Artificial Intelligence Review (2023) 56:12725–12769

https://doi.org/10.1007/s10462-023-10433-3

Detection of cross‑site scripting (XSS) attacks using machine

learning techniques: a review

Jasleen Kaur1 · Urvashi Garg1 · Gourav Bathla2

Published online: 23 March 2023

© The Author(s), under exclusive licence to Springer Nature B.V. 2023

Abstract
With the rising demand for E-commerce, Social Networking websites, it has become
essential to develop security protocols over the World Wide Web that can provide secu-
rity and privacy to Internet users all over the globe. Several traditional encryption tech-
niques and attack detection protocols can secure the data transmitted over public networks.
However, hackers can effortlessly exploit them to acquire access to the users’ sensitive
information such as user ID, session ID, cookies, passwords, bank account details, con-
tact numbers, private PINs, database information, etc. Researchers have continuously
innovated new techniques to build a secure and robust system that cannot be easily hacked
and manipulated. Still, there is much scope for novelty to provide security against con-
temporary techniques used by intruders. The motivation of this survey is to observe the
recent developments in Cross-Site Scripting attacks and techniques used by researchers to
secure confidential information. Cross-Site Scripting (XSS) has been recognized as one of
the top 10 online application security risks by the Open Web Application Security Project
(OWASP) for decades. Therefore, dealing with this security flaw in web applications has
become essential to avoid further personal and financial damage to Internet users and busi-
ness organizations. There is a need for an extensive survey of recent XSS attack detection
techniques that can provide the right direction to researchers and security professionals.
We present a complete overview of recent machine learning and neural network-based XSS
attack detection techniques in this paper, covering deep neural networks, decision trees,
web-log-based detection models, and many more. This paper also highlights the research
gaps that must be addressed while designing attack detection models. Further, challenges
researchers face during the development of recent techniques are also discussed. Finally,
future directions are provided to reflect on new concepts that can be used in forthcoming
research works to improve XSS attack detection techniques.

Keywords Web vulnerabilities · Cyber-attacks · Web-security · Machine learning · XSS

attack · Deep learning · Neural networks

* Gourav Bathla
[email protected]
Extended author information available on the last page of the article

13
Vol.:(0123456789)
12726 J. Kaur et al.

Abbreviations
APT Advanced persistent threats
CBoW Continuous bag of words
CMDi Operating system command injection
DBN Deep belief network
DDoS Distributed denial of service
DDQN Dueling deep q networks
DPI Deep packet inspection
ECDSA Elliptic curve digital signature algorithm
FSM Finite state machine
HTTPS Hyper text transfer protocol secure
IDS Intrusion detection system
IoT Internet of things
kNN K- Nearest neighbor
LSTM-RNN Long short-term memory recurrent neural network
MLP Multilayer perceptron
MNN Modular neural network
OSI Open system interconnection
OSN Online social networks
OWASP Open web application security project
PCA Principle component analysis
pdAPSO Primal dual particle swarm optimization
PNN Probabilistic neural network
PSO Particle swarm optimization
QSSVM Quarter sphere support vector machines
RNN Recurrent neural network
RSMT Robust software modelling tool
SARSA State action reward state action
SCADA Supervisory control and data acquisition
SDN Software-defined networks
SQLi SQL injection
SVM Support vector machine
TF-IDF Term frequency-inverse document frequency
URI Uniform resource identifier
URL Uniform resource locator
WAF Web application firewalls
XSS Cross-site scripting

1 Introduction

The proliferation of online services such as medical care, railway services, airline booking,
online shopping, electronic banking, online payments, social networking sites, and many
more have paved the way for cybercriminals to hack online systems and steal users’ sensitive
information (Snehi & Bhandari 2021). The users have become highly dependent upon the
internet and trust most web applications without analyzing their credibility. The common
public is entirely unaware of the black ethics used by hackers to steal their private data. Since
it is impractical to spread cyber security awareness among all individuals, implementing

13
Detection of cross‑site scripting (XSS) attacks using machine… 12727

robust attack detection techniques in web systems is the only possible method to protect user
data and privacy. The attack detection and mitigation tools and techniques aim to find flaws
in the website. These flaws in a website’s design, development, or configuration are com-
monly referred to as web vulnerabilities (Mohammad & Reza 2017). User ignorance and
a lack of formal staff training are frequent factors in the success of web attacks (Gkioulos
& Chowdhury 2021). Most web vulnerabilities are located at the application layer of the
OSI network model (Shabut et al. 2016) (Khan et al. 2017), while human error in website
source code accounts for 93% of data breaches. All public and private organizations refer to
security reports offered by OWASP for security analysis. An OWASP organization aims to
improve the security standards of software(s) by working on open-source projects and host-
ing international conferences so that proper action can be taken to improve the web appli-
cations’ security (owasp 2017). Table 1 illustrates the top 10 web-application-related criti-
cal security risks according to OWASP. It is evident from Table 1 that the most prominent
internet threat among web applications is cross-site scripting (XSS) attack and thus needs
the utmost attention of researchers. Also, the universal cost of threat activities executed by
hackers on online platforms crossed $6 trillion at the end of 2021 (Nick 2021).
Several surveys have been conducted in the past to explore XSS attack detection
techniques (Sarmah et al. 2018) (Gupta & Gupta 2015) (Jagajeevan Rao et al. 2021)
(Kumar & Goyal 2019; Li, 2018; Pitropakis et al. 2019; Rodríguez et al. 2020; Shabut
et al. 2016). Nevertheless, these surveys have not covered the latest developments and
research challenges in detail. The primary objective of this work is to investigate the
XSS attack detection mechanisms, focusing on recent improvements in this field. Fur-
ther, XSS attack and detection are elaborated, which is the most dangerous attack exe-
cuted by the hacker by bypassing the Same Origin Policy. This is done to masquerade
as a victim user and perform all the actions on a website that a legitimate user can do.
Any website is vulnerable to web attack if the input validation and encoding are not
performed accurately. It is estimated that almost two-thirds of web applications (owasp
2017) across the globe are at risk of XSS attacks. A comprehensive study investigated
by the EDGESCAN organization in 2019 and 2021 claimed that (EOIN KEARY, n.d.)
XSS, Brute force and authentication, file path traversal, administrative interface, and
information disclosure vulnerabilities still come under the category of high-risk fac-
tors. Even though substantial research is being carried out to develop new technologies
and methods for detecting, preventing, and mitigating cyberattacks, many websites

