Applied Intelligence (2021) 51:7094–7108

[Link]

Unsupervised deep learning approach for network intrusion

detection combining convolutional autoencoder and one-class SVM
Adel Binbusayyis1 · Thavavel Vaiyapuri1

Accepted: 7 January 2021 / Published online: 24 February 2021

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC part of Springer Nature 2021

Abstract
With the rapid advancement in network technologies, the need for cybersecurity has gained increasing momentum in recent
years. As a primary defense mechanism, an intrusion detection system (IDS) is expected to adapt and secure the computing
infrastructures from the ever-changing sophisticated threat landscape. Many deep learning approaches have recently been
proposed; however, these techniques face significant challenges in identifying all types of attacks, especially rare attacks due
to network traffic imbalances and the lack of a sufficient number of abnormal traffic samples for model training. To overcome
these shortcomings and improve detection performance, this paper presents an unsupervised deep learning approach for
intrusion detection. Unlike the existing IDS model that extracts features and trains a classifier in two separate stages, a single-
stage IDS approach that integrates a one-dimensional convolutional autoencoder (1D CAE) and a one-class support vector
machine (OCSVM) as a classifier into a joint optimization framework is introduced in this paper for the first time. Using
only the normal traffic samples, the approach simultaneously optimizes the 1D CAE for compact feature representation and
the OCSVM for classification by defining a unified objective function combining reconstruction error with classification
error. Thus, the generated compact feature representation has not only reconstruction ability but also discriminative ability
for classification. An in-depth ablation analysis validates the design decisions and provides further insight of the proposed
approach. An extensive set of experiments on two benchmark intrusion datasets, NSL-KDD and UNSW-NB15, demonstrates
the generalization ability of the proposed model for unseen attacks and confirms it as a competitive approach over the
recent state-of-the-art intrusion detection baselines. Overall, the obtained results emphasize that the proposed approach has
potential to serve as a baseline for building an effective IDS.

Keywords Cybersecurity · Network intrusion detection · Deep learning · 1D convolutional autoencoder ·

Feature representation learning · One-class classifier · Joint optimization framework · OCSVM

1 Introduction go beyond crippling our economy and cause loss of life

[2, 3]. Accordingly, ensuring network security has become
The rapidly evolving trends in networking technology have a more urgent requirement than ever before with the
not only radically changed and enriched people’s lifestyles, focus of attention on our society and organizations. In
but have also dramatically transformed the business and an effort to address this situation, security tools such as
social world with vast opportunities for economic growth firewalls, antispam techniques, antiviruses, etc., are utilized
and progress in all walks of life in all countries [1]. With to safeguard the business and social networks against
this growing dependency on technology, cybercriminals cyberattacks. Undeniably, these tools offer first-line-of-
are constantly renovating to stay a step ahead with their defense security but fail to recognize new and sophisticated
tactics in exploring sophisticated cyberattacks that could attacks. Under such circumstances, IDSs with the potential
to adapt to the dynamically changing threat landscape are
 Adel Binbusayyis considered to be at the forefront of cybersecurity [4, 5]. In
[Link]@[Link] particular, IDS is effectively a device or software designed
with the goal to monitor and analyze network traffic for any
1 College of Computer Engineering and Sciences,
attempt that breaks security. Despite decades of significant
Prince Sattam bin Abdulaziz University, Al-Kharj, progress, existing IDSs still remain incompetent in detecting
Saudi Arabia unknown new attacks with a high detection rate (DR) and a

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Unsupervised deep learning approach... 7095

low false alarm rate (FAR), which is a most critical issue in challenges in detecting the new attack vectors. Second, the
modern cyber ecosystems that needs to be addressed with real network traffic data are inherently imbalanced with
utmost prudence [6, 7]. more normal traffic than abnormal traffic. The UDLNs
In light of the powerful capabilities of artificial intel- trained with such imbalanced data are biased toward normal
ligence (AI) technology, the interests of many researchers traffic and face challenges in achieving a high detection rate
are piqued to design IDS models that capitalize on machine for intrusions.
learning to realize intelligent detection of network attacks In light of the aforementioned data sparsity challenge,
[8, 9]. Unfortunately, the IDS model based on traditional this paper presents an innovative IDS approach integrating
machine learning techniques exploits shallow architectures the benefits of a one-class classifier and UDLN within a
that liaise heavily on human-engineered network traffic fea- joint framework by defining a unified objective function
tures rather than the entire raw data [10]. This impedes the that enables to gain an improved performance with regard to
real network application of such models owing to the sig- intrusion detection. In short, the major contributions of this
nificant requirement of intensive human effort for feature work are highlighted below,
engineering. Recently, the advent of deep learning has pro-
pelled AI to new heights and has also opened a promising (a) For the first time, this work proposes a joint optimiza-
path for automated feature extraction from large-scale raw tion framework to optimize CAE and one-class classi-
data. Deep learning is a branch of machine learning based on fier simultaneously for feature representation learning
artificial neural networks with multiple hidden layers. Anal- and intrusion detection respectively.
ogous to machine learning techniques, deep learning also (b) Unlike the existing works, this work combines the
employs supervised and unsupervised learners to automati- reconstruction and classification error to define an
cally extract complex features in a hierarchical manner and unified objective function to ensure that CAE learns
subsequently applies them to make decisions on unseen data the optimal feature representation and minimize the
[6]. The supervised deep learning networks (SDLN), such classification error to achieve higher accuracy for
as the convolutional neural network (CNN) and recurrent intrusion detection.
neural network, infer the mapping function from labeled (c) To address the class imbalance problem, the proposed
training data. On other hand, the unsupervised deep learn- model is trained only with the normal samples in an
ing networks (UDLN), such as the deep belief network unsupervised manner to improve the generalization
(DBN) and deep autoencoder (AE), aim to find the hid- ability of the proposed model.
den structure of unlabeled data. In recent years, SDLNs (d) The ablation experiments and comparative analysis
have proven their potential strides in enhancing the intel- on benchmark intrusion datasets demonstrate the
ligence of IDS to detect the rising cyberattacks. However, effectiveness of the proposed model against the state-
all these recent successes of SDLNs heavily depend on the of-the-art methods.
availability of a sufficient quantity of accurately labeled
training data. In a large-scale environment, labeling network
traffic data will become tedious over time and may some- 2 Literature review
times require domain knowledge from experts. Naturally,
this may lead to error-prone data labels, especially when The recent literature on cybersecurity reveals how the
the network enters a malicious state. Intuitively, unsuper- advancements in AI have led the intrusion detection
vised deep learning networks are gaining a resurgence of problem to a more challenging level of study. At the
interest and have become the current research hotspot in the same time, the corresponding computational solutions have
field of intelligent intrusion detection with vital practical also improved its level of efficiency. In particular, deep
importance [11]. learning has shown state-of-the-art results in a multitude
By virtue of this, there is recently a considerable number of applications in the field of network security. This has
of works on the application of UDLN to the field of prompted an increasing number of researchers to turn their
intrusion detection. Not withstanding that existing works attention to utilizing deep learning techniques to improve
are encouraging, some challenges remain in practically the detection accuracy of IDS.
applying these UDLNs for intrusion detection. First, in a Correspondingly, a series of deep learning approaches
real network scenario, considering the system uncertainty have been proposed in the recent literature. Neverthe-
and network topology complexity, it is evidently very less, most of the reported state-of-the-art deep learning
demanding to collect abnormal traffic data in large approaches adopt supervised models that have an insatiable
sizes. Under such circumstances, the UDLN trained with appetite for labeled data that severely hinders their applica-
insufficient abnormal traffic data will fail to learn more bility in real network environments. To this end, the unsu-
generalized features about attack instances and may face pervised deep learning models that are strong in extracting

