1
Review: A Deep Learning Approach to Network
Intrusion Detection
Mayra Macas, ID Student: 11721083
Abstract—One of the key challenges in artificial intelligence
(AI) is how to insert the informal knowledge into a computer.
Several AI projects have sought to hard-code knowledge about
the world in formal languages. Nevertheless, a major source
of difficulty in many real-world AI applications is that many
variation factors influence every single piece of data we are able
to observe. Most applications require us to disentangle the factors
of variation and discard the ones that we do not care about.
Of course, it can be very difficult to extract such high-level,
abstract features from raw data. When it is nearly as difficult
to obtain a representation as to solve the original problem,
representation learning does not, at first glance, seem to help us.
Deep learning (DL) solves this central problem in representation
learning by introducing representations that are expressed in
terms of other, simpler representations. Nowadays DL is gaining
much popularity and wide use in various computer science fields,
such as object recognition, speech recognition, computer vision
and so forth. Nevertheless, the application of DL in network
systems has just started to receive research attention. In this
study, we are concerned with the investigation of the various deep
learning techniques employed for network intrusion detection.
Index Terms—Machine learning, Deep Neural Network, Intru-
sion Detection Systems
I. I NTRODUCTION
Most organizations nowadays face a rapid evolution of
information and communication technologies (ICTs), convert-
ing them into modern organizations. However, the growing
dependence and reliance on these information systems (IT)
have also led to increased different types of cyber-terrorism
and cyber attackers that in many cases have been proven
difficult to be controlled and detected by IT technicians on
time. How to identify network attacks is a key problem that
should never been overlooked. As an important and active
security mechanism, Intrusion Detection has become the key
technology of network security. The objective of Intrusion
Detection Systems (IDS) is to identify unusual access or at-
tacks on internal network security. IDSs are usually hybrid and
combine anomaly detection and misuse detection modules [1].
The anomaly detection module classifies the abnormal network
traffic. The misuse detection module classifies attack patterns
with known signatures or extracts new signatures from the
attack labeled data coming from the anomaly module [2].
Nevertheless, IDS produces high false-positive rates in the
identification of novel attacks.
Within this context, the scientific community has studied
and designed models based on deep learning (DL). DL has
become practical because of the improvement in CPU and
M. Macas is with the Department of Computer Science, Zhejiang Univer-
sity, China,e-mail: (mayramacas11@zju.edu.cn).
neural network algorithms aspects [3]. The use of DL for
attack detection in the cyberspace could be a resilient mech-
anism to small mutations or novel attacks because of its
high-level feature extraction capability [4]. The self-taught
and compression capabilities of deep learning architectures
are key mechanisms for hidden pattern discovery from the
training data so that attacks are discriminated from benign
traffic [5] [6] [7]. Existing DL approaches utilized in intrusion
detection have been categorized into unsupervised deep learn-
ing methods and supervised deep learning methods. The differ-
ence between supervised and unsupervised methods is whether
training data are labeled [8]. Specifically, unsupervised deep
learning methods using unlabeled data include Auto-encoders
[4] [5] [9] [14], restricted Boltzmann machine (RBM) [15],
deep belief network (DBN) [10] and recurrent neural network
(RNN) [11]. The supervised deep learning methods include
convolutional neural network (CNN) which is combined with
multi-layer perceptron (MLP) in [13]. With this motivation,
the deep learning based approaches can help to overcome
the challenges of developing an efficient Intrusion Detection
System.
The remainder of this analysis is organized as follows. Sec-
tion 2 provides an overview the most popular DL algorithms
employed for network intrusion detection, while Section 3 con-
stitutes a literature survey concentrating on defining existing
methodologies with the dataset used for evaluation. In Section
4 we conclude the paper.
II. BACKGROUND
Deep Learning (DL) models are based on Artificial Neural
Networks (ANNs). These computational models are more
or less designed in analogy with biological brain modules.
