Shallow and Deep Networks Intrusion Detection

System: A Taxonomy and Survey

Elike Hodo, Xavier Bellekens, Andrew Hamilton, Christos Tachtatzis and Robert
Atkinson
Department of Electronic & Electrical Engineering
University of Strathclyde
Division of Computing and Mathematics
University of Abertay Dundee

E-mail: [email protected]
E-mail: [email protected]

Abstract— Intrusion detection has attracted a considerable interest from researchers

and industries. The community, after many years of research, still faces the problem of
building reliable and efficient IDS that are capable of handling large quantities of data, with
changing patterns in real time situations. The work presented in this manuscript classifies
intrusion detection systems (IDS). Moreover, a taxonomy and survey of shallow and deep
networks intrusion detection systems is presented based on previous and current works. This
taxonomy and survey reviews machine learning techniques and their performance in
detecting anomalies. Feature selection which influences the effectiveness of machine learning
(ML) IDS is discussed to explain the role of feature selection in the classification and training
phase of ML IDS. Finally, a discussion of the false and true positive alarm rates is presented
to help researchers model reliable and efficient machine learning based intrusion detection
systems.

Keywords— Shallow network, Deep networks, Intrusion detection, False positive

alarm rates and True positive alarm rates

1.0 INTRODUCTION
Computer networks have developed rapidly over the years contributing significantly to social
and economic development. International trade, healthcare systems and military capabilities
are examples of human activity that increasingly rely on networks. This has led to an
increasing interest in the security of networks by industry and researchers. The importance of
Intrusion Detection Systems (IDS) is critical as networks can become vulnerable to attacks
from both internal and external intruders [1], [2].

An IDS is a detection system put in place to monitor computer networks. These have been in
use since the 1980’s [3]. By analysing patterns of captured data from a network, IDS help to
detect threats [4]. These threats can be devastating, for example, Denial of service (DoS)
denies or prevents legitimate users resource on a network by introducing unwanted traffic [5].
Malware is another example, where attackers use malicious software to disrupt systems [6].

1
Intrusion detection systems evolved as a response to these situations. Many IDS have been
developed but have experienced the problem of false positive or false negative alarms. These
are the false recognition of an attack by the IDS and increases the difficultly for network
administrators to handle intrusion reports. Researchers in this field aim at developing IDS
that have a high accuracy of detection and low false alarm rate [7]. Another problem of some
existing IDS is their inability to detect unknown attack types. These IDS rely on the
signatures of known attacks.

Human independent IDS that incorporate machine learning techniques have been developed
as a solution to these problems. Machine learning IDS learns from normal traffic and
abnormal traffic by training on a dataset to predict an attack by using classification. Several
machine learning techniques have been successfully implemented as classifiers on IDS but
present numerous flaws such as low throughput and high false detection rates [7].

The taxonomy (or classification) presented within this manuscript is highly relevant in the
present study due to the highly diverse types of systems and attacks [8], [9]. The taxonomy
will aid constructing two objectives: a clear description of the current state of IDS and the
guide lines in which to explore the complexity of it [10], [11].

The paper aims to provide a clear description and guidelines needed to understand intrusion
detection systems based on results from existing works in various papers. Its organisation
gives a broader view of IDS and aims at addressing the concerns and solutions of shallow and
deep learning IDS. The paper also compares previous works and their performance metrics.

The second objective of the paper is to present a survey and the classification of Intrusion
Detection Systems, taxonomy of Machine Learning IDS and a survey on shallow and deep
networks IDS.

2.0 INTRUSION DETECTION SYSTEMS

Intrusion detection systems are strategically placed on a network to detect threats and monitor
packets. The IDS accomplishes this by collecting data from different systems and network
sources and analysing the data for possible threats [12]. The functions of the IDS include
offering information on threats, taking corrective steps when it detects threats and recording
all important events within a network [13]. Figure 2 shows a model of Intrusion detection
system

Figure 2 Intrusion detection system model

2
2.1 Classification of Intrusion detection system

Figure 2.1 Classification of Intrusion Detection Systems

Different researchers have developed different classification representations. H.Debar et

al[14], S.Axelsson et al.[15], S.Amer et al.[16], C.Xenakis et al.[17], Hung-Ten Lia et
al.[18], have previously presented intrusion detection surveys and taxonomies. This research
builds upon their work and introduces deep networks technique which all these referenced
works do not describe. With the increasing value of big data, deep networks is an important
element to capture in an IDS taxonomy The taxonomy presented within this work provides a
fine-grained overview of the different machine learning techniques for intrusion detection
systems as shown in Figure 2.1. The detection mainly depends on the source of data and
intrusion technique used. The Source of data is the nodes that gather the information for
analysis. The source can be either host or network based or both. These two sources are
further discussed in section 2.1.1. The detection technique also differentiates between the
anomaly based and signature based detection techniques. The different techniques that fall in
the class of anomaly and signature based are important to develop a taxonomy that gives a
broader view of techniques and the capabilities in detecting attacks. Section 2.1.2 and 2.1.3
explains anomaly based and signature based techniques in detail. These techniques operate
either as self-learning or programmed. In Self-learning class, the system automatically learns
whilst in programming class a user teaches the system. The classification presented Fig.2.1
will help draw conclusions on detection accuracy and false alarm rates of IDS techniques.

2.1.1 Host Based IDS (HIDS) Vs. Network Based IDS (NIDS)
Host based IDS were the first type of IDS to be implemented [19]. HIDS are software based
products installed on a host computer that analyse and monitor all traffic activities on the
system application files and operation system [20], [21]. The traffic activities gathered by the
system application files and system [20] are called the audit trails [22]–[24]. HIDS has an

3
advantage of being able to detect threats from within by scanning through the traffic activities
before sending and receiving data [25]. Its main disadvantage is that only monitors the host
computer, meaning it has to be installed on each host [26].

Network based IDS are found at specific points on the network to capture and analyse the
stream of packets going through network links, unlike the HIDS whose approach is to analyse
each host separately [27] [28], [29]. It has the advantage of having a single system
monitoring an entire network, saving time and cost of installing software on each host. NIDS
main disadvantage is its vulnerability to any intrusion originating from the network targeting
a system within the network.

Table I. Performance comparison of HIDS and NIDS [30]

Performance in terms Host-Based IDS Network-Based IDS

of:
Intruder deterrence Strong deterrence for Strong deterrence for outside intruders
inside intruders
Threat response Weak real time response Strong response time against outside intruders
time but performs better for a
long term attack
Assessing damage Excellent in determining Very weak in determining extent of damage
extent of damage
Intruder prevention Good at preventing inside Good at preventing outside intruders
intruders
Threat anticipation Good at trending and Good at trending and detecting suspicious behavior
detecting suspicious patterns
behavior patterns

2.1.2 Anomaly Based Detection

Anomaly Based Detection is a behavioural based intrusion detection system. It observes
changes in normal activity within a system by building a profile of the system which is being
monitored [31], [32]. The profile is generated over a period of time when the system is
established to have behaved normally [33]. One advantage is that it offers the ability to detect
attacks which are new to the system [34]. Anomaly detection is categorised into two, based
on the way a normal profile of a system is specified.

1) Self-learning – The self-learning system operate by example with a baseline set for
normal operation. This is achieved by building a model for the underlying processes
with the observed system traffic built up over a period of time [15]. Self-learning
systems are sub-divided into the following main categories: time series model and
machine learning. Section 3 discusses machine learning technique in detail.
 Time series model takes into account the sequence of observation in order of
succession occurring in intervals of uniformity. If the probability of
occurrence of a new observation at a time is negligible then it is considered a
change in normal behaviour. Time series has an advantage of observing
trends of behaver over a period of time and flagging it, if it notices a change
in normal behaviour. It is an effective model when attacks are sequential over
a period of time [35]. This model has a disadvantage of being more costly

4
 computationally [36]. Auto regressive moving average (ARMA) is an
example of time series model used as IDS.
In Predictive Modelling for Intrusions in Communication Systems using
generalised autoregressive moving average (GARMA) and ARMA models
by [37] et al. fitted ARMA(1,1) and GARMA(1,2;𝛿,1) time series models to
4 types of attacks (DoS, probe, U2R and L2R). The parameter estimation was
done using Hannan-Rissanen algorithm, Whittle estimation and Maximum
likelihood estimation and the point forecast obtained through Whittle
estimation and maximum likelihood were close to the original value. The
time series models were able to forecast the attack but the performance of
GARMA (1, 2; 𝛿, 1) in attack detection was better.

2) Programmed – A programmed model is when a system needs either a user or an

external person to teach the system to detect changes in behaviour. The user decides
the extent of abnormal behaviour in the system and flags an intrusion threat [35]. The
programmed models are grouped into four categories: threshold, simple rule based,
and statistical models.
 Threshold models can be considered as the simplest programmed descriptive
statistical detector [15]. By correlating and analysing statistical data, a user
can program the system at a pre-defined threshold of alarm on a statistical
variable. Careful selection of the threshold is required to minimise the false
alarm rate. Instances where the threshold set too high leads to a risk of missing
an alarm when an intruder is performing a malicious action [38]. A typical
example is setting an alarm if three attempts to login to a system are
unsuccessful [39].
In anomalous payload-based network intrusion detection [40], Ke Wang et al.
presented a payload base anomaly detection by demonstrating the
effectiveness on 1999 DARPA dataset and a live dataset collected form the US
Columbia CS department network. In training phase, a profile byte frequency
distribution was computed and the standard deviation of application payload
connected to a single host and port. During the detection phase the
Mahalanobis distance was used to calculate the similarity of new data against
the pre-computed profile. The detector compared this measure against a
threshold and generated an alert when the distance of the new input exceeded
this threshold. It recorded nearly 100% in accuracy with 0.1% false positive
rate for port 80 traffic.
 Simple rule based is an approach for detection is to monitor events within a
system that trigger a rule either than what is considered as a normal behaviour
for the user. One limitation of this model is its failure to detect a threat that is
not programmed as a rule in the system [41]. RIPPER (repeated incremental
pruning to produce error reduction) is an example of rule based model that
builds a set of rules to detect normal behaviour and abnormal behaviour.
R.Naidu et al[42] used KDDCup ’99 dataset to compare the performance of
RIPPER rule, decision tree(C5) and support vector machine(SVM). The

5
 dataset was categorised into 3 (NORMAL, probe attacks and DoS attacks).
The RIPPER rule algorithm went through two stages during the experiment.
The first stage initialised the rule conditions and the second stage the rule
optimisation. The algorithm obtained conditions in each rule to classify the
testing data. The RIPPER rule recorded a total detection rate of 98.69%, C5
98.75% and SVM 98.63%.
 Statistics model collects data in a profile. Analysing the profile of normal
statistical behaviour gives a description shown by the patterns from data to
help make conclusions if an activity is normal or abnormal. The system then
develops a distance vector for the observed traffic and the profile. An alarm is
raised by the system when the distance is great enough [15],[14], [36], [43].
These models are sub categorised into four: mean/standard deviation,
multivariate, Markov process and operational model. Dorothy Denning [44]
discusses model based on the hypothesis that security violation can be
detected by monitoring a system’s audit records for changes in pattern. The
model includes profiles for representing the behaviour of subjects with respect
to objects in terms of metrics and statistical models and rules for acquiring
knowledge about this behaviour from audit records and detecting anomalous
behaviour.
a) Mean, standard deviation and any other form of correlations are known as
moments in statistics [35], [45]. A moment is said to be anomalous when
events fall either above or below a set interval. Decisions are made taking
system change into account by altering statistical rule set for the system
[35]. Its advantage over the operational model is its ability to detect attacks
without having prior knowledge of the normal activities to set its limits.
Rather it learns form observations to determine its normal activities [35]. It
is a complex model that has more flexibility than the threshold model. It
does not require prior knowledge to determine the normal behaviour to
help set pre-defined threshold. Varying the mean and standard deviations
slightly changes the computation by putting extra weights on the more
recent values [35]. A.Ashfag et al[46] proposed a standard deviation
normalised entropy of accuracy hybrid method of intrusion detection. Two
real traffic dataset were used. The endpoint dataset comprising 14months
of traffic traces on a diverse set of 13 endpoints. The dataset was reduced
to 6weeks of traffic for testing and training purposes. The endpoints were
infected with different malware attacks. The second was attack data
dataset was obtained from two international network locations at Lawrence
Berkeley National laboratory, USA on three distinct days. Using 9
prominent classifiers on the two datasets showed 3%-10% increase in
detection rate and 40% decrease in false alarm rate over the existing
classifiers can be achieved with the proposed hybrid technique.
b) Multivariate models are similar to the mean and standard deviation model
[35], [47]. The multivariate models are based on correlations between two
or more metrics. They use multiple variables to predict possible outcomes.
For example the number of CPU cycles can be compared to how long a

6
 login session is completed [47]. Theoretically this model could have a fine
distinction over one variable [47]. In an approach by W.Sha et al.[48],
multivariate time series and high-order Markov chain were taken into
account along the detailed design of training and testing algorithms. The
models were evaluated using DARPA dataset. Observing the multi order
Markov chain showed that the relative positions between results from
models of different orders provide a new effective indication for
anomalies. To improve sensitivity, a combining multiple sequences as a
multivariate one into a simple model was applied and proved that the
return of values of the system calls also play an important role in detection.
c) Markov process has two approaches: Markov chains and hidden Markov
models. The Markov model is a set of finite states interconnected going
through a stochastic process to determine the topology and capabilities of a
model [36]. Each stage in the process depends on the outcome of the
previous stage. The anomalies are detected [35] by comparison of the
associated probability recorded for the process with a fixed threshold. This
gives it an advantage of detecting unusual multiple occurrence of events
[47]. The hidden Markov model assumes the system to be a Markov
process where stochastic processes with finite states of possible outcomes
are hidden [36]. Ye Nong [49] presented an anomaly technique using
Markov chain model to detect intrusion. In this model, Markov chain was
used to represent a temporal profile of normal behaviour in a computer and
network system. The Markov chain model of the normal profile is learned
from the historic data of the system’s normal behaviour. The observed
behaviour of the system is analysed to see if the Markov chain model of
the normal behaviour supports the observed behaviour. A low probability
of support indicates an anomalous behaviour meaning an intrusion. The
technique was implemented on the Sun Solaris system and distinguished
normal activities from attacks perfectly.