Table 1  OWASP Top 10 Vulnerabilities 2020–2021 (owasp 2017)

CVE no Vulnerability name (2020) Vulnerability name (2021)

CVE1 Injection Injection

CVE2 Broken authentication Broken authentication and session management
CVE3 Sensitive data exposure Cross-site scripting
CVE4 XML external entities Insecure direct object references
CVE5 Broken access control Security misconfigurations
CVE6 Security misconfigurations Sensitive data exposure
CVE7 Cross-site scripting Missing function-level access control
CVE8 Insecure deserialization Cross-site request forgery
CVE9 Using components with known vulnerabilities Using components with known vulnerabilities
CVE10 Insufficient logging and monitoring Invalidated redirects and forwards

13
12728 J. Kaur et al.

continue to be targeted by hackers. (Tran et al. 2017). This explains why code-related
vulnerabilities must be recognized to prevent cyber-criminal groups from exploiting
websites. Moreover, according to EdgeScan report 2021, XSS (Reflected/ DOM) is
still at the top, where its percentage of occurrence is 37.20% at the application layer
(Keary Eoin, n.d.), as depicted in Fig. 1.
When an XSS flaw exists in a website, it can accept any untrusted malicious string as
input. The untrusted input is processed either on the client side or the server side, depending
on the vulnerability. In this way, an innocent web user becomes a victim of a web attack by
visiting a vulnerable website, downloading malicious files from an untrusted link, or clicking
on content that looks mischievous. Moreover, there is a significant possibility that if a website
is prone to XSS attack, it is also prone to SQLI, broken authentication, session management
attack, and DDoS attacks. The XSS attack targets HTML webpages by inserting malicious
code or a malicious URL into vulnerable content, like how the SQLI attack targets SQL-
based programs by bypassing SQL queries. Cybercriminals are frequently exploiting XSS
vulnerability to steal user sessions to take over their accounts and perform illegal operations
by masquerading as authenticated users. Also, it has become the most prominent way to get
into the company’s social network and hack private information about the company (Rod-
ríguez et al. 2020). The most notable web applications such as Facebook, Twitter, YouTube,
and TikTok have also suffered from this attack in the past (Fang et al. 2018).
The attacker performs the following steps for the successful execution of the XSS
attack:

• Firstly, the attacker leverages online vulnerability scanning tools to find out websites
with XSS flaws to execute the attack vector.
• Secondly, once a vulnerable website search is successful, the attacker passes malicious
scripts into the input field such as the comment section, search bar, and URL, and ana-
lyzes the processed results produced by the vulnerable webpage.
• The script injected by the attacker can be a JavaScript code, HTML, PHP, or VB
Script. The script is either loaded into the web system database or executed by the web
browser.
• Finally, by overcoming the SOP, the attacker obtains access to the victim’s confidential
credentials through a susceptible website. Figure 2 depicts a sample scenario of any
web-based attack on the web user.

This work aims to review the causes of web attacks and find out research gaps that can
be fulfilled to design an optimized solution for XSS attack detection. The articles selected

Fig. 1  Percentage of occurrence

of attacks

13
Detection of cross‑site scripting (XSS) attacks using machine… 12729

Fig. 2  The general scenario of the XSS attack illustration

for this survey were crawled from Google search, Scholar, Science Direct, Elsevier, Sco-
pus, IEEE explore, Springer, etc. We passed five different types of keywords to search the
paper: (i) XSS vulnerability, (ii) site-injection attacks, (iii) machine learning and cyber
security, (iv) detection of XSS attacks and machine learning, and (v) cyber-security chal-
lenges. The following are the primary research questions of our study:

RQ1: What new advances have been made in the realm of XSS attack detection?
RQ2: Why are traditional techniques not effective at combating XSS attacks?
RQ3: How can XSS attacks be detected and prevented using Machine Learning
techniques?
RQ4: What are the most popular and effective cyber-attack mitigation techniques used
by active researchers?
RQ5: What are the challenges and opportunities for young researchers in the field of
cyber security?

1.1 Motivation and contribution

The key motivation of this extensive survey is to identify the root cause of XSS attacks in
today’s contemporary era where web applications are already built on secure platforms.
During the analysis of existing research works, aforementioned five research questions
were found and our survey paper focuses on finding the answers to these research ques-
tions. The questions have been answered at the end of the manuscript in Sect. 7. Further,
in the past we have observed that researchers have not mentioned recent XSS attack detec-
tion techniques. In this survey, we have explored recent techniques and developments. The
key motivation is to analyze why accuracy is not adequate and what can be the scope of
improvement Fig. 3.
The remaining sections are organized as follows: Sect. 2 gives an outline of the SOP
policy, Sect. 3 describes cybersecurity threats and attacks, Sect. 4 discusses in detail the
significance of machine learning in combating cyber threats, and Sect. 5 elaborates on XSS
attacks, and their types. Section 6 uncovers recent developments in XSS attack detection
using machine learning and Deep Learning. Section 7 discusses the various challenges and
opportunities in this research field, and finally, Sect. 8 wraps up our study by outlining
future directions. Figure 3 depicts the organization of our survey. In the Introduction sec-
tion, the background of cyber-attacks, including a report on OWASP Top-10 vulnerabilities
and percentage occurrence of attacks, are listed. Same Origin Policy, the background of

13
12730 J. Kaur et al.

Fig. 3  Flowchart of the survey

Cybersecurity, machine learning techniques applied for XSS attack detection, and recent
developments are explained in different sections. Finally, the challenges and concepts cov-
ered in the paper are concluded.

2 Same origin policy

Same Origin Policy (SOP) is a vital concept in web applications that comes under the
web-security model enforced by web browsers. It defines a set of rules for the interac-
tion between two different web pages within a web application. Bypassing the SOP, the
attacker gains access to the victim’s confidential credentials via a susceptible website.
Web browsers implement SOP by restricting document A of website A from access-
ing another document B of website B that does not belong to the exact origin. Ori-
gin, in simple terms, refers to the three components: URI scheme, host, and port. URI
schemes- such as HTTP, HTTPS, FTP, etc.- specify the protocol to access the resources.
For example, HTTPS specifies the use of TLS protocol for communication over HTTP.
Host- host refers to a particular computer accessing the world wide web. For instance,
www.​wiki.​com is a hostname separated by dots. Every host will have its network
address and unique host address. Port- port is a programmable docking port through
which the web information flows. It is the communication point where the communica-
tion begins and ends.
The main idea behind the implementation of SOP is to prevent data access of web-
site A by malicious website B if both the websites have different origins, i.e., different
schemes, hosts, and ports. Therefore, SOP is critical in ensuring the confidentiality and
integrity of online applications. (Matt 2021). However, the case study by Schwenk (Guo
et al., 2017) stated that since SOP is not defined consistently in academic and non-aca-
demic areas, web vulnerabilities allow attackers to avoid SOP checks. Table 2 illustrates

13
Detection of cross‑site scripting (XSS) attacks using machine… 12731

a sample example of how a web page document is granted access by comparing the
three components of SOP.