Nowadays shallow ANNs are back in fashion under the term
DL. One of the milestones in this story occurred in 2006 when
Geoffrey Hinton’s lab was able to train efficiently a deep
network able to reconstruct high-dimensional input vectors
from a small central layer [16]. After this discovery by Hinton
et al. other deep networks using the same principle have
been introduced successfully in classification tasks [17]. Deep
networks IDS can be classified based on the way architectures
and techniques are being utilized. In this section we focus
in the most popular DL algorithms employed for network
intrusion detection. Figure 1 shows a classification of Deep
networks IDS
A. Supervised Networks
Supervised learning is the most common type of learning
not only in Deep Learning but also in Machine Learning

Fig. 1: Classification of deep learning IDS
in general. In this type of learning, the desired output of
the training data is known in advance, and it is supplied to
the model during training. Therefore, each training sample is
represented as a pair consisting of an input vector denoted
by x, and a desired output value denoted by y. The algorithm
should infer a function which can be used for mapping samples
with unknown outputs. Within a probabilistic framework, this
kind of models are also known as discriminative models
because they model the conditional probability distribution
P (y | x) which can be used for predicting y from x [19].
1) Deep Neural Networks (DNNs): An Artificial Neural
Network (ANN) is a generic term to encompass any structure
of interconnected neurons which sends information to each
other. Deep-learning networks are distinguished from the more
commonplace single-hidden-layer neural networks by their
depth; that is, the number of node layers through which data
passes in a multistep process of pattern recognition. Note
that a network with just one hidden layer cannot actually be
considered deep [18]. The main goal of using a neural net is
to arrive at the point of least error as fast as possible. In order
to achieve this during the implementation of a deep learning
concept for IDS, Roy et al. [12] used a multilayer feed-forward
network. The feed-forward network consisted of input layers,
about 400 hidden layer neurons, and output neurons. The
activation functions used were the rectifier activation function
and the softmax activation function. The authors proclaimed
that with the loss being set as Cross Entropy, this classification
model could be used to detect future intrusion attacks.
2) Deep Convolutional Neural Networks (DCNNs)): A con-
volutional 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. The
architecture of this type of network is similar to other networks
because they contain several hidden layers where each layer
applies an affine transformation to the input data followed by
a non-linearity. DCNNs leverage these three key ideas: local
receptive field, shared weights and pooling layers [19]. The
CNN has been found highly effective and been commonly
used in computer vision and image recognition. Therefore,
using CNNs for image conversion, feature learning in in-
trusion detection is practicable. Li et al. [13] used CNNs
for automatically learning the features of graphic NSL-KDD
transformation via a graphics conversion technique. Results
showed that the CNN model is sensitive to image conversion
of attack data and it can be used for intrusion detection.
2
B. Unsupervised Networks
Unlike the supervised models presented in Section 2.1,
this section is focused on other type of networks called
unsupervised models. These models aim to discover the hidden
structure of unlabeled data. From a probabilistic point of view,
this kind of learning is related to the problem of density
estimation, which deals with the estimation of the probability
distribution of the input data. One of the breakthroughs in the
development of DL techniques was the use of pre-training to
allow a more efficient training of deep networks [16]. The
idea is that each block of a deep network can be pre-trained
using an unsupervised model. Each block captures regularities
from its input distribution without requiring labeled data. This
process is done layer-wise, and the parameters learned in each
block serve as a good initialization of the weights of that block
in a deep neural network. This idea can also be regarded from
a probabilistic point of view. Consider random input-output
pairs (x, y) in a neural network. Learning a mapping function
between both of them involves modeling an approximation
of the probability distribution P (y | x) by maximizing its
likelihood. If the true P(x) and P (y | x) are related, learning
P (x) may facilitate the modeling of the real target distribution
P (y | x) [19].
1) Deep Boltzmann Machine (DBM): When DBM is
trained with a large supply of unlabeled data and fine-tuned
with labeled data, acts as a good classifier. 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. BM algorithm is simple to train but
turns out to be slow in the process. A reduction in the number
of hidden layers of a DBM to one forms a Restricted Boltz-
mann Machine (RBM). According to Salakhutdinov [20] Deep
Boltzmann machines are interesting for several reasons. First,
like deep belief networks, DBMs have the potential of learning
internal representations that become increasingly complex.