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

7096 A. Binbusayyis and T. Vaiyapuri

and representing the hidden essential features from unla- feature extraction and an SVM classifier for interpreting the
beled data have gained new inspiration and substantial encoded features to identify intrusion. The model efficiency
traction in recent years in the field of network security. in binary and multiclass classification was investigated
For example, Alom et al [12] proposed to employ DBN and compared with the state-of-the-art shallow machine
for feature learning and enhance the accuracy of intrusion learning models. The results showed the model capability
detection. Similarly, a study by Kang et al [13] presented an in improving the training and testing time of the SVM
NIDS for the in-vehicular network security, leveraging the with an improved accuracy rate compared to the previous
benefits of the unsupervised pretraining process for DBN. approaches. Moreover, the authors in [23] applied a sparse
The results of the study revealed the model efficiency in AE model exploiting the concepts of self-taught learning
detecting the intrusion activities. Additionally, a consider- to learn useful features for intrusion detection. In addition,
able number of studies have investigated the application of they combined the original features with extracted features
to improve the model generalization ability in recognizing
DBN in IDS design for feature extraction and have demon-
the network attacks. Furthermore, in [24], a two-stage
strated its effects on the performance of intrusion detection
framework that combines a sparse AE with long short-
[14, 15]. Nonetheless, the recent past studies have primar-
term memory is investigated for building an efficient IDS.
ily focused on AE applications for developing efficient IDS
Here, the framework employs the sparse AE for learning
due to its simple implementation and attractive computation effective feature representation and the LSTM model for
cost. This paper will also focus on the application of autoen- classifying normal and malicious traffic. In a work by
coders since they are more related to the work presented here. Shuaixin.T [25], the viability of combining stacked AEs
Among them, a plethora of studies have attempted to and an SVM classifier configured with a piecewise radial
develop variants of AE with enhanced discriminative and basis function to improve the classification performance of
robust feature representation for intrusion detection. For the SVM for intrusion detection is examined. Similarly, the
instance, Hassan et al [16] contributed a variant of the authors in [26] combined the advantage of stacked AEs
sparse AE optimizing the hyperparameters to exhibit bet- with a CNN to considerably achieve the high-performance
ter capability in extracting useful features and classifying demand of network IDS. Likewise, the authors in [27] have
malicious attacks. Furthermore, the proposed reconstruction studied the effectiveness of a stacked sparse AE model
error function establishes a balance between feature repre- for extracting useful features of intrusion behavior. The
sentation and network regularization by applying a sparsity study results indicated that the model can extract the more
constraint in the output layer. Sheng et al [17] designed a discriminative features and accelerate the detection process.
new framework for IDS with a discriminative encoder and Relatedly, the authors in [28] proposed new interesting
generator that works as a generative adversarial network online deep learning systems that apply an AE as function
during the training process and as an AE during the testing approximation in the Q-network of RL to achieve a higher
process to reconstruct the test samples. The reconstruction detection accuracy rate for network intrusion detection.
error with added extra loss empowers AE encoding ability From the above literature review, it is evident that
to discriminate the malicious network activities. Aygun et despite the significant performance gain achieved with the
al [18] enhanced the AE discriminative ability utilizing a application of AEs in IDS design, there is still room for
improvement. The causes of weakness include the shortage
stochastically determined threshold for reconstruction error
of intrusion network traffic and the imbalance among the
to reach an improved accuracy compared to deterministic
normal and abnormal network traffic. On these grounds,
AE variants on NSL-KDD intrusion datasets. In the same
the existing approaches are prone to overfit and show poor
manner, the authors in [19] presented a statistical analysis to generalization performance toward unseen cyberattacks.
extract more optimized and correlated features to improve Thus, research on unsupervised deep learning approaches
the accuracy of AE. Similarly, an IDS is proposed in [20] for IDS is still in its infancy in terms of development.
employing an ensemble variant of AE to collectively dif- Hence, the proposed research is expected to make a valuable
ferentiate the abnormal traffic patterns from normal traffic contribution to the existing knowledge pool.
behavior. Moreover, Shone et al [21] recently introduced
a nonsymmetric variant of a deep AE for unsupervised
learning and achieved promising results 3 Background
Another set of previous studies has suggested that
the application of AE for extracting intrinsic features of 3.1 Autoencoder
network traffic data can improve the detection accuracy
and performance of the classifier model in IDS. A work The autoencoder (AE) was introduced by [29] as an
by M. Al-Qatf et al [22] developed a deep learning model unsupervised neural network to learn robust feature
based on a self-taught learning framework for intrusion representation by reconstructing the given input as network
detection. Here, the authors have used a sparse AE for output. The basic structure of an AE consists of one hidden

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Unsupervised deep learning approach... 7097

layer. Here, the process between the input layer and the achieve a high detection rate on the attack class. However,
hidden layer is called the encoder. The task of the encoder is since the OCC scenario does not have access to any a
to map the given input data vector X = (x1 , x2 , x3 , ....xn ) to priori information about attacks, the training procedure
a lower representation H = (h1 , h2 , h3 , ...hr ) at the hidden may suffer from complications in making decisions on
layer, which is regarded as high-level features of the input. how to fit the boundary around the data in all directions
This is formulated as follows [30]: without overfitting. Numerous solutions were proposed in
the literature to address this problem.
H = f (W X + b) (1)

