Received November 5, 2021, accepted November 24, 2021, date of publication November 30, 2021,

date of current version December 13, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3131949

Applications of Generative Adversarial Networks

in Anomaly Detection: A Systematic
Literature Review
MIKAEL SABUHI 1 , (Graduate Student Member, IEEE),
MING ZHOU1 , (Graduate Student Member, IEEE),
COR-PAUL BEZEMER 1 , (Member, IEEE),
AND PETR MUSILEK 1,2 , (Senior Member, IEEE)
1 Department of Electrical and Computer Engineering, University of Alberta, Edmonton, AB T6G 2R3, Canada
2 Department of Applied Cybernetics, University of Hradec Králové, 500 03 Hradec Kralove, Czech Republic
Corresponding author: Mikael Sabuhi ([email protected])
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada under Project ALLRP549804-19
and Project CRD543899-19, and in part by the Government of Alberta under Major Innovation Fund Project RCP-19-001-MIF.

ABSTRACT Anomaly detection has become an indispensable tool for modern society, applied in a wide
range of applications, from detecting fraudulent transactions to malignant brain tumors. Over time, many
anomaly detection techniques have been introduced. However, in general, they all suffer from the same
problem: lack of data that represents anomalous behaviour. As anomalous behaviour is usually costly (or
dangerous) for a system, it is difficult to gather enough data that represents such behaviour. This, in turn,
makes it difficult to develop and evaluate anomaly detection techniques. Recently, generative adversarial
networks (GANs) have attracted much attention in anomaly detection research, due to their unique ability
to generate new data. In this paper, we present a systematic review of the literature in this area, covering
128 papers. The goal of this review paper is to analyze the relation between anomaly detection techniques
and types of GANs, to identify the most common application domains for GAN-assisted and GAN-based
anomaly detection, and to assemble information on datasets and performance metrics used to assess them.
Our study helps researchers and practitioners to find the most suitable GAN-assisted anomaly detection
technique for their application. In addition, we present a research roadmap for future studies in this area.
In summary, GANs are used in anomaly detection to address the problem of insufficient amount of data for
the anomalous behaviour, either through data augmentation or representation learning. The most commonly
used GAN architectures are DCGANs, standard GANs, and cGANs. The primary application domains
include medicine, surveillance and intrusion detection.

INDEX TERMS Anomaly detection, data augmentation, generative adversarial networks, outlier detection,
representation learning.

I. INTRODUCTION the impact, it is important to be able to detect such anomalies

In modern society, many systems depend on and generate as early as possible. For example, cancer is an anomaly in
enormous amounts of data. This data is important for many human tissue. Breast cancer is the second leading cause of
decision-making processes. Normally, systems operate under cancer death in women [1]. According to a recent study by the
the expected conditions. However, in rare cases, anomalies American Cancer Society [2], breast cancer alone accounts
may occur. Such anomalies can have a disastrous impact on for 30% of female cancers. Early detection and treatment
the system itself or on its environment. Therefore, to lower of breast cancer would highly increase the chance of sur-
vival [3]. Similarly, with an increasing need to ensure public
The associate editor coordinating the review of this manuscript and safety in crowded areas, the development of real-time video
approving it for publication was Baker Mohammad . surveillance systems becomes unavoidable. It is critical to

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 161003
 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

seamlessly monitor the crowd to immediately detect anoma- • RQ3: Which GAN architecture is used most often in
lous (or abnormal) movements to help prevent theft [4], anomaly detection systems? We identified 21 archi-
vandalism [5], and terrorist attacks [6]. tectures of GANs that are used for anomaly detec-
The process of finding the anomalous behaviour of a tion. Among these architectures, deep convolutional
system is referred to as anomaly detection. The primary GANs (DCGANs) (32%), standard GANs (23%), and
objective of anomaly detection is to differentiate between conditional GANs (16%) are the most commonly
the expected and unexpected behaviour of a system. Con- used.
sidering the importance of anomaly detection, it has received • RQ4: Which type of data instance and datasets are
widespread attention in research. Despite the progress in this most commonly used for anomaly detection with
research area, there is still an important open challenge: the GANs? 50% of the proposed GAN-based anomaly
acquisition of data about anomalies that can be used to test detection techniques use image datasets for anomaly
anomaly detection techniques. detection purposes. Before being fed to the anomaly
A recent trend in anomaly detection is the use of detection algorithms, the data are usually preprocessed.
generative adversarial networks (GANs). Proposed by Ian The most common preprocessing methods are resizing
Goodfellow et al. [7] in 2014, GANs are a type of unsuper- (23%), normalization (19%), and cropping (13%).
vised generative model which gained much attention from • RQ5: Which metrics are used to evaluate the per-
the research community. A well-trained GAN can generate formance of GANs in generating data and anomaly
realistic-looking data by sampling from a learned data distri- detection? Only 21% of the studied papers evaluated
bution. A GAN consists of a generator and a discriminator the GAN’s performance in generating synthetic data,
model. These two models are pitted against each other in either in data augmentation or representation learning.
a two-player zero-sum game situation, iteratively improving Structural similarity indices (SSIM) (26%) and peak
their capabilities to generate and discriminate data. signal-to-noise ratio (PSNR) (26%) are the most com-
The ability of GANs to generate data makes them attrac- monly used metrics. Visual inspection to evaluate the
tive for anomaly detection research from two perspectives. quality of the generated data was reported in 5% of the
First, they can potentially help generate the hard-to-acquire studied papers. To evaluate the performance of GANs
anomalous data points. Second, they can be used to learn the in anomaly detection applications, 53% of the primary
distribution of the data for the normal operating condition of studies used the area under the receiver operating char-
a system and act as an anomaly or outlier detector. acteristic curve (AUROC).
In this paper, we conduct a systematic literature review of • RQ6: Which anomaly detection techniques are used
the applications of GANs for anomaly detection. We address along with GANs? GAN-based anomaly detection is
the following research questions (RQs): mostly done in a semi-supervised manner. DCGANs
• RQ1: What is the role of GANs in anomaly detection? and standard GANs are the most popular archi-
We identified two roles that GANs play in anomaly tectures in semi-supervised anomaly detection using
detection: data augmentation and representation learn- GANs.
ing. In contrast to the remarkable ability of GANs to In supervised learning-based anomaly detection, GANs
generate realistic-looking data, most of the reviewed are used to augment the dataset for the anomalous
papers use them for representation learning rather than class. However, the studied papers report only minor
data augmentation. The reason for this inclination is improvements in the performance of anomaly detection
that, despite the improvement in the anomaly detection techniques after augmenting the dataset with GANs.
accuracy after data augmentation, the reported improve- Only a few primary studies focused on pure unsuper-
ments are not substantial. When GANs are used for vised anomaly detection based on GANs, most using the
data augmentation in anomaly detection, we refer to it standard version of GANs. Similar to semi-supervised
as GAN-assisted anomaly detection. The other role of techniques, unsupervised GAN-based anomaly detec-
GANs in anomaly detection is representation learning. tion techniques are mostly compared with autoencoder-
In this case, the examined papers use the data from the based approaches.
normal class for training a GAN to learn the distribution The findings presented in this survey will help researchers
of the normal data. A score is assigned to the new data and practitioners to find the most suitable GAN-based
by defining a score function, and the anomalous data in anomaly detection techniques for their applications.
the test stage is identified based on a specific threshold. The rest of this paper is organized as follows. Section II
We refer to these techniques as GAN-based anomaly provides a brief introduction to GANs. Section III describes
detection. the methodology used for conducting this systematic liter-
• RQ2: What are the application domains of anomaly ature review. Section IV presents the results of the review.
detection with GANs? The primary application areas Section V discusses the open challenges and provides direc-
where GANs are used for anomaly detection are tions for future research. Section VI identifies the threats
medicine (19%), surveillance (15%) and intrusion to validity of the review, and Section VII concludes the
detection (13%). paper.

161004 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

FIGURE 1. The building blocks of GANs. The classification error is used to update the parameters of the
discriminator and generator models (shown by dashed lines).

II. GENERATIVE ADVERSARIAL NETWORKS of being distinguished from the real data by the
In 2014, Goodfellow et al. [7] introduced a framework for discriminator.
estimating generative models based on an adversarial process. The adversarial training of the generator and the discrim-
This framework consists of two deep neural network-based inator model is a zero-sum game problem: when one model
models: a generative model G and a discriminator model D. gets better the other one gets worse in equal proportions [8].
Model G learns the training data distribution and uses it to For all zero-sum games, there is a point where neither of the
generate new samples. Model D determines whether a sample players can improve their situation. This point is referred to
comes from the training data or was generated by the gener- as the Nash equilibrium. The goal of a GAN is to reach this
ative model. The power of GANs comes from the adversarial equilibrium, as then the fake data produced by the generator
process, in which the two models are competing against each model is indistinguishable from the real data by the discrimi-
other to improve their accuracy in the designated task. nator model. The output of the discriminator is then a random
The diagram in Figure 1 shows the building blocks of a guess on whether the input data is real or fake.
GAN [8]:
III. METHODOLOGY
• The Real Data (X), or the training dataset, contains the
instances that the generator G should learn to generate, The planning, conducting, and reporting of this systematic lit-
usually in the form of a batch. erature review (SLR) were based on the guidelines proposed
• Random Noise Vector (z) is the raw input to the genera- by Kitchenham [10]. The planning stage of the SLR includes
tor. It is a vector of random numbers which the generator three steps: identification of the need for the systematic
uses to generate fake examples. review, development of the review protocol, and evaluation
• The Generator model (G) is trained to learn the distri- of the protocol [10]. In the conducting stage, based on the
bution of the input data. This model uses the input (z) to review protocol that was developed during planning stage,
generate fake examples (G(z)) that are indistinguishable we search for and select the primary studies, extract data
from the real data. from the primary studies, and synthesize the data. The set of
• The Discriminator model (D) tries to distinguish the data primary studies contains all individual studies that contribute
that is generated by the generator from the real data. to the SLR [10]. In the last stage, we conclude the systematic
The inputs to this model are the real data (X ) and the review by reporting the collected data and findings. Figure 2
generated data (G(z)). The output of this model is a summarizes the required steps for each stage of the review.
binary decision for each data instance, i.e. real/fake. In the following, each step is explained in more detail.
• Iterative Training: The GAN is trained using the classifi-
cation error of the discriminator. The error is used to tune A. THE NEED FOR A SYSTEMATIC REVIEW
the parameters (weights and biases) of the discriminator, Recently, GANs have become a hot research topic in many
and then the parameters of the generator. Backpropaga- application domains. One of these domains is anomaly detec-
tion [9] is commonly used as the training algorithm. This tion. The ability of GANs to generate realistic looking data
iterative training consists of two loops: and to perform representation learning makes them attractive
for anomaly detection research. Basically, GANs are trained
– An inner loop where the discriminator’s parameters in an unsupervised manner to learn the distribution of the
are tuned to maximize the classification accuracy of data. However, they are highly flexible and can be used
predicting correct labels for real data and generated in semi-supervised fashion as well (e.g. [11]). In addition,
data. GANs are implicit density models which do not require
– An outer loop where the generator’s parameters are any explicit hypothesis on the distribution of the data [12].
tuned to generate data that has a minimal chance Considering all these advantages, GANs can be leveraged to

VOLUME 9, 2021 161005

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

FIGURE 2. The steps of our systematic literature review, based on Kitchenham’s guidelines [10].

address some existing problems in anomaly detection, such Motivation: The use of GANs in anomaly detec-
as the lack of a sufficient amount of data for anomalous tion may be more common in certain domains. Here,
behaviour of the system. Therefore, a study summarizing we look into which domains and which types of GANs
existing research on applications of GANs in anomaly detec- work together well.
tion would be of a high value to the research community. 3) RQ3: Which GAN architecture is used most often in
When we started this systematic literature review, we iden- anomaly detection systems? (Section IV-C)
tified only one survey paper [13] reviewing applications of Motivation: There exist many architectures of GANs.
GANs in anomaly detection. However, this paper only cov- Each one attempts to handle a specific type of data or
ered 11 papers on anomaly detection with GANs. In addition, to address an existing problem in the previous archi-
it did not follow a systematic approach to conduct the review. tectures. Some architectures may be better suitable
This confirmed the need for a systematic literature review on for anomaly detection than others. Therefore, we look
applications of GANs in anomaly detection, which covers into which architectures of GANs are commonly
a vast number of papers. To reduce researcher bias [10], used.
we followed a systematic approach for designing, executing 4) RQ4: Which type of data instance and datasets are most
and reporting our findings. commonly used for anomaly detection with GANs?
(Section IV-D)
B. DEVELOPING THE REVIEW PROTOCOL Motivation: Identifying which datasets are used to eval-
To reduce the possibility of researcher bias in a system- uate anomaly detection with GANs in certain domains
atic manner, a review protocol is required to specify the can reveal the ‘‘standard benchmarks’’ in specific
method for conducting the systematic review [10]. This domains and which domains require benchmarks in
protocol includes definition of the following elements: general.
1) research questions, 2) search strategy, 3) study selection 5) RQ5: Which metrics are used to evaluate the perfor-
criteria (including study quality assessment), 4) data extrac- mance of GANs in generating data and anomaly detec-
tion strategy, and 5) synthesis of the extracted data. tion? (Section IV-E)
Motivation: Evaluating GANs in anomaly detection
systems is not a straightforward task as their goal is
1) OUR RESEARCH QUESTIONS
to create realistic looking data that is different enough
In this systematic literature review, we address the following from known anomalies, yet still representative of real
research questions (RQs): anomalies. Therefore, one cannot just compare the
1) RQ1: What is the role of GANs in anomaly detection? generated data with the real data. We study which
(Section IV-A) approaches are commonly used to evaluate the qual-
Motivation: It is important to learn how GANs are used ity of the generated data, and support practitioners in
in anomaly detection. One intuitive way is to generate deciding which metrics to use for evaluating data in
anomalous data to address the problem of the imbal- specific anomaly detection problems.
anced dataset. Still, there might be more opportunities. 6) RQ6: Which anomaly detection techniques are used
Moreover, we will investigate what are the alternative, along with GANs? (Section IV-F)
non-GAN approaches to handle these identified roles. Motivation: GANs are often used together with more
2) RQ2: What are the application domains of anomaly traditional anomaly detection techniques, especially
detection with GANs? (Section IV-B) when they are used in a supervised manner. In this

161006 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

question, we identify the anomaly detection techniques

that are based on or assisted by GANs.

2) SEARCH STRATEGY
To find relevant papers for this systematic review,
we searched the IEEE Xplore,1 ACM Digital Library,2 Sci-
ence Direct3 and Scopus4 digital libraries. The focus of
this study is the application of GANs in anomaly detection.
Therefore, we combined the keywords related to anomaly
detection with keywords and abbreviations for generative
adversarial networks. To find closely relevant papers for this
study, we searched the title and the abstract of the papers
for the following query: (‘‘anomaly’’ OR ‘‘anomalies’’ OR
‘‘anomalous’’ OR ‘‘outlier’’ OR ‘‘abnormal’’) AND (‘‘gener-
ative adversarial network’’ OR ‘‘generative adversarial net-
works’’ OR ‘‘GAN’’ OR ‘‘GANs’’). The list of primary studies
was collected on 3rd June, 2020.
We conducted a pilot study to ensure that the well-known
primary studies were included in the query results. During
this study, we searched for the matched papers and their
shared references on Google Scholar to ensure that the most
cited papers were covered by the query. After several iter-
ations of improving the query, we were confident that it
returned important and well-known studies.

3) STUDY SELECTION CRITERIA

We defined the following criteria for the inclusion of a paper
in our study. The last two criteria in the list are included to
assess the quality of the study.
• The paper must be in the specified digital libraries.
• The primary study should focus on anomaly detection
while leveraging GANs.
• The developed methods should be evaluated on at least
FIGURE 3. The procedure for searching, selecting, and extracting data
one real dataset, not only on simulated data, to ensure from the primary studies for conducting the SLR.
the practical relevance of the study.
• The primary study should be available online to ensure the subject. This refinement was accomplished by comparing
accessibility. the data extraction forms of the two first authors and address-
• The article should be written in the English language. ing the potential ambiguities in the data extraction form.
All types of papers, including journal, conference, work-
shop, and symposium papers are considered in this review. 5) DATA SYNTHESIS
The procedure for searching (described in Section III-B2) and During the data synthesis step, we aggregate the data col-
selecting the primary studies is shown in Figure 3. The final lected from the data extraction forms to answer the research
list of papers used for data extraction and synthesis consists questions. Putting all this data together gives invaluable infor-
of 128 primary studies. Not every selected primary study mation concerning the current best practices and architectures
provides answers to all six research questions. for anomaly detection with GANs.