Second, high-level representations can be built from a large
supply of unlabeled sensory inputs and very limited labeled
data can then be used to only slightly fine-tune the model for
a specific task at hand. Finally, unlike deep belief networks,
the approximate inference procedure, in addition to an initial
bottom-up pass, can incorporate top-down feedback, allowing
deep Boltzmann machines to better propagate uncertainty,
and hence deal more robustly with ambiguous inputs. Fiore
et al. [15] used Restricted Boltzmann Machine in order to
implement semi-supervised anomaly detection systems where
the classifier was trained with normal traffic data only, so that
knowledge about anomalous behaviors was constructed and
evolved in a dynamic way. The results revealed that when the
classifier was tested in a network widely different from the
one where training data were taken from, the results decayed.
This suggests the need for further investigation over the nature
of anomalous traffic and the intrinsic differences with normal
traffic.
2) Deep Belief Networks (DBNs): A DBN is a deep gener-
ative model composed of a visible layer and multiple hidden
layers of latent variables. There are connections between the
layers but not between units within each layer. DBN could

be used as a feature extraction method for dimensionality
reduction. On the other hand, when associating class labels
with feature vectors, DBN is used for classification [19]. Gao
et al. [10] applied DBN in the intrusion detection domain. Ac-
cording to the authors, DBN not only learns high dimensional
representations but also performs classification tasks in an
efficient manner. An unsupervised greedy learning algorithm
can be used to pre-train and fine-tune a DBN in order to learn a
similarity representation over the nonlinear, high-dimensional
input data, something that largely facilitates the classification
tasks. Therefore, the output of the intrusion detection model
could be improved when using DBN.
3) Auto-Encoders: Auto-Encoders [19] are shallow neural
networks that are trained to reconstruct their input using an
intermediate representation (code). Auto-Encoders and RBMs
are very similar models because they share the same ar-
chitecture: an input layer that represents the data, a hidden
layer that represents the code to be learnt, and the weighted
connections between them. However, The Auto-Encoder is
trained by minimizing the reconstruction error, so an extra
layer is added on the top to represent a reconstruction of the
original data. Both sets of weights are tight, which means that
they are actually the same weights. These models also have the
capability of extracting good representations from unlabeled
data that work well to initialize deeper networks [18]. One
of the main problems of the Auto-Encoder model is that it
could potentially learn a useless identity transformation when
the representation size (the hidden layer) is larger than the
input layer (the so-called over-complete case). To overcome
this potential limitation, there is a simple variant of the model
called Denoising Auto-Encoders (DAE) [21] that works well
in practice. The basic idea is to feed the encoder/decoder
system with a corrupted version of the original input, In order
to force the system to reconstruct the clean version. This
small change allows the DAE to learn useful representations,
even in the over-complete case. The Auto-Encoder model
is widely used in intrusion detection and according to Yu
et al. [9] it is possible to apply unsupervised deep learning
techniques to automatically learn essential features from raw
Network Traffic counts. To that end, the authors employed a
variant of DAE based deep learning architecture called Stacked
Denoising Autoencoders (SDA) to detect traffic generated by
botnets. The SDA approach also exhibited the best perfor-
mance in binary classification on the datasets. While SDA
and DBN have similar training principles, SDA appears to be
comparable and superior to DBN. The authors claimed that
SDA employing the denoising criterion can learn significant
higher level representation (features) from raw traffic data, and
that deep learning approaches have remarkable capabilities for
the intrusion detection task.
4) Recurrent Neural Networks (RNNs): The RNNs can be
used for either unsupervised or supervised learning. When
used in unsupervised learning mode, prediction of the data
sequence from previous data samples is possible, but this
concept sometimes causes difficulty in the training phase. [18].