Similarly, the process between the hidden layer and the 4 Methodology
output layer is called the decoder. In this process, the
decoder aims to reconstruct the input data vector X from the Figure 1 illustrates the proposed unsupervised deep learning
lower hidden representation H, formulated as follows, approach for IDS. As shown in the figure, the proposed
approach includes two essential components viz., AE for
Z = g(W  H + b ) (2) normal traffic feature representation learning and a one-
class classifier for intrusion detection. The two subsections
In the above equations, f and g are nonlinear activation that follow elaborate the technical details of these two
functions such as sigmoid, tanh, and the rectified linear components. Subsequently, the designed unified objective
unit (ReLU) function. The W and b represent the weight function to achieve joint optimization in the proposed
and bias vector of the encoder, respectively. Likewise, W  approach is presented.
and b represent weight and bias vector of the decoder,
respectively. These parameters of AE denoted as θ = 4.1 Autoencoder for feature representation
{W, W  , b, b } are optimized during the training process by
minimizing the reconstruction error defined by either using This study adopts the convolutional AE (CAE) proposed
an L1 or L2 loss function. by Masci et al [33] as an unsupervised feature representa-
tion learning method. It essentially combines the advantage
3.2 One-class classifier of a CNN and an AE to attain strong feature represen-
tation. Compared to other variants of the AE, the CAE
The one-class classifier (OCC) is a promising area of accomplishes strong feature representation considering the
machine learning in which extensive research has been relationships among the features that are more appropriate
devoted to anomaly detection [31]. OCC aims merely at for intrusion detection. Furthermore, CAE enables weight
discriminating a class of interest from all other classes [32]. sharing among the inputs and ensures to preserve the spa-
This class is labeled as normal. Whereas, all other classes tial locality of the features. By doing so, the number of
that deviate from normal are termed as attacks. Although parameters to be trained is reduced. This in turn reduces the
OCC seems to resemble binary classification, a significant memory requirement and computational efficiency of the
difference lies in its training process, which is based on CAE. Thus, the CAE is regarded as special type of AE with
the assumption that only normal samples are available for a convolutional layer rather than a fully connected layer for
training. Accordingly, OCC learns to derive a decision the encoding process and a deconvolutional layer for the
boundary only around normal samples as accurately as decoding process.
possible such that it encloses all normal samples while Taking inspiration from [34], this work adopts a 1D
minimizing the probability of accepting attack samples. CAE with a hypothesis that the application of a 1D
Since the OCC training process considers only normal CAE will enable achieving further higher efficiency with
samples, it is also known as learning in the absence of the sequential form of network traffic data compared to
counterexamples. a 2D CAE. Accordingly, the encoding process for the
After the training phase, at prediction time, the OCC convolutional layer with feature filters maps 1D network
uses the decision boundary to determine if the newly traffic data X to produce a hidden representation with the
arriving sample belongs to the normal class or not. The k th feature map and is represented as Hk = f (X ∗ Wk + bk ).
new sample is classified as normal if it falls within the Here, ∗ denotes the 1D convolution operation over the input
boundary. Conversely, if the new samples fall outside the vector X.
boundary, they are rejected and treated as the attack class. Moreover, to learn high-level network traffic features,
An ideal OCC algorithm should not overfit based on the a deep learning architecture of the 1D CAE is designed
provided training samples; rather, it should generalize from employing a series of convolutional layers. Under such a
training samples and gain a good discrimination ability to scenario, the hidden feature representation is obtained by

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

7098 A. Binbusayyis and T. Vaiyapuri

Fig. 1 Illustration of the

proposed intrusion detection
system architecture

recursive formulation, expressed as follows, in Section 3.2 is well suited to resolve the problem using
  only normal traffic data without the requirement of intrusion
H l = f H l−1 ∗ W l + bl (3) traffic samples. In this direction, one-class support vector
where l ∈ {1, 2, 3, ..., L} if L convolutional layers are machine (OCSVM), an improved version of traditional
employed and H 0 represents the original input vector X. SVM is extensively used in anomaly detection and has
As a result, H L represents the final low-dimensional hidden demonstrated promising results with imbalanced data [35,
representation of X. 36]. Taking inspiration from these literature, this work
Similarly, the designed 1D CAE employs a series adopts OCSVM to address the network traffic imbalance
of deconvolutional layers in the decoding process to and make the feature space more discriminative for intrusion
reconstruct the original input network traffic data from H L . detection. The basic idea of this approach is to deem
This can be expressed with recursive formulation as follows, all attack samples to lie on the origin and to use the
  advantages offered by the SVM to map the given training
Z l = g Z l−1 ∗ W (L−l+1)T + bl (4) samples to a new feature space where they become linearly
separable. Intuitively, they reduce the problem of fitting a
Thus, in the above equation, Z 1 represents the input of the nonlinear boundary around the data to a linear boundary
first decoding layer which is H L , the output of the last or a hyperplane in the new feature space that separates all
encoding layer. As result, the output of the last decoding training samples from the origin with the maximum possible
layer Z L represents the reconstruction of the original input margin. The hyperplane is represented as follows for a given
vector X. input vector X,
Moreover, in this work, the 1D CAE is designed to
measure the mean squared error (MSE) as the reconstruction f (x) = ω.ϕ(x) − ρ (6)
error between the original input network traffic vector X and Here, ω represents the weight coefficients, ρ is the distance
the reconstructed network traffic vector Z. Therefore, the from the origin to the hyperplane, and ϕ(.) is a feature map
objective function of the designed 1D CAE is formulated as obtained by applying certain kernel functions. The kernel
follows, function on any two samples from the input vector is defined
1 as follows,
Lr (θ ) = min X − Z2
θ 2N
K(xi , xj ) = ϕ(xi ).ϕ(xj ). (7)
1
= min X − g(f (X))2 (5) Here, the problem of finding an optimal hyperplane can be
θ 2N
formulated as the following optimization problem,
4.2 One-class classifier for attack detection 
N
2 ω −ρ+
1 2 1
minω,ρ,ξ νN ξi
i=1 (8)
As discussed in introduction section, collection of intrusion
s.t ω.ϕ(x) − ρ + ξi ≥ 0, ξi ≥ 0, ∀i
network traffic data in real practice is a major issue.
Therefore, most of the available training dataset are Here, N represents the number of training samples, ν ∈
imbalanced with a small amount of intrusion traffic data. (0, 1] represents the regularization term that controls the
Under such circumstances, the design of OCC discussed fraction of outliers in the training set and ξi indicates the

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Unsupervised deep learning approach... 7099

slack variable used to model the classification error with basis (RBF). Here, the RBF that usually leads to better
respect to the i th sample. Solving the above optimization performance is chosen. The RBF is defined as follows,
problem using the Lagrange multiplier method, the decision
function can be represented as K(x1 , x2 ) = exp(−γ .x1 − x2 2 ) (13)
N 
 Here, the hyperparameter of the kernel function has to be
f (x) = sign αi K(xi , x) − ρ , tuned such that K(hL L
i , hj ) always produce small values
i=1
 when both hL L
i and hi are not normal traffic samples. On
1, if x belongs to the target class other hand, the larger or very much smaller values of the
= . (9)
0, otherwise kernel parameter may lead to underfitting or overfitting,
respectively.
In the above equation, αi is obtained by solving its dual form
Similarly, setting the hyperparameter ν in (12) to 0 will
as follows,
make the last term vanish, meaning that the OCSVM will

N be forced to find a hyperplane that separates all training
maxα − 12 αi αj K(xi , xj )
i,j =1
samples as normal from the origin. Conversely, if ν is
(10)

N set to 1, then the OCSVM will tend to find a hyperplane
s.t αi = 1, 0 ≤ αi ≤ 1
νN , ∀i that rejects all training samples as attacks. Intuitively, the
i=1
hyperparameter ν has to be tuned to avoid the distortion
Having solved αi , ρ can be calculated as follows by of the hyperplane (decision boundary) by outlier data and
selecting any sample from the training set that satisfies improve the model generalization ability for unseen data.
0 ≤ αj ≤ νN1
and that the sample is a support vector. Taking into consideration, the sensitivity to false positive