2.1.3 Signature Based Detection

Signature based detection defines a set of rules used to match the patterns in the network
traffic. If a mismatch is detected it raises an alarm [50]. It has an advantage of being able to
detect attacks giving a low false positive detection ratio [51] It has a drawback of being able
to detect only attacks known to the database [19]. Signature based detection systems are
programmed with distinct decision rules. The rules set for detection are coded in a straight
forward manner to detect intrusion. They are programmed in four categories: state modelling,
expert system, and string matching.

 State modelling is the encoding of attacks as a number of different states in a finite

automaton. Each of these attacks has to be observed in the traffic profile to be
considered as an intrusion. They occur in sub-classes as time series models [15]: the
first is state transition which was proposed by Porras et al. which uses a state
transition diagram to represent intrusion [52]. The approach in [52]models intrusion
as a series of state transitions which are described as signature action and states
descriptions. State diagrams were written to correspond to the states of an actual

7
 computer system. The basis of a rule based expert system to detect intrusions was
formed by these diagrams. The intrusions are in the form of a simple chain
transitioning from start to end. The second is the petri-net where states form a petri-
net. They form a tree structure where the transition states are not in any order
[15][53].
 Expert systems contain a set of rules used to describe attacks scenarios known to the
system. The given rules that describe the attack scenarios are often forward-chaining
systems. A production based expert system tool has been used since they best handle
systems with new events entering into the system. The size of the rule based increases
as the execution speed increases since the rule will go through a longer list. It is
however vulnerable to attacks not known to the set rules [2]. P.Zhisong et al.[54],
presented an intrusion detection system model based on neural network and expert
system. The aim of the experiment was to take advantage of classification abilities of
neural network for probe and Dos attacks and expert based for U2R and R2L attacks.
KDD Cup’99 dataset was employed in the experiment. Expert system was able to
improve the detection accuracy using the detection rules: ALERT UDP ENET any <-
> HNET 31337 MESSAGE:"BO access" DATA: |ce63 d1d2 16e7 13cf 3ca5 a586|".
 String matching is a process of knowledge acquisition just as Expert system but has a
different approach in exploiting the knowledge [14]. It deals with matching the
patterns in the audit event generated by the attack but not involved in the decision
making process [47]. This technique has been used effectively commercially as an
IDS [55], [56]. It is noted that not all signature based IDS can be represented by a
simple pattern which gives it a limitation [57]. T.Sheu et al.[58]proposed an efficient
string matching algorithm with compact memory as well as high worst-case
performance. A magic number heuristic based on the Chinese remainder theorem was
adopted. The algorithm significantly reduced the memory requirements without
bringing complex processes. The latency of off-chip memory references was
drastically reduced. It was concluded the algorithm gives a cost effective and an
efficient IDS.

2.1.4 Discussion
A summary table of intrusion techniques, source of data, applications and data used in
selected reviewed papers between 2014 and 2016 is shown in Table II.

Table II. Summary table of selected reviewed papers.

Authors/Year Year Intrusion Source of Application Data Set

Published Technique Data
Chetna and Harsh 2014 Anomaly NIDS Cloud computing Real Life
[59]
M. Ali Alheeti et al. 2016 Anomaly NIDS Transport Real Life
[60]
J. Hong et al. 2014 Anomaly NIDS Substation Real Life
[61]
R. Mohan et al. [62] 2015 Signature HIDS/NIDS Cloud computing Real Life
S. Vasudeo et al. 2015 Anomaly/ HIDS/NIDS Computer systems Real Life
[63] Signature
W.Haider et al. [64] 2015 Anomaly HIDS Cloud computing ADFA-LD
and
KDD98

8
 H. Toumi et al. [65] 2015 Signature NIDS Cloud computing Real Life
M. Guerroumi et al. 2015 Signature HIDS Internet of things Real Life
[66] (IoT)
N. Dipika et al. [67] 2015 Anomaly NIDS Information KDD99
systems Cup
W.Haider et al. [68] 2015 Anomaly HIDS Cyber space KDD98
and UNM
N.Aissa et al. [69] 2015 Anomaly NIDS Computer systems KDD99
H. Omessaad [70] 2015 Signature HIDS Cloud computing Real Life
X. Lin et al. [71] 2015 Signature NIDS Transport Real Life
S. Banerjee et al. 2015 Signature/ NIDS Computer systems Real Life
[72] Anomaly
P. Satam [73] 2015 Anomaly NIDS Telecommunication Real Life

Recent works have shown that works are still on-going using the Intrusion technique and
source of data techniques. S.Vasudeo et al. [63] presented a hybrid of Signature and Anomaly
with a hybrid of data source on a real life dataset. S.Banerjee et al.[72]also combined
signature and anomaly technique on real life dataset. Other referenced works the table applied
single techniques on real life datasets and KDD Cup datasets. The application of these
techniques cut across all areas to determine their effectiveness in detecting intrusion.

3.0 MACHINE LEARNING TECHNIQUES

Anomaly detection systems are human-independent. They detect anomalies by revealing
abnormal characteristics in a system over a period of time [44]. The effectiveness of this
technique is its capabilities to differentiate between normal and abnormalities within a
network.

Machine Learning (ML) can provide IDS methods to detect current, new and subtle attacks
without extensive human-based training or intervention. It is defined as a set of methods that
can automatically detect patterns to predict future data trends [74],[75]. Whilst a large
number of machine learning techniques exist, the fundamental operation of all of them relies
upon optimal feature selection. These features of are the metrics which will be used to detect
patterns and trends. For example, one feature of a network is the packet size: machine
learning techniques may monitor the packet size over time and generate distributions from
which conclusions may be drawn regarding an intrusion. This section reviews the feature
selection of machine learning IDS and classification of machine learning techniques used as
IDS in Figure 3.

3.1 IDS feature selection

Machine learning classification involves two phases: the classification phase and training
phase. The training phase learns the distribution of the features and during the classification
phase the learned features is applied as a normal profile where any abnormality will be
detected [76], [77]. A.K. Jain et al. [78] developed a model of statistical pattern recognition
as shown in Fig.3. The test and training data are normalised in the processing unit as well as
removing noise form the data. In the training phase, the feature extraction unit extracts a
representation feature set from the processed training data which are used in training a
classifier. In the classification phase, the trained classifier is applied to assign the test data to
the selected features from the training phase [77].

9
 Figure 3 Model of Machine learning classification process

A high quality of training data is required to achieve the best performance of ML IDS. The
training data thus contain both normal and abnormal patterns [76]. Features are the important
information extracted form from raw data and are important in classification and detection
which influence the effectiveness of ML IDS. In [79], C. Kruegel et al. created attacks
containing system call sequences similar to normal system call. Kruegel et al. addressed the
issue by analysing on the system call arguments instead of finding the relations between
sequences of actions [76]. Table III. Shows the features extracted by Kruegel et al. for four
different arguments.

Table III. Extracted features by Kruegel et al. [76] for four different arguments

No. Feature Description

1 String length
2 String character
3 Structural inference
4 Token finder

In [80], V. Mahoney et al. extracted features by analysing each header-field of network

packets and flows. The proposed approach det6ected anomalies in a network based on 33
fields of IP, UDP, TCP, ICMP and Ethernet protocols. Table IV shows some of the extracted
features.

Table IV. Extracted features based on analysis of each header field

No. Feature Description No. Feature Description

1 IP Source 10 IP protocol

2 IP Destination 11 IP Length

3 TCP Source Port 12 IP Type of Service

4 TCP Destination Port 13 IP Header Length

5 UDP Source Port 14 IP Time to Live

6 UDP Destination Port 15 Ether Source

7 UDP Length 16 Ethernet Size

10
8 ICMP Code 17 Ethernet Destination

9 ICMP Type 18 Ethernet Protocol

In [81], W. Lee et al. extracted features from TCP/IP connections. Experiment was conducted
on 1998 DARPA dataset where a set of basic features from domain knowledge was extracted.
Most of the basic features could only be extracted after the TCP connection was terminated
leading to a delay in the detection phase [76]. The 1998 DARPA dataset is categorised into
five. These are NORMAL and four different types of attack. The attacks are: Dos attacks,
probe attack, User to root (U2R) attack and Remote to local (R2L) attacks. Dos attack which
denies or prevents legitimate users resource on a network or system by introducing useless or
unwanted traffic. Probe is an attack which scans through a computer network to make a
profile of information for future attacks. U2R is attacks use a local account from a remote
machine to gain access to the targets system due to vulnerabilities in its operating system.
R2L attacks are initiated to gain unauthorised access to root privileges from outside. Table
V. illustrates the basic features of individual TCP connections by W. Lee et al.

Table V. Basic features extracted from individual TCP connections

No. Feature Name Feature Description Type

1 Duration Length of connection in seconds continuous

2 protocol_type Type of the protocol used, e.g.. up,tcp, etc. discrete

3 Service Network service on the destination, e.g.http, discrete

telnet, etc.

4 scr_bytes Number of data bytes from source to destination continuous

5 dst_bytes Number of data bytes form destination to source continuous

6 Flag Normal or error status of network discrete

7 Land 1 if connection is from/to same host/port; 0 discrete

otherwise

8 wrong_fragment Number of wrong fragments continuous

9 Urgent Number of wrong fragments continuous

Table VI. Traffic features extracted from a 2second time window

No. Feature Name Feature Description Type

10 Count Number of connections to the same host as the continuous
current connection in the past 2seconds

11 serror_rate Percentage of connections having “SYN” error continuous

12 rerror_rate Percentage of connections having “REJ” error continuous

13 same_srv_rate Percentage of connections having same service continuous

14 diff_srv_rate Percentage of connections having different service continuous

11
15 srv_count Number of connections to the same service as the continuous
current 1connection in the past 2seconds

16 srv_serror_rate Percentage of connections having “SYN” errors continuous

17 srv_rerror_rate Percentage of connections having “REJ” errors continuous

18 srv_diff_host_rate Percentage of connections to different host

Table VII. Traffic features extracted from a window of 100 connections

No Feature Name Feature Description Type

19 dst_host_count Count for destination host continuous

Percentage of connections having “SYN” errors continuous
20 dst_host_serror_rate
Percentage of connections having “REJ” errors continuous
21 dst_host_rerror_rate
Percentage of connections having same service continuous
22 dst_host_same_srv_rate
Percentage of connections having different continuous
23 dst_host_diff_srv_rate service

24 dst_host_srv_count Number of connections to the same destination continuous

port

25 dst_host_srv_serror_rate Percentage of connections having “SYN” errors continuous

26 dst_host_srv_rerror_rate Percentage of connections having “REJ” errors continuous

27 dst_host_srv_diff_host _rate Percentage of connections to different host continuous

28 dst_host_same_src_port _rate Percentage of connections to the same source port continuous

Table VIII. Traffic features extracted within a connection suggested by domain knowledge

No. Feature Name Feature Description Type

29 Hot Number of “hot” indicators continuous

30 num_failed_logins Number of failed login attempts continuous

31 logged_in 1 if successfully login; 0 otherwise discrete

32 num_compromised Number of “compromised” conditions continuous

33 root_shell 1 if root shell is obtained; 0 otherwise discrete

34 su_attempted 1 if “su root” command attempted; 0 otherwise discrete

35 num_root Number of “root” accesses continuous

12
36 num_file_creations Number of file creation operations continuous

37 num_shells Number of shell prompts continuous

38 num_access_files Number of operations in access control files continuous

num_outbound_cmd Number of outbound commands in an ftp session continuous

39

40 is_hot_login 1 if the login belongs to the “hot” list; 0 otherwise discrete

41 is_guest_login 1 if the login is a “guest” login; 0 otherwise discrete

W. Lee et al. discovered many attacks such as R2L and U2R were in the payloads of packets.
This prompted a proposal to combine features extracted from the payload with domain
knowledge. They called the features “time based traffic” features for connection records.
Features 10 to 18 in table VI are extracted using a 2second time window with 11 to 15 having
same host connection and 16 to 18 having same service connection, 19 to 28 in Table VII are
extracted from a window of 100 connections with 20 to 24 same host connection and 25 to 28
same service connection. 29 to 41 in Table VIII are features extracted from connections
suggested by domain knowledge [77]. Both tables provide a detailed insight mapping the
most current and meaningful features for machine learning intrusion detection systems.