The RNN model architecture is a feedback loop linking layer
to layer with the ability to store data of previous input
increasing in that way the reliability of the model [19]. The
3
depth of an RNN can be as large as the length of the input data
sequence. RNNs are very powerful for modeling sequence data
(e.g., speech or text). With these precedents, Yin et al. [11]
proposed a deep learning approach for intrusion detection us-
ing recurrent neural networks (RNN). The authors proclaimed
that the RNN mode could be suitable for the development of
a classification model with higher accuracy and performance
than the traditional machine learning classification methods in
binary and multiclass classification.
III. L ITERATURE REVIEW
Based on the background presented above, this section
focuses mainly on providing information about intrusion de-
tection technologies in conjunction with machine learning
technique. Figure 2 shows deep learning algorithms to Net-
work Intrusion Detection.
Fig. 2: Classification of selected DL algorithms to Network
Intrusion Detection
In the early 2000s, several studies presented two issues
in developing an effective and flexible Network Intrusion
Detection System (NIDS) for unknown future attacks. The
first issue was that it was difficult to extract features from the
network traffic dataset for anomaly detection. The second issue
was that a labeled (trained) traffic dataset from real network
traffic was not available. In order to overcome those Fiore
et al. [15] used Restricted Boltzmann Machine in order to
implement semi-supervised anomaly detection systems where
the classifier was trained with normal traffic data only, so
that knowledge about anomalous behaviors was constructed
and evolved in a dynamic way. In the experimental phase,
the accuracy of RBM was tested in classifying normal data
and data infected by bot. A second experiment trained RBM
with the KDD Cup 99 dataset and tested against real world
data. To randomize the order of test data, the experiment was
repeated 10 times. The experiment confirmed that testing a
classifier in two different networks training data affects the
performance. The results revealed that when the classifier was
tested in a network widely different from the one where train-
ing data were taken from, the results decayed. This suggests
the need for further investigation over the nature of anomalous

traffic and the intrinsic differences with normal traffic. The
Deep Belief Networks (DBN) learning method consists of the
unsupervised training of each layer of Restricted Boltzmann
Machine (RBM), feeding the output vector of the last layer
of RBM as the input vector of BP neural network, and then
performing a supervised training to the BP neural network.
Gao et al. [10] applied DBN in the intrusion detection domain.
This deep learning model not only learns high-dimensional
representations but also performs classification tasks in an
efficient manner. The unsupervised greedy learning algorithm
can be used to pre-train and fine-tune a DBN in order to learn a
similarity representation over the nonlinear, high-dimensional
input data, something that largely facilitates the classification
tasks. In the experiment phase, the KDD Cup 1999 dataset was
used. It was distributed in 494021 records of train data and
11850 records of test data. The model was denoted as DBN i ,
where i signifies the number of Restricted Boltzmann Machine
(RBM) layers. Accuracy rates of 74.2%, 82.65%, 90.07%,
93.4%, 86.82%, and 82.30% were achieved by the DBN 1 ,
DBN 2 , DBN 3 , DBN 4 ; SVM and ANN, respectively. The
detection rates of the shallow DBN 1 and DBN 2 were not
better than those of the SVM and the ANN, but the deep
DBN 3 with its additional one layer RBM outscored them
both. Therefore, the performance of the DBN model using
RBM network with three or more layers outperforms SVM
and ANN.
Meanwhile, Li et al. [4] demonstrated that the use of the
AutoEncoder deep learning method Is effective for achieving
data dimension reduction and it can improve the detection ac-
curacy. In the experimental model the authors applied the Au-
toEnconder in order to convert complicated high-dimensional
data into low dimensional codes with a nonlinear mapping,
thereby reducing the dimensionality of data, extracting its
main features, and then applying the DBN learning method
to detect malicious code. The comparative analysis between
single DBN and the DBN + AutoEnconder Yielded accuracy
rates of 89.75% and 91.4%, respectively. Consequently, the
hybrid malicious code detection model was superior to that
of the single DBN. Javaid et al. [6] applied deep learning
techniques such as sparse AutoEncoder and SMR. Based on
those techniques, this research introduced two main steps in
terms of feature extraction and supervised classification. The
first step was to collect unlabeled network traffic data. Then,
the next step was to apply the extracted features to a labeled
traffic dataset, which could be collected in a confined, isolated,
and private network environment for supervised classification.