N rate and generalization ability of OCSVM, this work
ρ= αj K(xi , xj ) (11) focuses to tune the hyperparameters ν and γ with regard
j =1 to the most robust discriminative features learnt by CAE.
This is achieved by designing a new objective function
Thus, from the decision function of OCSVM, it is evident
that enables to integrate 1D CAE for feature representation
that OCSVM can effectively detect malicious activities
learning and OCSVM for classification within a joint
solely with the knowledge of normal network traffic
optimization framework. The subsection following briefs
samples with an optimal hyperplane. However, the problem
how the hyperparameters are tuned during model training
of finding an optimal hyperplane, the decision boundary of
process in an unsupervised manner defining a new objective
the OCSVM, is a pivotal issue and is strongly influenced by
function.
hyperparameter selection [37].
To address this issue and smoothly integrate the OCSVM
4.3 Designed objective function
within the proposed IDS architecture, this work applies the
hinge loss function, H inges (z) = max(0, s −z) [38], where
All the existing AE IDS models perform feature represen-
s represents the position of the hinge point to penalize the
tation learning and classification by learning independently
samples classified with an insufficient margin. Accordingly,
through pretraining and the fine-tuning process without
the optimization problem of the OCSVM given in (8)
joint optimization [6]. In that case, the learned features do
is transformed as follows by introducing the hinge loss
not guarantee strong discriminative ability for the intru-
function on the feature representation hLi learned from the sion detection task. To circumvent this problem, our work
original training samples [39] :
intends to combine the reconstruction loss term in (5) with
the structural and empirical risk term of the classifier given
1 
N
1
Lc(ν, γ ) = min ω2 −ρ+ max(0, ρ−ω.ϕ(hL
i )) in (12) and define the following objective function
ω,ρ,ν 2 νN
i=1
(12) L = L r + Lc (14)

The first and second terms in the above equation represent This objective function clearly reveals that our work aims
the structural risk loss function, and the last term denotes to guide the proposed approach to learn strong feature
the empirical risk based on the hinge loss function. representation for an improved effective intrusion detection
Furthermore, in the above equation, the mapping ϕ(.) is by integrating the feature representation and classification
usually implicit and indefinite. Therefore, the inner product process into a joint optimization framework. In doing so, the
of the mapped data is generally specified by the kernel proposed approach reduces the reconstruction loss and at the
function in practice. The most commonly used kernel same time ensures that the classification hyperplane margin
functions include linear, sigmoid, polynomial and radial is maximized for improving the detection accuracy of the

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

7100 A. Binbusayyis and T. Vaiyapuri

proposed approach. Algorithm 1 summarizes the working regarded as one of the most valuable benchmark resources
procedure of the proposed approach. in the field cybersecurity research for IDS evaluation. Each
sample in NSL-KDD contains 41 features and 1 class label
to characterize whether the network traffic is normal or
belongs to the attack category. The distributions of normal
traffic samples in the training and testing sets with regard to
attacks are given in Table 1.

5.1.2 UNSW-NB15 dataset

The UNSW-NB15 is a modernized dataset recently devel-

oped by ACCS with a hybrid representation of real normal
and synthesized contemporary attack behaviors from net-
work traffic flow [41]. This dataset includes 9 families of
attacks, namely, DoS, Analysis, Generic, Fuzzers, Back-
doors, Exploits, Shellcode, Reconnaissance, and Worms.
The dataset consists of 175,341 training samples and 82,332
testing samples, each characterized with 42 features and a
class label to discriminate the network traffic as normal or
malicious activities. The distributions of samples against
normal and attack classes is shown in Table 2.

5.2 Data preprocessing

Data preprocessing is essentially crucial for providing

quality input for model training and for boosting the
5 Experimental setup detection ability of the IDS. It includes two main operations,
namely, data encoding and normalization.
This section first describes the experimental datasets
(a) Data Encoding: In this work, the label encoding
and then details the methods used for preprocessing the
method is used to map all nonnumeric or nominal
datasets. Subsequently, the structure of the CAE network
features to numeric values. This method maps a
is described, followed by the training details. Finally,
nominal feature with C different values to an integer
the implementation details and the metrics used for
in the range of 0 to C-1. For example, the NSL-
experimental evaluation are presented.
KDD dataset includes three nominal features, namely,
protocol type, service type, and TCP status flag with
5.1 Datasets
3, 70 and 11 distinct nominal values, respectively.
After label encoding, the feature protocol type with
A number of datasets are available publicly for IDS
three values is mapped as follows: tcp:0, udp:1 and
research evaluation. Nonetheless, these datasets suffer from
icmp:2.
absences of traffic diversity and lack a sufficient number of
(b) Normalization: Generally, the machine learning
sophisticated attack styles. Therefore, in order to conduct
algorithms are biased by input features with large
a fair and effective evaluation of the proposed model, an
numeric values. To combat this effect, min-max
old benchmark NSL-KDD dataset and a new contemporary
normalization is applied to adjust the value range of all
UNSW-NB15 dataset are considered in this work. A brief
input features within the range [0,1].
description of these two intrusion datasets is given below,

5.1.1 NSL-KDD dataset Table 1 Data distributions in NSL-KDD

The NSL-KDD dataset is an improved version of the Class Training set Testing set
KDD’99 dataset, presented by Tavallaee et in 2009, that
Normal 67,343 9,710
resolves the redundancy in the KDD ’99 dataset [40].
Attack 58,630 12,833
This dataset contains an optimal ratio of 125,973 training
Total 125,973 22,543
samples to 22,543 testing samples. Thus, NSL-KDD is

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Unsupervised deep learning approach... 7101

Table 2 Data distributions in UNSW-NB15 for regularization with a rate of 0.2 after each max pooling
layer and as the last layer of the encoder. A mirrored encoder
Class Training set Testing set
structure with deconvolution and upsampling operations is
Normal 56,000 37,000 employed on the decoder side to reconstruct the original
Attack 119,341 45,332 network traffic input of dimension 1 × 41 from 10 low-
Total 175,341 82,332 dimensional high-level features resulting from the encoder.