4) DATA EXTRACTION STRATEGY C. EVALUATION OF THE REVIEW PROTOCOL

To facilitate the data extraction, we devised a data extraction The protocol is a critical part of the SLR. It was evaluated
form to collect the required information for each RQ from the by the last two authors and, after several iterations, the final
primary studies. This form (shown in Table 1) was refined version of the protocol was approved and used throughout the
through several iterations with randomly selected papers on conducting stage of the SLR.
1 https://ieeexplore.ieee.org
2 https://dl.acm.org D. CONDUCTING THE SLR
3 https://sciencedirect.com The conducting stage of the SLR includes the following four
4 https://scopus.com steps: searching for primary studies, selecting the primary

VOLUME 9, 2021 161007

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 1. The data extraction form.

studies, extracting the primary studies, and synthesizing the extracting data from all primary studies and addressing the
data. discrepancies in the data extraction forms, we created a
We identified 362 papers that matched our search query spreadsheet for each data extraction form.
(see Section III-B2): 49 papers in Science Direct, 24 in ACM We summarize and report the extracted data in the follow-
Digital Library, 104 in IEEE Xplore, and 185 in Scopus. ing section.
We organized the papers for further analysis using Mende-
IV. RESULTS
ley as a reference manager.
This systematic literature review covers 128 primary studies
We filtered out irrelevant and duplicated papers according
that describe applications of GANs in anomaly detection.
to the study selection criteria introduced in Section III-B3.
As shown in Figure 4(a) these primary studies were published
Figure 3 shows the procedure for selecting the primary stud-
between 2017 and early 2020. The number of studies per year
ies. We filtered the papers in two steps. In the first step, the
is increasing, suggesting that interest in this research area is
first two authors independently read the abstract and the title
growing rapidly. Figure 4(b) shows that the majority of the
of the primary studies and decided if they were related to
reviewed papers (63%) appeared in conference proceedings,
anomaly detection with GANs. The Cohen’s kappa coeffi-
29% of the papers were published in journals, and 4% in
cient [14] for this binary classification (relevant vs. irrelevant)
workshops and 4% in symposia.
was 0.83, which shows a satisfactory agreement between the
researchers. There were 32 papers from the initial list of
papers on which the first two authors disagreed. For those A. RQ1: WHAT IS THE ROLE OF GANs IN ANOMALY
papers, the third author was asked to make the final decision DETECTION?
regarding inclusion or exclusion. After the first phase of filter- We identified four types of GAN applications in anomaly
ing papers, we ended up with 145 papers. The second filtering detection: (1) generating abnormal data instances, (2) gener-
step was performed while reading the full text: the first two ating normal and abnormal instances, (3) learning the normal
authors decided to include or exclude the paper in the data behaviour of a system, (4) learning both normal and abnormal
extraction step. In this phase, the first two authors made the behaviour of a system. Applications 1 and 2 can be classified
same decision regarding the excluded papers and excluded as data augmentation with GANs and 3 and 4 as representa-
17 papers. Finally, 128 primary studies were included and tion learning with GANs.
analysed in this SLR. Generative models, such as GANs, are mainly designed
Based on the data extraction strategy introduced in for data augmentation, i.e., to generate new data and use
Section III-B4, we examined these 128 primary studies to it to augment the existing data. They can also be used for
collect the data that contributes to addressing the RQs of representation learning, i.e., to learn representations of the
this SLR. The primary studies were randomly divided into data to support information extraction for use when building
10 batches. For each batch, the first two authors extracted the classifiers or other predictors [15]. In this case, the generator
data from the primary studies and filled in the data extraction and the discriminator of a GAN can be used to learn the
forms. After extracting the data from each batch, the data distribution of a specific class of data, i.e. normal or abnormal
extraction forms were randomly distributed between the first data. In turn, the learned distribution can be used to identify
two authors, and then the disagreements were identified and nonconforming or irregular data. Table 2 shows the two main
discussed in a meeting. If they failed to reach a consensus, roles of GANs in anomaly detection, along with the types
one of the last two authors made the final decision. After of data used in each role. Most of the primary studies opted

161008 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

FIGURE 4. Distribution of the primary studies according to the publication types and years.

to use GANs to learn the representation of the data rather

than augmenting the datasets. Moreover, this representation
learning is mainly performed on normal data. The rationale
behind this preference is that, due to the data imbalance,
it is usually easier to learn a model of normality rather than
abnormality. In addition, by learning only the normal data
distribution, the need for data from the abnormal condition
of the target system is eliminated.

1) REPRESENTATION LEARNING WITH GENERATIVE

ADVERSARIAL NETWORKS
The main goal of GANs is to learn a generative model that
produces realistic-looking data by sampling from the learned
distribution. This generative power of GANs was highlighted
by Goodfellow et al. [7] and Radford et al. [144]. Represen-
tation learning with GANs for anomaly detection exploits
the ability of GANs in learning the distribution of a specific
class of data, as shown in Figure 5. Several anomaly detection
techniques are proposed that use this representation learning
FIGURE 5. Overview of representation learning with GANs (GAN-based
ability of GANs (shown in Table 3). We will explain the anomaly detection). In the training stage, a GAN is trained to learn the
concept of anomaly detection using representation learning distribution of the normal data (shown as the blue area) and also to
minimize the output of an anomaly score [A(x)] for every generated data.
through an examples of a well-known GAN-based anomaly In the testing stage, since the GAN is only trained on the normal data
detection techniques (AnoGAN). All other anomaly detection distribution, the anomaly score will be higher than a threshold (At ) for
abnormal data, indicating anomalies.
techniques that rely on representation learning through a
GAN are variations to some extent of the AnoGAN technique.
Schlegl et al. [104] introduced the first GAN-based the normal data by training GANs only on normal data. From
anomaly detection technique, called AnoGAN, taking an anomaly detection view, learning the representation of the
advantage of the representation learning ability of GANs. normal data is useful as one can decide for new (potentially
Schlegl et al. put forth AnoGAN, which employs the DCGAN anomalous) data points how likely they are part of that normal
architecture, to learn the distribution of normal anatomical data. During the training of a GAN, the generator learns the
variability. The idea comes from the concept of a smooth mapping from latent space to data space G(z) = z → x
transition in the latent space of the data, i.e., that sampling (i.e., the representation of the data). However, the inverse
from two close points in the latent space should lead to similar mapping, which is necessary to decide whether a data point is
data in the data space [145]. Schlegl et al. hypothesize that the anomalous, is not straightforward to obtain [104]. To address
latent vector of the GANs represents the distribution of the this problem, Schlegl et al. proposed an additional step after
trained data. Therefore, one can learn the representation of training the GAN on normal data. For an image x, they

VOLUME 9, 2021 161009

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 2. List of studies using normal, abnormal, and normal and abnormal data together for different tasks of GANs.

TABLE 3. Representation learning with GANs.

proposed to find a point z in the latent space that corresponds artificially generated instances (oversampling) using suitable
to an image G(z), which is the most similar to the image x on techniques.
the learned manifold χ. Schlegel et al. proposed an iterative The problem of the imbalanced dataset is more critical in
process to find the most similar image G(z0 ) to x using anomaly detection since it is hard and expensive to collect
residual and discrimination loss. The similarity of images x data on anomalous behaviour of the system under study.
and G(z) depends on how closely x follows the distribution of Often, there are very few or no examples of anomalous data
the data learned by the generator (pg ). After identifying the available. In this situation, GANs can help by generating
most similar image, AnoGAN computes an anomaly score more samples for the anomalous class, as shown in Figure 6.
that is related to the similarity of x and G(z). Finally, based on Table 2 summarizes the main roles of GANs in anomaly
a threshold for the anomaly score, AnoGAN decides whether detection and the type of data used for that purpose. Data
x is an anomaly. augmentation with GANs is mostly used to generate data that
represents anomalous behaviour of the system. There was no
2) DATA AUGMENTATION WITH GENERATIVE ADVERSARIAL primary study augmenting only the normal condition of the
NETWORKS system under study for anomaly detection. This is due to the
Machine learning techniques, especially deep learning meth- fact that there is usually an abundance of data for the normal
ods [146], require a massive amount of data to perform well in condition. However, some studies augmented both normal
their designated task [147]. Data augmentation, also known as and abnormal data, e.g., using Cycle-GANs, by learning the
oversampling, is carried out to compensate for an insufficient transformation from abnormal to normal and from normal
amount of data in the dataset to prevent model overfitting. to abnormal to generate new data. After augmenting the
It can also be used to address the problem of data imbalance, dataset, it is ready to be used for anomaly detection, usually
which occurs when the sizes of the classes in a dataset differ performed by a classifier (as discussed in Section IV-F). Most
considerably. For instance, in a binary classification task, the primary studies that use GANs for data augmentation report
class with fewer samples is called the minority class, and a slight improvement in classification accuracy compared
the other class is called the majority class. The correspond- either to traditional techniques or without data augmentation.
ing training process would be biased towards the majority For instance, Madani et al. [135] report that, using data aug-
class, hence a classifier trained using this dataset would mentation with GANs, the test accuracy for cardiovascular
have a better accuracy for this class [148]. To address the abnormality detection improved from 81.93% to 84.19%.
imbalanced dataset problem, one can either randomly remove In comparison, using traditional augmentation methods, they
samples from the majority class to balance the class size achieved only 83.12% test accuracy. This improvement is sig-
(undersampling), or augment the minority class by adding nificant when dealing with large amounts of data, especially

161010 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

the more difficult to learn instances are synthesized more

frequently. In addition, several other traditional techniques
for data augmentation, such as adding Gaussian noise to
the dataset [118], Gaussian smoothing, unsharp masking,
minimum filtering [138], and affine transforms [141] were
compared to GANs. Most of these studies show that data
augmentation with GANs results in training datasets that
improve the anomaly detection.

B. RQ2: WHAT ARE THE APPLICATION DOMAINS OF

ANOMALY DETECTION WITH GANs?
Table 5 shows the different domains where GANs were
applied in the primary studies. The table reveals that
a vast number of primary studies (24 papers) perform
anomaly detection in medical applications, closely followed
by surveillance and intrusion detection with 19 and 17 papers,
respectively.

FIGURE 6. Overview of data augmentation with GANs (GAN-assisted 1) MEDICAL ANOMALY DETECTION
anomaly detection). In the training stage, a GAN learns to generate
realistic-looking samples for the minority (abnormal) class. In the data Anomaly detection in medicine deals with analyzing patients’
augmentation stage, the generator of the trained GAN is used to generate health conditions using medical records and images [153].
samples to augment the minority class.
Specific applications include retinal optical coherence
tomography (OCT) anomaly detection [17], [26], [29], [104],
in medical applications. However, in some studies, it has seizure detection [18], cardiovascular disease detection [30],
been reported that GAN did not meet their expectations in lung nodule detection [42], abnormal chest X-ray identifi-
improving the classification accuracy after augmenting the cation [59], [97], [135], [138], polyp detection [80], [123],
data, e.g. [141]. The list of different GANs used for data metastatic bone tumor detection [78], lesion detection [101],
augmentation is shown in Table 4. [137], laparoscopy anomaly detection [85], breast cancer
In the examined primary studies, we identified several tra- detection [132], [143], MRI quality control [98], diabetic
ditional techniques for addressing the problem of imbalanced retinopathy detection [133], brain tumor detection [134] and
and insufficient amounts of data. The effects of adopting these hemorrhage detection [105]. One of the challenges in this
techniques are compared to the GANs in terms of improving domain is the difficulty of obtaining expert labels for medical
the classification accuracy. For example, random undersam- data, such as clinical images, since annotation is an exhaus-
pling was evaluated in two primary studies [110], [126], tive and time-consuming task.
where samples were randomly removed from the majority
class. Using this approach, some important and critical data 2) SURVEILLANCE ANOMALY DETECTION
may be lost that could otherwise be beneficial for learning To improve public safety, surveillance cameras are widely
a robust decision boundary [149]. Random oversampling used in public places such as streets, stores, and banks. The
was investigated in four primary studies [53], [110], [111], goal of video surveillance is to identify suspicious activ-
[126]. In this case, some samples from the minority class ity, unusual traffic patterns, or accidents by automatically
are copied to increase its size. However, this approach is analyzing the behaviour of the surveillance target. In video
likely to cause over-fitting [126]. All these studies confirmed surveillance anomaly detection, this can be accomplished by
the superiority of GANs in data augmentation compared to identifying the out-of-ordinary behaviours that differ from
random over/undersampling. dominant (normal) behaviours in the scene [153]. Automated
Chawla et al. [150] proposed synthetic minority video surveillance can reduce the dependence on human
oversampling (SMOTE) to improve the random oversam- workers and reduce the risk of late detection of anomalous
pling by synthesizing new samples from the neighbour- behaviour. Most primary studies in this application domain
hood of the minority class samples. This improvement is leverage GANs for video anomaly detection to find irregular-
accomplished by interpolating between several minority ities in the crowds. However, traffic anomaly detection [24]
class instances. SMOTE and its variants (e.g., borderline- and threat object recognition with X-ray imaging [91] have
SMOTE [151]) were compared with GANs in several stud- also been studied.
ies [53], [110], [111], [117], [126], [132]. The ADAptive
SYNthetic (ADASYN) sampling approach for imbalanced 3) INTRUSION DETECTION
datasets [152] was compared with GANs in two studies [53], Intrusion detection systems are defined as software and/or
[126]. ADASYN uses a weighted distribution for different hardware components that monitor and analyze events in
minority class instances based on their difficulty level, i.e., computer systems to identify signs of intrusion [154]. Any

VOLUME 9, 2021 161011

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 4. Different types of GANs used for data augmentation.