3.2 Classification of Machine Learning Techniques

This section gives an extensive classification of ML techniques. Each of the techniques is
described in detail and how these techniques have been applied as IDS. Figure 3.1 shows the
classification of Machine learning IDS.

13
 Figure 3.1 ML based techniques

 Bayesian Networks are graphical modelling tools used to model the probability of
variables of interest. They are directed acrylic graphs where each node represents a
discrete random variable of interest. Each node contains the states of the random
variable in tabular form representing the conditional probability table (CPT) which
specifies conditional probabilities of the domain variable with other connected
variables. The CPT of each node contains probabilities of the node being in a specific
state given the domain variable states. The existence of the relationship between the
nodes of the domain variable and connected variables in a Bayesian networks show
the direction of causality between other connected variables. Thus the connected
variables are causally dependent on the ones represented by the domain variable.
Given a set of discrete random variable 𝑋 = {𝑥1 , 𝑥2 … 𝑥𝑛 }, the joint probability of the
variable can be computed based on Bayes Rule as:

𝑃(𝑥1 , 𝑥1 … 𝑥𝑛 ) = ∏𝑛𝑡=1 𝑃(𝑥𝑖 \𝑝𝑎 (𝑥𝑖 )) (1)

14
 where 𝑝𝑎 (𝑥𝑖 ) represents the specific values of the variables in the domain variable
node of 𝑥𝑡 [82], [83]. This technique has generally been used as an intrusion detection
system.
A. Onik et al.[84] Conducted an experiment using Bayesian networks on NSL-KDD
dataset containing 25,192 records with 41 features. Apart from the normal class label
it contained 4 more class of attacks known as Dos, U2R, R2L and probe attacks. The
filter approach of feature selection was used to reduce the dataset features from 41 to
16 important features. Bayesian model was built and proved to predict attacks with
superior overall performance accuracy rate of 97.27% keeping the false positive rate
at a lower rate of 0.008. The model as compared to Naïve Bayes, K-means clustering,
decision stamp and RBF network recorded 84.86%, 80.75%, 83.31% and 91.03%
respectively in terms of accuracy.
An experiment by M. Bode et al.[85] Analysed the network traffic in a cyber situation
with Bayesian network classifier on the KDD Cup’99 dataset with 490,021 records.
The data set was made up of 4 types of attacks (DoS, U2R, R2L and probe). In this
experiment, they adopted the risk matrix to analyse the risk zone of the attacks. The
risk analysis adopted showed DoS was most frequent attack in occurrences. The
results showed Bayesian network classifier is a suitable model resulting in same
performance level classifying the DoS attacks as association rule mining. Bayesian
network classifier outperformed Genetic Algorithm in classifying probe and U2R
attacks and classified Dos equally.
 Genetic Algorithm (GA) is an adaptive search method in a class of evolutional
computation using techniques inspired from convolutional biological process. The
principle is based on a stochastic global [25] search method initialising with a random
generation of chromosomes. The chromosomes are called population. They evolve
through selection, crossover and mutation as shown in Figure 3.2. Each chromosome
represents a problem to be solved and encoded as strings. The positions of the
chromosomes are commonly represented as binary (0, 1) or as a list of integers. These
positions sometimes referred to as genes keep changing at each initialisation. The
solution created during each generation is based on an evaluation function. The
selection is thus based on the chromosome fitness level [86], [87].

Figure 3.2 Flow chart of Genetic Algorithm

15
 Tao Xia et al. [88] developed a hybrid method based information theory and GA.
Information theory was used to filter out the most important features out of 41
features in KDD’99 dataset with 494021 records. Linear rule was used initially to
classify normal and abnormal before GA to obtain the appropriate classification. In
the detection of Dos, U2R, R2L and probe attacks, the information theory and GA
hybrid method recorded 99.33%, 63.64%, 5.86% and 93.95% as detection rates
respectively. The detection rates were compared to Ctree and C5. Ctree recorded
98.91%, 88.13%, 7.41% and 50.35% whilst C5 recorded 97.1%, 13.2%, 8.4% and
83.3% respectively for the attacks.

In layered approach for intrusion detection systems based GA, M. Padmadas et al.[89]
Proposed a method to overcome the weakness in a single layer intrusion detection
system. The layered approach is based on GA with the four layers corresponding to
four groups of attacks (probe, DoS, U2R and R2L). Each layer is trained separately
with a number of features where the layer acts as a filter to block any malicious
activity. In layered approach there is no mathematical approach to calculate the filter
parameters for the attacks. This paper presented GA approach in calculating the filter
parameters making the system more secure. The model efficiently detected R2L attack
and recorded an accuracy of 90% in detection.

 Support Vector Machines (SVM) is a machine learning algorithm that learns to

classify data using points labelled training examples falling into one or two classes.
The SVM algorithm builds a model that can predict if a new example falls into one
category or the other [90],[25], [91] . Figure 3.3 shows a hyperplane defined by
(𝑤, 𝑏), where 𝑤 is a weight and 𝑏 bias constructed in a finite space of the training
sample 𝑁 with points:
{(𝑥1 , 𝑦1 )(𝑥2 , 𝑦2 … (𝑥𝑁 , 𝑦𝑁 ))} (2)

Where 𝑥𝑖 ∈ 𝑅 𝑑 𝑎𝑛𝑑 𝑦𝑖 ∈ {1, −1}. This is conducted since in general the larger the
margin, the lower the generalisation error of the classifier [92].

Figure 3.3 Maximum-margin hyper plane and margins for an SVM trained with
samples from two classes.

16
 B. Senthilnayaki et al.[93] Used GA pre-processed KDD Cup 99 dataset in a pre-
processing module for data reduction since it was complex to process the dataset with
all 41 features. GA was used to select 10 features out of 41 features present in the
KDD Cup 99 dataset and applied SVM for classification. The experiment was carried
out with 100,000 records from the dataset out of which 95% was used as training data
and remaining 10% as test data. The classification process continued till a 10 fold
cross validation was done for results verification. The SVM classified four different
attacks (DoS, probe, U2R, R2L attacks). The performance of SVM classifier was
compared with four other classifiers as shown in table IX.

Table IX. SVM performance analysis SVM, GMM, Naive, MLP and linear algorithm

Classified Algorithm Accuracy% Accuracy%

(41 feature) (10 features)
Gmm 30.5 45.07
Naive 45.17 55.75
MLP 55.24 67.87
Linear 65.47 76.75
SVM 82.45 95.26

Another experiment by L.Teng et al. [94]proposed a novel method integrating PCA

and SVM by optimising the kernel parameters using automatic parameter selection
technique. The experiment was performed on KDD Cup’99 dataset containing five
categories of traffic (normal, DoS attack, R2L attack, U2R attack and probe attack).
Each network record had 41 features of which 7 were discrete and 34 continuous
features. C parameter for RBF kernel of SVM was optimised by the proposed
automatic parameter reduction along with cross validation to reduce the training and
testing time to give better accuracy in detecting attacks.

 K-Nearest Neighbour (K-NN) is a fundamental technique for sample classification.

This technique is known to be non-parametric and highly efficient in classification
[95]. It evaluates the class labels of the test samples [96] based on the majority of test
sample neighbours. The parameter 𝑘 is determined by the user. Based on the test
sample, 𝑘 numbers of training points are determined by taking the closest distance to
the test sample. The prediction of the test sample is the 𝑘 nearest neighbours [97].
A hybrid method of intrusion detection system by Y.Canbay et al.[96]proposed
combination of K-NN and GA algorithm. They tested the hybrid algorithm on KDD
Cup 99 dataset labelled out of five classes; normal, probe attack, Dos attack, R2L
attack and U2R attack. The dataset was reduced to 1000,2000,3000,4000 and 5000
records respectively with 19 features. GA was used to select the 𝑘 nearest neighbour
for K-NN classifier. Three experiments were performed to conclude on the hybrid
method. Each experiment used 10 fold cross validation with different 𝑘 values. The
values in in terms of accuracy were compared with the conventional K-NN. The
proposed hybrid method proved better than conventional K-NN in all values of 𝑘 used
in the experiment.

17
 On the same dataset Q.Zeng et al. [98]compared the performance of K-NN and SVM
model and RIPPER method in detecting attacks. The multi attribute decision was
adopted in this experiment. In the classification of an unknown document vector 𝑋,
𝑘–nearest neighbour algorithm ranks the document’s neighbour among the training
document vectors and uses the class labels of the 𝑘 most similar neighbours to predict
the class of the new document. The similarity in each neighbour to 𝑋 is used to
determine the classes of the neighbour where the similarity is measured by the
Euclidean distance between two document vectors. With this adoption they
categorised each new program behaviour in the dataset into either normal or attack
class. Each system call was treated as a word and each process as a document. 𝑘 was
varied between 15 and 35 till an optimal value of 19 found. The classification was
performed with K-NN and SVM model. The Hit rate was compared to the RIPPER
method. The results showed 97.26% accuracy rate and 6.03% false alarm rate for K-
NN and SVM model whilst the RIPPER method gave 87.26% accuracy rate and 8.6%
false positive rate.
 Decision Tree (DT) algorithm learns and models a data set in classification problems.
It classifies new data set according to what it has learnt from previous data set [99]. It
uses a well-defined criterion in the selection of best features of each node tree during
their construction. A decision tree model has a root node linking to different nodes as
attribute data deciding the path for each node [100]. Decisions are made by
comparison of previous data and marked as leaves [101]. A common decision tree
approach is the C4.5 algorithm.
In network intrusion detection system using decision tree (J48) by S.Sahu et al. [102]
a labelled data set called Kyoto 2006+ was used. The data consist of 24 features, 14 of
which was extracted form KDD Cup 99 dataset and an additional ten important
features. The Perl language was used to extract 15 features from the dataset for the
experiment. The sample dataset contained 134665 records; 44257 normal, 86649
known attacks and 3759 unknown attacks. Decision tree (J48) built using WEKA
3.6.10 tool was used to classify normal, known attacks and unknown attacks in the
network packets. The results showed the decision tree generated classified 97.23%
correctly and 2.67% incorrectly. The simulation results showed decision tree can
classify unknown attacks as well.
T.Komviriyavut et al.[101]presented two intrusion detection techniques which were
decision tree (C4.5) and RIPPER rules to test an online dataset (RLD09 dataset).
RLD09 dataset was collected from actual environment and refined to have 13
features. The dataset is categorised into three; normal, DoS attack and Probe attack.
The experimental data set had 3,000 records of three types of unknown probe attacks
each with 1,000 attack records. Initial experiment with known attacks on dataset
showed a total detection rate of about 98% for both decision tree and RIPPER rule.
A second experiment on the same data set with three unknown probe attacks (advance
port scan, Xmas tree and ACK scan) was performed. The decision tree maintained its
detection rate of 98% whilst the RIPPER rule degraded to about 50%in average.
 Fuzzy logic (FL) concept is derived [103] from the fuzzy sets theory which deals
with approximately reasoning with uncertainty and imprecision [104] by human
being. The features this technique which handles real life uncertainty makes it

18
 attractive for anomaly detection. Intrusion detection involves the classification of a
normal class and an abnormal class. The well-defined nature of the two classes makes
this computation paradigm a helpful one.
A.Toosi et al.[105]introduced a new approach using different soft computing
techniques into classification system to classify abnormal behaviour form normal
depending on the attack type. This work investigated neuro-fuzzy networks, fuzzy
inference and GA on KDD Cup 99 10% dataset. The dataset had a distribution of
normal and four different attacks (probe, DoS, R2L and U2R). A set of parallel neuro-
fuzzy classifiers were used initially for classification. The fuzzy inference system was
based on the output of the neuro-fuzzy classifier determining normal or abnormal
activity. The best results were attained by optimising the structure of the fuzzy
decision tree with GA.

