. M. (2005). How do we evaluate artificial immune
systems? In Evolutionary Computation. MIT Press.
Gonzalez, F. and Dasgupta, D. (2002). An immunogenetic
technique to detect anomalies in network traffic. In In
Proceedings of the genetic and evolutionary compua-
tion conference. Morgan Kaufmann.
IJCCI2013-InternationalJointConferenceonComputationalIntelligence
134
Gonzalez, F. and Dasgupta, D. (2003). Anomaly detec-
tion using real-valued negative selection. In Genetic
Programming and Evolvable Machines. Kluwer Aca-
demic.
Gonzalez, F., Dasgupta, D., and Gomez, J. (2003a). The
effect of binary matching rules in negative selection.
In In Proceedings of the Genetic and Evolutionary
Computation Conference (GECCO 2003). Springer-
Verlag.
Gonzalez, F., Dasgupta, D., and Kozma, R. (2002). Com-
bining negative selection and classification techniques
for anomaly detection. In Proceedings of the 2002
Congress on Evolutionary Computation. IEEE Com-
puter Society.
Gonzalez, F., Dasgupta, D., and Nino, L. F. (2003b). A
randomized real-valued negative selection algorithm.
In In Proceedings of the 2nd International Conference
on Artificial Immune Systems. Springer-Verlag.
Greensmith, J., Aickelin, U., and Twycross, J. (2004). De-
tecting danger: Applying a novel immunological con-
cept to intrusion detection systems. In Proceedings of
the 6th International Conference in Adaptive Comput-
ing in Design and Manufacture. Springer-Verlag.
Hart, E. and Timmis, J. (2008). Application areas of ais:
The past, the present and the future. In Applied Soft
Computing. Elsevier Science.
Hofmeyr, S. A. and Forrest, S. (1999). Immunity by de-
sign: An artificial immune system. In In Proceedings
of the Genetic and Evolutionary Computation Confer-
ence. Morgan Kaufmann.
Hofmeyr, S. A. and Forrest, S. (2000). Architecture for an
artificial immune system. In Evolutionary Computa-
tion Journal. MIT Press.
Jerne, N. (1974). Towards a network theory of the immune
system. In Annals of Immunology. Inst. Pasteur.
Ji, Z. and Dasgupta, D. (2004). Real-valued negative se-
lection algorithm with variable-sized detectors. In In
proceeding of Genetic and Evolutionary Computation.
Springer-Verlag.
Ji, Z. and Dasgupta, D. (2005). A boundary-aware negative
selection algorithm. In In Proceedings of the inter-
national conference on artificial intelligence and soft
computing. ACRA Press.
Ji, Z. and Dasgupta, D. (2007). Revisiting negative selection
algorithms. In MIT Evolutionary Computation. MIT
Press.
Ji, Z. and Dasgupta, D. (2009). V-detector: An efficient
negative selection algorithm with ’probably adequate’
detector coverage. In Information Sciences. Elsevier
Science.
Kim, J. and Bentley, P. J. (2001). An evaluation of negative
selection in an artificial immune system for network
intrusion detection. In Proceedings of the Genetic
and Evolutionary Computation Conference. Morgan
Kaufmann.
Kim, J., Greensmith, J., Twycross, J., and Aickelin, U.
(2005). Malicious code execution detection and re-
sponse immune system inpired by the danger theory.
In Proceedings of the Adaptive and Resilient Comput-
ing Security Workshop (ARCS-05).
L. O. V. B. Oliveira, R. L. M. Motay, D. A. C. B. (2012).
Clonal selection classifier with data reduction: Clas-
sification as an optimization task. In IEEE World
Congress on Computational Intelligence. IEEE.
Li, J., Gao, H., and Wang, S. (2012). A novel clone selec-
tion algorithm with reconfigurable search space ability
and its application. In Fourth International Confer-
ence on Natural Computation. IEEE Computer Soci-
ety.
Liu, L. and Xu, W. (2008). A cooperative artificial immune
network with particle swarm behavior for multimodal
function optimization. In IEEE Congress on Evolu-
tionary Computation. IEEE Press.
Matzinger, P. (2002). The danger model: A renewed sense
of self. In Science.
Powers, S. T. and He, J. (2006). Evolving discrete-valued
anomaly detectors for a network intrusion detection
system using negative selection. In In the 6th Annual
Workshop on Computational Intelligence (UKCI ’06).
University of Leeds.
Roper, M. (2009). Artifical immune systems, danger the-
ory, and the oracle problem. In Testing: Academic and
Industrial Conference - Practice and Research Tech-
niques. IEEE Computer Society.
Stibor, T., Mohr, P., Timmis, J., and Eckert, C. (2005a). Is
negative selection appropriate for anomaly detection?
In Proceedings of the 2005 conference on Genetic and
evolutionary computation. ACM.
Stibor, T., Timmis, J., and Eckert, C. (2005b). A compara-
tive study of real-valued negative selection to statisti-
cal anomaly detection techniques. In Proceedings of
the 4th international conference on Artificial Immune
Systems. Springer-Verlag.
Timmis, J. and Neal, M. (2001). A resource limited artifi-
cial immune system for data analysis. In Knowledge
Based Systems. Elsevier.
Timmis, J., Neal, M., and Hunt, J. (2000). An artificial
immune system for data analysis. In Biosystems. El-
sevier.
X. Yuel, F. Zhang, L. X. and Wangl, D. (2010). Opti-
mization of self set and detector generation base on
real-value negative selection algorithm. In 2010 In-
ternational Conference on Computer and Communi-
cation Technologies in Agriculture Engineering. IEEE
Xplore.
Zeng, J., Liu, X., Li, T., Liu, C., Peng, L., and Sun, F.
(2009). A self-adaptive negative selection algorithm
used for anomaly detection. In Progress in Natural
Science. Elsevier.
Zhong, Y. and Zhang, L. (2012). An adaptive artifi-
cial immune network for supervised classification
of multi-/hyperspectral remote sensing imagery. In
IEEE Transactions on Geoscience and Remote Sens-
ing. IEEE.
Zhu, Y. and Tan, Y. (2011). A danger theory inspired learn-
ing model and its application to spam detection. In
Proceedings of the second international conference on
advances in swarm intelligence. Springer-Verlag.
AReviewofArtificialImmuneSystems
135