TABLE 5. The application domains of GANs for anomaly detection.

malicious intrusion or attack on network vulnerabilities, medical and trajectory anomaly detection. Wang et al. [67]
computers or information systems may result in a serious study the application of GANs in fraud and intrusion detec-
predicament and violate the confidentiality, integrity and tion, and Liu et al. [120] evaluate their approach in medicine,
availability of the systems [155]. The examined primary stud- image recognition, aviation, human activity, spam identifica-
ies are mainly focused on network intrusion detection [83], tion, and waveform anomaly detection.
[87], [106], [110], [117], [119], [121], [128], [129], [131].
Other applications of GANs in intrusion detection are smart- 5) SYSTEM HEALTH ANOMALY DETECTION
phone lock pattern intrusion detection [112], presentation
System health monitoring is a way to identify anomalous
attack detection [82], phishing detection [114], cognitive
behaviour in large (often industrial) systems. In industrial
radio intrusion detection [76], cyber-physical system intru-
processes, the anomalous behaviour can represent, for exam-
sion detection [93], and IoT security [19], [46].
ple, wear or damage to the industrial equipment after contin-
4) GENERAL APPROACHES
uous use. It is critical that such degradations in a system’s
performance are detected before they escalate and cause
Some primary studies do not focus on a single application
loss of revenue or endanger human life. Examples of indus-
domain. Instead, they evaluate the proposed approaches in
trial applications of system health anomaly detection with
different application domains (shown as Various in Table 5).
GANs include industrial process anomaly detection [16],
For example, three primary studies [38], [50], [51] investigate
[28], electrical insulator anomaly detection [116], rolling
their proposed GAN-based anomaly detection for intrusion
bearing anomaly detection [53], steam turbine anomaly
detection and image recognition. Two evaluated studies [79],
detection [58], magnetic flux leakage detection [139], fused
[107] apply GAN-based anomaly detection in image recog-
magnesium furnace anomaly detection [125], railway turnout
nition and video surveillance applications. Other primary
anomaly detection [92] and communication system anomaly
studies evaluate their anomaly detection approach in intrusion
detection [100].
detection, medical and image recognition domains [62], [71].
Zhu et al. [52] investigate the application of their pro-
posed GAN-based anomaly detection in medicine and on 6) IMAGE RECOGNITION
trajectory anomaly detection. Oh et al. [111] evaluate their Image anomaly detection refers to finding images with abnor-
proposed technique for image recognition in addition to mal patterns that do not comply with other images in the

161012 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

same set. Most primary studies in this application domain use advertisement fraud [113], stock market manipulation [45],
public image datasets, such as MNIST or CIFAR-10, to prove credit card fraud [126], health care insurance providers
the concept of their proposed anomaly detection techniques. fraud [130], and satellite image forgery [103].
However, Moussa and Lim [136] evaluate the application
of GANs for object recognition in images, such as finding 11) OTHER DOMAINS OF ANOMALY DETECTION WITH GANs
an airplane in the picture. Bergmann et al. [66] propose a There are several additional application domains for anomaly
dataset of high-resolution color images of different object and detection using GANs that are less common: trajectory
texture categories suitable for anomaly detection. They eval- anomaly detection [61], [115], human mobility anomaly
uate several anomaly detection techniques, including GANs, detection [115], climate change [142], text anomaly detec-
to process their dataset. The proposed dataset aims to provide tion [68], and software systems anomaly detection [34], [94].
more challenging images than the commonly used datasets
mentioned above. C. RQ3: WHICH GAN ARCHITECTURE IS USED MOST
OFTEN IN ANOMALY DETECTION SYSTEMS?
7) MANUFACTURING ANOMALY DETECTION Many types of GANs have been proposed to tackle
This anomaly detection application refers to the quality the deficiencies of the first type of GAN proposed by
inspection of manufactured products to identify defective Goodfellow et al. [7] or to handle specific tasks. In most
products. These defects reveal themselves as irregularities cases, they modify the GAN architecture or the cost function
on metal or wood surfaces, electronic parts, and so on. For of the generator and discriminator. According to the GAN
example, an application of visual surface defect detection is Zoo GitHub repository,5 more than 500 types of GAN were
studied in four primary studies [33], [44], [69], [124] and identified from 2014 to 2018.
industrial quality inspection is investigated in three We identified 21 different types of GANs used for anomaly
studies [40], [77], [99]. detection purposes (see Table 6 for a list of primary studies
using each of these architectures). DCGANs, standard GANs,
8) ANOMALY DETECTION IN AUTONOMOUS SYSTEMS and cGANs are the most commonly used GAN architec-
Autonomy is defined as self-governance or freedom from tures. These were among the first proposed GAN architec-
external influences [156]. An autonomous system is referred tures, and there are many new ones which are not (yet)
to as a system that can perceive the environment, make deci- used for anomaly detection purposes. The correspondence
sions based on the sensed information, and then react to inter- between the identified GAN architectures and their appli-
nal/external changes using actuators. However, a fault may cation domains is shown in Table 7. DCGANs, standard
occur in each of these steps. For example, in an autonomous GANs, and BiGANs have been used in various application
robot, faults can occur in sensors, software, or after phys- domains, indicating their flexibility. A variety of GAN archi-
ical damage to the actuator. This domain includes driving tectures have also been used for applications in medicine,
anomaly detection [23], [84], [118] to assist the driver or to intrusion detection, and system health. However, some of
identify abnormalities in the driver’s behaviour. Autonomous the application domains are not well researched regarding
surveillance with moving agents is addressed in three primary GAN architectures, such as text anomaly detection and fraud
studies [31], [72], where an autonomous moving agent, such detection.
as a patrol robot, scans the environment to find abnormal Since the anomaly detection techniques examined in
activity. Two primary studies focused on controller anomaly this review are either based on or assisted by GANs,
detection [56], [73], to identify abnormal decision making by any deficiency in the networks used for anomaly detec-
a controller in a closed-loop control system. Sun et al. [60] tion directly impacts the performance of the corresponding
study autonomous vehicle anomaly detection. anomaly detection techniques. Therefore, the improvements
in anomaly detection techniques using GANs are strongly
9) POWER/ENERGY ANOMALY DETECTION correlated with the advances in the GAN architecture and
This application domain is concerned with identifying abnor- training strategies. There are many studies in the literature
malities in the power/energy consumption and power/energy describing the challenges of existing GANs and available
infrastructure. Examples include catenary support component solutions [12], [177]–[179]. The most crucial problem with
anomaly detection [32], [35], [49], power plant anomaly GANs is the problem of mode collapse. When this happens,
detection [48], [127] and power consumption anomaly detec- the generator of a GAN always generates samples from
tion [102]. a highly concentrated distribution (partial collapse) [12],
or simply a single sample (complete collapse) [180], [181].
10) FRAUD DETECTION Therefore, the generated data lacks the expected diver-
Fraud is defined as exploiting one’s occupation for per- sity. There have been several treatments proposed to lessen
sonal enrichment by willful misuse or application of their the effect of mode collapse during GAN training, such as
employer’s resources or assets without authorization [157]. WGAN [160], and Unrolled GAN [182]. Another challenge
Fraud detection refers to uncovering these illegal activi-
ties. Examples of applications in this domain include click 5 https://github.com/hindupuravinash/the-gan-zoo

VOLUME 9, 2021 161013

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 6. Type of GAN.

TABLE 7. Type of GAN used in each application domain. AS: autonomous systems, CC: climate changes, FD: fraud detection, HI: hyperspectral images, IR:
image recognition, ID: intrusion detection, MA: manufacturing, ME: medical, PE: power/energy, SS: software systems, SU: surveillance,SH: system health,
TE: text, TD: trajectory detection, VA: various.

of training of GANs is the instability of the training process The architectures of the top five GAN variants are shown
and its failure to converge to a Nash equilibrium. Methods in Figure 7. We elaborate on the most used architectures,
proposed to address this problem include Two Time-Scale and what makes them suitable for anomaly detection in the
Update Rule (TTUR) [183], WGAN [160] and feature match- following subsections. In addition, we discuss if and how they
ing [182]. deal with the challenges mentioned above.

161014 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

FIGURE 7. The architectures of the most commonly used generative adversarial networks: (a) the standard GAN, where G and D denote the generator and
discriminator; note that the architecture of the Wasserstein GAN is the same; (b) conditional GAN; the only difference is the addition of auxiliary data,
shown as c; (c) deep convolutional GAN (DCGAN), where Deconv denotes a deconvolutional layer, Conv a convolutional layer, and FC a fully connected
layer; (d) BiGAN, with E denoting the encoder; (e) VAE-GAN, where a variational autoencoder is combined with a GAN; (f) Cycle-GAN, which contains two
mapping functions G : X → Y and F : Y → X ; the associated adversarial discriminators are shown as DY and DX .

1) STANDARD GENERATIVE ADVERSARIAL NETWORKS does not provide any inference model to directly capture the
(GAN) inverse mapping. Hence, further training is needed to attain
In the first GAN architecture, proposed by this inference model, adding to the computational cost of
Goodfellow et al. [7], the generator and discriminator models the GAN training. Moreover, as standard GAN uses MLP
are defined by fully connected multilayer perceptrons. For in the generator and discriminator models, it is not suitable
the generator model, to learn the distribution of the generator for high dimensional data such as images. This is because
pg over data x, a prior on input noise variable pz (z) must MLPs are fully connected networks that require optimization
be defined. This mapping is represented as G(z; θg ), where of many parameters. Therefore, more efficient GAN architec-
G is a differentiable function represented by a multilayer tures (such as DCGANs) are preferred for images and other
perceptron with parameters θg . For the discriminator model, high-dimensional data.
another multilayer perceptron D(z; θd ) is defined. Its single
scalar output represents the probability that x comes from the 2) CONDITIONAL GENERATIVE ADVERSARIAL NETWORKS
data rather than pg . The training goal for the discriminator (cGAN)
model D is to maximize the probability of assigning the Mirza and Osindero proposed the conditional GAN [158]
correct label to both training examples and samples from the as an extension to the standard GAN that can control what
generator model G. Simultaneously, G is trained to minimize type of data is generated. For example, a condition can be
log(1 − D(G(z)). D and G are pitted against each other specified to generate only data of a certain class or type.
following a two-player minimax game. The conditional model of GAN can be obtained if both the
The Standard GAN optimizes the Jensen-Shannon (JS) generator and the discriminator are conditioned on some
divergence to learn the distribution of the data. Consequently, additional information y fed through additional input layers.
it suffers from an unstable, weak signal when the discrimina- There is no limitation on the type of the data; for example,
tor is approaching a local optimum, known as the problem it can contain class labels or data from different sources [158].
of gradient vanishing [176]. This can also lead to mode The conditional data generation is advantageous for anomaly
collapse. Another problem of the standard GAN is that it detection purposes since cGANs can better generate data
VOLUME 9, 2021 161015
 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

from different sources, i.e. multimodal data generation, or it clipping used in its discriminator made it difficult to con-
can be used for multimodal anomaly detection. verge. Gulrajani et al. [161] proposed an improved version
WGAN-GP introducing a gradient penalty to the discrimina-
3) DEEP CONVOLUTIONAL GENERATIVE ADVERSARIAL tor model of WGAN instead of weight clipping. This results
NETWORKS (DCGAN) in better convergence, training speed, and sample quality by
Striving to bridge the rift between the success of Convo- forcing the discriminator to learn relatively smoother decision
lutional Neural Networks (CNNs) for supervised learning boundaries [179]. This improved version of WGAN is already
and unsupervised learning, Radford et al. [144] introduced used by several studies for anomaly detection (see Table 6).
DCGANs, which integrate convolutional neural networks
into the standard GAN. DCGANs provide a better network 6) VARIATIONAL AUTOENCODER GANs (VAE-GANs)
topology for more stable GAN training. The optimization and Larsen et al. [162] proposed a new GAN architecture that
training processes are the same as for the standard GANs. combines a variational autoencoder (VAE) with a GAN.
However, Radford et al. proposed several improvements to A VAE-GAN replaces the decoder of a VAE with a generator
the CNNs and Standard GANs. of a GAN and modifies the loss function to be computed
These modifications are: (1) using all convolutional by a discriminator (see Figure 7e). The rationale behind this
nets [184] in the generator and discriminator, (2) removing modification is to use the discriminator model to evaluate
fully connected layers on top of the convolutional layer, and the similarity of the reconstructed image and the original
(3) using batch normalization [185]. These changes result in a one [98]. A VAE-GAN outperforms VAEs in terms of visual
better model and training stability with deeper gradient flow fidelity, and Larsen et al. demonstrate how to use the learned
through the network, preventing mode collapse. feature representation in the GAN discriminator as the basis
DCGANs were originally designed for image processing for the VAE reconstruction objective. They replaced the
since they employ CNNs. The CNNs allow DCGANs to learn element-wise errors with feature-wise errors, leading to a
a hierarchy of representations from object parts to scenes in better performance in learning the distribution of the data.
both the generator and discriminator, which makes DCGANs
well suited for image anomaly detection. 7) CYCLE-GAN
Zhu et al. [163] proposed the Cycle-GAN to learn the map-
4) BI-DIRECTIONAL GENERATIVE ADVERSARIAL NETWORKS ping (G) between a source domain (X ) to a target domain (Y )
(BiGAN) in the absence of paired examples. The goal of a Cycle-GAN
The bi-directional GAN [159] adds an autoencoder that learns is to learn a mapping G : X → Y in a way that the dis-
the mapping of data x to the latent representation z (infer- tribution of images from G(X ) is indistinguishable from the
ence), which makes it well suited for anomaly detection. distribution of Y , using adversarial loss. Since this mapping
BiGANs do not make any assumptions about the nature or is extremely under-constrained, a Cycle-GAN uses an inverse
structure of data. As a result, they provide a general, robust mapping F : Y → X and introduces the cycle consistency
approach for unsupervised representation learning capable of loss to enforce F(G(X )) ≈ X (and vice versa). A Cycle-
capturing semantic attributes of the data [159]. Donahue et al. GAN uses two different GANs coupled together to perform
empirically show that, despite their generality, BiGANs are this transformation. It uses cycle-consistency loss to preserve
competitive with the state-of-the-art approaches to perform the original image after a cycle of translation between two
self-supervised and weakly supervised feature learning tasks. domains. In the training stage, the first discriminator tells
Comparing BiGANs with the standard GAN, the inference whether the original image belongs to the source domain. The
mechanism, i.e., feature learning, of BiGANs makes it suit- same thing happens for the target domain discriminator. In the
able for anomaly detection techniques since they can be investigated primary studies, the Cycle-GAN is only used as
immediately used to generate anomaly scores. a data augmentation technique.

5) WASSERSTEIN GENERATIVE ADVERSARIAL NETWORKS D. RQ4: WHICH TYPE OF DATA INSTANCE AND DATASETS
(WGAN) ARE MOST COMMONLY USED FOR ANOMALY DETECTION
In an attempt to alleviate the problem of mode collapse and WITH GANs?
the challenges of standard GANs to converge to the Nash We identified six types of input data used for anomaly detec-
equilibrium, Arjovsky et al. [160] suggested using the Earth- tion with GANs. As shown in Table 8, image is by far the most
Mover (EM) distance or Wasserstein-1 distance instead of common type, appearing in 50% of the examined papers. The
JS divergence used in the standard GAN. Unlike DCGAN, two most common application domains for anomaly detection
WGAN attempts to enhance the stability of GANs by modi- with GANs are related to images: medicine and surveillance.
fying the adversarial cost function. Arjovsky et al. show that Tabular data is second (26%), followed by video, time series,
these distances provide gradients that are more useful for text, and frequency data.
updating the generator than the JS divergence function [160]. Data preprocessing is a key element that determines the
Although WGAN better handles the problem of mode col- success or failure of many deep learning models [41],
lapse compared to standard GANs and DCGANs, the weight [110], [112]. We identified 22 types of data preprocessing

161016 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 8. List of different data types.

TABLE 9. List of different preprocessing types with corresponding application to different data types.

techniques, summarized in Table 9. Owing to images images, different types of moving objects, and the presence
being the most common data type, resizing, normaliza- of one or more anomalies in the scene.
tion, and cropping are the most prevalent preprocessing The MNIST and CIFAR-10 datasets both appear in 7% of
techniques. These techniques make data more uniform by the examined studies. The MNIST database of handwritten
changing its range and scale. A normalized dataset also digits has a training set of 60,000 examples and a test set
speeds up learning. Preprocessing is commonly applied to of 10,000 examples. The CIFAR-10 dataset is a collection
image, tabular and time series data, but rarely to other of 60,000 colour images arranged in 10 object classes of
types of data. It is also worth noting that some stud- equal size. When MNIST and CIFAT-10 are used in anomaly
ies [76], [117] first transform tabular or frequency data detection studies, one class is simulated as abnormal and
to images before applying other types of preprocessing removed from the training class, while the remaining classes
techniques. are treated as normal.
Tables 10-a and 10-b show the datasets used in the primary The UMN crowd dataset [189] is used in 4% of the
studies as well as their associated application domains. The examined papers. It contains normal and abnormal crowd
‘‘custom dataset’’ stands for a dataset that was either con- behaviour captured in indoor and outdoor scenes of the Uni-
structed by the study authors or that contains proprietary data versity of Minnesota. The dataset contains 11 videos with
not released to the public domain. These tables show that a total of 7,736 frames that were captured under several
the majority of the utilized datasets are custom, while UCSD scenarios at three different indoor and outdoor scenes.
pedestrian [186], MNIST [187], and CIFAR-10 [188] are the From the perspective of application domain, we found
most commonly used publicly available datasets. that studies on fraud detection use the Credit Card Fraud
The UCSD anomaly detection dataset was acquired by Detection dataset, real world credit (RWC) dataset, Talking-
a stationary camera that captures pedestrian walkways. Data AdTracking, UCI dataset and a custom dataset. For
It includes two subsets: Ped1 with 34 training and 36 testing surveillance anomaly detection purposes, 74% studies use the
video sequences, and Ped2 with 16 training and 12 test- CUHK avenue, ShanghAaiTech, UCSD, and UMN datasets.
ing video sequences. In the normal setting, the video in Among all datasets identified in the primary studies, twelve
this dataset contains only pedestrians. Abnormal events are used for intrusion detection, seven for manufacturing
occur when either nonpedestrian entities are in the walkway, anomaly detection, fifteen for medical anomaly detection,
or there are anomalous pedestrian motion patterns, such as and twelve for image anomaly detection. While most datasets
people walking across a walkway or in the grass that sur- are not used across all domains, eighteen are used in multiple
rounds it. This dataset is challenging due to the low-resolution domains as highlighted in Table 10-a and Table 10-b. For

VOLUME 9, 2021 161017

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 10. a: The list of identified datasets from primary studies. AS: autonomous systems, CC: climate changes, FD: fraud detection, HI: hyperspectral
images, IR: image recognition, ID: intrusion detection, MA: manufacturing, ME: medical, PE: power/energy, SS: software systems, SU: surveillance, SH:
system health, TE: text, TD: trajectory detection, VA: various. Note that the datasets highlighted in bold are applied in more than one domain. b: The list of
identified datasets from primary studies. AS: autonomous systems, CC: climate changes, FD: fraud detection, HI: hyperspectral images, IR: image
recognition, ID: intrusion detection, MA: manufacturing, ME: medical, PE: power/energy, SS: software systems, SU: surveillance, SH: system health, TE:
text, TD: trajectory detection, VA: various. Note that the datasets highlighted in bold are applied in more than one domain.