3 Cyber security threats and attacks

Cyber security is a set of security rules, protections, tools, policies, risk-management

strategies, and best practices that protect cyberspace, organizations, and users’ assets
(von Solms & van Niekerk 2013). Cybersecurity aims to preserve the integrity, con-
fidentiality, and availability of user information. Web applications are often versatile
in nature because of the different third-party components used, such as AJAX and
HTML5, and third-party sign-in options via Google, Facebook, etc. All these web appli-
cations are developed using PHP, Ruby, ASP.Net, or Java, where PHP is used by 44%
of business organizations for their website development. With the growing demand for
PHP web development, there has been a significant decline in vulnerabilities since 2016
(Gupta B.B. & Chaudhary Pooja, 2020). All thanks to the implementation of security
patches in PHP scripts by web designers and security professionals that work together
to deal with web attacks. However, designing a flawless web application is still a dream
that needs to be accomplished (Kumar & Goyal 2019). The most popular attack types
leveraged by the attackers for privacy violations are discussed briefly in the following
subsections.

3.1 Ransomware attack

Ransomware has grown in popularity as one of the most dangerous cyber threats to busi-
nesses. (Brewer 2016). The pandemic year 2020 has seen a vast increase in ransomware
attacks on corporate networks (CISCO 2021). When a ransomware infection compro-
mises the victim’s system, the victim is unable to access vital data stored in his system
and is demanded a ransom amount of money to regain data access or get the decryption
key. The victim is given some time to pay in bitcoin. If the victim cannot make the pay-
ment, the ransom increases after the deadline. Data access is denied either by encrypt-
ing the data or by adopting another criminal tactic. In 2013, CryptoLocker ransomware
targeted the Police department and affected almost 250 systems (Kharraz et al. 2015),
while in 2016, the healthcare industry was targeted by SamSam ransomware. In addi-
tion, WannaCry Ransomware appeared in 2017 and affected more than two lakhs com-
puter systems across 150 countries worldwide (Furnell & Emm 2017). Ransomware is
usually spread over the network using phishing emails or exploit kits, allowing infected

Table 2  Verification of web access by SOP

URL Decision Reason

https://​www.​demos​ite1.​com Accessible Protocol, host, port match

http://​www.​demos​ite2.​com Not accessible Protocol mismatch
https://​www.​demos​ite2x.​com Not accessible Host mismatch
https://​www.​demos​ite2.​com:​8081 Not accessible Port mismatch

13
12732 J. Kaur et al.

files to be run on the device. Most crypto-malware add persistence mechanism so that
if the system is rebooted during the malware encryption process, it will restart from the
previous point until it completes the process. The best practice to defend against a ran-
somware attack is to save the backup of all critical files and documents.

3.2 Injection attack

Injection attacks are the oldest and are still executed by cybercriminals to obfuscate
users into revealing their sensitive information. Since this type of attack can allow
attackers to execute remote commands or even install malware onto the system, they
pose a severe threat to web security and give an open challenge to security researchers.
OS command injection, SQL injection, XSS injection, Buffer Overflow, LDAP injec-
tion, CRLF injection, XPath injection, and many more are examples of injection attacks.
OWASP defines an injection attack as an injection flaw that allows a hacker to send
untrusted data or malware to the program known as an interpreter in the network (owasp
2017). The program further executes the hostile data as a part of code or query that
results in alteration of the program. This alteration can lead to further cyber threats,
compromising the whole system of victims, data loss, and data integrity (Kaur & Garg
2022).

3.3 Broken authentication and session management attack

The exploitation of a website through broken authentication happens when the user cre-
dentials are not appropriately secured. In other words, when a user visits the website,
the server provides a unique long session number, commonly referred to as session iden-
tifier or session id, to the user for identification on the server. This session id is stored
on the user’s system for further communication. Whenever the user sends any request to
the server, the browser sends the session ID and the request back to the server. Moreo-
ver, this session id is used throughout the user’s entire session for communication pur-
poses. However, for security reasons, the session id is refreshed every specified interval,
for example, 10 s. The problem occurs when a website cannot update the session id after
some specified time interval or the session id is visible in the website URL due to poor
session management or authentication management. In such a case, the attacker can use
this vulnerability as an opportunity to hijack the account by using the unsecured ses-
sion-id (J. Kaur & Garg 2022). To make predicting the next session ID difficult for the
user, the session ID is randomly produced as a long set of letters and numbers. Depend-
ing on the requirements of the web application, the session id is stored on the server
or the client’s web browser (Wikipedia 2021), (Hassan et al., 2018). Most developers
choose to store the session inside the authentication cookie on the client side due to the
complexity of server-side session management. However, if data quality, legitimacy, and
secrecy are not assured, this typical practice creates more hazards. Thus, it allows hack-
ers to compromise the system through the existing loopholes in the website.

3.4 Cross‑site scripting attack

Another prominent type of code injection-based online assault is a Cross-Site Scripting

attack, commonly referred to as an XSS attack. XSS exposure has affected 80% of web