5.4 Training process

Fundamentally, the training process plays a key role in

5.3 CAE model configuration optimizing the hyperparameters of a model and directly
influences the performance of the model. Therefore, in
Notably, the structure of the deep learning CAE network has this work, model training was conducted with utmost
a great impact on the classification performance. Therefore, care and meticulous planning to learn the robust feature
in this work, the configuration adopted for 1D CAE was representation that can ensure a strong reconstruction ability
determined by conducting a range of experiments with for the input network traffic and a discriminative ability
various structural combinations. Figure 2 illustrates the of the OCSVM. Accordingly, the training process uses
structure used for the 1D CAE network. It consists of the minibatch gradient descent optimization algorithm to
three convolutional layers with an ReLU as the activation minimize the reconstruction loss and classifier loss jointly.
function on the encoder side. The number of filters in In this approach, the loss computed over a batch of training
these layers is reduced by half from low-level to high-level samples is used in backpropagation to balance the tradeoff
convolutional layers. For example, the first convolutional between the robustness and efficiency of the proposed
layer consists of 32 filters, the second layer consists of deep network architecture. Moreover, the Adam update
16 layers and last with 8. Nonetheless, the size of kernels rule [42] with a minibatch size of 32 and a learning
is kept the same for all convolution layers and is set to rate of 0.001 is adopted considering its fast convergence
1 × 3. This pyramid architecture not only reduces the rate and fewer memory requirements for computing the
number of trainable parameters but also enables learning gradients of trainable parameters and to achieve an optimal
the most essential features from the input network traffic by network architecture. To obtain stabilized results, the
eliminating the redundant and irrelevant features. training process is terminated when the number of epochs
A max pooling layer with a pool size of 2 is added exceeds 15 or the loss value of the deep network falls
after each convolutional layer except the last to extract the below the threshold value of 0.005. Notably, to keep
most essential features from the network traffic data and the backpropagated gradient values and activation values
improve the representation capability of the model. With within a reasonable range, all the trainable parameters of
two max pooling layers, the dimension of the input network the CAE are initialized using the Xavier algorithm [43].
traffic flow is reduced to 10. Furthermore, all convolutional Additionally, the grid search (GS) algorithm is adopted
layers are followed by a batch normalization layer to to initialize the OCSVM parameter ν over a subset of
stabilize the training process and accelerate the network values {0.1, 0.05, 0.07} on the given training samples. For
convergence. In addition, to avoid model overfitting and initializing the parameter γ , GS works on given training
improve computational efficiency, a dropout layer is added samples and the user provided subset of three values

Fig. 2 Structure of the developed Convolutional Autoencoder Network

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

7102 A. Binbusayyis and T. Vaiyapuri

{a, b, c}. However, later during each iteration of the training (c) F1-measure (F1) : Also termed the F1-Score, is
process, the parameter γ is tuned for the learned feature considered a more effective measure than accuracy
representation using grid search with a new subset of four to evaluate the performance of an intrusion detection
values that are determined based on their initial optimal model, especially for imbalanced datasets. It is an
value. For example, if the outcome of GS is 50 for the user harmonic average of the detection rate and precision as
given subset of values: 50, 100, 150, then the computed new follows
subset will be as follows 10, 20, 30, 40.
Under the above settings, the proposed network architec- 2 × (DR × P recision)
ture is trained with normal samples in the training datasets. F1 = (17)
DR + P recision
During each iteration of the training process, as the normal
samples in the training dataset flow through the CAE net- Here, precision measures the proportions of
work, the essential features are extracted automatically with detected intrusion traffic that are actually correct. It is
a reduction in dimension to a size of 10 x 8 through three expressed as follows,
encoding layers. Then, the original input is reconstructed
TP
from the extracted features through three decoding layers. P recision = (18)
Next, the features extracted by the encoder part of the CAE T P + FP
are mapped to the kernel space using an RBF to train the (d) False alarm rate (FAR) : Also termed the false positive
OCSVM. At the end of each iteration, the reconstruction rate, measures the proportion of normal network traffic
loss and classifier loss are computed to update the trainable flows that are incorrectly classified. It is computed as
parameter and obtain an optimal network architecture. follows,
FP
5.5 Implementation details FAR = (19)
FP + T N
All the experiments are conducted on a personal computer
with the specifications as follows: an Intel Core i7-8565H
@ 1.8GHz with 128 GB RAM and the Windows 10
6 Experimental results and discussion
operating system. The proposed model is implemented in
This section describes the three sets of analyses designed
the Jupyter development environment using Python 3 as
to demonstrate the supremacy of the proposed approach. In
the programming language. More specifically, the Python
particular, these experiments aim to achieve the following
libraries, Keras and TensorFlow are used to implement
various deep learning tasks [44]. Additionally, the Python (a) Validate the design decision of the proposed approach
Scikit-learn library is used to implement various evaluation on the benchmark dataset NSL-KDD
measures and data preprocessing tasks. (b) Investigate whether the performance of the proposed
approach is stable with the recent intrusion dataset
5.6 Evaluation metrics (c) Compare the performance of the proposed approach
with recent state-of-the-art methods published in
The effectiveness of the proposed IDS approach is literature.
measured by analyzing four evaluation metrics that are most
commonly used in the field of intrusion detection. The 6.1 Ablation analysis
relevant definitions of these four metrics are as follows,
At the outset, two sets of ablation experiments are
(a) Accuracy (ACC): Measures the proportion of network
conducted to analyze the design decision of the proposed
traffic flows that are correctly classified and is
approach and CAE architecture correspondingly on the
computed as follows,
standard benchmark intrusion dataset, NSK-KDD, in terms
TP +TN of ACC, DR and FAR. The subsections below describe these
ACC = (15) experiments in detail.
T P + T N + FP + FN

(b) Detection rate (DR): Also called Recall or Sensitivity, 6.1.1 Experiment 1
measures the proportions of intrusion traffic flow that
are correctly classified as given below, As a first step, the design decision of the proposed approach
is analyzed investigating how different components in the
TP proposed approach can contribute jointly to the overall
DR = (16)
T P + FN performance. For this purpose, an analysis study is

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Unsupervised deep learning approach... 7103

the CAE is first trained to learn the essential feature

representation. Then, the features extracted by the
CAE are used to train the OCSVM for classification.
For a fair comparison, the ablation experiments are
conducted using the same parameters and environmental
setup as the proposed model, and the results are reported in
Table 3. Observation of these ablation results demonstrates
the significance and relevance of all components in
the proposed approach against the achieved performance
benefits. Particularly, it can be seen that the baseline variant
OC induces a high FAR. This illustrates the significant role
of the CAE in learning the most essential high-level features
from the normal network traffic flow to deliver an improved
performance in terms of DR, FAR and ACC.
Similarly, it is obvious from the experimental results of
the variant, CAE+softmax, that without the OC classifier,
the overall performance drops significantly. This clearly
reflects that the OC classifier is a more efficient component
that uses its kernel tricks to contribute to a compact
representation of the normal samples in the proposed
approach, thereby ensuring the overall improvement in the
intrusion detection task.
Furthermore, since timeliness is another critical metric
required in modern IDS, the training and testing time of
the proposed method is compared on NSL-KDD datasets.
Fig. 3 Structure of the CAE+Softmax variant To establish a fair comparison, all computation time
are calculated under the same operating environment as
discussed in Section 5.5. The results of this comparison are
conducted developing three variants of proposed approach
shown in Table 3. From these results, it can be seen that
as follows,
the proposed approach takes longer training time of 1123s
(a) OC: This variant is created by removing the CAE than other ablations to discover the optimal hyperplane that
component to evaluate the effectiveness of the can effectively compact the features representation learnt by
proposed approach for feature representation learning. CAE. This might be due to the joint framework adopted to
(b) CAE+softmax: This variant is created by replacing train both CAE and OCSVM simultaneously to minimize
the decoder of the trained CAE and OCSVM the reconstruction and classifier loss. Nevertheless, it is easy
components with a softmax layer as shown in Fig. 3 to to observe that the proposed method ranks second among
evaluate the effectiveness of the proposed approach for all the ablations and takes 0.35ms to detect an intrusion
one-class unsupervised classification. traffic instance during the testing period. In turn, this
(c) CAE+OC: This version indeed is developed to implies that in comparison to offline training time, the test
evaluate the effectiveness of the joint optimization time is more crucial element for an IDS and the proposed
framework with the CAE and OCSVM. To this end, method is well-designed to reduce the detection time while