161018 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 10. (Continued.) a: The list of identified datasets from primary studies. AS: autonomous systems, CC: climate changes, FD: fraud detection, HI:
hyperspectral images, IR: image recognition, ID: intrusion detection, MA: manufacturing, ME: medical, PE: power/energy, SS: software systems, SU:
surveillance, SH: system health, TE: text, TD: trajectory detection, VA: various. Note that the datasets highlighted in bold are applied in more than one
domain. b: The list of identified datasets from primary studies. AS: autonomous systems, CC: climate changes, FD: fraud detection, HI: hyperspectral
images, IR: image recognition, ID: intrusion detection, MA: manufacturing, ME: medical, PE: power/energy, SS: software systems, SU: surveillance, SH:
system health, TE: text, TD: trajectory detection, VA: various. Note that the datasets highlighted in bold are applied in more than one domain.

instance, KDD-cup99 10% is used in both image recogni- the similarity of two samples through the ratio of the total
tion [50] and intrusion detection [51] while MNIST is used number of pixels divided and the mean squared error between
in image recognition [44], [47], [51], [63], [86], [96], [108], the original and generated images. A higher value of the
[115] and manufacturing anomaly detection [71]. PSNR indicates that the generated sample is closer to the
original.
E. RQ5: WHICH METRICS ARE USED TO EVALUATE THE The Fréchet inception distance (FID), adopted by two
PERFORMANCE OF GANs IN GENERATING DATA AND papers [82], [124], is a widely used evaluation method
ANOMALY DETECTION? for evaluating the diversity and similarity of generated
We found that 27 out of 128 primary studies evaluated images [183]. By calculating and comparing the feature vec-
the quality of the generated samples using 9 different per- tors of a collection of real and generated images, FID can
formance evaluation metrics. Most studies evaluated data measure the distance between the real and generated distri-
quality quantitatively, while six papers implemented visual bution.
inspection to evaluate the quality of the generated sam- We also studied whether performance metrics were used
ples [17], [40], [60], [69], [135], [139]. During the inspec- with specific input data types. As shown in Table 11,
tions, the generated samples were examined by application we observed that most performance metrics have been
domain experts, or simply the authors of the individual stud- applied to image and video data, while only two papers
ies. Quantitative evaluation was performed using eight per- utilized metrics for time series data [53], [60]. It is also worth
formance metrics, most commonly the structural similarity mentioning that frame data is usually extracted from the video
index measure (SSIM) and peak signal-to-noise ratio (PSNR) before being fed into the GAN. Therefore, the performance
that were each used in 26% of the studies that evaluated metrics are essentially used only to evaluate the image data.
performance. We also examined the relationship between the performance
SSIM, adopted by seven papers [26], [55], [56], [63], metrics and application domains. Table 12 shows that most
[65], [98], [99], quantifies the relative perceptual similarity metrics were adopted in surveillance anomaly detection to
between two images. This metric ranges from -1 to 1, with evaluate the generated samples, while most domains such as
1 indicating a perfect pixel match between the original and fraud detection, power/energy anomaly detection, and soft-
generated samples, -1 corresponding to inverted images, and ware systems did not evaluate the quality of the generated
0 marking no similarity [190]. Seven papers [21], [22], [24], samples at all. In addition, SSIM, PSNR, FID, and visual
[41], [44], [85], [137] used PSNR as a metric to measure inspection were used in various domains, while other metrics
the quality of the generated images. This metric evaluates were only applied to one specific domain.

VOLUME 9, 2021 161019

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 11. The performance metrics that are used to evaluate generated samples with correspondent input types. Note that there are no performance
metrics used for evaluating tabular, text and frequency data.

TABLE 12. The performance metrics that are used to evaluate generated samples with correspondent application domains. AS: autonomous systems, CC:
climate changes, FD: fraud detection, HI: hyperspectral images, IR: image recognition, ID: intrusion detection, MA: manufacturing, ME: medical, PE:
power/energy, SS: software systems, SU: surveillance, SH: system health, TE: text, TD: trajectory detection, VA: various.

The actual anomaly detection performance in the stud- is available. A training dataset is used to train the model
ies is evaluated using different (usually more traditional) for predicting the class labels, and then the predictive model
metrics. The area under the receiver operating characteristic is evaluated on an unseen test dataset. Dependence of the
curve (AUROC) is the most frequently used metric to eval- supervised anomaly detection techniques on the data labels
uate anomaly detection (in 68 primary studies). Precision, makes them more vulnerable to the problem of an imbalanced
F1-score, accuracy, recall, sensitivity, equal error rate (EER), dataset. Therefore, these anomaly detection techniques are
and specificity are also frequently used as metrics in this area. biased toward classifying the majority class.
In 26.3% of the investigated primary studies, GANs are
F. RQ6: WHICH ANOMALY DETECTION TECHNIQUES ARE applied to address the problem of the imbalanced dataset for
USED ALONG WITH GANs? anomaly detection by augmenting it with the data generated
This section discusses the different types of anomaly detec- by GANs. The list of primary studies that use GANs for data
tion techniques that either use GANs or are compared with augmentation is shown in Table 2. The list of papers that used
GANs. Based on the labelled data availability, anomaly detec- GANs along with supervised anomaly detection techniques,
tion techniques are divided into three classes: supervised, i.e. GAN-assisted approaches, is shown in Table 13.
semi-supervised, and unsupervised anomaly detection. Dur-
ing data synthesis for RQ6, we noticed that not all primary 2) SEMI-SUPERVISED ANOMALY DETECTION
studies use consistent definitions for these classes. Therefore, Anomaly detection techniques trained in a semi-supervised
we use Chandola et al.’s [153] definition of supervised, semi- manner assume that the labelled data is available only for the
supervised, and unsupervised anomaly detection. In addition, normal class. The main benefit of semi-supervised anomaly
as there is a wide variety of anomaly detection techniques, detection techniques is that they do not require data for the
we only considered the techniques used in more than three anomalous class. In the reviewed primary studies, GANs
primary studies. are mostly used in a semi-supervised manner. By training a
GAN to learn the distribution of the normal class, a deviation
1) SUPERVISED ANOMALY DETECTION from the normal distribution is identified using an anomaly
Anomaly detection techniques trained in a supervised fashion scoring technique. The list of GAN-based semi-supervised
assume that labelled data for the normal and anomaly classes anomaly detection techniques is shown in Table 14. From all

161020 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 13. GAN-assisted supervised anomaly detection techniques.

TABLE 14. Semi-supervised GAN-based anomaly detection techniques.

TABLE 15. Semi-supervised anomaly detection techniques compared to GANs.

GAN-based techniques, AnoGAN [104] is the GAN-based to the performance of GANs in anomaly detection such
technique most often used as a baseline for comparison with as standard AEs [194], Variational AEs (VAEs), convolu-
newly proposed methods. There are several other techniques tional AEs (CVAEs) [195], Denoising AEs (DAEs) [196],
that can be used for anomaly detection purposes in a semi- and Adversarial AEs (AAEs) [167]. Moreover, some of the
supervised manner, as listed in Table 15. In investigated primary studies compared the proposed anomaly detection
primary studies, the performance of these anomaly detection techniques with a Long Short Term Memory (LSTM) based
techniques is evaluated and compared to the GAN-based approach [197].
techniques. The experimental results of the primary studies
show that the GAN-based anomaly detection techniques have 3) UNSUPERVISED ANOMALY DETECTION
as good as or superior performance over these competing These types of anomaly detection techniques do not require
techniques. a labelled dataset. This is based on the central assumption
The table shows that several papers used Mixture of that normal instances are far more frequent than anomalies in
Dynamic Texture (MDT) [191] as an anomaly detec- the test data [153]. However, if this assumption is not valid,
tion technique in crowded scenes. Mehran et al. [189] use the anomaly detection will significantly suffer from false
a generative probabilistic model called Social Force for alarms. Assuming that the unlabeled dataset contains very
semi-supervised anomaly detection. Two primary studies few anomalous instances and the model is robust against these
investigate sparse dictionary learning-based anomaly detec- few anomalies, we can adapt a semi-supervised anomaly
tion techniques, detection at 150 FPS [192] and sparse recon- detection technique to work in an unsupervised manner by
struction [193]. In the primary studies, the performance training the model on a portion of the unlabeled dataset. A list
of different types of autoencoders (AEs) were compared of GAN-based unsupervised anomaly detection techniques

VOLUME 9, 2021 161021

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

TABLE 16. Unsupervised anomaly detection techniques based on GANs.

TABLE 17. Unsupervised anomaly detection techniques compared to GANs.

is presented in Table 16. In addition, Table 17 presents a of resource consumption [22], [28], [83], [89], [127]. For
list of unsupervised anomaly detection techniques that have instance, the effects of selecting GAN hyperparameters on the
been considered for anomaly detection and compared to anomaly detection performance have not been studied in the
GANs in the literature. The results of the investigated primary literature. There is also a need to consider the use of emerging
studies demonstrate that their proposed GAN-based anomaly GAN optimization and training methods, e.g. [180], [182], for
detection techniques achieve superior performance compared better training stability and faster convergence.
to the techniques presented in Table 17.
From Table 17, we can observe that one-class classi- B. FUTURE DIRECTION 2: ACCOUNTING FOR CHANGING
fiers [198] have been of great interest from an unsuper- BEHAVIOUR OF A SYSTEM
vised anomaly detection perspective. Isolation forest [199] is In most industrial anomaly detection applications, behaviour
another unsupervised technique competing with GANs in this of the target system varies over time. Therefore, it is crucial
area. Several variants of principal component analysis [200] to examine the temporal behaviour of the system to find
have also been compared often to GANs in terms of perfor- anomalies. RQ4 showed that only 7% of the primary stud-
mance. Several linear models have also been applied, namely ies used GANs for anomaly detection in time series data.
REAPER [201] and Low Rank Representation [202]. Other Therefore, more studies are required on anomaly detection
techniques compared to GANs for anomaly detection include using GANs for time series data to make them suitable for
Gaussian mixture models [203], R-graph [204], local outlier industrial applications, especially for multivariate time series
factor [205], deep structured energy-based models [206], and data. In many industrial applications, data is collected online.
deep autoencoding Gaussian mixture models [207]. Huang and Lei [110] suggest to take advantage of this data via
online training of GAN-based anomaly detection techniques.
V. FUTURE RESEARCH DIRECTIONS This approach might be adaptable for more real-time anomaly
Generative adversarial network-based anomaly detection is in detection tasks.
its early stage of development with many research opportu-
nities. However, most of these opportunities lie in the field C. FUTURE DIRECTION 3: IMPROVING SUPPORT FOR
of GANs itself. In this section, we present possible directions MULTIMODAL, DISCRETE AND NOISY DATA
for the future work of applying GANs in anomaly detection. Another open challenge of using GANs for anomaly detec-
tion is the lack of studies on multimodal anomaly detection
A. FUTURE DIRECTION 1: SPEEDING UP THE GAN using GANs. In real-world cases, data often comes from mul-
TRAINING PROCESS tiple sources with different types. For instance, Qiu et al. [23]
Training GANs is a computationally demanding task. propose a GAN-based driving anomaly detection technique
As reported in almost all primary studies, it takes a long time using physiological and CAN-bus data. Qiu et al. suggest
and powerful GPUs to train GANs to the point of satisfactory incorporating other information such as results obtained from
performance. Consequently, future studies need to explore vision-based object detection systems applied to the road.
GAN architectures that are lightweight and efficient in terms Many other GAN-based anomaly detection approaches could

161022 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

benefit from using multimodal data. In addition, GANs were detection, e.g. [55], [67], [89]. Future studies may exam-
initially created to generate continuous data. As a result, they ine the effect of using improved GAN architectures, e.g.,
have limited ability to deal with discrete data, as it hinders the to improve the resolution of images using SRGAN [208],
backpropagation process [68]. Fadhel and Nyarko [68] point ESRGAN [209], or BigGAN [210], on the performance of
out this problem and study GAN architectures suitable for anomaly detection techniques.
discrete data. Despite the promising results they report, this
study is the only example of anomaly detection for discrete VI. THREATS TO VALIDITY
data in our review. Finally, Lei et al. [22] used optical flow as One threat to the validity of our review is that of miss-
foreground shape information for video anomaly detection. ing papers. The source of this threat is selecting the search
Lei et al. point out that when the optical flow is inaccurate, keywords and the search engines. To mitigate this threat,
it will affect the robustness of their proposed GAN-based we iteratively added keywords to our search query until no
anomaly detection technique. Thus, a potential direction for relevant new papers were found.
future research is the study of the effects of measurement Moreover, the list of papers was finalized on the 3rd of
inaccuracies and noise in the data on the performance of June, 2020, and no papers that were published afterwards
GAN-based anomaly detection techniques and the develop- were added. In a fast-moving field in which many papers
ment of solutions to alleviate their impact. are published, such as anomaly detection using GANs, it is
impossible to include all literature up to the date of submis-
D. FUTURE DIRECTION 4: SEARCHING FOR BETTER sion of the review. Due to the amount of necessary work
ANOMALY SCORING METHODS to conduct the SLR, there will always be a (considerable)
As mentioned earlier, GAN-based anomaly detection tech- amount of time between the cutoff date for the data collection
niques require an anomaly scoring method to distinguish and the submission of the review. Hence, it is possible that
between normal and abnormal samples. The selection of there are new GAN-based anomaly detection techniques that
anomaly metrics for anomaly scores is still a challenging address some of the issues or challenges identified in this
task. Further investigation is needed to improve the scoring review.
methods and to identify the best fit for each application Some recent works on the applications of GANs in
domain or for a specific case [28], [63], [99]. anomaly detection which were published after our data col-
lection ended include work in the following areas: medical
E. FUTURE DIRECTION 5: IMPROVING THE applications [211]–[213], surveillance [214]–[216], intrusion
PERFORMANCE EVALUATION OF GANs detection [217]–[219], hyperspectral imaging [220]–[223],
It is essential to evaluate the performance of GANs in gener- and manufacturing [224]. In the next literature review on
ating data before using them for anomaly detection, either in the applications of GANs in anomaly detection, these studies
a GAN-assisted or GAN-based setup. By doing so, one can need to be covered as well. Our paper is the first to systemat-
ensure that GAN has learned the distribution of the data cor- ically review the applications of GANs in anomaly detection
rectly. The results of RQ5 revealed that most primary studies (up to June 3, 2020) and should be used as the first building
do not evaluate the performance of their GAN-generated data. block towards building a concise survey of the available work
Additionally, almost all cases that assess data performance on this topic.
use image data. For other types of data, such as tabular, Also, the data synthesis of the RQs was divided between
text and time series, there is no performance indicator of the two first authors. To reduce bias in the data synthesis step,
the generated data quality. Therefore, additional research is the first two authors met regularly to address disagreements.
needed to identify the most suitable metrics for assessing the If a disagreement could not be resolved, one of the last two
performance of GANs for each data type. authors made the final decision.