The primary goal of this research was to evaluate the perfor-
mance of deep learning based on accuracy. To evaluate the
classification accuracy, this research evaluated the NSL-KDD
training dataset using 2 − class (normal and attack), 5 − class
(normal and four different attack categories), and 23 − class
(normal and twenty-two different attacks). According to the
results, deep learning showed better accuracy performance for
2−class compared to SMR; however, there was no significant
improvement for 5 − class and 23 − class. The accuracy
for 2 − class was 88.39%, which was much higher than
SMR (78.06%). It should be noted that The highest accuracy
achieved using the NB-Tree methodology was 82%.
4
Ashfaq et al. [7] improved the classifiers performance
for the intrusion detection systems (IDS) through a novel
Fuzziness based semi-supervised learning approach by uti-
lizing unlabeled samples assisted with a supervised learning
algorithm. A single hidden layer feed-forward neural network
(SLFN) was trained to output a fuzzy membership vector,
and the sample categorization (low, mid, and high fuzziness
categories) on the unlabeled samples was performed using the
fuzzy quantity. The classifier was trained after incorporating
each category separately into the original training set. During
the experiment phase, the authors performed necessary scaling
to normalize the data. Two subsets were extracted from the
training file. The division of training samples and unlabeled
samples was done following a 10 : 90 ratio, where the 10%
of the NSL-KDD dataset was used as labeled data and the
remaining 90% as unlabeled data. According to the results,
the accuracy obtained by the proposed algorithm was the
highest compared to those obtained with J48, Naive Bayes, NB
tree, Random forests, Random tree, Multi-layer perceptron,
and Support Vector Machine (SVM). Another cybersecurity
work utilizing the NSL-KDD dataset is Diro et al. [5] was
on NSL-KDD dataset. This work uses a self-taught deep
learning scheme in which unsupervised feature learning has
been applied on training data. It utilizes a novel model of
parallel training and parameters sharing by local fog nodes
and it detects network attacks in the distributed fog-to-things
networks following a deep learning approach. The outputs
of the model training on the distributed fog nodes are the
attack detection models and their associated local learning
parameters. These local parameters are sent to the coordinating
fog node for global update and re-propagation. This sharing
scheme results in better learning as it enables the sharing
of the best parameters avoiding in this way local overfitting.
The results of the experiment give rise to two conclusions.
First, the distributed model has a better performance than the
centralized model, since the increased number of nodes in the
distributed network of Fog systems, leads to a 3% increase
in the overall accuracy of detection from around 96% to over
99%. Second, the detection rate exhibits as well that deep
learning is better than classic machine learning for both binary
and multi-classes.
Yu et al. [9] demonstrated the application of unsupervised
deep learning techniques to automatically learn essential fea-
tures from raw Network Traffic counts. The authors Imple-
mented a Stacked Denoising Autoencoders (SDA) based deep
learning architecture to detect traffic generated by botnets.
The procedure of SDA was split into two stages. The first
stage, unsupervised layer-wise pre-training, is a greedy layer-
wise training process. In the second stage, supervised fine-
tuning stage, a logistic regression layer for classification was
added on top of the stacked denoising autoencoders. In the
experiment phase, the UNB ISCX IDS 2012 dataset was
used. Two datasets of different sizes were constructed. One
consisting of 43% of the UNB ISCX IDS 2012 and another
using the original dataset in its entirety. The experiment was
divided into three parts: binary classification using the SDA-
based deep neural network, multi-classification using the SDA-
based deep neural network, and classification using different

deep learning architectures. In the binary classification using
the SDA-based deep neural network, the values of accuracy
rate achieved were over 99%. These results indicate that the
SDA approach could learn highly significant features from the
raw payloads of the network application layer. In the second
part of the experiment, the multi-classification (8classes)
using the SDA-Based deep neural network, the SDA approach
performed equally well as in the binary classification task on
both the 43% dataset and the whole dataset. Specifically, the
SDA approach achieved 98.11% accuracy rate on the whole
dataset, and 97.96% accuracy rate on the 43% dataset. The
SDA approach performed better in the larger dataset for the
multi-classification task, a fact also observed in the binary clas-
sification. In the last part of the experiment, the classification
using different deep learning architectures, the authors used
2 − class and 8 − class on the 43% and the whole dataset.