Table 3 Ablation analysis results on NSL-KDD dataset for different variants of proposed approach

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

7104 A. Binbusayyis and T. Vaiyapuri

maintaining the detection accuracy with optimal number

of batch normalization, maxpooling and dropout layer to
demonstrate comparably better detection time efficiency.
Overall, it can be observed that the proposed approach
excels in all three variants in terms of all three evaluation
metrics, which confirms the contribution of the joint
optimization framework to the success of the proposed
approach. The outcome is attributed to the ability of
the approach to effectively leverage the benefits of the
CAE and OC classifier for representation learning and the
one-class classification, respectively, through simultaneous
optimization of reconstruction loss and classifier loss.
Fig. 4 Illustration of the reconstructed network traffic data by the CAE
6.1.2 Experiement 2 component when the hidden dimension is 10

This set of experiments aims to study the relations between half for each successive convolutional layer to better extract
intrusion detection performance and other factors, such as the essential features through the convolution process.
the hidden state dimension and number of filters, exploring Table 5 summarizing the corresponding results evidently
different network design for CAE. indicates the improvement in overall performance with the
Successively, a range of experiments were conducted increasing number of filters, which is also in accordance to
varying the final hidden state dimension to 20, 10 and the claim in the literature. Nonetheless, it can be observed
5 but keeping other factors fixed. The results reported in that this increase in performance is not significant in
Table 4 clearly demonstrate the impact of the hidden state contrast to the increase in the computational complexity
dimension on the performance of the proposed approach. when the number of filters in the first convolution layer is 64
Although the hidden state dimension at 20 delivered higher and is reduced by half successively. Therefore, the number
DR and ACC values on the training set than its counterparts, of filters for the first convolution layer in the CAE was
it fails to maintain a higher performance on the testing chosen as 32 and was reduced by half for each successive
set. On the other hand, the hidden state dimension of 10 convolutional layer in our subsequent experiments.
displays higher performance on both training and testing
sets. In addition, Fig. 4 illustrates the ability of CAE to 6.2 Performance analysis
learn the robust feature representation that can reconstruct
the original input with a small variation when the hidden In the literature, it is stated that the change of datasets
state dimension is 10. Therefore, in further experiments, the considerably affects and varies the performance of the
CAE network of the proposed model was designed with the detection process [53]. Accordingly, to investigate the stable
feature representation of dimension 10. performance of the proposed approach on different datasets,
Next, considering the claim in the previous literature this experiment is conducted choosing a most recent
[45] that the number of filters has a great impact on benchmark dataset, UNSW-NB15, that includes many new
model performance, a series of ablation experiments were modern attack styles.
conducted to determine the optimal number of filters for the The confusion matrix delivered by the proposed approach
CAE component in the proposed approach. In this analysis, on UNSW-NB15 training and testing datasets is shown
the number of filters in the first convolutional layer was in Fig. 5. The evaluation metrics computed using these
varied among 64, 32, 16 and 8. However, due to successive confusion matrices are presented in Fig. 6. The figure
feature downsampling, the number of filters is reduced by demonstrates that the proposed approach is very effective

Table 4 Ablation analysis results on NSL-KDD dataset for different hidden state dimensions in the CAE component of proposed approach

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Unsupervised deep learning approach... 7105

Table 5 Ablation analysis results on NSL-KDD dataset for different number of filters in the CAE component of proposed approach

in achieving a DR of 97.7, FAR of 5.55, ACC of 96.7 and provided in their published papers are used to maintain a fair
F1 of 97.57 on the training dataset. Comparably, a DR of comparison and the results of this comparison are presented
97.8, FAR of 1.8, ACC of 97.6 and F1 of 98.1 on the testing in Table 6. Here, for clarity purposes, the highest score is
dataset clearly reveals the efficacy of the proposed approach highlighted in bold for each metric on both NSL-KDD and
to generalize even a complex dataset such as UNSW-NB15, UNSW-NB15 datasets.
and at the same time, the findings confirm that the proposed Now, observing the results on NSL-KDD, it can be
approach is very competitive for modern attack detection. realized that the proposed approach outperforms all the
It is can be noted that, similar to the results on recent IDS approaches for all metrics except for the model
NSL-KDD, the performance improvement of the proposed introduced in [52] with the few-shot supervised learning
approach on the UNSW-NB15 dataset also remains at approach (FSL-IDS) in terms of accuracy. However, while
a promising level. This consistent performance of the the FSL-IDS model shows slightly higher accuracy than
proposed approach is evidently attributed to the joint the proposed approach, its probability for FAR is the worst
optimization of feature representation and classification at 7.21%. This indicates that the proposed approach is
learning for the intrusion detection task. competitively effective in generating a lower FAR rate
than all other recent approaches when applied to intrusion
6.3 Comparative analysis detection. As FAR is one of the extremely important metrics
that should be kept low for an ideal IDS model, it is evident
The effectiveness of the proposed approach is further that the proposed approach is well designed to exhibit the
highlighted by a comparison with recent and relevant best performance benefits with regard to all metrics.
deep learning approaches from the literature of intrusion Similarly, comparing the results on UNSW-NB15 pre-
detection. Since it is impractical to compare all latest sented in Table 6, it can be seen that proposed approach dis-
approaches, only those approaches that have used both plays very competitive results compared to all other recent
NSL-KDD and UNSW-NB15 datasets are considered to IDS approaches under study. Nevertheless, one exception
provide a meaningful comparison. Additionally, the results is observed with the multilayered echo state machine (ML-
ESM) model proposed in [51]. Comparing the performance
of ML-ESM on NSL-KDD and UNSW-NB15, it can be
observed that ML-ESM presents the highest performance
on the UNSW-NB15 dataset but fails to deliver consistent

Fig. 5 Confusion matrix of the proposed approach on UNSW-NB15. Fig. 6 Performance analysis of the proposed approach on UNSW-
a Training Set. b Testing Set NB15