13
Detection of cross‑site scripting (XSS) attacks using machine… 12733

applications over the globe (Dora and Nemoga, 2021) (Gao et al., 2010). Due to this reason,
it has become the most dangerous web vulnerability and a recent open challenge among
security researchers and web designers (Geetha and Thilagam, 2020). Hence, research on
security in the XSS attack detection model needs to be overlooked (Morzeux and Http-
ParamsDataset, 2021) (Gkioulos and Chowdhury, 2021). An XSS vulnerability is the root
cause of online data breaches and privacy violations in big organizations. In 2018, the larg-
est airline in the UK- British Airways, suffered a data breach caused due to XSS vulner-
ability in the online payment module where data of 3,80,000 users was stolen (Gupta and
Pooja, 2020). In 2020, Facebook was found vulnerable to an XSS attack, and the flaw was
in its login button where window.postmessage () function allowed cross-communication
between window objects (Leyden 2020). XSS vulnerability earned $20,000 for security
researchers via a Bug Bounty Program. In 2021, more than 98.2 million individuals from
the healthcare industry and financial, professional, and retail services were affected by data
breaches only in the first half of the year, as reported by CRN news on July 2021 (Novin-
son Michael 2021). In March 2021, Gmail’s AMP (Accelerated Mobile Page) feature was
found to be vulnerable to XSS by a group of security researchers (XSS in Gmail’s AMP For
Email Earns Researcher $5,000 | The Daily Swig, n.d.). AMP was introduced by Google in
2019 to make email content dynamic and more interactive for customers. Further in 2022,
XSS vulnerability was exposed in Chrome browser extension- Screencastify, WordPress
5.8.3, Microsoft Teams, Google Cloud, and Google Play (Screencastify Chrome Extension
Flaws Allow Webcam Hijacks, n.d.),(XSS Vulnerabilities in Google Cloud, Google Play
Could Lead to Account Hijacks | The Daily Swig, n.d.), (WordPress Stored XSS Vulnerabil-
ity—Update Now, n.d.), (Microsoft Teams Security Vulnerability Left Users Open to XSS
via Flawed Stickers Feature | The Daily Swig, n.d.).
Any website is vulnerable to an XSS attack when it accepts untrusted content and
inserts it into the benign website without proper input encoding or input validation.
The attacker-supplied code can be written in any technology that a web browser sup-
ports such as HTML/JavaScript, VBScript, ActiveX, Java, Flash, or any other browser-
based technology (Rodríguez et al. 2020). Upon injecting the malicious query into the
web browser, the injected script will execute in the confidential zone of the website
and thus allowing admin privileges to transmit sensitive information to the attacker.
For instance, if a hacker inserts a malicious JavaScript code into the comments sec-
tion or any other section that accepts some input from the user, and the browser further
executes the malicious code/URL with the dynamic content, then it means that web-
site is vulnerable to an XSS attack. XSS vulnerability with high-risk factors can even
disclose the user’s session cookies, thereby allowing the attacker to compromise the
entire user’s account. If a web application is exposed to an XSS attack, it is also vul-
nerable to DDoS (Distributed Denial of Service) and ATO (Account Takeover) attacks.
Thus, a single XSS flaw can result in malware installation, session hijacking, sensitive
data disclosure, and social engineering attacks. Table 3 lists some of the popular XSS
payloads utilized by hackers to compromise a system (Conti et al., 2018) (Rodríguez
et al. 2020).
XSS attacks can be classified into four different categories depending upon the
nature of the threat vector

13
12734 J. Kaur et al.

Table 3  XSS Payloads Sr. no Types of XSS payloads

1 Steal cookies without page redirection

2 Accessing page HTML
3 Creating alert boxes and commands
4 Reauthentication dialogs
5 Injection of Java payload
6 Keyloggers
7 Webpage redirection
8 Tracking browser details
9 Web-page defacement
10 Auto-run exploits when connected to BeEf
11 Network enumeration
12 Ping sweep network
13 Rewriting of links
14 Stealing clipboard
15 Detection of TOR
16 Metasploit AutoPWN

3.4.1 Persistent/ stored XSS attack

A website is vulnerable to persistent XSS attacks when the malicious script injected by the
hacker gets stored on the web server and database; hence, the name stored XSS attacks.
Since the attack vector injected into the website gets stored in the database, it can affect
millions of users visiting the webpage simultaneously. Moreover, the malicious script gets
loaded into the system automatically even if the user does not click on any link. Therefore,
it is also considered the most dangerous type of vulnerability. Figure 4 illustrates the sam-
ple scenario where a victim accesses a vulnerable webpage with a malicious script. As it is
crystal clear from the diagram below, a web user loses access to his confidential data when
he pays a visit to a vulnerable web page, and a malicious script gets loaded into his system
implicitly.

3.4.2 Non‑Persistent/reflected XSS attack

A reflected XSS attack is another common attack performed by the hacker. A website is
prone to this attack when it echoes back the content or a string inserted by the user in
its input fields, such as the search bar, comment section, and log-in text box. When the
hacker identifies a reflected XSS flaw, he creates a malicious payload and inserts it into the
webpage as a link or shares it directly with the user via email. Once the user clicks on the
malicious link, the hacker can steal the user’s cookies and login credentials. Both reflected
and stored XSS are server-side vulnerabilities. Figure 5 illustrates the basic scenario of the
reflected XSS attack where the user clicks on an untrusted malicious link, thinking it came
from a trusted source. Within a single click, the victim gets directed to a malicious web
page and compromises his system unintentionally.

13
Detection of cross‑site scripting (XSS) attacks using machine… 12735

Fig. 4  Persistent XSS attack

Fig. 5  Reflected XSS attack

3.4.3 DOM‑based XSS attack

It is a client-side vulnerability where the malicious script is loaded in the web browser’s
DOM (Document Object Model). Whenever a victim visits a web page with a DOM
vulnerability, injected code can alter the content of DOM and the values of the object’s
properties. Figure 6 illustrates the sample scenario of a DOM-based XSS attack on a web
application. This vulnerability is exploited through client-side scripts such as JavaScript,
VB Script, AJAX, ActiveX, JQuery, or any other technology supported by web brows-
ers. There is a common myth among people that AJAX (Asynchronous JavaScript and
XML) and JQuery are secure concepts and thus enhance web security. However, with
the usage of image tags and AJAX requests, it is possible to execute direct attacks on
third-party applications via DOM-XSS (Hickling 2021). The attacker can thus retrieve
cookies and sensitive information like user credentials and IP addresses. One thumb rule
to detect DOM-XSS flaw is to inject malicious script into the browser and check whether
the script is visible in the server code. If the malicious payload is visible in the browser’s

13
12736 J. Kaur et al.

Fig. 6  DOM XSS Attack

page source but not in the server code, it concludes that there is a DOM-XSS flaw in the
web page.

3.4.4 Mutation‑based XSS attack

In 2013, Dr. Mario found a new type of XSS attack called mutation XSS (Heiderich et al.,
2013). Due to the obvious variances in browser interpretation of HTML standards, M-XSS
is self-referential, and the iteration is very uncertain. InnerHTML verifies the automated
alteration of HTML text inside the browser, allowing the attacker to circumvent the DOM.
Google was formerly vulnerable to a mutated XSS attack, which was disclosed to Google
by security researcher Masato Kinugawa in February 2019 (Nidecki Tomasz Andrzej
2019). Figure 7 depicts an example of a mutation-based XSS attack scenario. In this attack,
when the hacker inserts malicious HTML content into the web application, the web appli-
cation filters the injected content and passes it to the browser. This further modifies the
content of innerHTML while parsing the markups. However, in this process, the browser
mutates the contents, and unknowingly it leads to an XSS attack. Owing to this, MXSS is
challenging to track.