F. FUTURE DIRECTION 6: EMPLOYING IMPROVED GAN VII. CONCLUSION

ARCHITECTURES FOR ANOMALY DETECTION This systematic literature review presents an extensive study
We observed in RQ3 that the ‘older’ GAN architectures are by on the applications of GANs for anomaly detection, covering
far the most popular in anomaly detection studies. However, 128 primary studies. We define and answer several RQs to
many improved GAN architectures were proposed recently, capture the current best practices and available techniques to
which could improve anomaly detection as well. For exam- employ GANs for anomaly detection purposes. We also iden-
ple, it is desirable to generate high-resolution images with tify the existing challenges and provide six future research
GANs. However, it is a challenging task. High-resolution directions in this area.
images make it easier for the discriminator to tell apart the The results reveal that GANs are used for GAN-assisted
generated images from the training samples [164]. Also, (data augmentation) and GAN-based (representation learn-
high-resolution images require more memory storage, which ing) anomaly detection. In both cases, the problem of insuf-
leads to smaller minibatches and may compromise train- ficient amount of data for the anomalous behaviour of the
ing stability [165]. Several primary studies highlighted the system is addressed. In a GAN-assisted approach, the goal
need to generate high-resolution images for better anomaly is to augment the minority class using the generative ability

VOLUME 9, 2021 161023

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

of GANs. In GAN-based anomaly detection, the goal is to use [6] N. Mould, J. L. Regens, C. J. Jensen, and D. N. Edger, ‘‘Video surveillance
the representation learning ability of GANs, eliminating the and counterterrorism: The application of suspicious activity recognition
in visual surveillance systems to counterterrorism,’’ J. Policing, Intell.
need for minority class data. The most commonly used GAN Counter Terrorism, vol. 9, no. 2, pp. 151–175, Jul. 2014.
architectures in the primary studies are DCGANs, standard [7] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
GANs, and cGANs. GANs are applied for anomaly detection S. Ozair, A. Courville, and Y. Bengio, ‘‘Generative adversarial nets,’’ in
Proc. Adv. Neural Inf. Process. Syst., 2014, pp. 2672–2680.
in a wide range of application domains. The primary areas in [8] J. Langr and V. Bok, GANs in Action: Deep Learning With Generative
which GANs are used for anomaly detection are medicine, Adversarial Networks. Shelter Island, NY, USA: Manning Publications,
surveillance and intrusion detection. However, their applica- 2019.
[9] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, ‘‘Learning repre-
tion in many other domains, such as anomaly detection in sentations by back-propagating errors,’’ Nature, vol. 323, pp. 533–536,
sensor networks, smart grids, and cloud computing shows that Oct. 1986.
GANs are a suitable solution for anomaly detection. [10] B. Kitchenham, Procedures for Performing Systematic Reviews, vol. 33.
Keele, U.K.: Keele Univ., 2004, pp. 1–26.
We identified six important directions for future research. [11] J. T. Springenberg, ‘‘Unsupervised and semi-supervised learning with
Some of these directions are related to fundamental GAN categorical generative adversarial networks,’’ 2015, arXiv:1511.06390.
research. For example, our study reveals that only 21% of the [12] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and
A. A. Bharath, ‘‘Generative adversarial networks: An overview,’’ IEEE
primary studies evaluated the quality of the data that was gen- Signal Process., vol. 35, no. 1, pp. 53–65, Jan. 2017.
erated with GANs. Hence, an important direction for future [13] B. J. B. Rani, ‘‘Survey on applying GAN for anomaly detection,’’ in Proc.
research is to investigate how the performance of GANs can Int. Conf. Comput. Commun. Informat. (ICCCI), Jan. 2020, pp. 1–5.
[14] M. L. McHugh, ‘‘Interrater reliability: The Kappa statistic,’’ Biochem.
be evaluated, as better performing GANs will also result in Med., vol. 22, no. 3, pp. 276–282, 2012.
better performing anomaly detection approaches. Another [15] Y. Bengio, A. Courville, and P. Vincent, ‘‘Representation learning: A
important future research direction is speeding up the training review and new perspectives,’’ IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 35, no. 8, pp. 1798–1828, 2013.
process of GANs. In addition, we identified several important [16] S. Plakias and Y. S. Boutalis, ‘‘Exploiting the generative adversarial
future research directions for anomaly detection researchers. framework for one-class multi-dimensional fault detection,’’ Neurocom-
In particular, GAN-assisted anomaly detection approaches puting, vol. 332, pp. 396–405, Mar. 2019.
[17] T. Schlegl, P. Seeböck, S. M. Waldstein, G. Langs, and
should improve their support for multimodal, discrete and U. Schmidt-Erfurth, ‘‘f-AnoGAN: Fast unsupervised anomaly detection
noisy data, and account for the changing behaviour of a with generative adversarial networks,’’ Med. Image Anal., vol. 54,
system. Finally, researchers should investigate how recent pp. 30–44, May 2019.
[18] S. You, B. H. Cho, S. Yook, J. Y. Kim, Y.-M. Shon, D.-W. Seo, and
improvements to GAN architectures can help improve their I. Y. Kim, ‘‘Unsupervised automatic seizure detection for focal-onset
role in anomaly detection. seizures recorded with behind-the-ear EEG using an anomaly-detecting
This systematic review of GAN literature has examined generative adversarial network,’’ Comput. Methods Programs Biomed.,
vol. 193, Sep. 2020, Art. no. 105472.
the fundamentals and recent advances in the area of GAN [19] Y. Zhang, Q. Ai, H. Wang, Z. Li, and X. Zhou, ‘‘Energy theft detection
applications in anomaly detection. With hundreds of new in an edge data center using threshold-based abnormality detector,’’ Int.
articles published every year, it can be expected that there J. Electr. Power Energy Syst., vol. 121, Oct. 2020, Art. no. 106162.
[20] H. Wu, J. Shao, X. Xu, F. Shen, and H. T. Shen, ‘‘A system for spatiotem-
will be an influx of new architectures and improved learning poral anomaly localization in surveillance videos,’’ in Proc. 25th ACM
methods. They promise to provide a powerful tool to generate Int. Conf. Multimedia, Oct. 2017, pp. 1225–1226.
realistic data across a broad range of problem domains. In the [21] A. Jamadandi, S. Kotturshettar, and U. Mudenagudi, ‘‘PredGAN: A
deep multi-scale video prediction framework for detecting anomalies in
context of anomaly detection, this is extremely valuable as
videos,’’ in Proc. 11th Indian Conf. Comput. Vis., Graph. Image Process.,
anomalous data is scarce and expensive to acquire. To ensure Dec. 2018, pp. 1–8.
the quality of the generated data, it is necessary to expand the [22] Z. Lei, F. Deng, and X. Yang, ‘‘Spatial temporal balanced generative
adversarial AutoEncoder for anomaly detection,’’ in Proc. Int. Conf.
variety of standard datasets for use in GAN training, as well
Image, Video Signal Process. (IVSP), 2019, pp. 1–7.
as to develop new performance metrics that are currently [23] Y. Qiu, T. Misu, and C. Busso, ‘‘Driving anomaly detection with
missing for many important domains. conditional generative adversarial network using physiological and
CAN-bus data,’’ in Proc. Int. Conf. Multimodal Interact., Oct. 2019,
pp. 164–173.
REFERENCES [24] K.-T. Nguyen, D.-T. Dinh, M. N. Do, and M.-T. Tran, ‘‘Anomaly
[1] Y. Sun, Z. Zhao, Z. Yang, F. Xu, H. Lu, Z. Zhu, W. Shi, J. Jiang, P. Yao, detection in traffic surveillance videos with GAN-based future
and H. Zhu, ‘‘Risk factors and preventions of breast cancer,’’ Int. J. Biol. frame prediction,’’ in Proc. Int. Conf. Multimedia Retr., Jun. 2020,
Sci., vol. 13, no. 11, p. 1387, 2017. pp. 457–463.
[2] R. L. Siegel, K. D. Miller, and A. Jemal, ‘‘Cancer statistics, 2020,’’ CA A, [25] L. Xu and Z. Xu, ‘‘One-class classification with deep adversarial learn-
Cancer J. Clinicians, vol. 70, no. 4, pp. 7–30, 2020. ing,’’ in Proc. 3rd Int. Conf. Comput. Sci. Artif. Intell., Dec. 2019,
[3] O. Ginsburg, C. H. Yip, A. Brooks, A. Cabanes, M. Caleffi, pp. 103–106.
J. A. D. Yataco, B. Gyawali, V. McCormack, M. M. de Anderson, [26] K. Zhou, S. Gao, J. Cheng, Z. Gu, H. Fu, Z. Tu, J. Yang, Y. Zhao,
R. Mehrotra, and A. Mohar, ‘‘Breast cancer early detection: A and J. Liu, ‘‘Sparse-GAN: Sparsity-constrained generative adversarial
phased approach to implementation,’’ Cancer, vol. 126, pp. 2379–2393, network for anomaly detection in retinal OCT image,’’ in Proc. IEEE 17th
May 2020. Int. Symp. Biomed. Imag. (ISBI), Apr. 2020, pp. 1227–1231.
[4] R. Mandal and N. Choudhury, ‘‘Automatic video surveillance for theft [27] K. Jiang, W. Xie, Y. Li, J. Lei, G. He, and Q. Du, ‘‘Semisupervised
detection in ATM machines: An enhanced approach,’’ in Proc. 3rd spectral learning with generative adversarial network for hyperspectral
Int. Conf. Comput. Sustain. Global Develop. (INDIACom), Mar. 2016, anomaly detection,’’ IEEE Trans. Geosci. Remote Sens., vol. 58, no. 7,
pp. 2821–2826. pp. 5224–5236, Jul. 2020.
[5] M. Ghazal, C. Vazquez, and A. Amer, ‘‘Real-time automatic detection of [28] X. Yang and D. Feng, ‘‘Generative adversarial network based anomaly
vandalism behavior in video sequences,’’ in Proc. IEEE Int. Conf. Syst., detection on the benchmark Tennessee Eastman process,’’ in Proc. 5th
Man Cybern., Oct. 2007, pp. 1056–1060. Int. Conf. Control, Autom. Robot. (ICCAR), Apr. 2019, pp. 644–648.

161024 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

[29] C. Zhang, Y. Wang, X. Zhao, Y. Guo, G. Xie, C. Lv, and B. Lv, [49] Y. Lyu, Z. Han, J. Zhong, C. Li, and Z. Liu, ‘‘A GAN-based anomaly
‘‘Memory-augmented anomaly generative adversarial network for retinal detection method for isoelectric line in high-speed railway,’’ in Proc.
OCT images screening,’’ in Proc. IEEE 17th Int. Symp. Biomed. Imag. IEEE Int. Instrum. Meas. Technol. Conf. (I2MTC), May 2019, pp. 1–6.
(ISBI), Apr. 2020, pp. 1971–1974. [50] Y. Hou, Z. Chen, M. Wu, C.-S. Foo, X. Li, and R. M. Shubair, ‘‘Maha-
[30] F. Luer, D. Mautz, and C. Bohm, ‘‘Anomaly detection in time series lanobis distance based adversarial network for anomaly detection,’’
using generative adversarial networks,’’ in Proc. Int. Conf. Data Mining in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP),
Workshops (ICDMW), Nov. 2019, pp. 1047–1048. May 2020, pp. 3192–3196.
[31] W. Lawson, E. Bekele, and K. Sullivan, ‘‘Finding anomalies with gener- [51] P. C. Ngo, A. A. Winarto, C. K. L. Kou, S. Park, F. Akram, and
ative adversarial networks for a patrolbot,’’ in Proc. IEEE Conf. Comput. H. K. Lee, ‘‘Fence GAN: Towards better anomaly detection,’’ in Proc.
Vis. Pattern Recognit. Workshops (CVPRW), Jul. 2017, pp. 12–13. IEEE 31st Int. Conf. Tools With Artif. Intell. (ICTAI), Nov. 2019,
[32] P. Yang, W. Jin, and P. Tang, ‘‘Anomaly detection of railway catenary pp. 141–148.
based on deep convolutional generative adversarial networks,’’ in Proc. [52] G. Zhu, H. Zhao, H. Liu, and H. Sun, ‘‘A novel LSTM-GAN algorithm
IEEE 3rd Adv. Inf. Technol., Electron. Autom. Control Conf. (IAEAC), for time series anomaly detection,’’ in Proc. Prognostics Syst. Health
Oct. 2018, pp. 1366–1370. Manage. Conf. (PHM-Qingdao), Oct. 2019, pp. 1–6.
[33] W. Zhai, J. Zhu, Y. Cao, and Z. Wang, ‘‘A generative adversarial network [53] W. Jiang, Y. Hong, B. Zhou, X. He, and C. Cheng, ‘‘A GAN-based
based framework for unsupervised visual surface inspection,’’ in Proc. anomaly detection approach for imbalanced industrial time series,’’ IEEE
IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), Apr. 2018, Access, vol. 7, pp. 143608–143619, 2019.
pp. 1283–1287. [54] Y. Yang, Z. Fu, and S. M. Naqvi, ‘‘Enhanced adversarial learning based
[34] T. Kumarage, S. Ranathunga, C. Kuruppu, N. D. Silva, and M. Ranawaka, video anomaly detection with object confidence and position,’’ in Proc.
‘‘Generative adversarial networks (GAN) based anomaly detection in 13th Int. Conf. Signal Process. Commun. Syst. (ICSPCS), Dec. 2019,
industrial software systems,’’ in Proc. Moratuwa Eng. Res. Conf. (MER- pp. 1–5.
Con), Jul. 2019, pp. 43–48. [55] H. Cheng, H. Liu, F. Gao, and Z. Chen, ‘‘ADGAN: A scalable GAN-
[35] Y. Lyu, Z. Han, J. Zhong, C. Li, and Z. Liu, ‘‘A generic anomaly based architecture for image anomaly detection,’’ in Proc. IEEE 4th Inf.
detection of catenary support components based on generative adversarial Technol., Netw., Electron. Autom. Control Conf. (ITNEC), Jun. 2020,
networks,’’ IEEE Trans. Instrum. Meas., vol. 69, no. 5, pp. 2439–2448, pp. 987–993.
May 2020. [56] N. Patel, A. N. Saridena, A. Choromanska, P. Krishnamurthy, and
[36] T. Jiang, Y. Li, W. Xie, and Q. Du, ‘‘Discriminative reconstruction con- F. Khorrami, ‘‘Learning-based real-time process-aware anomaly moni-
strained generative adversarial network for hyperspectral anomaly detec- toring for assured autonomy,’’ IEEE Trans. Intell. Vehicles, vol. 5, no. 4,
tion,’’ IEEE Trans. Geosci. Remote Sens., vol. 58, no. 7, pp. 4666–4679, pp. 659–669, Dec. 2020.
Jul. 2020. [57] T. Jiang, J. Zeng, K. Zhou, P. Huang, and T. Yang, ‘‘Lifelong disk failure
prediction via GAN-based anomaly detection,’’ in Proc. IEEE 37th Int.
[37] T. Golda, N. Murzyn, C. Qu, and K. Kroschel, ‘‘What goes around
Conf. Comput. Design (ICCD), Nov. 2019, pp. 199–207.
comes around: Cycle-consistency-based short-term motion prediction for
anomaly detection using generative adversarial networks,’’ in Proc. 16th [58] Z. J. Que, Y. Xiong, and Z. G. Xu, ‘‘A semi-supervised approach
IEEE Int. Conf. Adv. Video Signal Based Surveill. (AVSS), Sep. 2019, for steam turbine health prognostics based on GAN and PF,’’ in
pp. 1–8. Proc. IEEE Int. Conf. Ind. Eng. Eng. Manage. (IEEM), Dec. 2019,
pp. 1476–1480.
[38] C. Wang, Y.-M. Zhang, and C.-L. Liu, ‘‘Anomaly detection via minimum
[59] Y.-X. Tang, Y.-B. Tang, M. Han, J. Xiao, and R. M. Summers, ‘‘Abnormal
likelihood generative adversarial networks,’’ in Proc. 24th Int. Conf.
chest X-ray identification with generative adversarial one-class classi-
Pattern Recognit. (ICPR), Aug. 2018, pp. 1121–1126.
fier,’’ in Proc. IEEE 16th Int. Symp. Biomed. Imag. (ISBI), Apr. 2019,
[39] T. Ganokratanaa, S. Aramvith, and N. Sebe, ‘‘Anomaly event detec-
pp. 1358–1361.
tion using generative adversarial network for surveillance videos,’’ in
[60] Y. Sun, W. Yu, Y. Chen, and A. Kadam, ‘‘Time series anomaly detection
Proc. Asia–Pacific Signal Inf. Process. Assoc. Annu. Summit Conf.
based on GAN,’’ in Proc. 6th Int. Conf. Social Netw. Anal., Manage.
(APSIPA ASC), Nov. 2019, pp. 1395–1399.
Secur. (SNAMS), Oct. 2019, pp. 375–382.
[40] Y. Lai, J. Hu, Y. Tsai, and W. Chiu, ‘‘Industrial anomaly detection
[61] P. R. Roy and G.-A. Bilodeau, ‘‘Adversarially learned abnormal trajectory
and one-class classification using generative adversarial networks,’’ in
classifier,’’ in Proc. 16th Conf. Comput. Robot Vis. (CRV), May 2019,
Proc. IEEE/ASME Int. Conf. Adv. Intell. Mechatronics (AIM), Jul. 2018,
pp. 65–72.
pp. 1444–1449.
[62] H. Zenati, M. Romain, C.-S. Foo, B. Lecouat, and V. Chandrasekhar,
[41] F. Dong, Y. Zhang, and X. Nie, ‘‘Dual discriminator generative adver- ‘‘Adversarially learned anomaly detection,’’ in Proc. IEEE Int. Conf. Data
sarial network for video anomaly detection,’’ IEEE Access, vol. 8, Mining (ICDM), Nov. 2018, pp. 727–736.
pp. 88170–88176, 2020.
[63] J. Bian, X. Hui, S. Sun, X. Zhao, and M. Tan, ‘‘A novel and effi-
[42] Y. Kuang, T. Lan, X. Peng, G. E. Selasi, Q. Liu, and J. Zhang, ‘‘Unsuper- cient CVAE-GAN-based approach with informative manifold for semi-
vised multi-discriminator generative adversarial network for lung nodule supervised anomaly detection,’’ IEEE Access, vol. 7, pp. 88903–88916,
malignancy classification,’’ IEEE Access, vol. 8, pp. 77725–77734, 2020. 2019.
[43] M. Yan, X. Jiang, and J. Yuan, ‘‘3D convolutional generative adversarial [64] M. Ravanbakhsh, E. Sangineto, M. Nabi, and N. Sebe, ‘‘Training
networks for detecting temporal irregularities in videos,’’ in Proc. 24th adversarial discriminators for cross-channel abnormal event detection
Int. Conf. Pattern Recognit. (ICPR), Aug. 2018, pp. 2522–2527. in crowds,’’ in Proc. IEEE Winter Conf. Appl. Comput. Vis. (WACV),
[44] R. Skilton and Y. Gao, ‘‘Visual detection of generic defects in indus- Jan. 2019, pp. 1896–1904.
trial components using generative adversarial networks,’’ in Proc. [65] T. Ganokratanaa, S. Aramvith, and N. Sebe, ‘‘Unsupervised anomaly
IEEE/ASME Int. Conf. Adv. Intell. Mechatronics (AIM), Jul. 2019, detection and localization based on deep spatiotemporal translation net-
pp. 489–494. work,’’ IEEE Access, vol. 8, pp. 50312–50329, 2020.
[45] T. Leangarun, P. Tangamchit, and S. Thajchayapong, ‘‘Stock price manip- [66] P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, ‘‘MVTec AD—
ulation detection using generative adversarial networks,’’ in Proc. IEEE A comprehensive real-world dataset for unsupervised anomaly detec-
Symp. Ser. Comput. Intell. (SSCI), Nov. 2018, pp. 2104–2111. tion,’’ in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR),
[46] A. Ferdowsi and W. Saad, ‘‘Generative adversarial networks for dis- Jun. 2019, pp. 9592–9600.
tributed intrusion detection in the Internet of Things,’’ in Proc. IEEE [67] C. Wang, Y. Dai, and W. Dai, ‘‘Deep embedding GAN-based model
Global Commun. Conf. (GLOBECOM), Dec. 2019, pp. 1–6. for anomaly detection on high-dimensional sparse data,’’ in Proc. Chin.
[47] K. Gray, D. Smolyak, S. Badirli, and G. Mohler, ‘‘Coupled IGMM-GANs Control Conf. (CCC), Jul. 2019, pp. 8718–8722.
for improved generative adversarial anomaly detection,’’ in Proc. IEEE [68] M. B. Fadhel and K. Nyarko, ‘‘GAN augmented text anomaly detection
Int. Conf. Big Data (Big Data), Dec. 2018, pp. 2538–2541. with sequences of deep statistics,’’ in Proc. 53rd Annu. Conf. Inf. Sci. Syst.
[48] Y. Choi, H. Lim, H. Choi, and I.-J. Kim, ‘‘GAN-based anomaly detection (CISS), Mar. 2019, pp. 1–5.
and localization of multivariate time series data for power plant,’’ in [69] K. Liu, A. Li, X. Wen, H. Chen, and P. Yang, ‘‘Steel surface defect
Proc. IEEE Int. Conf. Big Data Smart Comput. (BigComp), Feb. 2020, detection using GAN and one-class classifier,’’ in Proc. 25th Int. Conf.
pp. 71–74. Autom. Comput. (ICAC), Sep. 2019, pp. 1–6.