 Clustering is the division of data with no apparent distinguished differences into

groups by functions of two or more parameters. Each group is called a cluster and
have no similarity. The greater the dissimilarity the better the grouping. Their class
labels are always unknown. K-means is an example of clustering algorithm used as
IDS [106],[107]. K-means algorithm makes use of the Euclidean distance to calculate
the data and the cluster centre. The aim is to have a minimum distance within a cluster
and achieve a maximum distance between clusters by the minimisation of an objective
function [108].
Li Jun Tao et al.[109]proposed a k-means clustering with dynamic adjustable number
of cluster. They introduced an algorithm that uses an improved Euclidean distance
formula to calculate the distance between the data and cluster centre by automatically
adjusting the number of cluster when the distance is more than the threshold. This
experiment was performed on KDD Cup’99 dataset containing four attacks (DoS,
Probe, R2L and U2R). The total number of dataset records was 72471. The improved
k-means algorithm recorded 77 clusters with a detection rate of 90% and false
positive rate of 15%. The results compared to the traditional k-means were 108
clusters, detection rate of 85% and false positive rate of 42%.

3.3 Analytical Comparison of ML Techniques

The various ML techniques discussed in the above section are different in their capabilities to
classify attacks. Table X. analyses the advantages and disadvantages in the performance of
ML techniques.

Table X. Advantages and disadvantages of ML techniques

ML Technique Advantages Disadvantages

Bayesian Network The graphical representation gives it Slow in classifying data sets with many
an advantage of breaking complex features.
problems into different smaller
models.
Genetic Algorithm -Uses a technique which is inspired Gets stack in local optima
by convolutional biological process. (overfitting)
-It has the capability of solving
optimisation problems during
classification.
Support -The algorithm is simple to analyse -Selection of kernel function is not straight

19
Vector Machine mathematically. forward.
-All computations are performed in -Slow in training and requires more memory
space using kernels giving it an edge space.
to be used practically.
K-Nearest Neighbour Easy to implement and can solve -Slow in training and requires large memory
multi-class problems space.
-It is computationally complex because to
classify a test sample involve the
consideration of all training samples.
Decision tree -It has a unique structure therefore -If trees are not pruned back it causes
easy to interpret. overfitting.
-It has no limitation in handling high -Type of data must be considered when
dimensional data sets. constructing tree.
(i.e. Categorical or numerical)
Fuzzy Logic -It is based on human reasoning -It’s construction has a high level of
concepts which are not precise. generality there by high consumption of
-It gives a representation of resource.
uncertainty.
K-means Algorithm Simple to implement and effective -The outcome of clustering depends on how
cluster centres are initialised to specify k
value.
-The algorithm works for only numerical
data.

3.4 ML Hybrid Techniques

ML Intrusion detection systems have generally been used to detect attacks but in recent
times, using two or more different techniques to form a hybrid has improved the overall
performance. Table XI. Shows accuracy of detection, type of attacks and data set for selected
reviewed ML IDS papers from 2011-2016.

Table XI. ML IDS accuracy of detection and Data set used

Authors(s)/Year ML Technique Attack Type(s) Data Set Accuracy

S.Chordia et al. K-means +KNN+DT R2L,U2R,DoS and KDDCup99 96.55%
[107] 2015 probe
P. Jongsuebsuk et FL+GA Dos and Probe Real Life 97%
al. [110]
2013
B.Senthilnayaki et GA+SVM Dos KDDCup99 99.15%
al. [93] Probe 99.08%
2015 U2R 97.03%
R2L 96.50%
B. Masduki SVM R2L KDDCup99 96.08%
et al. [7]
2015
A.Enache et al. SVM+BAT Malicious NL-KDD 99.38%
[111] Algorithm
2015
S. Akbar et al. GA DoS,R2L,U2R and KDDCup99 92.6%
[112] Probe
2012
A. Aziz et al. DT Dos NSL-KDD 82%
[113] Probe 64%
2013
Chao Lin et al. Cluster centre+ K- Probe KDDCup99 99.98%
[114] NN Dos (6 dimensional data 99.99%
2015 set)

A.Aburomman et SVM+K-NN DoS,R2L,U2R and KDDCup99 87.44-91.69%

20
al. [115] Probe
2016
Shi-Jin Horng et SVM + hierarchical DoS,R2L,U2R and KDDCup99 95.72%
al. [116] clustering algorithm Probe
2011
E. Hodo et ANN DDoS/Dos Real Life 99.4%
al.[117] 2016

As demonstrated in Table XI, the majority of Intrusion Detection Systems using machine
learning have tested their work against the KDDCup99 dataset or the NSL-KDD, in contrast
only two recent studies tested their machine learning systems against real network data.

3.5 Binary Classification metrics

The effectiveness of prediction by the ML algorithm which is either 1 or 0 is based on
confusion matrix prediction outcome as shown Table IX. The outcomes are True Negative
(TN), True Positive (TP), False Positive (FP) and False Negative (FN) [118], [119].

Table IX. Confusion Matrix

Predicted Class

Negative Positive
Class(Normal) Class(Attack)

Negative Class(Normal) True False

Negative(TN) Positive(FP)
Actual Class
Positive Class(Attack) False True
Negative(FN) Positive(TP)

 True Negative (TN): a measure of the number of normal events rightly classified
normal.
 True Positive (TP): a measure of attacks classified rightly as attack.
 False Positive (FP): a measure of normal events misclassified as attacks.
 False Negative (FN): a measure of attacks misclassified as normal.

The following are the basic metrics used to calculate the performance of ML IDS:
𝑇𝑃
True negative rate (Specifity) = 𝐹𝑃+𝑇𝑁 (3)

𝑇𝑃
True positive rate (Sensitivity) = 𝑇𝑃+𝐹𝑁 (4)

𝑇𝑃
False positive rate (Fallout) = = 1 − 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑡𝑦 (5)
𝑇𝑃+𝐹𝑁

𝐹𝑁
False negative rate (Miss Rate) = = 1 − 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 (6)
𝐹𝑁+𝑇𝑃

𝑇𝑃
Precision = (7)
𝑇𝑃+𝐹𝑃

21
 𝑇𝑃
Recall = (8)
𝑇𝑃+𝐹𝑁

𝑇𝑃+𝑇𝑁
Overall accuracy= (9)
𝑇𝑁+𝑇𝑃+𝐹𝑁+𝐹𝑃

A commonly used ML IDS metric is detection rate. This is defined as the number of data
examples correctly classified divided by the test examples.

4.0 ARTIFICIAL NEURAL NETWORK AND DEEP NETWORKS IDS

This section reviews artificial neural network (ANN) and Deep learning which use the
computational intelligence approach to detect attacks.

4.1 Artificial Neural Network

Figure 4 Classification of Artificial Neural Network

Artificial neural network consist of information processing elements known to mimic neurons
of the brain. ANN is categorised into supervised and unsupervised learning. Figure 4 shows
the types of ANN.

4.1.1 Supervised Learning

In supervised learning, the neural network is provided with a labelled training set which
learns a mapping from inputs 𝑥 to outputs 𝑦 given a labelled set of inputs-output pairs

𝑑 = {(𝑥𝑖 , 𝑦𝑖 )}𝑁
𝑖=1 (10)

Where 𝑑 is called the training set and 𝑁 is the number of training examples. It is assumed that
𝑦𝑖 is a categorical variable from some infinite set 𝑦𝑖 ∈ {1 … 𝐶} [120]. Two types of
supervised learning algorithms are used to train a neural network for intrusion detection.

 Multilayer Perceptron (MLP) is a feedforward neural network. The structure of MLP

consists of one or more nodes as the inner layer between the input and output nodes as
shown in Figure 4.1. The most common technique used to train the MLP neural

22
network is the Back Propagation hence the name MLP-BP. The construction of the
MLP-BP neural network is by putting layers of non-linear elements to form complex
hypotheses. The more stages that are added (nodes) the more advance the hypotheses.
Each node takes an element of a feature vector. The output nodes give an output of
two classes (normal and attack). The interconnection between the nodes is associated
with scalar weights with an initial weight assigned to the connection. During training,
the weights are adjusted. Evaluating the hypotheses is done by setting the input modes
in a feed-back process and the values are propagated through the network to the
output. At this stage the gradient descents is used so as to push the error in the output
node back through the network by a back propagation process in order to estimate the
error in the hidden nodes. The gradient of the cost – function can thus be calculated
[99], [121].

Figure 4.1 Three layer neural network

In [122], P. Barapatre et al.[122]experimented with neural network having an input

layer, one hidden layer and one output layer. In this experiment, the input node had 41
features from KDD Cup’99 dataset. The output node was to classify normal or attack
present in dataset. If the output node is ‘0’ then it is labelled “normal” and if ‘1’ it is
labelled “attack”. The learning rate was initially set to 0.1 and the system retrained to
reduce learning rate. A sigmoid activation function was used and values
corresponding to the weights randomly initialised between -1 and -. The system was
trained for each category of attacks (DoS, Probe, U2R and R2L attacks) to determine
the performance of the algorithm on individual attacks. Table IIX shows the
experimental results. It was concluded MLP-BP neural network detected DoS and
Probe attacks more accurately than U2R attacks. It observed an increase in rate of
classification as rate of learning decreases giving rise to a slow convergence. Also
reducing the rate of learning and the sum squared error (SSE) values and re-training
the network showed an improvement in rate of detection.

23
 Table IIX Detection rate comparison for different attack types using MLP
Attack type Detection False positive rate False negative rate
rate % % %
DoS 99.75 4.78 0.24
Probe 98.16 1.33 1.83
U2R 87.09 31.03 12.9
DoS-Probe 99.33 18.5 0.66
R2L 98.99 9 1.01
Overall 81.96 8.51 18.03

 Radial Basis function (RBF) is another feed forward neural network. It classifies by
taking a measurement of the distance between the inputs and the centre of hidden
neurons [118]. Figure 4.2 is an RBF architecture showing the input nodes, one hidden
nodes and output. Each RBF has different parameters with an input vector. The
network output is thus a linear combination of the radial basis function’s output. The
input and hidden nodes weights are always 1 since the transfer function of the
network is a Radial basic function. This allows an adjustment on the weight between
the hidden nodes and the output [123].

Figure 4.2 Radial basis function architecture

C.Zhang et al.[124] compared the performance of RBF and MLP-BP in detection

detecting four different types of attacks on KDD Cup’99 dataset. Both training and
testing data contained 1000 records which were selected from the dataset. In this
experiment each sample was unique with 34 numerical features and 7 symbolic
features. The symbolic features were converted to ASCII numbers before being used
as training or testing data. The results showed RBF achieved better performance than
MLP-BP with a detection rate of 99.2% and false positive rate of 1.2%. MLP-BP
showed a detection rate of 93.7% and false positive rate of 7.2%.

24
 Ju Jiang et al. [125]applied RBF and Back propagation algorithm (BPL) to both
misuse and anomaly detection. The experiment was performed on KDD Cup ‘99
dataset containing 4,900,000 records categorised into 5 (normal, probe attacks, DoS
attacks, R2L attacks and U2R attacks). The misuse detection network had 41 input
features, 4 hidden nodes and 4 output nodes representing normal probe, DoS and R2L
classes respectively. The anomaly network also had 41input features, 1 hidden node
and 1 output node to classify normal and attacks behaviour. The experimental results
showed in misuse detection, RBF based IDS performed similarly to BPL based IDS.
However RBF used a shorter time in training as compared to BPL based IDS and
needed to adjust its decision thresholds. In anomaly detection, the BPL based IDS had
to adjust itself output threshold manually according to the characteristics of the
training dataset to achieve best performance. In anomaly detection RBF based IDS
out performed BPL based IDs.
Bi Jing et al. [123] compared RBF and MLP-BP using a processed KDD Cup ’99
dataset by converting al strings to numeric , reducing the dimension of the dataset
and determining the rational value domain. The features were reduced from 41 to 31
to train both RBF and MLP-BP. The RBF structure used had one hidden layer and
output neuron being the weighted sum of all the output items of the hidden layer. The
simulation results showed RBF network is better than MLP-BP in its property of
having a more regular out with shorter training time and better accuracy in attack
detection.

4.1.2 Unsupervised learning

In unsupervised learning, the neural network is only provided with input data without
conceptualising the output: it discovers patterns within the data autonomously. The data yet
to be discovered is called unlabelled data [120]. Two typical unsupervised learning are Self-
Organisation Maps (SOMs) and Adaptive Resonance theory (ART).