The following deep learning approaches were evaluated: SDA,
SAE, DBN, and AE-CNN achieving accuracy rates of 99.41%,
99.26%, 99.29% and 98.46% on the 43% dataset (2 − class).
Similarly, the accuracy rates achieved on the whole dataset
for the 2 − class case were 99.48%, 99.42%, 99.39%, and
98.54% respectively. Regarding the 8 − class case, accuracy
rates of 97.96%, 98.51%, 98.04% and 96.37% on the 43%
dataset were achieved by SDA, DBN, and AE-CNN, whereas
the corresponding rates on the whole dataset were 98.11%,
97.96%, 97.55% and 93.58% respectively. According to the
results SDA achieved better overall performance in almost all
experiments except multi-classification on the 43% dataset.
Furthermore, while other approaches got worse performance
as the dataset size increased, the SDA method achieved its
highest accuracy on the whole dataset. The SDA approach also
demonstrated the best performance in the binary classification
on both datasets. While SDA and DBN have similar training
principles, SDA appears to be comparable and superior to
DBN. And the AE-CNN method always yields worse perfor-
mance than others. The authors have demonstrated that SDA
employing the denoising criterion can learn significant higher
level representation (features) from raw traffic data, and that
deep learning approaches have remarkable capabilities for the
intrusion detection task.
Yin et al. [11] proposed a deep learning approach for
intrusion detection using recurrent neural networks (RNN).
The principal goal of this study was to demonstrate that RNN
is suitable for the development of a classification model with
higher accuracy and performance than the traditional machine
learning classification methods in binary and multiclass clas-
sification. To evaluate the classification accuracy, this research
used the NSL-KDD dataset in the experimental phase. In this
phase, the training of the RNN model consisted of two states:
Forward Propagation and Back Propagation. The experiment
consisted of two parts: the study of the performance of the
RNN-IDS model for binary classification (Normal, anomaly)
and five category classification such as Normal, DoS, R2L,
U2R and Probe. In the binary classification and the five-
category classification, they compared the performance of the
RNN model with that of an ANN, naive Bayesian, random
forest, support vector machine, random tree and multilayer
perceptron. The result in the binary classification showed that
5
the RNN model achieved a detection rate of 83.28% when
given 100 epochs. Meanwhile, the algorithms: J48, Naive
Bayesian, Random Forest, Multi-layer Perceptron, Support
Vector Machine and ANN obtained a detection rate of 81.2%.
In the five-category classification the RRN model obtained
an accuracy rate of 81.29%, which was better than the 79.9%
obtained by using J48, Naive Bayesian, Random Forest, Multi-
layer Perceptron, Support Vector Machine and ANN. In order
to increment the accuracy rate in the RNN model, they applied
the reduced-size technique on the dataset, reaching in that way
to 97.09% (gain rate: 17.19%).
Roy et al. [12] used a multilayer feed-forward network to
implement a deep learning concept for IDS. The feed-forward
network consisted of input layers, about 400 hidden layer
neurons, and output neurons. The activation functions used
were the rectifier activation function and the softmax activation
function. In the experimental phase, the authors compared the
performance of the Deep Neural Network (DNN) with that
of a Support Vector Machine (SVM). The KDD Cup 1999
dataset was divided into two parts: training and validation,
75% of the dataset were assigned to the training frame
and 25% of the dataset has been assigned to the validation
frame. The results of the experiment provided the following
information: The training and validation models have a very
high R2 (0.999944) value. This high value indicates that the
adopted model is highly accurate. In effect, the DNN has better
performance than SVM, since its accuracy rate was around
99.99% compared to 84.63%. With the loss being set as Cross
Entropy, this classification model could be used to detect future
intrusion attacks.