VOLUME 9, 2021 161025

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

[70] T. H. Emerson, J. A. Edelberg, T. Doster, N. Merrill, and C. C. Olson, [92] L. Xue and S. Gao, ‘‘Unsupervised anomaly detection system for railway
‘‘Generative and encoded anomaly detectors,’’ in Proc. 10th Workshop turnout based on GAN,’’ J. Phys., Conf. Ser., vol. 1345, no. 3, Nov. 2019,
Hyperspectral Imag. Signal Process., Evol. Remote Sens. (WHISPERS), Art. no. 032069.
Sep. 2019, pp. 1–5. [93] D. Li, D. Chen, B. Jin, L. Shi, J. Goh, and S.-K. Ng, ‘‘MAD-GAN:
[71] S. Mao, J. Guo, and Z. Li, ‘‘Discriminative autoencoding framework Multivariate anomaly detection for time series data with generative adver-
for simple and efficient anomaly detection,’’ IEEE Access, vol. 7, sarial networks,’’ in Artificial Neural Networks and Machine Learning—
pp. 140618–140630, 2019. ICANN 2019: Text and Time Series. Cham, Switzerland: Springer, 2019,
[72] M. Baydoun, M. Ravanbakhsh, D. Campo, P. Marin, D. Martin, pp. 703–716.
L. Marcenaro, A. Cavallaro, and C. S. Regazzoni, ‘‘A multi-perspective [94] B. Xia, J. Yin, J. Xu, and Y. Li, ‘‘LogGAN: A sequence-based generative
approach to anomaly detection for self—Aware embodied agents,’’ adversarial network for anomaly detection based on system logs,’’ in
in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), Proc. Int. Conf. Sci. Cyber Secur., Springer, 2019, pp. 61–76.
Apr. 2018, pp. 6598–6602. [95] L. Deecke, R. Vandermeulen, L. Ruff, S. Mandt, and M. Kloft, ‘‘Image
[73] K. Vatanparvar and M. A. Al Faruque, ‘‘Self-secured control with anomaly detection with generative adversarial networks,’’ in Proc. Joint
anomaly detection and recovery in automotive cyber-physical systems,’’ Eur. Conf. Mach. Learn. Knowl. Discovery Databases. Springer, 2018,
in Proc. Design, Autom. Test Eur. Conf. Exhib. (DATE), Mar. 2019, pp. 3–17.
pp. 788–793. [96] S. Akcay, A. Atapour-Abarghouei, and T. P. Breckon, ‘‘GANomaly:
[74] M. Ravanbakhsh, M. Nabi, E. Sangineto, L. Marcenaro, C. Regazzoni, Semi-supervised anomaly detection via adversarial training,’’ in
and N. Sebe, ‘‘Abnormal event detection in videos using generative adver- Computer Vision—ACCV 2018. Cham, Switzerland: Springer, 2019,
sarial nets,’’ in Proc. IEEE Int. Conf. Image Process. (ICIP), Sep. 2017, pp. 622–637.
pp. 1577–1581. [97] Y. Tang, Y. Tang, M. Han, J. Xiao, and R. M. Summers, ‘‘Deep adversarial
[75] H. Song, C. Sun, X. Wu, M. Chen, and Y. Jia, ‘‘Learning normal patterns one-class learning for normal and abnormal chest radiograph classifica-
via adversarial attention-based autoencoder for abnormal event detection tion,’’ Proc. SPIE, vol. 10950, Mar. 2019, Art. no. 1095018.
in videos,’’ IEEE Trans. Multimedia, vol. 22, no. 8, pp. 2138–2148, [98] M. Mostapha, J. Prieto, V. Murphy, J. Girault, M. Foster, A. Rumple,
Aug. 2020. J. Blocher, W. Lin, J. Elison, J. Gilmore, S. Pizer, and M. Styner,
[76] A. Toma, A. Krayani, M. Farrukh, H. Qi, L. Marcenaro, Y. Gao, and
‘‘Semi-supervised VAE-GAN for out-of-sample detection applied to
C. S. Regazzoni, ‘‘AI-based abnormality detection at the PHY-layer of
MRI quality control,’’ in Medical Image Computing and Computer
cognitive radio by learning generative models,’’ IEEE Trans. Cognit.
Assisted Intervention—MICCAI 2019. Cham, Switzerland: Springer,
Commun. Netw., vol. 6, no. 1, pp. 21–34, Mar. 2020.
[77] H. Yan, H.-M. Yeh, and N. Sergin, ‘‘Image-based process monitoring via 2019, pp. 127–136.
adversarial autoencoder with applications to rolling defect detection,’’ [99] E. A. Donahue, T. T. Quach, K. Potter, C. Martinez, M. Smith,
in Proc. IEEE 15th Int. Conf. Autom. Sci. Eng. (CASE), Aug. 2019, and C. D. Turner, ‘‘Deep learning for automated defect detection in
pp. 311–316. high-reliability electronic parts,’’ Proc. SPIE, vol. 11139, Sep. 2019,
[78] H. Watanabe, R. Togo, T. Ogawa, and M. Haseyama, ‘‘Bone metastatic Art. no. 1113907.
tumor detection based on AnoGAN using CT images,’’ in Proc. IEEE 1st [100] D. Zhang, Q. Niu, and X. Qiu, ‘‘Detecting anomalies in communication
Global Conf. Life Sci. Technol. (LifeTech), Mar. 2019, pp. 235–236. packet streams based on generative adversarial networks,’’ in Wireless
[79] M. Sabokrou, M. Khalooei, M. Fathy, and E. Adeli, ‘‘Adversarially Algorithms, Systems, and Applications. Cham, Switzerland: Springer,
learned one-class classifier for novelty detection,’’ in Proc. IEEE/CVF 2019, pp. 470–481.
Conf. Comput. Vis. Pattern Recognit., Jun. 2018, pp. 3379–3388. [101] C. Baur, B. Wiestler, S. Albarqouni, and N. Navab, ‘‘Deep autoencoding
[80] Y. Liu, Y. Tian, G. Maicas, L. Z. Cheng Tao Pu, R. Singh, J. W. Verjans, models for unsupervised anomaly segmentation in brain MR images,’’
and G. Carneiro, ‘‘Photoshopping colonoscopy video frames,’’ in Proc. in Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain
IEEE 17th Int. Symp. Biomed. Imag. (ISBI), Apr. 2020, pp. 1–5. Injuries. Cham, Switzerland: Springer, 2019, pp. 161–169.
[81] S. Ammar, T. Bouwmans, N. Zaghden, and M. Neji, ‘‘Deep detector [102] Z. Huang, W. Mao, M. Chen, Q. Wu, B. Xiong, and W. Xu, ‘‘An intelligent
classifier (DeepDC) for moving objects segmentation and classification operation and maintenance system for power consumption based on
in video surveillance,’’ IET Image Process., vol. 14, no. 8, pp. 1490–1501, deep learning,’’ IOP Conf. Ser., Mater. Sci. Eng., vol. 486, no. 1, 2019,
Jun. 2020. Art. no. 012107.
[82] S. Yadav, C. Chen, and A. Ross, ‘‘Relativistic discriminator: A one-class [103] S. K. Yarlagadda, D. Güera, P. Bestagini, F. M. Zhu, S. Tubaro, and
classifier for generalized iris presentation attack detection,’’ in Proc. IEEE E. J. Delp, ‘‘Satellite image forgery detection and localization using GAN
Winter Conf. Appl. Comput. Vis. (WACV), Mar. 2020, pp. 2635–2644. and one-class classifier,’’ Electron. Imag., vol. 2018, no. 7, pp. 1–214,
[83] M. Labonne, A. Olivereau, B. Polve, and D. Zeghlache, ‘‘Unsupervised 2018.
protocol-based intrusion detection for real-world networks,’’ in Proc. Int. [104] T. Schlegl, P. Seeböck, S. M. Waldstein, U. Schmidt-Erfurth, and
Conf. Comput., Netw. Commun. (ICNC), Feb. 2020, pp. 299–303. G. Langs, ‘‘Unsupervised anomaly detection with generative adversarial
[84] J.-J. Chung and H.-J. Kim, ‘‘An automobile environment detection system networks to guide marker discovery,’’ in Information Processing in Med-
based on deep neural network and its implementation using IoT-enabled ical Imaging. Cham, Switzerland: Springer, 2017, pp. 146–157.
in-vehicle air quality sensors,’’ Sustainability, vol. 12, no. 6, p. 2475, [105] A. Singanamalli, J. Mitra, K. Wallace, P. Venugopal, S. Smith, L. Mo,
Mar. 2020. L. Y. Leung, J. Morrison, T. Rasmussen, and L. Marinelli, ‘‘Blood flow
[85] W. Reiter, ‘‘Video anomaly detection in post-procedural use of laparo- anomaly detection via generative adversarial networks: A preliminary
scopic videos,’’ in Bildverarbeitung Für Die Medizin 2020. Wiesbaden, study,’’ Proc. SPIE, vol. 11315, Mar. 2020, Art. no. 1131522.
Germany: Springer Fachmedien Wiesbaden, 2020, pp. 101–106. [106] J.-T. Wang and C.-H. Wang, ‘‘High performance WGAN-GP based
[86] S. Anno and Y. Sasaki, ‘‘GAN-based abnormal detection by recogniz- multiple-category network anomaly classification system,’’ in Proc. Int.
ing ungeneratable patterns,’’ in Pattern Recognition. Cham, Switzerland: Conf. Cyber Secur. Emerg. Technol. (CSET), Oct. 2019, pp. 1–7.
Springer, 2020, pp. 401–411. [107] M. Sabokrou, M. Pourreza, M. Fayyaz, R. Entezari, M. Fathy, J. Gall, and
[87] Y. Sun, L. Guo, Y. Li, L. Xu, and Y. Wang, ‘‘Semi-supervised deep learn-
E. Adeli, ‘‘AVID: Adversarial visual irregularity detection,’’ in Computer
ing for network anomaly detection,’’ in Algorithms and Architectures for
Vision—ACCV 2018. Cham, Switzerland: Springer, 2019, pp. 488–505.
Parallel Processing. Cham, Switzerland: Springer, 2020, pp. 383–390.
[88] R. Anirudh, J. J. Thiagarajan, B. Kailkhura, and P. T. Bremer, ‘‘Mimic- [108] M. Kimura and T. Yanagihara, ‘‘Anomaly detection using GANs for
GAN: Robust projection onto image manifolds with corruption mimick- visual inspection in noisy training data,’’ in Computer Vision—ACCV
ing,’’ Int. J. Comput. Vis., vol. 128, no. 10, pp. 2459–2477, 2020. 2018 Workshops. Cham, Switzerland: Springer, 2019, pp. 373–385.
[89] N. Tuluptceva, B. Bakker, I. Fedulova, and A. Konushin, ‘‘Perceptual [109] W. Shin and S.-B. Cho, ‘‘CCTV image sequence generation and mod-
image anomaly detection,’’ in Proc. Asian Conf. Pattern Recognit., eling method for video anomaly detection using generative adversarial
Springer, 2019, pp. 164–178. network,’’ in Intelligent Data Engineering and Automated Learning—
[90] Y. Mu and B. Zhang, ‘‘Abnormal event detection and localization in IDEAL 2018. Cham, Switzerland: Springer, 2018, pp. 457–467.
visual surveillance,’’ in Communications, Signal Processing, and Sys- [110] S. Huang and K. Lei, ‘‘IGAN-IDS: An imbalanced generative adversarial
tems. Singapore: Springer, 2020, pp. 1217–1225. network towards intrusion detection system in ad-hoc networks,’’ Ad Hoc
[91] J. K. Dumagpi, W.-Y. Jung, and Y.-J. Jeong, ‘‘A new GAN-based anomaly Netw., vol. 105, Aug. 2020, Art. no. 102177.
detection (GBAD) approach for multi-threat object classification on [111] J.-H. Oh, J. Y. Hong, and J.-G. Baek, ‘‘Oversampling method using outlier
large-scale X-ray security images,’’ IEICE Trans. Inf. Syst., vol. E103.D, detectable generative adversarial network,’’ Expert Syst. Appl., vol. 133,
no. 2, pp. 454–458, 2020. pp. 1–8, Nov. 2019.