 Self-Organisation Maps
SOMs transform the input of a network into two dimensional feature maps based on
the topological properties of SOMs. The computation of feature maps is by Kohonen
unsupervised learning. The two dimensional feature maps are neurons represented by
coloured squares showing the weights corresponding to each neuron. The inputs are
grouped based on their similarity. The more features mapped, the bigger the coloured
square. The quality is determined by the back ground colouring of the clusters.
Anomaly events can thus be identified by analysing the normal and abnormal events
from the mapping [118], [126].
SOMs are widely used anomaly detection systems. P. Lichodajewski et al. in [127]
applied SOMs as a Host based intrusion detection system. In this work SOM was
trained with an explicit coding of data and gave a clear clustering of abnormal
behaviours.
In [128], H. Gunes Kayacik et al. investigated and demonstrated with hierarchical
SOM architecture with two basic feature sets , one limited to 6 basic features with the
others containing all 41-features. The results gave a false positive rate of 1.38% and
detection rate of 90.4%.

25
 V. Kumar et al in [129] presented a unified framework on SOM. Their approach
detected attacks on a mobile ad-hoc network (MANET) using different parameters.
Their experimental results were found to be better than other neural network
approaches in terms of detection rate and false alarm rate.

 Adaptive resonance Theory (ART)

ART as a basic theory is an unsupervised learning model but as a hybrid it performs
supervised learning. It functions as a pattern recognition and predictive tool. As an
unsupervised learning model, ART-1, ART-2, ART-3 and fuzzy Art makes the list.
The supervised models are made up of ARTMAP, Fuzzy ARTMAP and Gaussian
ARTMAP [118]. The model compares an input vector to a single neuron’s weight
(weight vector) [118], [126].
In [130], M. Chauhan et al. presented a novel technique by uploading the weights of
the network based on sign of an evaluation function [131]. An appropriate evaluation
function was selected by utilizing the probabilities of various states. It concludes that
if the weights in a modified ART network are oscillating, it implies an intrusion and
thus the ART network has the capability of detecting any intrusion.
A hybrid method by P. Somwang et al. in [132] used principal component analysis
(PCA) and fuzzy adaptive resonance theory to identify different attacks on KDD Cup
’99 data set. The results showed a high performance of detection rate of 96.13% and a
false alarm rate of 3.86% of anomaly intrusion detection.

4.1.3 Summary
I. Ahmad et al. in [133] evaluated SOMs, ART and MLP-BP in terms of main criteria and
sub-criteria using analytic Hierarchy. Evaluation based on main criteria investigated less
overhead, maturity, competency, performance and suitability. Sub-criteria investigated cost
effective, time saving, detection rate, minimum false positive, minimum false negative,
handling varied intrusion and handling coordinated intrusion. Different radar graphs with
distinct colours were used for the analysis.

Analysing MLP-BP and SOM showed MLP-BP had a better detection rate, minimum false
positive, minimum false negative, time saving and cost effective. In the case of less overhead
and capability of handling coordinated and varied intrusion, MLP-BP was not as good as
SOM. Analysing MLP-BP and ART showed the same results.

A comparison of ART and SOM showed better results for SOM in terms of less overhead and
handling coordinated intrusion. In another scenario ART was better than SOM in terms of
detection rate, minimum false negative, maturity, time saving and cost effective.

In conclusion, a hybrid ANN approach was found to be the most suitable intrusion detection
system in terms of detection rate, false positive, false negative, cost and time saving.

4.2 Deep Networks

This section focuses on networks that look like the multilayer perceptron (MLP) but have a
different architecture. The difference between the MLP and the deep networks is their
training procedures. This is a class of ML techniques where classification is conducted by
training data with many layers in hierarchical networks with unsupervised learning. Deep

26
networks are inspired by the architecture depth of the brain. In 2006, Hinton et al. from the
university of Toronto came up with Deep Belief Network (DBN) [134]. They trained data
with an algorithm that greedily trains layer by layer using unsupervised learning for each
layer of Restricted Boltzmann Machine (RBM) [135]. After this discovery by Hinton et al.
other deep networks have been introduced using the same principle have been successful in
classification task [136]. Deep networks IDS can be classified based on how the architectures
and techniques are being used. Figure 4.3 shows a classification of Deep networks IDS.

Figure 4.3 Classification of deep learning IDS

4.2.1 Generative Architecture

Generative models are also referred to as graphical model because they depict
independence/dependence for distribution. They are visualised as graphs that have nodes
representing random variables and arcs show the relationship between random variables that
can have millions of parameters to graphically represent the given system [136], [137]. The
joint statistical distribution of the variables can be written as products of the nodes and their
associated variables [136], [138]. The graphical models have hidden variables that cannot be
observed. Generative models are associated with supervised learning since their training does
not depend on the labels of the data. For classification purposes the models goes through a
pre-training stage (unsupervised learning). During this process, each of the lower layers are
trained separately from the other layers which allows the other layers to be trained in a
greedily layer by layer from bottom to up. All other layers are trained after pre-training [136].
The sub-classes of generative models are Recurrent Neural Network (RNN), Deep Auto-
Encoder, Deep Boltzmann Machine (DBM) and Deep Believe networks (DBN).

 Recurrent Neural Network (RNN)

RNN is a class of deep networks that are either considered supervised or unsupervised
learning with an input sequential data whose length could be as large as its depth
[139]. The RNN model architecture is a feedback loop linking layer by layer with the
ability to store data of previous input increasing the reliability of the model [140].
There are two types on the RNN in terms of architecture: Elman and Jordan RNNs.
The Elman model has a simple feedback looping layer by layer. The Jordan model has

27
 a feedback looping all neurons within a layer to the next layer. There exists also a
feedback connecting a neuron to itself. The ability of the Jordan RNN to store
information in the neurons allows it to train less input vector for classification of
normal and abnormal patterns with high accuracy [140]. Figures 4.4 and 4.5 are
simple architectures of the Elman RNN and Jordan RNN showing the context unit
known to store information of the previous output of hidden layer.

Figure 4.4 Jordan Network

Figure 4.5 Elman Network

A study by K. Jihyun et al. [141] applied long short term memory (LSTM)
architecture to RNN and trained the IDS using KDD Cup ‘99 dataset. By comparing
the accuracy with other IDS classifiers, LSTM-RNN recorded 96.93% accuracy with
a detection rate of 98.88%. Although the FAR was slightly higher than the others,
they concluded its overall performance was the best.

 Deep Auto-Encoder
These are energy based deep models classified as generative models in their original
form. It comes in different forms mostly also generative models. Other forms are
stacked auto encoder and de noising auto encoder a [138].
28
 An Auto-Encoder becomes deep when it has multiple hidden layers. It is made up of
an input layer unit representing the sample data, one or two hidden layer units where
the features are transformed and then mapped to the output layer unit for
reconstruction. Training the auto-encoder gives it a “bottleneck” structure where the
hidden layer becomes narrower than the input layer to prevent the model from
learning its identity function [120].

Figure 4.6 Deep auto encoder

Unfortunately, the deep auto encoder when trained with back propagation has not
been a success the evaluation gets stuck in local minima with a minimal gradient
signal when trained with back propagation. Pre-training a deep auto encoder using the
greedy layer wise approach by training each of the layers in turns has proven to
alleviate the backpropagation problems [120], [139],[142].
In [143], B. Abolhasanzadeh proposed an approach to detect attacks in big data using
deep auto encoder. The experiment was conducted on NSL-KDD data set to test the
method of applying bottle neck features in dimensionality reduction as part of
intrusion detection. The results in terms of accuracy rate out performed PCA, factor
analysis and Kernel/PCA. It was concluded; the results recorded in terms of accuracy
makes this approach promising one for real world intrusion detection.
 Deep Boltzmann Machine(DBM)
DBM is a unidirectional graphical model. Currently there exist no connection between
units on the same layer but between the input units and the hidden units. DBM when
trained with a large supply of unlabelled data and fine-tuned with labelled data acts
as a good classifier [136]. Its structure is an offspring of a general Boltzmann machine
(BM) which is a network of units based on stochastic decisions to determine their on
and off states [138]. BM algorithm is simple to train but turns to be slow in the
process. A reduction in the number of hidden layers of a DBM to one forms a
Restricted Boltzmann Machine (RBM) [139]. DBM when trained with a large supply
of unlabelled data and fine-tuned with labelled data acts a good classifier. Training a
stack of RBM with many hidden layers using the feature activation on one RBM as
the input for the next layer leads to the formation of Deep Believe Network (DBN).

29
 Figure 4.7 Left: A Boltzmann machine with stochastic on and off hidden features and
a variable layer representing a vector of stochastic on and off states. Right: An RBM
with no visible to visible units connected and no hidden to hidden units connected.
U. Fiore et al. in [144]explored RBM in anomaly detection by training a network with
real world data traces from a 24hour work station traffic. This experiment was to test
the accuracy of RBM to classify normal data and data infected by bot. a second
experiment trained RBM with KDD Cup ‘99 data set and tested against real world
data. To randomize the order of test data, the experiment was repeated 10times. The
experiment confirmed testing a classifier in two different networks training data
affects the performance. They suggested the nature of anomalous traffic and normal
traffic should be investigated.
 Deep Belief Networks (DBN)
DBN uses both unsupervised pre training and supervised fine-tuning techniques to
construct the models. Figure 4.8 shows a DBN which is made up of a stack of
Restricted Boltzmann Machines (RBMs) and one or more additional layers for
discrimination task. RBMs are probabilistic generative models that learn a joint
probability distribution of observed (training) data without using data labels. Once the
structure of a DBN is determined the goal for training is to learn the weights 𝑤
between layers. Each node is independent of other nodes in the same layer given all
nodes which gives it the characteristic allowing us to train the generative weights of
each RBM [120]. It then goes through a greedy layer by layer learning algorithm
which learns each stack of RBM’s layer at a time. In Figure 4.8 left, the top layers in
red form a RBM and the lower layers in blue form directed sigmoid believe network
[145].

Figure 4.8 Three layer Deep Belief network and three layers deep Boltzmann machine

30
 DBN has been trained as a classifier by N. Gao et al. [146] to detect intrusion by
comparing the performance to SVM and ANN. The classifiers were trained on KDD
data set. The authors proved that deep learning of DBN can successfully be used as an
effective ID. They concluded the greedy layer by layer learning algorithm when used
to pre-train and fine-tune a DBN gives a high accuracy in classification. The results
showed that DBN recorded the best accuracy of 93.49%, a TP value of 92.33 and FP
of 0.76%.
Z. Alom et al.[147] also exploited the DBNs capabilities to detect intrusion through
series of experiments. The authors trained DBN with NSL-KDD data to identify
unknown attack on it. They concluded by proposing DBN as a good IDS based on an
accuracy of 97.5% achieved in the experiment. This results was compared with
existing DBN-SVM and SVM classifiers which it out performed.

4.2.1 Discriminative Architecture

The discriminative architecture uses discriminative power for classification by characterising
the posterior distributions of classes conditioned on the input data. Recurrent neural network
and convolutional neural network are two types of discriminative architecture.

 Recurrent neural network (RNN) uses discriminative power for classification when
the model’s output is an explicit labelled data in sequence with the input data
sequence. To train RNN as a discriminative model, training data needs to be pre-
segmented and a post-processing to transform the output to a labelled data [138].
 Convolutional neural network
A convolutional neural network (CNN) is a type of discrimination deep architecture
with one or more convolutional and pooling layers in an array to form a multilayer
neural network [139],[148],[149]. In general, convolutional layers share many
weights followed by sampling of the convolutional layer’s output by the pooling layer
which results in some form of translational invariant properties [139].
CNN has fewer parameters as compared to other connected networks with the same
number of hidden units which gives it an advantage of easier training [148]. CNN
architecture is that of multi-layer perceptron [150], and are variant of MLP which are
inspired biologically. Hubel and Wiesel worked on the cat’s visual cortex and
deduced that visual cortex is made up of an arrangement of cells in a complex
manner. These cells are sensitive to small sub-regions of the visual field known as the
receptive field. These fields are positioned to shield the entire visual field to enable
the cell behave just like a local filter over the input space [151]. The series of layers
making up a CNN architecture are the convolutional layer, max pooling layer and the
fully connected layer [150], [152]. The convolutional layer is made up of neurons
forming a rect6angular grid where previous layers are made of neuros shaped as a
rectangular grid. The rectangular grid neurons are connected to each other with the
inputs from previous rectangular units through a set of weights known as filter banks
[149], [152]. These weights for the rectangular units do not change for every
rectangular grid of neuron to form convolutional layers. In architectures where the
convolutional layer is made up of different grids [152], each grid uses a different filter
bank [149], [153]. Each convolutional layer is followed by a pooling layer which

31
 merges subsets of the convolutional layer’s rectangular block by taking sub-samples
to give an output of the block. The pooling can be done in several ways such as [152]
computing the maximum or average or a learned linear summing of neurons in the
blocks. Some blocks turn to be shifted more than a row or a column and in turn feed
in an s input to neighbouring pooling units. This causes a reduction in the dimension
of the system design and thus causing variations to the input [149], [153]. The final
stage which has several convolutional and max-pooling layers non-linearly stacked in
the neural network form a fully connected layer of network [152]. The connectivity
which allows the set of weights of the filter banks to be trained easily [149], [152].