Fig. 7  Mutation-based XSS

Attack

13
Detection of cross‑site scripting (XSS) attacks using machine… 12737

4 Cyber security threats and attacks

To combat web security-related issues, researchers have raised an urgent need to utilize
machine learning techniques. This is so because machine learning techniques can detect
new attacks within a small timeframe without performing any code or file analysis (Khan
et al. 2017) (Thakkar & Lohiya 2021). In this section, we will delve deeper into the related
work done by several researchers. In the end, we will come out with the research gaps
that must be resolved to overcome the vulnerability issues. Artificial Intelligence (AI) has
become ubiquitous because of its ability to solve complex problems such as language anal-
ysis, speech recognition, machine translation, robotics, cyber defense, and many more. The
rising interest in AI is driven by its advanced computing power (Zhang et al. 2021b).
The authors (Jian-hua Li 2021) presented a detailed review of the role of machine learn-
ing in cyberspace. The author discussed the various machine learning and Deep Learning
algorithms used to combat security issues in web applications. Furthermore, the author
studied numerous attacks that an AI model itself may suffer. The researcher also focused
on developing a secure AI model by encrypting the deep neural network. Support Vec-
tor Machine (SVM) requires more training time, which is a limitation. However, it can be
trained with AVL Tree to reduce the training and attack detection time. Attribute-QSSVM
and Temporal Attribute-QSSVM have less computational complexity than SVM. Since
most of the data contain irrelevant fields and is noisy, reducing the amount of data for
yielding effective results by applying dimensionality reduction algorithms is of utmost
importance (Laghrissi et al. 2021). Traditional machine learning techniques such as Sup-
port Vector Machine (SVM), Decision Trees, Random Forests, K-nearest neighbors, and
feature selection techniques uses shallow learning and, therefore, are ineffective at per-
forming well in a real-time environment. The authors (Chauhan et al. 2019) represented a
comparison between shallow and Deep Learning methods for predicting the performance
of stroke patients. The comparison was made between CNN and traditional machine learn-
ing methods, including Ridge Regression (RR), PCA, and SVR (Support Vector Regres-
sion), a kernel-based machine learning regression model. Their results indicate that deep
Convolutional Neural Networks (CNN) did not perform better than the shallow learning
methods. However, their hybrid RR model (with CNN and PCA features as input) outper-
formed other methods (CNN, PCA+RR, SVR). Moreover, their observations conclude that
CNN yields the best results only when more data is available.
On the one hand, machine learning attack detection models can learn the general pat-
terns related to web attacks and can quickly analyze different types of attacks by differen-
tiating their properties using classification algorithms. On the other hand, the performance
of machine learning algorithms strongly depends upon the accuracy of feature extraction
(Onan et al. 2016) (Onan 2021a, 2021b; Onan et al. 2017) (Onan 2018b) and recogni-
tion, size of the training set, and training time. That means that ML algorithms will clas-
sify the data samples as per the selected features only, and features that are not mentioned
will escape attack detection (Li, 2018) (Fang et al. 2019). Despite its significant progress
in recent years, machine learning techniques can be attacked or deceived (Li, 2018). For
instance, training or testing samples can be modified, the working environment can be
manipulated, and an external model can be added to affect the system’s performance. Since
the AI model is evaluated upon the testing dataset, it is impossible to verify the accuracy
of the AI model in a real-world scenario. Therefore, most of the traditional ML algorithms
may generate false true positives and false true negatives. Another limitation of the ML
algorithm is that features are compiled manually, which is costly and time consuming

13
12738 J. Kaur et al.

Fig. 8  Classification of Machine Learning Techniques

(Onan 2022). Furthermore, many security mechanisms do not implement a detailed packet
analysis, leading to insufficient feature generation to detect zero-day attacks (Wang et al.
2022). However, Deep Learning can effortlessly produce positive results with a large
amount of data in relatively less time as the models can be trained without implement-
ing feature extraction (Onan, 2020). The next section presents several aspects of machine
learning that can be leveraged in security standards and protocols to safeguard websites
from cyber criminals. Figure 9 represents the existing ML techniques available for identi-
fying web attacks. ML techniques are of five types: (i) supervised learning models, which
include classification and regression problem solving, (ii) unsupervised learning models
mainly cover clustering algorithms such as K-means clustering, PCA, fuzzy-k cluster-
ing, (iii) semi-supervised which integrate benefits of supervised and unsupervised learn-
ing models while covering graph-based algorithms, self-training models, SGAN, etc. (iv)
Reinforcement learning models, and (v) Deep Learning models which include several neu-
ral network-based models. The following Fig. 8 represents some of the popular machine
learning techniques used by researchers to tackle with XSS attacks detection mechanism.

13
Detection of cross‑site scripting (XSS) attacks using machine… 12739

5 Machine learning techniques for XSS attack detection

Researchers have shifted their focus from traditional XSS attack detection techniques to
machine learning-based detection techniques because ML techniques have several advan-
tages over conventional mechanisms such as detecting malicious patterns automatically
and higher accuracy scores. Owing to the weaknesses in traditional XSS attack detec-
tion mechanisms, researchers are shifting their focus to machine learning based detection
approached. The following Table 4 represents the Retrospective analysis and inferences
of existing literature with a comparison of best practices followed by the researchers and
scope of improvements in their proposed approaches.
XSS attack mitigation techniques using machine learning can be classified into the fol-
lowing categories:

5.1 Supervised machine learning

Supervised learning is the first major category of machine learning prediction models
where the labeled dataset is fed to the input section of the learning algorithm. Several
techniques have been used from supervised learning to prevent and detect XSS vulner-
abilities in web applications. Supervised machine learning is one of the prominent areas
leveraged for the implementation of XSS attack detection techniques. Decision trees
such as Random Forests, ADTree, J48 along with boosting algorithms are implemented
by researchers. Apart from Decision Trees, other supervised models such as Support
Vector Machines, Principal Component Analysis, kNN, mutual information, Naïve
Bayes and SVM are also popular. Some of them are discussed in the upcoming para-
graphs below:
Kaur Gurpreet et al. (G. Kaur et al. 2018) introduced an ML-based model that can
identify the malicious attack vector before it is processed by the browser on the victim’s
system. They used the Linear Support Vector classification algorithm to identify blind
XSS and stored XSS attacks. During the features generation phase, authors leveraged
the JS events and scripts that the hackers inject to attack the website. The dataset used
for experimentation purposes was Linear separable while the experiment was simulated
on Mutillidae, a free susceptible website. With a recall value of 0.951 and a false posi-
tive rate of 0.111, the detection accuracy is 95.4 percent.
According to S. Sharma et al. (Machine Learning Based Intrusion Detection System
for Web-Based Attacks, 2020), feature set extraction is crucial in detecting web-based
attacks. Therefore, the authors proposed a feature set extraction approach that, if applied
with ML based intrusion detection model, can improve the results to a significant extent.
They performed the experiment using CSIC HTTP 2010 dataset and the Weka tool. The
experiment was performed in three steps. The first step involves preprocessing the data
of the dataset with a python script, while the second step focuses on extracting features
from the dataset that contains specific keywords such as GET, POST, DROP, DELETE,
MODIFY, UNION, DROP, etc., before feeding it into the Weka for data modeling. In
the final phase, the data is fed into the three ML models, namely, J48, OneR, and Naïve
Bayes in Weka, where J48 produced the best results compared to other classifiers. The
same results were reported by (Alam & Pachauri 2017) regarding the performance of
the J48 classifier, where they leveraged Weka tool to detect fraudulent activity in credit
card payments. J48 decision tree requires less time to construct the model, prediction
accuracy is higher, and the numbers of correctly classified instances are also more. A

13
 Table 4  Retrospective analysis and inferences of existing literature with a comparison:
12740

ML solution Reference Best practices followed Scope of improvement

13
Supervised learning (G. Kaur et al. 2018) Their technique can detect XSS attack before the mali- Focus area is only blind XSS attack while DOM XSS,
cious script is executed on the vulnerable web page reflected XSS are not considered. Research is carried
on using only JavaScript events feature set. Accuracy
of attack detection could have been improved
(R. Wang et al. 2015) Features are extracted based on keywords and their fre- Covers attacks detection only on social networking sites.
quency related to URL, JavaScript and HTML tags. It Mechanism to prevent overfitting problem and biasness
considered the spreading speed of attack is not covered
(Kascheev & Olenchikova, 2020) A comparison of four machine learning algorithms is Using t-SNE for data dimensionality results in slow
made: Decision tree, Naïve Bayes classifier, Logistic computation when applied over a dataset of more than
regression, and support vector machine, where the 10000 records. Therefore, it can be combined with
decision tree showed the best performance rates PCA to reduce the noise in dataset and to improve the
computational speed
(Machine Learning Based The authors proposed a feature set extraction approach that, Their model does not work on higher layers and consid-
Intrusion Detection System for if applied with ML based intrusion detection model, can ers only HTTP requests. Features such as JavaScript
Web-Based Attacks, 2020) improve the results to a significant extent. The authors keywords, HTML tags, and malicious URLs and IP
used three models i.e., J48, OneR, and Naïve Bayes in addresses could be used to improve the detection
Weka, where J48 produced the best results results in real time
(Munonye & Péter 2021) Researchers focused on fixing OAuth vulnerabilities that The technique considers only HTTP requests for feature
can assist in identification of XSS attacks. They utilized generation. Features such as JavaScript keywords, HTML
Principal Component Analysis and Factor analysis to tags, and malicious URLs and IP addresses could be used
extract features that are more likely to impact the output to improve the detection results in real time
(Rathore et al. 2017) 10 ML classifiers were evaluated to measure the The approach does handle complex dataset and extract-
performance of the detection model, tree classi- ing SNS features is more time-consuming. Considering
fiers (Random Forest and ADTree) showed the best JavaScript keywords can help improve results in real-
results. Features are extracted using URL, HTML and time. Overfitting and biasness problems can be handled
Social networking sites keywords while working with decision trees
J. Kaur et al.
 Table 4  (continued)
ML solution Reference Best practices followed Scope of improvement
(Khan et al. 2017) Researchers used lexical analysis LALR for generation Performing static analysis on JavaScript code is time
of JavaScript token and to detect obfuscated code consuming. Model may give inaccurate results with
which is used by hackers to avoid attack detection large dataset as Decision Tree can result in overfitting.
Large dataset could have been used as in real time,
dataset used is quite large
Unsupervised learning (Hoang 2020) PCA and CART decision tree algorithm is applied for Overfitting and biasness problems can be handled while
XSS detection. Experiment is executed on real web working with decision trees. Techniques to detect
logs collected from web servers obfuscated code can be included
(Fang et al., 2019) Researchers proposed four new attack escaping strate- Technique can be implemented on multiple datasets to
gies: replacement of sensitive words, encoding obfus- get better results in real-time scenarios
cation, positional morphological transformation, and
adding special character
(Tariq et al. 2021) They used a threat intelligence model, reinforcement When one of the patterns is altered, slight change in
learning, GA, and statistical inference, and a combi- accuracy is observed
nation of GA and statistical inference to guard against
any XSS assault
Detection of cross‑site scripting (XSS) attacks using machine…

(Praise et al., 2020) Signature-based pattern matching algorithm is inte- Attack escape detection strategy is not applied in the
grated with reinforcement learning to block malicious firewall implementation
attacks in cloud environments
Deep learning (Feng et al. 2019) Researchers implemented a plug-and-play ready-to-use KD99 dataset is used which is old dataset. The technique
device to detect DOS attacks, XSS attacks, and SQL can be implemented with latest datasets to verify the
attacks using DNN, CNN, and LSTM models accuracy. Also, the dataset does not consider XSS
attack. Techniques to detect obfuscated code can be
included
(Laghrissi et al. 2021) Comparison between three LSTM classifier-based KD99 dataset is used which is old dataset. The technique
Intrusion Detection Systems is done: LSTM, can be implemented with latest datasets to verify the
LSTM-PCA, and LSTM-MI, where the PCA-based accuracy. Also, the dataset does not consider XSS
model produced the best results for both binary and attack. Techniques to detect obfuscated code can be
multiclass (3 class: Normal, Remote to Local Attack, included
DoS attack) classifications. PCA and MI is applied to
reduce the data dimensions
12741