161026 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

[112] S.-Y. Shin, Y.-W. Kang, and Y.-G. Kim, ‘‘Android-GAN: Defend- [134] H. C. Shin, N. A. Tenenholtz, J. K. Rogers, C. G. Schwarz, M. L. Senjem,
ing against Android pattern attacks using multi-modal generative net- J. L. Gunter, K. P. Andriole, and M. Michalski, ‘‘Medical image synthesis
work as anomaly detector,’’ Expert Syst. Appl., vol. 141, Mar. 2020, for data augmentation and anonymization using generative adversarial
Art. no. 112964. networks,’’ in Simulation and Synthesis in Medical Imaging. Cham,
[113] G. S. Thejas, K. G. Boroojeni, K. Chandna, I. Bhatia, S. S. Iyengar, Switzerland: Springer, 2018, pp. 1–11.
and N. R. Sunitha, ‘‘Deep learning-based model to fight against ad click [135] A. Madani, M. Moradi, A. Karargyris, and T. Syeda-Mahmood, ‘‘Chest
fraud,’’ in Proc. ACM Southeast Conf., Apr. 2019, pp. 176–181. X-ray generation and data augmentation for cardiovascular abnormality
[114] A. AlEroud and G. Karabatis, ‘‘Bypassing detection of URL-based phish- classification,’’ Proc. SPIE, vol. 10574, Mar. 2018, Art. no. 105741M.
ing attacks using generative adversarial deep neural networks,’’ in Proc. [136] M. Moussa and C. H. Lim, ‘‘Interpreting abnormality of a complex static
6th Int. Workshop Secur. Privacy Anal., Mar. 2020, pp. 53–60. scene using generative adversarial network,’’ in Proc. Asia–Pacific Signal
[115] D. Smolyak, K. Gray, S. Badirli, and G. Mohler, ‘‘Coupled IGMM-GANs Inf. Process. Assoc. Annu. Summit Conf. (APSIPA ASC), Nov. 2019,
with applications to anomaly detection in human mobility data,’’ ACM pp. 2003–2007.
Trans. Spatial Algorithms Syst., vol. 6, no. 4, pp. 1–14, Aug. 2020. [137] L. Sun, J. Wang, Y. Huang, X. Ding, H. Greenspan, and J. Pais-
[116] L. Luo, W. Hsu, and S. Wang, ‘‘Data augmentation using generative ley, ‘‘An adversarial learning approach to medical image synthesis for
adversarial networks for electrical insulator anomaly detection,’’ in Proc. lesion detection,’’ IEEE J. Biomed. Health Informat., vol. 24, no. 8,
2nd Int. Conf. Manage. Sci. Ind. Eng., Apr. 2020, pp. 231–236. pp. 2303–2314, Aug. 2020.
[117] M. Salem, S. Taheri, and J. S. Yuan, ‘‘Anomaly generation using gen- [138] P. Ganesan, S. Rajaraman, R. Long, B. Ghoraani, and S. Antani, ‘‘Assess-
erative adversarial networks in host-based intrusion detection,’’ in Proc. ment of data augmentation strategies toward performance improvement of
9th IEEE Annu. Ubiquitous Comput., Electron. Mobile Commun. Conf. abnormality classification in chest radiographs,’’ in Proc. 41st Annu. Int.
(UEMCON), Nov. 2018, pp. 683–687. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Jul. 2019, pp. 841–844.
[118] H. Kim, J. Park, K. Min, and K. Huh, ‘‘Anomaly monitoring framework in [139] M. Fu, J. Liu, H. Zhang, and S. Lu, ‘‘Multisensor fusion for magnetic flux
lane detection with a generative adversarial network,’’ IEEE Trans. Intell. leakage defect characterization under information incompletion,’’ IEEE
Transp. Syst., vol. 22, no. 3, pp. 1603–1615, Mar. 2021. Trans. Ind. Electron., vol. 68, no. 5, pp. 4382–4392, May 2021.
[119] M. Usama, M. Asim, S. Latif, and J. Qadir, ‘‘Generative adversarial [140] D. Wang, R. Vinson, M. Holmes, G. Seibel, A. Bechar, S. Nof, and Y. Tao,
networks for launching and thwarting adversarial attacks on network ‘‘Early detection of tomato spotted wilt virus by hyperspectral imaging
intrusion detection systems,’’ in Proc. 15th Int. Wireless Commun. Mobile and outlier removal auxiliary classifier generative adversarial nets (OR-
Comput. Conf. (IWCMC), Jun. 2019, pp. 78–83. AC-GAN),’’ Sci. Rep., vol. 9, no. 1, pp. 1–14, Dec. 2019.
[120] Y. Liu, Z. Li, C. Zhou, Y. Jiang, J. Sun, M. Wang, and X. He, ‘‘Generative [141] S. Guan, ‘‘Breast cancer detection using synthetic mammograms from
adversarial active learning for unsupervised outlier detection,’’ IEEE generative adversarial networks in convolutional neural networks,’’
Trans. Knowl. Data Eng., vol. 32, no. 8, pp. 1517–1528, Aug. 2020. J. Med. Imag., vol. 6, no. 3, p. 1, Mar. 2019.
[121] S. Shin, I. Lee, and C. Choi, ‘‘Anomaly dataset augmentation using the [142] X. Li, Y. Liang, M. Zhao, C. Wang, and Y. Jiang, ‘‘Few-shot learning with
sequence generative models,’’ in Proc. 18th IEEE Int. Conf. Mach. Learn. generative adversarial networks based on WOA13 data,’’ Comput., Mater.
Appl. (ICMLA), Dec. 2019, pp. 1143–1148. Continua, vol. 60, no. 3, pp. 1073–1085, 2019.
[122] U. Muneeb, E. Koyuncu, Y. Keshtkarjahromi, H. Seferoglu, M. F. Erden,
[143] S. Guan and M. Loew, ‘‘Using generative adversarial networks and
and A. E. Cetin, ‘‘Robust and computationally-efficient anomaly detec-
transfer learning for breast cancer detection by convolutional neural
tion using powers-of-two networks,’’ in Proc. IEEE Int. Conf. Acoust.,
networks,’’ Proc. SPIE, vol. 10954, Mar. 2019, Art. no. 109541C.
Speech Signal Process. (ICASSP), May 2020, pp. 2992–2996.
[144] A. Radford, L. Metz, and S. Chintala, ‘‘Unsupervised representation
[123] Y. Shin, H. A. Qadir, and I. Balasingham, ‘‘Abnormal colon polyp image
learning with deep convolutional generative adversarial networks,’’ 2015,
synthesis using conditional adversarial networks for improved detection
arXiv:1511.06434.
performance,’’ IEEE Access, vol. 6, pp. 56007–56017, 2018.
[145] R. A. Yeh, C. Chen, T. Y. Lim, A. G. Schwing, M. Hasegawa-Johnson,
[124] S. Niu, B. Li, X. Wang, and H. Lin, ‘‘Defect image sample generation
and M. N. Do, ‘‘Semantic image inpainting with deep generative models,’’
with GAN for improving defect recognition,’’ IEEE Trans. Autom. Sci.
in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017,
Eng., vol. 17, no. 3, pp. 1611–1622, Jul. 2020.
[125] S. Lu and H. Gao, ‘‘Deep learning based fusion of RGB and infrared pp. 5485–5493.
images for the detection of abnormal condition of fused magnesium [146] G. Marcus, ‘‘Deep learning: A critical appraisal,’’ 2018,
furnace,’’ in Proc. IEEE 15th Int. Conf. Control Autom. (ICCA), Jul. 2019, arXiv:1801.00631.
pp. 987–993. [147] Z. Obermeyer and E. J. Emanuel, ‘‘Predicting the future-big data, machine
[126] M. Shao, N. Gu, and X. Zhang, ‘‘Credit card transactions data adversarial learning, and clinical medicine,’’ New England J. Med., vol. 375, no. 13,
augmentation in the frequency domain,’’ in Proc. 5th IEEE Int. Conf. Big p. 1216, 2016.
Data Anal. (ICBDA), May 2020, pp. 238–245. [148] H. Kaur, H. S. Pannu, and A. K. Malhi, ‘‘A systematic review on imbal-
[127] S. Jintao, G. Chaoyue, S. Hui, S. Jiangang, and L. Zhe, ‘‘Data expansion anced data challenges in machine learning: Applications and solutions,’’
for foreign object detection in power grid,’’ in Proc. IEEE PES Asia– ACM Comput. Surv., vol. 52, no. 4, pp. 1–36, Sep. 2019.
Pacific Power Energy Eng. Conf. (APPEEC), Dec. 2019, pp. 1–4. [149] T. R. Hoens and N. V. Chawla, Imbalanced Datasets: From Sampling to
[128] B. Mohammadi and M. Sabokrou, ‘‘End-to-end adversarial learning for Classifiers. Hoboken, NJ, USA: Wiley, 2013, ch. 3, pp. 43–59.
intrusion detection in computer networks,’’ in Proc. IEEE 44th Conf. [150] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer,
Local Comput. Netw. (LCN), Oct. 2019, pp. 270–273. ‘‘SMOTE: Synthetic minority over-sampling technique,’’ J. Artif. Intell.
[129] J. Yang, T. Li, G. Liang, W. He, and Y. Zhao, ‘‘A simple recurrent unit Res., vol. 16, no. 1, pp. 321–357, 2002.
model based intrusion detection system with DCGAN,’’ IEEE Access, [151] H. Han, W.-Y. Wang, and B.-H. Mao, ‘‘Borderline-SMOTE: A new
vol. 7, pp. 83286–83296, 2019. over-sampling method in imbalanced data sets learning,’’ in Advances in
[130] K. Naidoo and V. Marivate, ‘‘Unsupervised anomaly detection of health- Intelligent Computing. Berlin, Germany: Springer, 2005, pp. 878–887.
care providers using generative adversarial networks,’’ in Proc. 19th [152] H. He, Y. Bai, E. A. Garcia, and S. Li, ‘‘ADASYN: Adaptive synthetic
Conf. e-Bus., e-Services e-Soc. (I E), in Responsible Design, Implemen- sampling approach for imbalanced learning,’’ in Proc. IEEE Int. Joint
tation and Use of Information and Communication Technology, Skukuza, Conf. Neural Netw. (IEEE World Congr. Comput. Intell.), Jun. 2008,
South Africa. Springer, Apr. 2020, pp. 419–430. pp. 1322–1328.
[131] J. Lee and K. Park, ‘‘AE-CGAN model based high performance net- [153] V. Chandola, A. Banerjee, and V. Kumar, ‘‘Anomaly detection: A survey,’’
work intrusion detection system,’’ Appl. Sci., vol. 9, no. 20, p. 4221, ACM Comput. Surv., vol. 41, no. 3, pp. 1–58, Jul. 2009.
Oct. 2019. [154] A. Lazarevic, V. Kumar, and J. Srivastava, ‘‘Intrusion detection: A sur-
[132] S. K. Lim, Y. Loo, N.-T. Tran, N.-M. Cheung, G. Roig, and Y. Elovici, vey,’’ in Managing Cyber Threats: Issues, Approaches, and Challenges.
‘‘DOPING: Generative data augmentation for unsupervised anomaly Boston, MA, USA: Springer, 2005, pp. 19–78.
detection with GAN,’’ in Proc. IEEE Int. Conf. Data Mining (ICDM), [155] H.-J. Liao, C.-H. R. Lin, Y.-C. Lin, and K.-Y. Tung, ‘‘Intrusion detection
Nov. 2018, pp. 1122–1127. system: A comprehensive review,’’ J. Netw. Comput. Appl., vol. 36, no. 1,
[133] R. Zheng, L. Liu, S. Zhang, C. Zheng, F. Bunyak, R. Xu, B. Li, and pp. 16–24, 2013.
M. Sun, ‘‘Detection of exudates in fundus photographs with imbalanced [156] C. D. Nguyen, S. Miles, A. Perini, P. Tonella, M. Harman, and M. Luck,
learning using conditional generative adversarial network,’’ Biomed. Opt. ‘‘Evolutionary testing of autonomous software agents,’’ Auton. Agent
Exp., vol. 9, no. 10, pp. 4863–4878, Oct. 2018. Multi-Agent Syst., vol. 25, no. 2, pp. 260–283, 2012.