Figure 4.9 Image of CNN showing layers making up architecture [154]

5.0 Conclusion
Shallow and deep networks intrusion detection systems have gained a considerable interest
commercially and amongst the research community. With advancement in data sizes,
intrusion detection systems should have the characteristics to handle noisy data with high
accuracy in detection with high computational speed. This paper gives an overview of the
general classification of intrusion detection systems and taxonomy with recent and past
works. This taxonomy gives a clear description of intrusion detection system and its
complexity.
Current studies of deep learning intrusion detection systems have been reviewed in this paper
to help address the challenges in this new technique still in its early stages in intrusion
detection. In particular recent papers have been reviewed in this work considering all the
machine learning techniques including the single and hybrid techniques.
The scope of the work on classifying intrusion detection systems, reviewing the various
methods of detecting anomaly, performance of these methods were based on past and recent
works revealing the advantages and disadvantages of each of them.
The focus of the paper on shallow and deep networks described experiments comparing the
performance of these learning algorithms. The experiments demonstrated deep networks
significantly outperformed the shallow network in detection of attacks.
To the best of our knowledge CNN has not been exploited in the field of intrusion detection
but has proven to be a good classifier. DBN is also new in its exploitation in this field and
experimental works are still in progress to determine the reliability of these learning
algorithms to detect attacks.
Signature based technique have been in use commercially but have not been able to detect all
types of attacks especially if the IDS signature list did not contain the right signature.

32
Research work is in progress experimenting new approaches to test the reliability and
efficiency of knowledge based and behavioural approaches in intrusion detection.

REFERENCE

[1] N. Chand, P. Mishra, C. R. Krishna, E. S. Pilli, and M. C. Govil, “A comparative

analysis of SVM and its stacking with other classification algorithm for intrusion
detection,” in 2016 International Conference on Advances in Computing,
Communication, & Automation (ICACCA) (Spring), 2016, pp. 1–6.
[2] B. Mukherjee, L. T. Heberlein, and K. N. Levitt, “Network intrusion detection,” IEEE
Netw., vol. 8, no. 3, pp. 26–41, May 1994.
[3] SANS Institute, “The History and Evolution of Intrusion Detection.” [Online].
Available: https://www.sans.org/reading-room/whitepapers/detection/history-
evolution-intrusion-detection-344. [Accessed: 20-Feb-2016].
[4] R. P. R. I. Sravan Kumar Jonnalagadda, “A Literature Survey and Comprehensive
Study of Intrusion Detection,” Int. J. Comput. Appl., vol. 11, no. 81(16), pp. 40–47,
2013.
[5] R. Mitchell and I.-R. Chen, “A survey of intrusion detection techniques for cyber-
physical systems,” ACM Comput. Surv., vol. 46, no. 4, pp. 1–29, Mar. 2014.
[6] “Internet of Things: How Much are We Exposed to Cyber Threats? - InfoSec
Resources.” [Online]. Available: http://resources.infosecinstitute.com/internet-things-
much-exposed-cyber-threats/. [Accessed: 10-Dec-2015].
[7] B. W. Masduki, K. Ramli, F. A. Saputra, and D. Sugiarto, “Study on implementation
of machine learning methods combination for improving attacks detection accuracy on
Intrusion Detection System (IDS),” in 2015 International Conference on Quality in
Research (QiR), 2015, pp. 56–64.
[8] P. Sneath and R. R. Sokal, “Numerical taxonomy. The principles and practices of
numerical classification,” 1973. [Online]. Available:
http://www.brclasssoc.org.uk/books/Sneath/title_contents.pdf.
[9] M. A., “SIMPSON, GG Principles of animal taxonomy. Columbia University Press,
New York: 1990. Pp xii, 247; illustrated. Price: US $21.50. ISBN: 0-231-09650-X
(paperback reprint),” Arch. Nat. Hist., vol. 19, no. 1, pp. 124--124, 1992.
[10] S. Axelsson and Stefan, “The base-rate fallacy and its implications for the difficulty of
intrusion detection,” in Proceedings of the 6th ACM conference on Computer and
communications security - CCS ’99, 1999, pp. 1–7.
[11] J. McHugh, “Testing Intrusion detection systems: a critique of the 1998 and 1999
DARPA intrusion detection system evaluations as performed by Lincoln Laboratory,”
ACM Trans. Inf. Syst. Secur., vol. 3, no. 4, pp. 262–294, Nov. 2000.
[12] D. Rozenblum, “Understanding Intrusion Detection Systems,” SANS Inst., no. 122, pp.
11–15, 2001.
[13] “What it is Network intrusion detection system? | COMBOFIX.” [Online]. Available:
http://www.combofix.org/what-it-is-network-intrusion-detection-system.php.
[Accessed: 10-Dec-2015].

33
[14] H. Debar, M. Dacier, and A. Wespi, “A revised taxonomy for intrusion-detection
systems,” Ann. Des Télécommunications, vol. 55, no. 7–8, pp. 361–378.
[15] S. Axelsson and S. Axelsson, “Intrusion Detection Systems: A Survey and
Taxonomy,” 2000. [Online]. Available:
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.1.6603. [Accessed: 30-Sep-
2016].
[16] A. Hafez, S. and hamilton Jr, and A. John, “Intrusion Detection Systems (IDS)
Taxonomy-A Short Review,” This is a Paid Advert. STN 13-2 June 2010 Defensive
Cyber Secur. Policies Proced. 2, p. 23.
[17] C. Xenakis, C. Panos, and I. Stavrakakis, “A comparative evaluation of intrusion
detection architectures for mobile ad hoc networks,” Comput. Secur., vol. 30, no. 1, pp.
63–80, 2011.
[18] H.-J. Liao, C.-H. Richard Lin, Y.-C. Lin, and K.-Y. Tung, “Intrusion detection system:
A comprehensive review,” J. Netw. Comput. Appl., vol. 36, no. 1, pp. 16–24, Jan.
2013.
[19] H. Debar, M. Dacier, and A. Wespi, “Towards a taxonomy of intrusion-detection
systems,” Comput. Networks, vol. 31, no. 8, pp. 805–822, Apr. 1999.
[20] Sans Penetration Testing, “Host- vs. Network-Based Intrusion Detection Systems,”
2001. [Online]. Available: https://cyber-defense.sans.org/resources/papers/gsec/host-
vs-network-based-intrusion-detection-systems-102574. [Accessed: 24-Feb-2016].
[21] SANS Institute InfoSec Reading Room, “Application of Neural Networks to Intrusion
Detection,” 2001. [Online]. Available: https://www.sans.org/reading-
room/whitepapers/detection/application-neural-networks-intrusion-detection-336.
[Accessed: 24-Feb-2016].
[22] A. Mounji and B. Le Charlier, “Continuous assessment of a Unix configuration:
integrating intrusion detection and configuration analysis,” in Proceedings of SNDSS
’97: Internet Society 1997 Symposium on Network and Distributed System Security,
1997, pp. 27–35.
[23] S. S. Soniya and S. M. C. Vigila, “Intrusion detection system: Classification and
techniques,” in 2016 International Conference on Circuit, Power and Computing
Technologies (ICCPCT), 2016, pp. 1–7.
[24] R. P. R. I. Sravan Kumar Jonnalagadda, “A Literature Survey and Comprehensive
Study of Intrusion Detection,” Int. J. Comput. Appl., vol. 81, no. 16, pp. 40--47, 2013.
[25] Heba Fathy Ahmed Mohamed Eid, “Computational Intelligence in Intrusion Detection
System,” 2013. [Online]. Available:
http://scholar.cu.edu.eg/sites/default/files/abo/files/phd_thesis_computational_intellige
nce_in_intrusion_detection_system_2013.pdf. [Accessed: 24-Feb-2016].
[26] S. Mallissery, J. Prabhu, and R. Ganiga, “Survey on intrusion detection methods,” in
3rd International Conference on Advances in Recent Technologies in Communication
and Computing (ARTCom 2011), 2011, pp. 224–228.
[27] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection,” ACM Comput. Surv.,
vol. 41, no. 3, pp. 1–58, Jul. 2009.

34
[28] S. S. Tirumala, H. Sathu, and A. Sarrafzadeh, “Free and open source intrusion
detection systems: A study,” in 2015 International Conference on Machine Learning
and Cybernetics (ICMLC), 2015, vol. 1, pp. 205–210.
[29] S. Liu, J. Gong, J. Chen, Y. Peng, W. Yang, W. Zhang, and A. Jakalan, “A flow based
method to detect penetration,” in The 7th IEEE/International Conference on Advanced
Infocomm Technology, 2014, pp. 184–191.
[30] H. Kozushko, “Intrusion detection: Host-based and network-based intrusion detection
systems,” Sept., vol. 11, 2003.
[31] N. K. Mittal, “A survey on Wireless Sensor Network for Community Intrusion
Detection Systems,” in 2016 3rd International Conference on Recent Advances in
Information Technology (RAIT), 2016, pp. 107–111.
[32] J. Shun and H. a. Malki, “Network Intrusion Detection System Using Neural
Networks,” 2008 Fourth Int. Conf. Nat. Comput., vol. 5, pp. 242–246, 2008.
[33] A. G. Tokhtabayev and V. A. Skormin, “Non-Stationary Markov Models and Anomaly
Propagation Analysis in IDS,” in Third International Symposium on Information
Assurance and Security, 2007, pp. 203–208.
[34] C. Modi, D. Patel, B. Borisaniya, H. Patel, A. Patel, and M. Rajarajan, “A survey of
intrusion detection techniques in Cloud,” J. Netw. Comput. Appl., vol. 36, no. 1, pp.
42–57, Jan. 2013.
[35] A. Qayyum, M. H. Islam, and M. Jamil, “Taxonomy of statistical based anomaly
detection techniques for intrusion detection,” in Proceedings of the IEEE Symposium
on Emerging Technologies, 2005., 2005, pp. 270–276.
[36] P. García-Teodoro, J. Díaz-Verdejo, G. Maciá-Fernández, and E. Vázquez, “Anomaly-
based network intrusion detection: Techniques, systems and challenges,” Comput.
Secur., vol. 28, no. 1–2, pp. 18–28, Feb. 2009.
[37] T. R. Pillai, S. Palaniappan, A. Abdullah, and H. M. Imran, “Predictive modeling for
intrusions in communication systems using GARMA and ARMA models,” in 2015 5th
National Symposium on Information Technology: Towards New Smart World
(NSITNSW), 2015, pp. 1–6.
[38] H. Debar, M. Becker, and D. Siboni, “A neural network component for an intrusion
detection system,” in Proceedings 1992 IEEE Computer Society Symposium on
Research in Security and Privacy, 1992, pp. 240–250.
[39] A. M. Richard Heady,George Luger, “The Architecture of a Network Level Intrusion
Detection System,” Department of Computer Science ,College of
Engineering,University of New Mexico, 1990. [Online]. Available:
https://www.researchgate.net/profile/Mark_Servilla/publication/242637613_The_archi
tecture_of_a_network_level_intrusion_detection_system/links/5564805a08ae06101ab
df482.pdf. [Accessed: 19-Feb-2016].
[40] K. Wang and S. J. Stolfo, “Anomalous Payload-Based Network Intrusion Detection,”
Springer Berlin Heidelberg, 2004, pp. 203–222.
[41] T. F. Lunt, “A survey of intrusion detection techniques,” Comput. Secur., vol. 12, no.
4, pp. 405–418, Jun. 1993.