Li et al. [13] employed a visual conversion of the NSL-KDD
dataset format to evaluate the performance of Convolutional
Neural Networks (CNN) in detecting novel attacks. In the
experimental phase, the data were categorized into five classes:
Normal, Dos, Probe, U2R, and R2L. The NSL-KDD feature
attributes were classified into three groups: basic features,
traffic features, and content features. Each sample of the NSL-
KDD dataset contained 41 features (integer, float, symbolic,
or binary). To convert NSL-KDD data format into the visual
image type, they mapped various types of features into binary
vector space and then transformed the binary vector into the
image. After that, the NSL-KDD data form turned into a binary
vector with 464 dimensions. Then they turned every 8 bits
into a grayscale pixel. The binary vector with 464 dimensions
turned into an 8 ∗ 8 grayscale image with vacant pixels padded
by 0. The NSL-KDD dataset was divided into training and
test sets. In order to test intrusion detection and identify the
ability to discover new attacks, there were 17 additional attack
types were incorporated in the test set. The results showed
that the proposed method had good performance on the NSL-
KDD Test21 dataset in comparison to J48, Naive Bayes, NB
Tree, Random forest, Random tree, Multi-layer perceptron,
and SVM. The CNN model obtained 81.57% of accuracy
rate, whereas, the other algorithms achieved accuracy rates of
63.97%, 55.77%, 66.16%, 62.26%, 58.51%, 57.34%, 42.29%
and 81.57% respectively.
One of the most recent works in cyber-security is of Shone
et al [14]. In this study the authors proposed a non-symmetric

deep auto-encoder (NDAE) for unsupervised feature learning.
The classifier model was constructed by stacked NDAEs on
KDD Cup99 and NSL-KDD datasets. Fundamentally, this
involves the proposed shift from the encoder-decoder paradigm
(symmetric) towards utilizing just the encoder phase (non-
symmetric). The reasoning behind this is that given the correct
learning structure, it is possible to reduce both computa-
tional and time overheads, with minimal impact on accuracy
and efficiency. NDAE was used as a hierarchical unsuper-
vised feature extractor. The model was realized by stacking
NDAEs in order to create a deep learning hierarchy. Stacking
the NDAEs offered a layer-wise unsupervised representation
learning algorithm, which allowed the model to learn the
complex relationships between different features. It also had
feature extraction capabilities. Hence, it was able to refine the
model by prioritizing the most descriptive features. The model
was implemented in GPU-enabled TensorFlow and evaluated
using the benchmark KDD Cup99 and NSL-KDD datasets.
The authors combined the deep learning power of stacked
NDAEs with a shallow learning classifier. Random Forest (RF)
was used as shallow learning classifier in order to increase
the classification power of stacked autoencoders. The model
trained the RF classifier using the encoded representations
learned by the stacked NDAEs to classify network traffic into
normal data and known attacks. In the evaluation phase, they
compared the stacked NDAE model against the mainstream
DBN technique. These comparisons have demonstrated that
the model offers up to a 5% improvement in accuracy and up
to 98.81% of training time reduction
IV. C ONCLUSION
Motivated by the challenges and the current state-of-the-
art for deep learning applications in a multitude of areas, in
this work we have surveyed and classified the deep learning
techniques employed for network intrusion detection. The re-
sults of the examined models demonstrate that the application
of deep learning in cybersecurity and in particular in IDS is
promising. In spite of the significant advancements, however,
there is still much room for future research on a variety of
open issues, such as new approaches to testing the reliability
and efficiency of knowledge-based and behavioral approaches
in intrusion detection.
R EFERENCES
[1] H. Debar, M. Dacier and A. Wespi Towards a taxonomy of intrusion-
detection systems. Computer Networks, 31(8), 805-822
[2] S. Axelsson (2000).Intrusion detection systems: A survey and taxonomy
(Vol. 99). Technical report.