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

7106 A. Binbusayyis and T. Vaiyapuri

Table 6 Comparative analysis of proposed approach against recent IDS approaches

Recent IDS approaches NSL-KDD dataset UNSW-NB-15 dataset

DR FAR ACC F1 DR FAR ACC F1

ICVAE-DNN [46] 77.43 2.74 85.97 86.27 95.68 19.01 89.08 90.61
SAVAER [47] 95.98 4.70 89.36 90.08 91.94 5.67 93.01 93.54
DNN with one Layer [48] 96.90 NA∗ 80.1 80.7 72.5 NA 78.4 82.0
MDPCA +DBN [49] 70.51 2.62 82.08 81.75 96.22 17.15 90.21 91.54
Hybrid ML [50] 86.8 11.7 85.79 NA 91.3 8.9 91.27 NA
Multilayer ESM [51] 83.0 3.3 NA NA 98 5.10 NA NA
Few shot Learning [52] 92.25 7.21 92.34 92.26 NA 8.01 92.11 NA
Proposed CAE+OCSVM 97.11 2.43 91.58 92.87 96.49 5.51 94.28 95.06

* denotes that the corresponding metric is not available/provided in the published paper

results on the NSL-KDD dataset. This reveals its setback UNSW-NB15, in terms of DR, FAR, ACC and F1. The
in stability with regard to different datasets. Moreover, this comprehensive ablation analysis results not only confirm
confirms the superiority of proposed approach against other the design decision rationale of the proposed approach but
recent approaches regardless of data distributions. The rea- also demonstrate that the approach can show considerable
son is possibly attributed to the introduced joint optimiza- performance improvement for effective intrusion detection.
tion framework that enables the CAE to generate a feature Furthermore, the experimental results on UNSW-NB15
representation with the potential ability for not only recon- have proven the potential efficacy of the proposed approach
struction but also for enhancing the classifier discriminative with conformance to our initial discussion, namely, that
ability for the intrusion detection task simultaneous optimization of feature representation and
In summary, it can be concluded that the superior classifier learning in an unsupervised manner serves as an
performance of the proposed approach demonstrates that effective approach in detecting unseen modern attack styles.
it has great potential to be a used as promising tool for The comparative analysis and discussion also evidently
intrusion detection. signify the advantage of the joint optimization framework
on the generalization ability of the proposed approach
and indicate that the proposed approach is a competitive
7 Conclusion candidate for intrusion detection among the latest state-of-
the-art IDS approaches. In conclusion, it is anticipated that
In this research, a novel unsupervised approach for intrusion the proposed approach will serve as a future benchmark
detection combining the benefits of deep learning and an for building a promising tool to safeguard the network
OC classifier is introduced and discussed. To the best environment against intrusion detection.
of our knowledge, the proposed approach is the first
attempt to integrate the 1D CAE and OCSVM using a
joint optimization framework. The novelty of the proposed
References
approach is twofold, as briefly described below.
(a) It bridges the gap between the feature representation 1. Kagermann H (2015) Change through digitization—value creation
and classifier learning that exists in traditional IDS in the age of industry 4.0. In: Management of permanent change.
Springer, pp 23–45
approaches by combining both reconstruction loss and 2. Kamasa J (2020) Securing future 5g-networks. Policy Perspectives
classification loss into a unified objective function. 8:4
(b) In contrast to traditional IDS approaches, the proposed 3. Bartock M, Cichonski J, Souppaya M (2020) 5g cybersecurity:
approach simultaneously learns the robust feature preparing a secure evolution to 5g. Technical report, National
Institute of Standards and Technology
representation from network traffic data and optimizes
4. Binbusayyis A, Vaiyapuri T (2019) Identifying and benchmarking
the OC classifier competitively to gain superior key features for cyber intrusion detection: an ensemble approach.
detection accuracy for intrusion. IEEE Access 7:106495–106513
5. Benmessahel I, Xie K, Chellal M (2018) A new evolutionary
The effectiveness of the proposed approach is evaluated neural networks based on intrusion detection systems using
on two benchmark intrusion datasets, NSL-KDD and multiverse optimization. Appl Intell 48(8):2315–2327

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Unsupervised deep learning approach... 7107