35
[42] R. China Appala Naidu and P. S. Avadhani, “A comparison of data mining techniques
for intrusion detection,” in 2012 IEEE International Conference on Advanced
Communication Control and Computing Technologies (ICACCCT), 2012, pp. 41–44.
[43] D. Anderson, T. Frivold, A. Tamaru, A. Valdes, and B. Release, “Next Generation
Intrusion Detection Expert System (NIDES), Software Users Manual.” [Online].
Available: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.26.5048.
[Accessed: 19-Feb-2016].
[44] D. E. Denning, “An Intrusion-Detection Model,” IEEE Trans. Softw. Eng., vol. SE-13,
no. 2, pp. 222–232, Feb. 1987.
[45] K. M. P. V. Jyothsna, V. V. Rama Prasad, “A Review of Anomaly based Intrusion
Detection Systems,” Int. J. Comput. Appl., vol. 28, no. 7, pp. 26–35, 2011.
[46] A. B. Ashfaq, M. Javed, S. A. Khayam, and H. Radha, “An Information-Theoretic
Combining Method for Multi-Classifier Anomaly Detection Systems,” in 2010 IEEE
International Conference on Communications, 2010, pp. 1–5.
[47] B. A. Kuperman, “A categorization of computer security monitoring systems and the
impact on the design of audit sources,” Purdue Univ. West Lafayette, 2004.
[48] W. Sha, Y. Zhu, T. Huang, M. Qiu, Y. Zhu, and Q. Zhang, “A Multi-order Markov
Chain Based Scheme for Anomaly Detection,” in 2013 IEEE 37th Annual Computer
Software and Applications Conference Workshops, 2013, pp. 83–88.
[49] Y. Nong, “A markov chain model of temporal behavior for anomaly detection,” in
Proceedings of the 2000 IEEE Systems, Man, and Cybernetics Information Assurance
and Security Workshop,vol. 166, 2000, p. 169.
[50] H. Zhengbing, L. Zhitang, and W. Junqi, “A Novel Network Intrusion Detection
System (NIDS) Based on Signatures Search of Data Mining,” in First International
Workshop on Knowledge Discovery and Data Mining (WKDD 2008), 2008, pp. 10–16.
[51] H. E. Poston, “A brief taxonomy of intrusion detection strategies,” in 2012 IEEE
National Aerospace and Electronics Conference (NAECON), 2012, pp. 255–263.
[52] P. A. Porras and R. A. Kemmerer, “Penetration state transition analysis: A rule-based
intrusion detection approach,” in [1992] Proceedings Eighth Annual Computer
Security Application Conference, 1992, pp. 220–229.
[53] T. Verwoerd and R. Hunt, “Intrusion detection techniques and approaches,” Comput.
Commun., vol. 25, no. 15, pp. 1356–1365, 2002.
[54] ZhiSong Pan, Hong Lian, GuYu Hu, and GuiQiang Ni, “An integrated model of
intrusion detection based on neural network and expert system,” in 17th IEEE
International Conference on Tools with Artificial Intelligence (ICTAI’05), 2005, p. 2
pp.-pp.672.
[55] SANS Institute InfoSec Reading Room, “Intrusion Detection Systems: An Overview
of RealSecure,” 2001. [Online]. Available: http://www.sans.org/reading-
room/whitepapers/detection/intrusion-detection-systems-overview-realsecure-342.
[Accessed: 04-Mar-2016].
[56] “IBM X-Force: Ahead of the Threat - Resources,” 07-Dec-2011. [Online]. Available:
http://www-03.ibm.com/security/xforce/resources.html#all. [Accessed: 04-Mar-2016].

36
[57] Sandeep Kumar, “CLASSIFICATION AND DETECTION OF COMPUTER
INTRUSIONS,” 1995. [Online]. Available:
http://web.mst.edu/~tauritzd/compsec/papers/kumar95classification.pdf. [Accessed:
04-Mar-2016].
[58] T.-F. Sheu, N.-F. Huang, and H.-P. Lee, “NIS04-6: A Time- and Memory- Efficient
String Matching Algorithm for Intrusion Detection Systems,” in IEEE Globecom 2006,
2006, pp. 1–5.
[59] C. Vaid and H. K. Verma, “Anomaly-based IDS implementation in cloud environment
using BOAT algorithm,” in Proceedings of 3rd International Conference on
Reliability, Infocom Technologies and Optimization, 2014, pp. 1–6.
[60] K. M. A. Alheeti, A. Gruebler, K. D. McDonald-Maier, and A. Fernando, “Prediction
of DoS attacks in external communication for self-driving vehicles using a fuzzy petri
net model,” in 2016 IEEE International Conference on Consumer Electronics (ICCE),
2016, pp. 502–503.
[61] J. Hong, C.-C. Liu, and M. Govindarasu, “Detection of cyber intrusions using
network-based multicast messages for substation automation,” in ISGT 2014, 2014, pp.
1–5.
[62] R. Mohan, V. Vaidehi, and S. S. Chakkaravarthy, “Complex Event Processing based
Hybrid Intrusion Detection System,” in 2015 3rd International Conference on Signal
Processing, Communication and Networking (ICSCN), 2015, pp. 1–6.
[63] S. H. Vasudeo, P. Patil, and R. V. Kumar, “IMMIX-intrusion detection and prevention
system,” in 2015 International Conference on Smart Technologies and Management
for Computing, Communication, Controls, Energy and Materials (ICSTM), 2015, pp.
96–101.
[64] W. Haider, J. Hu, X. Yu, and Y. Xie, “Integer Data Zero-Watermark Assisted System
Calls Abstraction and Normalization for Host Based Anomaly Detection Systems,” in
2015 IEEE 2nd International Conference on Cyber Security and Cloud Computing,
2015, pp. 349–355.
[65] H. Toumi, M. Talea, K. Sabiri, and A. Eddaoui, “Toward a trusted framework for
cloud computing,” in 2015 International Conference on Cloud Technologies and
Applications (CloudTech), 2015, pp. 1–6.
[66] M. Guerroumi, A. Derhab, and K. Saleem, “Intrusion Detection System against Sink
Hole Attack in Wireless Sensor Networks with Mobile Sink,” in 2015 12th
International Conference on Information Technology - New Generations, 2015, pp.
307–313.
[67] D. Narsingyani and O. Kale, “Optimizing false positive in anomaly based intrusion
detection using Genetic algorithm,” in 2015 IEEE 3rd International Conference on
MOOCs, Innovation and Technology in Education (MITE), 2015, pp. 72–77.
[68] W. Haider, J. Hu, and M. Xie, “Towards reliable data feature retrieval and decision
engine in host-based anomaly detection systems,” in 2015 IEEE 10th Conference on
Industrial Electronics and Applications (ICIEA), 2015, pp. 513–517.
[69] N. B. Aissa and M. Guerroumi, “A genetic clustering technique for Anomaly-based
Intrusion Detection Systems,” in 2015 IEEE/ACIS 16th International Conference on
Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed
37
 Computing (SNPD), 2015, pp. 1–6.
[70] O. Hamdi, M. Mbaye, and F. Krief, “A cloud-based architecture for network attack
signature learning,” in 2015 7th International Conference on New Technologies,
Mobility and Security (NTMS), 2015, pp. 1–5.
[71] Xiaodong Lin and Rongxing Lu, “Vehicular Ad Hoc Network Security and Privacy,”
21-Aug-2015. [Online]. Available:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=7159960. [Accessed: 03-
Jun-2016].
[72] S. Banerjee, R. Nandi, R. Dey, and H. N. Saha, “A review on different Intrusion
Detection Systems for MANET and its vulnerabilities,” in 2015 International
Conference and Workshop on Computing and Communication (IEMCON), 2015, pp.
1–7.
[73] P. Satam, “Cross Layer Anomaly Based Intrusion Detection System,” in 2015 IEEE
International Conference on Self-Adaptive and Self-Organizing Systems Workshops,
2015, pp. 157–161.
[74] K. Patel and Kayur, “Lowering the barrier to applying machine learning,” in
Proceedings of the 28th of the international conference extended abstracts on Human
factors in computing systems - CHI EA ’10, 2010, p. 2907.
[75] E. Alpaydın, “Introduction To Machine learning,” 2010. [Online]. Available:
https://www.lri.fr/~xlzhang/KAUST/CS229_slides/chapter18_RL.pdf. [Accessed: 20-
Jan-2015].
[76] A. A. Ghorbani, W. Lu, and M. Tavallaee, “Network Intrusion Detection and
Prevention,” … and Autonomous System, 2010. [Online]. Available:
http://link.springer.com/10.1007/978-0-387-88771-5.
[77] H. T. Nguyen, K. Franke, and S. Petrovic, “Feature Extraction Methods for Intrusion
Detection Systems,” in Threats, Countermeasures, and Advances in Applied
Information Security, vol. 3, IGI Global, 2012, pp. 23–52.
[78] A. K. Jain, P. W. Duin, and J. Jianchang Mao, “Statistical pattern recognition: a
review,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 1, pp. 4–37, 2000.
[79] C. Kruegel, D. Mutz, F. Valeur, and G. Vigna, “On the Detection of Anomalous
System Call Arguments,” Springer Berlin Heidelberg, 2003, pp. 326–343.
[80] M. V. Mahoney and P. K. Chan, “An Analysis of the 1999 DARPA/Lincoln
Laboratory Evaluation Data for Network Anomaly Detection,” vol. 2820, 2003, pp.
220–237.
[81] W. Lee and S. J. Stolfo, “A framework for constructing features and models for
intrusion detection systems,” ACM Trans. Inf. Syst. Secur., vol. 3, no. 4, pp. 227–261,
Nov. 2000.
[82] M. Khosravi-Farmad, A. A. Ramaki, and A. G. Bafghi, “Risk-based intrusion response
management in IDS using Bayesian decision networks,” in 2015 5th International
Conference on Computer and Knowledge Engineering (ICCKE), 2015, pp. 307–312.
[83] R. Stuart, Artificial Intelligence: A Modern Approach, 2nd ed. Upper Saddle
River,New Jersey: Prentice Hall Pa, 2009.

38
[84] A. R. Onik, N. F. Haq, and W. Mustahin, “Cross-breed type Bayesian network based
intrusion detection system (CBNIDS),” in 2015 18th International Conference on
Computer and Information Technology (ICCIT), 2015, pp. 407–412.
[85] M. A. Bode, S. A. Oluwadare, B. K. Alese, and A. F.-B. Thompson, “Risk analysis in
cyber situation awareness using Bayesian approach,” in 2015 International Conference
on Cyber Situational Awareness, Data Analytics and Assessment (CyberSA), 2015, pp.
1–12.
[86] Hartmut Pohlheim, “Competition and Cooperation in Extended Evolutionary
Algorithms,” 2001. [Online]. Available:
http://www.pohlheim.com/Papers/conf_gecco2001/PohlheimH_CompetitionExtEA_G
ECCO2001.pdf. [Accessed: 10-Mar-2016].
[87] W. Li, “Using Genetic Algorithm for Network Intrusion Detection,” Proc. United
States Dep. Energy Cyber Secur. Grou, pp. 1--8, 2004.
[88] Tao Xia, Guangzhi Qu, S. Hariri, and M. Yousif, “An efficient network intrusion
detection method based on information theory and genetic algorithm,” in PCCC 2005.
24th IEEE International Performance, Computing, and Communications Conference,
2005., pp. 11–17.
[89] M. Padmadas, N. Krishnan, J. Kanchana, and M. Karthikeyan, “Layered approach for
intrusion detection systems based genetic algorithm,” in 2013 IEEE International
Conference on Computational Intelligence and Computing Research, 2013, pp. 1–4.
[90] G. Wang, D.-Y. Yeung, and F. H. Lochovsky, “A kernel path algorithm for support
vector machines,” in Proceedings of the 24th international conference on Machine
learning - ICML ’07, 2007, pp. 951–958.
[91] C. J. C. Burges, “A Tutorial on Support Vector Machines for Pattern Recognition,”
Data Min. Knowl. Discov., vol. 2, no. 2, pp. 121–167.
[92] F. Zhu, N. Ye, D. Pan, and W. Ding, “Incremental Support Vector Machine Learning:
An Angle Approach,” in 2011 Fourth International Joint Conference on
Computational Sciences and Optimization, 2011, pp. 288–292.
[93] B. Senthilnayaki, K. Venkatalakshmi, and A. Kannan, “Intrusion detection using
optimal genetic feature selection and SVM based classifier,” in 2015 3rd International
Conference on Signal Processing, Communication and Networking (ICSCN), 2015, pp.
1–4.
[94] L. Teng, S. Teng, F. Tang, H. Zhu, W. Zhang, D. Liu, and L. Liang, “A Collaborative
and Adaptive Intrusion Detection Based on SVMs and Decision Trees,” in 2014 IEEE
International Conference on Data Mining Workshop, 2014, pp. 898–905.
[95] K. Shi, L. Li, H. Liu, J. He, N. Zhang, and W. Song, “An improved KNN text
classification algorithm based on density,” in 2011 IEEE International Conference on
Cloud Computing and Intelligence Systems, 2011, pp. 113–117.
[96] Y. Canbay and S. Sagiroglu, “A Hybrid Method for Intrusion Detection,” in 2015
IEEE 14th International Conference on Machine Learning and Applications (ICMLA),
2015, pp. 156–161.
[97] H. Zhang and G. Chen, “The Research of Face Recognition Based on PCA and K-
Nearest Neighbor,” in 2012 Symposium on Photonics and Optoelectronics, 2012, pp.