[3] T. A. Tang, L. Mhamdi, D. McLernon, S. A. R. Zaidi and M. Ghogho
(2016, October). Deep Learning Approach for Network Intrusion De-
tection in Software Defined Networking. In Wireless Networks and
Mobile Communications (WINCOM), 2016 International Conference on
(pp. 258-263). IEEE.
[4] Y. Li, R. Ma and R Jiao (2015).A hybrid malicious code detection method
based on deep learning. methods. methods, 9(5).
[5] A. A. Diro and N. Chilamkurti Distributed attack detection scheme using
deep learning approach for Internet of Things. Future Generation
Computer Systems.
6
[6] A. Javaid, Q. Niyaz and M. Alam (2016, May).A deep learning approach
for network intrusion detection system. In Proceedings of the 9th EAI
International Conference on Bio-inspired Information and Communica-
tions Technologies (formerly BIONETICS) (pp. 21-26). ICST (Institute
for Computer Sciences, Social-Informatics and Telecommunications En-
gineering).
[7] R. A. R.. Ashfaq, X. Z. Wang, J. Z. Huang, H. Abbas and Y. L. He
(2017).Fuzziness based semi-supervised learning approach for intrusion
detection system. Fuzziness based semi-supervised learning approach
for intrusion detection system. Information Sciences, 378, 484-497.
[8] B. Krawczyk (2016).Learning from imbalanced data: open challenges
and future directions. Progress in Artificial Intelligence, 5(4), 221-232.
[9] Y. Yu, J. Long and Z. Cai (2017, October).Session-Based Network
Intrusion Detection Using a Deep Learning Architecture. In Modeling
Decisions for Artificial Intelligence (pp. 144-155). Springer, Cham.
[10] N. Gao, L. Gao, Q. Gao and H. Wang (2014, November).An intrusion
detection model based on deep belief networks. In Advanced Cloud and
Big Data (CBD), 2014 247 on (pp. 247-252). IEEE.
[11] C. Yin, Y. Zhu, J. Fei and X. He Deep Learning Approach for Intrusion
Detection Using Recurrent Neural Networks. IEEE Access, 5, 21954-
21961.
[12] S.S. Roy, A. Mallik, R. Gulati, M.S. Obaidat and P. V. Krishna (2017,
January) A Deep Learning Based Artificial Neural Network Approach for
Intrusion Detection. In International Conference on Mathematics and
Computing (pp. 44-53). Springer, Singapore.
[13] Z. Li, Z. Qin, K. Huang, X. Yang and S. Ye (2017, November).Intrusion
Detection Using Convolutional Neural Networks for Representation
Learning. In International Conference on Neural Information Processing
(pp. 858-866). Springer, Cham.
[14] N. Shone, T. N. Ngoc, V. D. Phai and Q. Shi (2018).A deep learning
approach to network intrusion detection. IEEE Transactions on Emerging
Topics in Computational Intelligence, 2(1), 41-50.
[15] U. Fiore, F. Palmieri, A. Castiglione and A. De Santis (2013). Network
anomaly detection with the restricted Boltzmann machine. Neurocom-
puting, 122, 13-23.
[16] G. E. Hinton, S. Osindero and Y. W. Teh (2006).A fast learning
algorithm for deep belief nets. Neural computation, 18(7), 1527-1554.
[17] G. E. Hinton and R. R. Salakhutdinov Reducing the dimensionality of
data with neural networks. science, 313(5786), 504-507.
[18] Y. LeCun, Y. Bengio and G. Hinton (2015).Deep learning. nature,
521(7553), 436.
[19] I. Goodfellow, Y. Bengio, A. Courville and Y. Bengio(2016)Deep
learning (Vol. 1). Cambridge: MIT press.
[20] R. Salakhutdinov (2010). Learning deep Boltzmann machines using
adaptive MCMC. In Proceedings of the 27th International Conference
on Machine Learning (ICML-10) (pp. 943-950).
[21] P. Vincent,H. Larochelle, Y. Bengio and P. A. Manzagol (2008, July).
Extracting and composing robust features with denoising autoencoders.
In Proceedings of the 25th international conference on Machine learning
(pp. 1096-1103). ACM.