VOLUME 9, 2021 161027

 M. Sabuhi et al.: Applications of GANs in Anomaly Detection

[157] Y. Kou, C.-T. Lu, S. Sirwongwattana, and Y.-P. Huang, ‘‘Survey of fraud [184] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller, ‘‘Striving
detection techniques,’’ in Proc. IEEE Int. Conf. Netw., Sens. Control, for simplicity: The all convolutional net,’’ 2014, arXiv:1412.6806.
Mar. 2004, pp. 749–754. [185] S. Ioffe and C. Szegedy, ‘‘Batch normalization: Accelerating
[158] M. Mirza and S. Osindero, ‘‘Conditional generative adversarial nets,’’ deep network training by reducing internal covariate shift,’’ 2015,
2014, arXiv:1411.1784. arXiv:1502.03167.
[159] J. Donahue, P. Krähenbühl, and T. Darrell, ‘‘Adversarial feature learning,’’ [186] V. Mahadevan, W. Li, V. Bhalodia, and N. Vasconcelos, ‘‘Anomaly detec-
2016, arXiv:1605.09782. tion in crowded scenes,’’ in Proc. IEEE Comput. Soc. Conf. Comput. Vis.
[160] M. Arjovsky, S. Chintala, and L. Bottou, ‘‘Wasserstein GAN,’’ 2017, Pattern Recognit., Jun. 2010, pp. 1975–1981.
arXiv:1701.07875. [187] Y. LeCun, C. Cortes, and C. Burges. MNIST Handwritten Digit
[161] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. C. Courville, Database. Florham Park, NJ, USA. Accessed: 2010. [Online]. Available:
‘‘Improved training of Wasserstein GANs,’’ in Adv. neural Inf. Process. http://yann.lecun.com/exdb/mnist/
Syst., 2017, pp. 5767–5777. [188] A. Krizhevsky and G. Hinton, ‘‘Learning multiple layers of features from
[162] A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther, ‘‘Autoen- tiny images,’’ M.S. thesis, Dept. Comput. Sci., Univ. Toronto, Toronto,
coding beyond pixels using a learned similarity metric,’’ in Proc. Int. ON, Canada, 2009.
Conf. Mach. Learn., 2016, pp. 1558–1566. [189] R. Mehran, A. Oyama, and M. Shah, ‘‘Abnormal crowd behavior detec-
[163] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, ‘‘Unpaired image-to-image tion using social force model,’’ in Proc. IEEE Conf. Comput. Vis. Pattern
translation using cycle-consistent adversarial networks,’’ in Proc. IEEE Recognit., Jun. 2009, pp. 935–942.
[190] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, ‘‘Image quality
Int. Conf. Comput. Vis. (ICCV), Oct. 2017, pp. 2223–2232.
assessment: From error visibility to structural similarity,’’ IEEE Trans.
[164] A. Odena, C. Olah, and J. Shlens, ‘‘Conditional image synthesis with
Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004.
auxiliary classifier GANs,’’ in Proc. Int. Conf. Mach. Learn., 2017,
[191] W. Li, V. Mahadevan, and N. Vasconcelos, ‘‘Anomaly detection and
pp. 2642–2651.
localization in crowded scenes,’’ IEEE Trans. Pattern Anal. Mach. Intell.,
[165] T. Karras, T. Aila, S. Laine, and J. Lehtinen, ‘‘Progressive grow- vol. 36, no. 1, pp. 18–32, Jan. 2014.
ing of GANs for improved quality, stability, and variation,’’ 2017, [192] C. Lu, J. Shi, and J. Jia, ‘‘Abnormal event detection at 150 FPS in MAT-
arXiv:1710.10196. LAB,’’ in Proc. IEEE Int. Conf. Comput. Vis., Dec. 2013, pp. 2720–2727.
[166] J. Su, ‘‘O-GAN: Extremely concise approach for auto-encoding genera- [193] Y. Cong, J. Yuan, and J. Liu, ‘‘Sparse reconstruction cost for abnormal
tive adversarial networks,’’ 2019, arXiv:1903.01931. event detection,’’ in Proc. CVPR, Jun. 2011, pp. 3449–3456.
[167] A. Makhzani, J. Shlens, N. Jaitly, I. Goodfellow, and B. Frey, ‘‘Adversarial [194] M. Sakurada and T. Yairi, ‘‘Anomaly detection using autoencoders with
autoencoders,’’ 2015, arXiv:1511.05644. nonlinear dimensionality reduction,’’ in Proc. MLSDA 2nd Workshop
[168] G. Mariani, F. Scheidegger, R. Istrate, C. Bekas, and C. Malossi, Mach. Learn. Sensory Data Anal. (MLSDA), 2014, pp. 4–11.
‘‘BAGAN: Data augmentation with balancing GAN,’’ 2018, [195] J. An and S. Cho, ‘‘Variational autoencoder based anomaly detection
arXiv:1803.09655. using reconstruction probability,’’ Special Lecture IE, vol. 2, pp. 1–18,
[169] J. Zhao, M. Mathieu, and Y. LeCun, ‘‘Energy-based generative adversar- Dec. 2015.
ial network,’’ 2016, arXiv:1609.03126. [196] D. Xu, E. Ricci, Y. Yan, J. Song, and N. Sebe, ‘‘Learning deep represen-
[170] T. Nguyen, T. Le, H. Vu, and D. Phung, ‘‘Dual discriminator gener- tations of appearance and motion for anomalous event detection,’’ 2015,
ative adversarial nets,’’ in Proc. Adv. Neural Inf. Process. Syst., 2017, arXiv:1510.01553.
pp. 2670–2680. [197] M. Du, F. Li, G. Zheng, and V. Srikumar, ‘‘DeepLog: Anomaly detection
[171] J. Su, ‘‘GAN-QP: A novel GAN framework without gradient vanishing and diagnosis from system logs through deep learning,’’ in Proc. ACM
and Lipschitz constraint,’’ 2018, arXiv:1811.07296. SIGSAC Conf. Comput. Commun. Secur., Oct. 2017, pp. 1285–1298.
[172] P. Perera, R. Nallapati, and B. Xiang, ‘‘OCGAN: One-class novelty [198] B. Schölkopf, R. C. Williamson, A. J. Smola, J. Shawe-Taylor, and
detection using GANs with constrained latent representations,’’ in Proc. J. C. Platt, ‘‘Support vector method for novelty detection,’’ Adv. Neural
IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2019, Inf. Process. Syst., vol. 12, 1999, pp. 582–588.
pp. 2898–2906. [199] F. T. Liu, K. M. Ting, and Z.-H. Zhou, ‘‘Isolation forest,’’ in Proc. 8th
[173] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, ‘‘Image-to-image translation IEEE Int. Conf. Data Mining, Dec. 2008, pp. 413–422.
with conditional adversarial networks,’’ in Proc. IEEE Conf. Comput. Vis. [200] M. E. Tipping and C. M. Bishop, ‘‘Probabilistic principal component
Pattern Recognit. (CVPR), Jul. 2017, pp. 1125–1134. analysis,’’ J. Roy. Statist. Soc. B, vol. 61, no. 3, pp. 611–622, 1999.
[174] A. Jolicoeur-Martineau, ‘‘The relativistic discriminator: A key element [201] G. Lerman, M. B. McCoy, J. A. Tropp, and T. Zhang, ‘‘Robust compu-
missing from standard GAN,’’ 2018, arXiv:1807.00734. tation of linear models by convex relaxation,’’ Found. Comput. Math.,
[175] L. Yu, W. Zhang, J. Wang, and Y. Yu, ‘‘SeqGAN: Sequence generative vol. 15, no. 2, pp. 363–410, Apr. 2015.
adversarial nets with policy gradient,’’ in Proc. 31st AAAI Conf. Artif. [202] G. Liu, Z. Lin, and Y. Yu, ‘‘Robust subspace segmentation by low-rank
Intell., 2017, pp. 2852–2858. representation,’’ in Proc. 27th Int. Conf. Mach. Learn. (ICML), 2010,
pp. 663–670.
[176] Y. Zhang, Z. Gan, K. Fan, Z. Chen, R. Henao, D. Shen, and
[203] X. Yang, L. J. Latecki, and D. Pokrajac, ‘‘Outlier detection with globally
L. Carin, ‘‘Adversarial feature matching for text generation,’’ 2017,
optimal exemplar-based GMM,’’ in Proc. SIAM Int. Conf. Data Mining.
arXiv:1706.03850.
Philadelphia, PA, USA: SIAM, 2009, pp. 145–154.
[177] Z. Pan, W. Yu, X. Yi, A. Khan, F. Yuan, and Y. Zheng, ‘‘Recent progress
[204] C. You, D. P. Robinson, and R. Vidal, ‘‘Provable self-representation based
on generative adversarial networks (GANs): A survey,’’ IEEE Access,
outlier detection in a union of subspaces,’’ in Proc. IEEE Conf. Comput.
vol. 7, pp. 36322–36333, 2019.
Vis. Pattern Recognit. (CVPR), Jul. 2017, pp. 3395–3404.
[178] J. Gui, Z. Sun, Y. Wen, D. Tao, and J. Ye, ‘‘A review on genera- [205] M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, ‘‘LOF: Identi-
tive adversarial networks: Algorithms, theory, and applications,’’ 2020, fying density-based local outliers,’’ ACM SIGMOD Rec., vol. 29, no. 2,
arXiv:2001.06937. pp. 93–104, Jun. 2000.
[179] A. Jabbar, X. Li, and B. Omar, ‘‘A survey on generative adversarial [206] S. Zhai, Y. Cheng, W. Lu, and Z. Zhang, ‘‘Deep structured energy based
networks: Variants, applications, and training,’’ 2020, arXiv:2006.05132. models for anomaly detection,’’ 2016, arXiv:1605.07717.
[180] M. Arjovsky and L. Bottou, ‘‘Towards principled methods for training [207] B. Zong, Q. Song, M. R. Min, W. Cheng, C. Lumezanu, D. Cho, and
generative adversarial networks,’’ 2017, arXiv:1701.04862. H. Chen, ‘‘Deep autoencoding Gaussian mixture model for unsupervised
[181] S. Arora, R. Ge, Y. Liang, T. Ma, and Y. Zhang, ‘‘Generalization and anomaly detection,’’ in Proc. Int. Conf. Learn. Represent., 2018, pp. 1–19.
equilibrium in generative adversarial nets (GANs),’’ in Proc. Int. Conf. [208] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, A. Acosta,
Mach. Learn., 2017, pp. 224–232. A. Aitken, A. Tejani, J. Totz, Z. Wang, and W. Shi, ‘‘Photo-realistic
[182] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, single image super-resolution using a generative adversarial network,’’
and X. Chen, ‘‘Improved techniques for training GANs,’’ 2016, in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017,
arXiv:1606.03498. pp. 4681–4690.
[183] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, [209] X. Wang, K. Yu, S. Wu, J. Gu, Y. Liu, C. Dong, Y. Qiao, and C. Change
‘‘GANs trained by a two time-scale update rule converge to a local Loy, ‘‘ESRGAN: Enhanced super-resolution generative adversarial net-
Nash equilibrium,’’ in Proc. Adv. Neural Inf. Process. Syst., 2017, works,’’ in Proc. Eur. Conf. Comput. Vis. (ECCV) Workshops, Sep. 2018,
pp. 6626–6637. pp. 63–79.

161028 VOLUME 9, 2021

M. Sabuhi et al.: Applications of GANs in Anomaly Detection

[210] A. Brock, J. Donahue, and K. Simonyan, ‘‘Large scale GAN training for MING (CHLOE) ZHOU (Graduate Student Mem-
high fidelity natural image synthesis,’’ 2018, arXiv:1809.11096. ber, IEEE) received the B.Sc. degree in electri-
[211] C. Han, L. Rundo, K. Murao, T. Noguchi, Y. Shimahara, Z. Á. Milacski, cal engineering from the University of Alberta,
S. Koshino, E. Sala, H. Nakayama, and S. Satoh, ‘‘MADGAN: Unsuper- Edmonton, AB, Canada, in 2020, where she is cur-
vised medical anomaly detection GAN using multiple adjacent brain MRI rently pursuing the M.Sc. degree with the Depart-
slice reconstruction,’’ BMC Bioinf., vol. 22, no. S2, pp. 1–20, Apr. 2021. ment of Electrical and Computer Engineering. She
[212] C. Han, L. Rundo, R. Araki, Y. Nagano, Y. Furukawa, G. Mauri,
is also a member of the ENergydigiTizAtIon Lab
H. Nakayama, and H. Hayashi, ‘‘Combining noise-to-image and image-
(ENTAIL). Her research interests include anomaly
to-image GANs: Brain MR image augmentation for tumor detection,’’
IEEE Access, vol. 7, pp. 156966–156977, 2019. detection, smart grids, and time series forecasting.
[213] C. Baur, R. Graf, B. Wiestler, S. Albarqouni, and N. Navab,
‘‘SteGANomaly: Inhibiting CycleGAN steganography for unsupervised
anomaly detection in brain MRI,’’ in Medical Image Computing and
Computer Assisted Intervention—MICCAI 2020. Cham, Switzerland:
Springer, 2020, pp. 718–727.
[214] M. Pourreza, B. Mohammadi, M. Khaki, S. Bouindour, H. Snoussi, and
M. Sabokrou, ‘‘G2D: Generate to detect anomaly,’’ in Proc. IEEE Winter
Conf. Appl. Comput. Vis. (WACV), Jan. 2021, pp. 2003–2012.
[215] X. Feng, D. Song, Y. Chen, Z. Chen, J. Ni, and H. Chen, ‘‘Convolutional
transformer based dual discriminator generative adversarial networks for
video anomaly detection,’’ in Proc. 29th ACM Int. Conf. Multimedia,
Oct. 2021, pp. 5546–5554.
[216] K. Doshi and Y. Yilmaz, ‘‘Any-shot sequential anomaly detection in
surveillance videos,’’ in Proc. IEEE/CVF Conf. Comput. Vis. Pattern COR-PAUL BEZEMER (Member, IEEE) received
Recognit. Workshops (CVPRW), Jun. 2020, pp. 934–935. the B.Sc., M.Sc., and Ph.D. degrees in computer
[217] P. F. de Araujo-Filho, G. Kaddoum, D. R. Campelo, A. G. Santos, science from the Delft University of Technology,
D. Macedo, and C. Zanchettin, ‘‘Intrusion detection for cyber–physical The Netherlands, in 2007, 2009, and 2014, respec-
systems using generative adversarial networks in fog environment,’’ IEEE tively. He is currently an Assistant Professor with
Internet Things J., vol. 8, no. 8, pp. 6247–6256, Apr. 2021. the Electrical and Computer Engineering Depart-
[218] Y. Zhong, Y. Zhu, Z. Wang, X. Yin, X. Shi, and K. Li, ‘‘An adversarial
ment, University of Alberta. He also heads the
learning model for intrusion detection in real complex network envi-
Analytics of Software, GAmes And Repository
ronments,’’ in Wireless Algorithms, Systems, and Applications. Cham,
Switzerland: Springer, 2020, pp. 794–806. Data (ASGAARD) Lab. Before that, he was a Post-
[219] M. H. Shahriar, N. I. Haque, M. A. Rahman, and M. Alonso, ‘‘G-IDS: doctoral Research Fellow with the Software Anal-
Generative adversarial networks assisted intrusion detection system,’’ ysis and Intelligence Lab (SAIL), Queen’s University, Kingston, Canada. His
in Proc. IEEE 44th Annu. Comput., Softw., Appl. Conf. (COMPSAC), work has been published at premier software engineering venues, such as
Jul. 2020, pp. 376–385. the TSE and EMSE journals and the ESEC-FSE, ICSME, and ICPE confer-
[220] S. Arisoy, N. M. Nasrabadi, and K. Kayabol, ‘‘GAN-based hyperspectral ences. His research interests include software engineering and performance
anomaly detection,’’ in Proc. 28th Eur. Signal Process. Conf. (EUSIPCO), engineering-related topics. He is one of the vice-chairs of the SPEC Research
Jan. 2021, pp. 1891–1895. Group on DevOps Performance.
[221] Y. Li, T. Jiang, W. Xie, J. Lei, and Q. Du, ‘‘Sparse coding-inspired GAN
for hyperspectral anomaly detection in weakly supervised learning,’’
IEEE Trans. Geosci. Remote Sens., early access, Aug. 12, 2021, doi:
10.1109/TGRS.2021.3102048.
[222] C. Zhao, C. Li, S. Feng, and N. Su, ‘‘Hyperspectral anomaly detection
using bilateral-filtered generative adversarial networks,’’ in Proc. IEEE
Int. Geosci. Remote Sens. Symp. (IGARSS), Jul. 2021, pp. 4408–4411.
[223] T. Jiang, W. Xie, Y. Li, J. Lei, and Q. Du, ‘‘Weakly supervised discrim-
inative learning with spectral constrained generative adversarial network
for hyperspectral anomaly detection,’’ IEEE Trans. Neural Netw. Learn.
Syst., early access, May 31, 2021, doi: 10.1109/TNNLS.2021.3082158.
[224] C. Cooper, J. Zhang, R. X. Gao, P. Wang, and I. Ragai, ‘‘Anomaly
detection in milling tools using acoustic signals and generative adversarial
networks,’’ Proc. Manuf., vol. 48, pp. 372–378, Jan. 2020. PETR MUSILEK (Senior Member, IEEE) received
the Ing. degree (Hons.) in electrical engineering
and the Ph.D. degree in cybernetics from the
Military Academy in Brno, Czech Republic, in
1991 and 1995, respectively. In 1995, he was
MIKAEL SABUHI (Graduate Student Member, appointed the Head of the Computer Applications
IEEE) received the B.Sc. and M.Sc. degrees in Group, Institute of Informatics, Military Medi-
control systems from the Iran University of Sci- cal Academy, Hradec Kralove, Czech Republic.
ence and Technology, in 2016 and 2018, respec- From 1997 to 1999, he was an NATO Science Fel-
tively. He is currently pursuing the Ph.D. degree in low with the Intelligent Systems Research Labora-
software engineering and intelligent systems with tory, University of Saskatchewan, Canada. In 1999, he joined the Department
the University of Alberta. He is also a member of Electrical and Computer Engineering, University of Alberta, Canada,
of the ENergy digiTizAtIon Lab (ENTAIL) and where he is currently a Full Professor. Since 2016, he has been serving as
the Analytics of Software, GAmes And Reposi- the Director of the Computer Engineering Program and an Associate Chair
tory Data (ASGAARD) Lab, working on anomaly (Undergraduate). He is also an Associate Chair (Research and Planning).
detection and empirical performance analysis of cloud software systems. His research interests include artificial intelligence and energy systems.
While he has different artificial intelligence-related research interests, his He developed a number of innovative solutions in the areas of renewable
main focus is on applications of deep learning architectures for anomaly energy systems, smart grids, wireless sensor networks, and environmental
detection with a particular emphasis on applications of generative adversarial monitoring and modeling.
networks (GANs) for anomaly detection.

VOLUME 9, 2021 161029