13
 Table 4  (continued)
12742

ML solution Reference Best practices followed Scope of improvement

(Abaimov & Bianchi 2019) The concept of the pre-processing module, based on Domain specific semantic knowledge is required. Tech-

13
the type of keyword or symbol, removes noisy data niques to detect obfuscated code can be included
from input queries so that ambiguous data value
pairs from the dataset can be removed. This directly
impacts the neural network’s performance, thereby
increasing the detection accuracy from 75 to 95%
(Yan et al. 2018) Abstract Syntax Tree is used to extract features of Techniques to detect obfuscated code can be included
JavaScript code. Chi-square test is done to select the
top n-grams
(Pan et al. 2019) Their method uses RSMT to detect many sorts of Technique can be implemented on multiple datasets to
assaults with distinct features without manual extrac- get better results in real-time scenarios. Techniques to
tion. Web security domain knowledge is not required. detect obfuscated code can be included
Autoencoders is used to evaluate the model
(Mokbal et al. 2019) Their proposed approach is integrated with the Multiple datasets can be used
dynamic feature extractor and Back-Propagation
algorithm for XSS attack detection
(Abdullah Alqarni et al. 2022) Decoder is used to decode any encoded malicious Word2vec is not suggested for XSS attack detection as
script obfuscated the attacker it is time consuming, therefore, it should be avoided.
The dataset used in this study is small and unbalanced,
i.e., 100000 benign samples, while malicious samples
are only 1000. Attention mechanism can be included to
yield maximum results
(Luo et al. 2021) A combination of three deep learning-based models Using MLP as an ensemble model could result in com-
i.e., MRN, CNN, and LSTM is utilized to analyze the plex and time-consuming computations
URL requests for anomalies. Search is carried on real
world dataset collected from security company with
good amount of records
(Tian et al. 2020) For data discrimination, the FastText deep learning Word2vec is not suggested for XSS attack detection as it
neural model has been used with some modifications. is time consuming, therefore, it should be avoided
The model is trained over three datasets: HTTP CSIC
2010, FWAF, and HttpParams dataset
J. Kaur et al.
 Table 4  (continued)
ML solution Reference Best practices followed Scope of improvement

(Kaur & Singh 2019) Researchers leveraged Misuse Detection Engine Techniques to detect obfuscated code can be included
(MDE), where the main job is to differentiate the
common attacks from the network traffic
Detection of cross‑site scripting (XSS) attacks using machine…
12743

13
12744 J. Kaur et al.

study by K. Munonye and M. Peter (Munonye & Péter 2021) was solely focused on
fixing OAuth vulnerabilities. They utilized Principal Component Analysis and Factor
analysis to extract features that are more likely to impact the output. To identify the
various challenges related to the domain and OAuth workflow, the Finite State Machine
model was implemented. Features were extracted manually because of the unavailability
of the existing dataset for the research. For detecting the successful workflows, Gradient
Boost Classifier (GBC) was used, which produced 0.82 accuracy and 0.71 value for the
ROC curve.
R Wang et al. in (R. Wang et al. 2014) developed an approach that detects XSS
worms in Online Social Network web pages. The benign and malicious web pages
are quite different and can be differentiated by counting the frequencies of scripting
functions in both. In their model, the researchers opted for the ADTree decision tree
and AdaBoost.M1 (Adaptive Boosting) algorithms for the classification because of
the higher accuracy of ADTree than other decision trees and AdaBoost.M1 produces
a strong classifier. A feature extractor model is also developed to automatically cap-
ture the features from the webpages because feature extraction and database creation
play a crucial role in generating a classification model. Benign and malicious samples
are collected from DMOZ and XXSed Database. The researchers extracted four groups
of features from web pages, namely, (1) keyword features, (2) JavaScript features, (3)
HTML tag features, and (4) URL features, where each of these features further covers
sub-features. Researchers in this paper used the Weka tool to generate and evaluate the
classification models and the tenfold cross-validation method for evaluation. However,
the technique proposed produces a high false-positive rate of 4.20% and gives a shallow
detection rate. ADTree generates precision value 0.938, recall value 0.936, F-Measure
0.936 while AdaBoost.M1 gave higher values i.e., precision = 0.941, recall = 0.939,
F-measure = 0.939.
S. Kascheev and T. Olenchikova in (Kascheev & Olenchikova 2020) proposed super-
vised machine learning algorithms to detect XSS attacks using the cross-validation method
to train 70% of the data and generate 30% of test samples. The model’s performance is
evaluated based on correct responses, accuracy, completeness, and F-measure. Their data-
set contains 40,000 malicious codes and 200,000 lines of benign code. First, each request
is decoded into Unicode characters, and regular expressions are used to extract the que-
ry’s parameters. However, abnormal queries with many parameters are removed from the
dataset. The source data set is represented using Word2Vec. To reduce the data dimen-
sion and restore the data nesting structure, T-distributed Stochastic Neighbor Embedding
(t-SNE) is used. A comparison of four machine learning algorithms is made: Decision tree,
Naïve Bayes classifier, Logistic regression, and support vector machine, where the decision
tree showed the best performance rates. Metrics used are accuracy, precision, recall, and
F-measure.
R. Banerjee et al. (Banerjee et al. 2020) investigated the use of four ML algorithms,
namely SVM, KNN, Random Forest, and Logistic Regression, for the detection of XSS
attacks. The true and false values in the dataset were mapped by employing the Logistic
Regression model. The experiment was performed using python and scikit library on the
dataset having 24 attributes based on URL and JavaScript features. Of the four classifi-
ers, Random Forest Classifier had shown promising results with high accuracy and a low
false positive rate. S. Rathore et al. (Rathore et al. 2017) proposed XSSClassifier to detect
an XSS attack on social websites such as Facebook and Twitter. Their approach used a
machine learning classifier for attack detection. First, the features are extracted from URLs,
social networking sites (SNSs), and HTML content. During the evaluation phase, it was

13
Detection of cross‑site scripting (XSS) attacks using machine… 12745

Fig. 9  XSS Vulnerability Interceptor

found that the detection model generates higher accuracy when fed with SNS features to
the classifier. While a total of 10 ML classifiers were evaluated to measure the performance
of the detection model, tree classifiers (Random Forest and ADTree) showed the best
results with 97.2% accuracy and a false positive rate of 0.87 in real-time.
Nayeem Khan et al. in (Khan et al. 2017) also used four different machine learning clas-
sifiers: SVM, KNN, J48, and Naïve Bayes for XSS attack detection, where J48 outper-
formed with 99.22% accuracy when the dataset was split into training and testing data.
Moreover, an interceptor plugin is implemented that ensures secure communication
between the browser and server, as shown in Fig. 9.
The HTTP GET response is sent to the server via the interceptor. If any malicious activ-
ity is found in JavaScript code, the web page is blocked automatically before being sent
to the browser. While the approach is browser and platform-independent and efficiently
detects XSS attacks in real-time, it is also lightweight in nature and requires minimal runt-
ime overhead. Features to be fed into the classifier are extracted through static analysis.
To break down the website’s source code into a list of tokens, a yacc parser is used in the
detection model. Furthermore, lexical analysis is performed to remove any obfuscated code
used by hackers to hide malicious activity. Lexical analysis is a security scanning compo-
nent that breaks down a source code into tokens before parsing (C.C. Michael & Steven
Lavenhar, 2013).