6. Aldweesh A, Derhab A, Emam AZ (2020) Deep learning network intrusion detection. IEEJ Trans Electron Inform Syst
approaches for anomaly-based intrusion detection systems: a 140(6):592–599
survey, taxonomy, and open issues. Knowl-Based Syst 189: 25. Shuaixin T (2020) An intrusion detection method based on stacked
105124 autoencoder and support vector machine. In: J phys conf series,
7. Binbusayyis A, Vaiyapuri T (2020) Comprehensive analysis and vol 1453, pp 1–17
recommendation of feature evaluation measures for intrusion 26. Yu Y, Long J, Cai Z (2017) Network intrusion detection through
detection. Heliyon 6(7):e04262 stacking dilated convolutional autoencoders. Secur Commun
8. Truong TC, Zelinka I, Plucar J, Čandı́k M, Šulc V (2020) Netw, 2017
Artificial intelligence and cybersecurity: past, presence, and 27. Yan B, Han G (2018) Effective feature extraction via stacked
future. In: Artificial intelligence and evolutionary computations in sparse autoencoder to improve intrusion detection system. IEEE
engineering systems. Springer, pp 351–363 Access 6:41238–41248
9. Kaja N, Shaout A, Ma D (2019) An intelligent intrusion detection 28. Kim C, Park JS (2019) Designing online network intrusion
system. Appl Intell 49(9):3235–3247 detection using deep auto-encoder q-learning. Comput Electric
10. Maza S, Touahria M (2019) Feature selection for intrusion Eng 79:106460
detection using new multi-objective estimation of distribution 29. Rumelhart DE, Hinton GE, Williams RJ (1988) Learning
algorithms. Appl Intell 49(12):4237–4257 representations by back-propagating errors neurocomputing:
11. Aleesa AM, Zaidan BB, Zaidan AA, Sahar NM (2020) Review foundations of research
of intrusion detection systems based on deep learning techniques: 30. Wang Y, Yao H, Zhao S (2016) Auto-encoder based dimensional-
coherent taxonomy, challenges, motivations, recommendations, ity reduction. Neurocomputing 184:232–242
substantial analysis and future directions. Neural Comput Appl 31. Tajoddin A, Abadi M (2019) Ramd: registry-based anomaly
32(14):9827–9858 malware detection using one-class ensemble classifiers. Appl
12. Alom MZ, Bontupalli VR, Taha TM (2015) Intrusion detec- Intell 49 (7):2641–2658
tion using deep belief networks. In: 2015 National aerospace 32. Khan SS, Madden MG (2014) One-class classification: taxonomy
and electronics conference (NAECON). IEEE, pp 339– of study and review of techniques. Knowl Eng Rev 29(3):
344 345–374
13. Kang M-J, Kang J-W (2016) Intrusion detection system using 33. Masci J, Meier U, Cireşan D, Schmidhuber J (2011) Stacked
deep neural network for in-vehicle network security. PloS One convolutional auto-encoders for hierarchical feature extraction. In:
11(6):e0155781 International conference on artificial neural networks. Springer,
14. Ni G, Gao L, Gao Q, Wang H (2014) An intrusion detection model pp 52–59
based on deep belief networks. In: 2014 second international 34. Chen S, Yu J, Wang S (2020) One-dimensional convolutional auto-
conference on advanced cloud and big data. IEEE, pp 247–252 encoder-based feature learning for fault diagnosis of multivariate
15. Zhang X, Chen J (2017) Deep learning based intelligent processes. J Process Control 87:54–67
intrusion detection. In: 2017 IEEE 9th international conference on 35. Tan FHS, Park JR, Jung K, Lee JS, Kang D-K (2020) Cascade of
communication software and networks (ICCSN). IEEE, pp 1133– one class classifiers for water level anomaly detection. Electronics
1137 9(6):1012
16. Musafer H, Abuzneid A, Faezipour M, Mahmood A (2020) 36. Tian Y, Mirzabagheri M, Tirandazi P, Mojtaba S, Bamakan
An enhanced design of sparse autoencoder for latent features H (2020) A non-convex semi-supervised approach to opinion
extraction based on trigonometric simplexes for network intrusion spam detection by ramp-one class svm. Inform Process Manag
detection systems. Electronics 9(2):259 57(6):102381
17. Mao S, Guo J, Li Z (2019) Discriminative autoencoding 37. Wang S, Liu Q, En Z, Porikli F, Yin J (2018) Hyperparameter
framework for simple and efficient anomaly detection. IEEE selection of one-class support vector machine by self-adaptive data
Access 7:140618–140630 shifting. Pattern Recognit 74:198–211
18. Can Aygun R, Gokhan Yavuz A (2017) Network anomaly 38. Xiao Y, Wang H, Xu W (2017) Ramp loss based robust one-class
detection with stochastically improved autoencoder based models. svm. Pattern Recogn Lett 85:15–20
In: 2017 IEEE 4th international conference on cyber security and 39. Platt JC, Shawe-Taylor J, Smola AJ, Williamson RC, et al.
cloud computing (CSCloud). IEEE, pp 193–198 (1999) Estimating the support of a high-dimensional distribution.
Technical Report MSR-t R-99–87 Microsoft Research (MSR)
19. Ieracitano C, Adeel A, Morabito FC, Hussain A (2020) A novel
40. Tavallaee M, Bagheri E, Lu W, Ghorbani AA (2009) A detailed
statistical analysis and autoencoder driven intelligent intrusion
analysis of the kdd cup 99 data set. In: 2009 IEEE symposium on
detection approach. Neurocomputing 387:51–62
computational intelligence for security and defense applications.
20. Mirsky Y, Doitshman T, Elovici Y, Shabtai A (2018) Kitsune: an
IEEE, pp 1–6
ensemble of autoencoders for online network intrusion detection. 41. Moustafa N, Slay J (2015) Unsw-nb15: a comprehensive data set
arXiv:1802.09089 for network intrusion detection systems (unsw-nb15 network data
21. Shone N, Ngoc TN, Vu DP, Qi S (2018) A deep learning approach set). In: 2015 military communications and information systems
to network intrusion detection. IEEE Trans Emerg Topics Comput conference (MilCIS). IEEE, pp 1–6
Intell 2(1):41–50 42. Kingma DP, Ba J (2014) Adam: a method for stochastic
22. Al-Qatf M, Yu L, Al-Habib M, Al-Sabahi K (2018) Deep learning optimization. arXiv:1412.6980
approach combining sparse autoencoder with svm for network 43. Glorot X, Bengio Y (2010) Understanding the difficulty of
intrusion detection. IEEE Access 6:52843–52856 training deep feedforward neural networks. In: Proceedings of the
23. Qureshi AS, Khan A, Shamim N, Durad MH (2019) Intrusion thirteenth international conference on artificial intelligence and
detection using deep sparse auto-encoder and self-taught learning. statistics, pp 249–256
Neural Comput Applic 32:1–13 44. Géron A (2019) Hands-on machine learning with Scikit-Learn,
24. Kherlenchimeg Z, Nakaya N (2020) A deep learning approach Keras, and tensorflow: Concepts, tools, and techniques to build
based on sparse autoencoder with long short-term memory for intelligent systems. O’Reilly Media

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

7108 A. Binbusayyis and T. Vaiyapuri

45. Agrawal A, Mittal N (2020) Using cnn for facial expression Adel Binbusayyis is currently
recognition: a study of the effects of kernel size and number of an Assistant Professor with the
filters on accuracy. Visual Comput 36(2):405–412 College of Engineering and
46. Yang Y, Zheng K, Wu C, Yang Y (2019) Improving the Computer Science, Prince Sat-
classification effectiveness of intrusion detection by using tam Bin Abdulaziz Univer-
improved conditional variational autoencoder and deep neural sity, where he is a special-
network. Sensors 19(11):2528 ist in cybersecurity and tech-
47. Yang Y, Zheng K, Wu B, Yang Y, Wang X (2020) Net- nology transfer. He is also
work intrusion detection based on supervised adversarial varia- the Vice-Dean of e-learning
tional auto-encoder with regularization. IEEE Access 8:42169– with the Deanship of Informa-
42184 tion Technology and Distance
48. Vinayakumar R, Alazab M, Soman KP, Poornachandran P, Al- Learning, Prince Sattam Bin
Nemrat A, Venkatraman S (2019) Deep learning approach for inte- Abdulaziz University. He is
lligent intrusion detection system. IEEE Access 7:41525– also an Advisor of Vice Rector
41550 with Prince Sattam Bin Abdu-
49. Yang Y, Zheng K, Wu C, Niu X, Yang Y (2019) Building an laziz University, where he is
effective intrusion detection system using the modified density responsible for monitoring the performance executions of the univer-
peak clustering algorithm and deep belief networks. Appl Sci sity strategic goals.
9(2):238
50. Tama BA, Comuzzi M, Rhee K-H (2019) Tse-ids: a two-
stage classifier ensemble for intelligent anomaly-based intrusion Thavavel Vaiyapuri is cur-
detection system. IEEE Access 7:94497–94507 rently an Assistant Professor
51. Tchakoucht TA, Ezziyyani M (2018) Multilayered echo-state with the College of Computer
machine: a novel architecture for efficient intrusion detection. Engineering and Sciences,
IEEE Access 6:72458–72468 Prince Sattam Bin Abdulaziz
52. Yu Y, Bian N (2020) An intrusion detection method using few- University. Her research inter-
shot learning. IEEE Access 8:49730–49740 ests include the fields of data
53. Fu A, Dong C, Wang L (2015) An experimental study on stability science, security, computer
and generalization of extreme learning machines. Int J Machine vision, and high-performance
Learn Cybern 6(1):129–135 computing. With nearly 20
years of research and teaching
experience, she has published
more than 50 research publi-
cations in impacted journals
and international conferences.
Publisher’s note Springer Nature remains neutral with regard to She is also a member of the
jurisdictional claims in published maps and institutional affiliations. IEEE Computer Society, and also a Fellow of HEA, U.K.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

onlineservice@[Link]