39
 1–4.
[98] Q. Zeng and S. Wu, “Anomaly Detection Based on Multi-Attribute Decision,” in 2009
WRI Global Congress on Intelligent Systems, 2009, pp. 394–398.
[99] K. A. Jalill, M. H. Kamarudin, and U. T. Mara, “Comparison of Machine Learning
Algorithms Performance in Detecting Network Intrusion,” pp. 221–226, 2010.
[100] F. Sebastiani and Fabrizio, “Machine learning in automated text categorization,” ACM
Comput. Surv., vol. 34, no. 1, pp. 1–47, Mar. 2002.
[101] T. Komviriyavut, P. Sangkatsanee, N. Wattanapongsakorn, and C. Charnsripinyo,
“Network intrusion detection and classification with Decision Tree and rule based
approaches,” in 2009 9th International Symposium on Communications and
Information Technology, 2009, pp. 1046–1050.
[102] S. Sahu and B. M. Mehtre, “Network intrusion detection system using J48 Decision
Tree,” in 2015 International Conference on Advances in Computing, Communications
and Informatics (ICACCI), 2015, pp. 2023–2026.
[103] S. Rajasekaran and G. A. V. Pai, “Neural Neworks, Fuzzy Logic and Genetic
Algorithm: Synthesis and Applications (WITH CD),” 2003. [Online]. Available:
https://books.google.com/books?hl=en&lr=&id=bVbj9nhvHd4C&pgis=1. [Accessed:
29-Mar-2016].
[104] M. Wahengbam and N. Marchang, “Intrusion Detection in MANET using fuzzy
logic,” in 2012 3rd National Conference on Emerging Trends and Applications in
Computer Science, 2012, pp. 189–192.
[105] A. N. Toosi and M. Kahani, “A new approach to intrusion detection based on an
evolutionary soft computing model using neuro-fuzzy classifiers,” 2007.
[106] H. Om and A. Kundu, “A hybrid system for reducing the false alarm rate of anomaly
intrusion detection system,” in 2012 1st International Conference on Recent Advances
in Information Technology (RAIT), 2012, pp. 131–136.
[107] S. Gupta, “An effective model for anomaly IDS to improve the efficiency,” in 2015
International Conference on Green Computing and Internet of Things (ICGCIoT),
2015, pp. 190–194.
[108] A. Ben Ayed, M. Ben Halima, and A. M. Alimi, “Survey on clustering methods:
Towards fuzzy clustering for big data,” in 2014 6th International Conference of Soft
Computing and Pattern Recognition (SoCPaR), 2014, pp. 331–336.
[109] Li Jun Tao, Liu Yin Hong, and Hao Yan, “The improvement and application of a K-
means clustering algorithm,” in 2016 IEEE International Conference on Cloud
Computing and Big Data Analysis (ICCCBDA), 2016, pp. 93–96.
[110] P. Jongsuebsuk, N. Wattanapongsakorn, and C. Charnsripinyo, “Network intrusion
detection with Fuzzy Genetic Algorithm for unknown attacks,” in The International
Conference on Information Networking 2013 (ICOIN), 2013, pp. 1–5.
[111] A.-C. Enache and V. Sgarciu, “Anomaly Intrusions Detection Based on Support
Vector Machines with an Improved Bat Algorithm,” in 2015 20th International
Conference on Control Systems and Computer Science, 2015, pp. 317–321.

40
[112] S. Akbar, J. A. Chandulal, K. N. Rao, and G. S. Kumar, “Improving network security
using machine learning techniques,” in 2012 IEEE International Conference on
Computational Intelligence and Computing Research, 2012, pp. 1–5.
[113] A. S. A. Aziz, A. E. Hassanien, S. E.-O. Hanaf, and M. F. Tolba, “Multi-layer hybrid
machine learning techniques for anomalies detection and classification approach,” in
13th International Conference on Hybrid Intelligent Systems (HIS 2013), 2013, pp.
215–220.
[114] W.-C. Lin, S.-W. Ke, and C.-F. Tsai, “CANN: An intrusion detection system based on
combining cluster centers and nearest neighbors,” Knowledge-Based Syst., vol. 78, pp.
13–21, Apr. 2015.
[115] A. A. Aburomman and M. Bin Ibne Reaz, “A novel SVM-kNN-PSO ensemble method
for intrusion detection system,” Appl. Soft Comput., vol. 38, pp. 360–372, Jan. 2016.
[116] S.-J. Horng, M.-Y. Su, Y.-H. Chen, T.-W. Kao, R.-J. Chen, J.-L. Lai, and C. D.
Perkasa, “A novel intrusion detection system based on hierarchical clustering and
support vector machines,” Expert Syst. Appl., vol. 38, no. 1, pp. 306–313, Jan. 2011.
[117] E. Hodo, X. Bellekens, A. Hamilton, P.-L. Dubouilh, E. Iorkyase, C. Tachtatzis, and
R. Atkinson, “Threat analysis of IoT networks Using Artificial Neural Network
Intrusion Detection System,” in 2016 3rd International Symposium on Networks,
Computers and Communications (ISNCC), 2016, pp. 1–6.
[118] S. X. Wu and W. Banzhaf, “The use of computational intelligence in intrusion
detection systems: A review,” Appl. Soft Comput., vol. 10, no. 1, pp. 1–35, Jan. 2010.
[119] S. Dong, D. Zhou, and W. Ding, “The Study of Network Traffic Identification Based
on Machine Learning Algorithm,” 2012 Fourth Int. Conf. Comput. Intell. Commun.
Networks, pp. 205–208, Nov. 2012.
[120] K. Murphy, “Machine learning: a probabilistic perspective,” Chance encounters:
Probability in …, 2012. [Online]. Available:
http://link.springer.com/chapter/10.1007/978-94-011-3532-0_2. [Accessed: 06-Jan-
2015].
[121] M. A. Alsheikh, S. Lin, D. Niyato, and H.-P. Tan, “Machine Learning in Wireless
Sensor Networks: Algorithms, Strategies, and Applications,” IEEE Commun. Surv.
Tutorials, vol. 16, no. 4, pp. 1996–2018, 2014.
[122] P. Barapatre, N. Z. Tarapore, S. G. Pukale, and M. L. Dhore, “Training MLP neural
network to reduce false alerts in IDS,” in 2008 International Conference on
Computing, Communication and Networking, 2008, pp. 1–7.
[123] J. Bi, K. Zhang, and X. Cheng, “Intrusion Detection Based on RBF Neural Network,”
in 2009 International Symposium on Information Engineering and Electronic
Commerce, 2009, pp. 357–360.
[124] C. Zhang, J. Jiang, and M. Kamel, “Comparison of BPL and RBF Network in Intrusion
Detection System,” in Rough Sets, Fuzzy Sets, Data Mining, and Granular Computing,
Berlin, Heidelberg: Springer Berlin Heidelberg, 2003, pp. 466–470.
[125] M. Kamel, “RBF-based real-time hierarchical intrusion detection systems,” in
Proceedings of the International Joint Conference on Neural Networks, 2003., 2003,
vol. 2, pp. 1512–1516.

41
[126] Y. Fu, Y. Zhu, and H. Yu, “Study of Neural Network Technologies in Intrusion
Detection Systems,” in 2009 5th International Conference on Wireless
Communications, Networking and Mobile Computing, 2009, pp. 1–4.
[127] A. N. Z.-H. Peter Lichodzijewski, “Host-Based Intrusion Detection Using Self-
organizing Maps,” in IJCNN ’02, 2002, pp. 1714–1719.
[128] M. I. H. H. Gunes Kayacik, A. Nur Zincir-Heywood, “A hierarchical SOM-based
intrusion detection system,” vol. 20, pp. 439–451, 2007.
[129] V. Dinesh Kumar and S. Radhakrishnan, “Intrusion detection in MANET using Self
Organizing Map (SOM),” in 2014 International Conference on Recent Trends in
Information Technology, 2014, pp. 1–8.
[130] M. Chauhan, A. Pratap, and A. Dixit, “Designing a technique for detecting intrusion
based on modified Adaptive Resonance Theory Network,” in 2015 International
Conference on Green Computing and Internet of Things (ICGCIoT), 2015, pp. 448–
451.
[131] S. Haykin, “Neural Networks-A comprehensive foundation,” 1999. [Online].
Available:
https://cdn.preterhuman.net/texts/science_and_technology/artificial_intelligence/Neura
l Networks - A Comprehensive Foundation - Simon Haykin.pdf. [Accessed: 22-Jan-
2015].
[132] P. Somwang and W. Lilakiatsakun, “Intrusion detection technique by using fuzzy ART
on computer network security,” in 2012 7th IEEE Conference on Industrial
Electronics and Applications (ICIEA), 2012, pp. 697–702.
[133] A. S. A. Iftikhar Ahmad, Azween B Abdullah, “Evaluating Neural Network Intrusion
Detection Approaches Using Analytic Hierarchy Process,” in 2010 IEEE International
Symposium on Information Technology, 2010.
[134] and Y. T. G. E. Hinton, S. Osindero, “A fast learning algorithm for deep belief nets,”
Neural Comput., vol. 18, no. 7, pp. 1527–1554, 2006.
[135] Y. Freund and D. Haussler, “Unsupervised learning of distributions on binary vectors
using two layer networks,” Santa Cruz, 1994.
[136] Y. Bengio, “Learning Deep Architectures for AI,” Foundations and Trends® in
Machine Learning, 2009. [Online]. Available:
http://www.nowpublishers.com/product.aspx?product=MAL&doi=2200000006.
[Accessed: 10-Jul-2014].
[137] M. I. Jordan, “Learning in Graphical Models.,” 1998. [Online]. Available:
https://www.cs.cmu.edu/~tom/10-702/zoubin-varintro.pdf. [Accessed: 02-Aug-2016].
[138] Li Deng, “A tutorial survey of architectures, algorithms, and applications for deep
learning,” APSIPA Trans. Signal Inf. Process., vol. 3, no. e2, pp. 1--29, 2014.
[139] L. Deng and D. Yu, “Foundations and Trends® in Signal Processing,” 2014. [Online].
Available: https://www.microsoft.com/en-us/research/publication/deep-learning-
methods-and-applications/. [Accessed: 02-Aug-2016].
[140] L. O. Anyanwu, J. Keengwe, and G. A. Arome, “Scalable Intrusion Detection with
Recurrent Neural Networks,” in IEEE 2010 Seventh International Conference on

42
 Information Technology: New Generations, 2010, pp. 919–923.
[141] J. Kim, J. Kim, H. L. T. Thu, and H. Kim, “Long Short Term Memory Recurrent
Neural Network Classifier for Intrusion Detection,” in 2016 International Conference
on Platform Technology and Service (PlatCon), 2016, pp. 1–5.
[142] A. Ng, “Stacked Autoencoders - Ufldl.” [Online]. Available:
http://ufldl.stanford.edu/wiki/index.php/Stacked_Autoencoders. [Accessed: 18-Feb-
2015].
[143] B. Abolhasanzadeh, “Nonlinear dimensionality reduction for intrusion detection using
auto-encoder bottleneck features,” in 2015 7th Conference on Information and
Knowledge Technology (IKT), 2015, pp. 1–5.
[144] U. Fiore, F. Palmieri, A. Castiglione, and A. De Santis, “Network anomaly detection
with the restricted Boltzmann machine,” Neurocomputing, vol. 122, pp. 13–23, Dec.
2013.
[145] R. Salakhutdinov and G. Hinton, “Deep Boltzmann Machines,” Artif. Intell., vol. 5, no.
2, pp. 448–455, 2009.
[146] N. Gao, L. Gao, Q. Gao, and H. Wang, “An Intrusion Detection Model Based on Deep
Belief Networks,” in 2014 Second International Conference on Advanced Cloud and
Big Data, 2014, pp. 247–252.
[147] M. Z. Alom, V. Bontupalli, and T. M. Taha, “Intrusion detection using deep belief
networks,” in 2015 National Aerospace and Electronics Conference (NAECON), 2015,
pp. 339–344.
[148] “Unsupervised Feature Learning and Deep Learning Tutorial.” [Online]. Available:
http://ufldl.stanford.edu/tutorial/supervised/ConvolutionalNeuralNetwork/. [Accessed:
13-Nov-2015].
[149] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp.
436–444, May 2015.
[150] M. Dalto, “Deep neural networks for time series prediction with applications in ultra-
short-term wind forecasting.” [Online]. Available:
http://www.fer.unizg.hr/_download/repository/KDI-Djalto.pdf. [Accessed: 19-Feb-
2015].
[151] L. Lab, “Deep Learning Tutorial,” 2015. [Online]. Available:
http://deeplearning.net/tutorial/deeplearning.pdf. [Accessed: 20-Aug-2016].
[152] “Convolutional Neural Networks - Andrew Gibiansky.” [Online]. Available:
http://andrew.gibiansky.com/blog/machine-learning/convolutional-neural-networks/.
[Accessed: 24-Nov-2015].
[153] Y. LeCun, “Learning Invariant feature Hierarchies,” 2012. [Online]. Available:
http://yann.lecun.com/exdb/publis/pdf/lecun-eccv-12.pdf. [Accessed: 13-Nov-2015].
[154] “Convolutional Neural Networks (LeNet) — DeepLearning 0.1 documentation.”
[Online]. Available: http://deeplearning.net/tutorial/lenet.html. [Accessed: 29-Apr-
2016].

43