5.2 Unsupervised machine learning

Unlike supervised machine learning, unsupervised learning is trained on the dataset which
is unlabeled. Therefore, these models carry the ability to predict unknown hidden patterns
and cluster the data without any human supervision. K-means clustering algorithms, PCA,
dimensionality reduction, and autoencoders are some of the unsupervised learning models.
Now let us discuss further some of the popular unsupervised machine learning models to
detect an anomaly or malicious behavior of web applications in the below paragraphs.
(Wang et al. 2015) proposed an improved version of a model motivated by the n-gram
model for identifying XSS in online social networks. They utilized a set of correlated
data values for the feature extraction purpose. They extracted features based on keywords
related to URL, JavaScript, and HTML tags and developed their classifier using the Alter-
nating Decision Tree. Taking into account the spreading rate of XSS attacks in online
social networks, they considered the spread speed of suspicious identified features. The
detection model is iterated twice in the implementation stage, where they used their pro-
posed model in the second iteration to detect malicious samples. In terms of precision and
memory, their strategy was successful.

13
12746 J. Kaur et al.

Laghrissi F. et al. in (Laghrissi et al. 2021) presented a comparison between three LSTM
classifier-based Intrusion Detection Systems (IDS): LSTM, LSTM-PCA, and LSTM-MI,
where the PCA-based model produced the best results for both binary and multiclass (3 class:
Normal, Remote to Local Attack, DoS attack) classifications. KD99 dataset was used and
divided into three categories in Google Colab: testing data, training data, and validation data
(6:2:2). For evaluation, F-1 score, accuracy (training + testing), sensitivity, and precision are
used. LSTM produced more effective results for binary classification than multiclass classifi-
cation, while MI performed better with four feature sets. LSTM-PCA produced a recall value
of 0.99 for both binary and multiclass, precision of 1.0 for binary and 0.99 for multiclass,
F1 score of 0.99 for both types of classes, training accuracy of 99.49 for binary and 99.39
for multiclass, and testing accuracy as 99.44 for binary and 99.36 for multiclass. X.D. Hoang
(Hoang 2020) proposed a web logs-based machine learning detection model for common web
attacks like SQLi, XSS, Path traversal, and CMDi. The HTTP Param dataset contains 20000
payload requests which are used as a training set and 11067 as a testing set. Under the detec-
tion stage, URIs are first extracted from the weblogs to add in input further to the detection
model using the n-gram approach (n=3). In the second step, URIs are pre-processed to con-
vert them into vector representation by incorporating the TF-IDF method. Finally, the PCA
method is implemented to reduce the number of features. The CART decision tree model is
implemented for training purposes as it is the fastest and best match for real-time detection
systems. For detection purposes, the classifier produced from the training stage is utilized to
classify the URI vector and generates either normal or attack results. The experiments con-
clude that the overall detection rate is 98.52%. To explain, the results for average payload,
SQLi, and path traversal only are acceptable; however, for XSS and CMDi attacks, the detec-
tion rate (accuracy) is meager i.e., 85.88 and 73.33%, respectively, because of the amount of
training data used. Table 5 below represents a confusion matrix that is commonly used to eval-
uate the performance of XSS attack detection model that predicts whether an instance belongs
to a positive (malicious) or negative (genuine) class.

5.3 Reinforcement machine learning

Reinforcement machine learning has shown favorable results in the cybersecurity field. It uses
trial-and-error strategy where the learning model performs a sequence of steps and receives
a reward based on the outcomes. Algorithms such as SARSA, Q-Learning, and deep Q net-
works come under this category.
Yong Fang et al. (Fang et al. 2019) developed the RLXSS approach for XSS attack detec-
tion using reinforcement learning. The authors came up with the idea of implementing the
adversarial and retraining models, where the former is supposed to deal with adversarial
attacks, and the latter would re-identify the malicious samples fed by the former model. For
optimization reasons, adversarial samples collected from the adversarial model were put into
the retraining model. Apart from this, four new attack escaping strategies were also proposed:
replacement of sensitive words, encoding obfuscation, positional morphological transforma-
tion, and adding special characters. Ultimately, this technique converts the XSS escape attack
into a fleeing tactic. Their methodology can detect adversarial assaults that are undetectable
by white-box and black-box testing. Both white-box and black-box situations are considered
when determining the reward function. The reward for the white-box environment is decided
based on confidence level, while for the black-box model the parameter for reward is whether
it escapes attack detection or not. The following reward function is applied:

13
Detection of cross‑site scripting (XSS) attacks using machine… 12747

Table 5  Confusion matrix Actual values

Genuine Malicious

Predicted values Genuine True negative False positive

(Type-I
error)
Malicious False negative True positive
(Type-II error)

{
result ∗ score, if blackbox = 1
rt = 1 − result
∗ score, if whitebox = 0 (1)
1 − threshold

where 1 indicates True value, 0 indicated False value, the result is the feedback value for a
reward under a black-box environment, the threshold is the confidence value under a white-
box environment and the score indicates success in escape. To evaluate their model, DR
(Detection rate) and ER (Escape rate) are used and their formulas are mentioned below:
Count of malicios samples detected
DR = (2)
Total count of malicious samples

ER = 1 − DR (3)
After the calculation of ER and DR, these evaluation metrics are further tested using
SafeDog, XSSChop, LSTM, ADTree and AdaBoost where the LSTM model gave the
higher results. However, the results were slightly lower than SafeDog and XSSChop.
Praise J et al. (2020) in (Praise et al. 2020) proposed an effective firewall filtering
method that utilizes machine learning for malicious activity. Signature-based pattern
matching algorithm is integrated with reinforcement learning to block malicious attacks in
cloud environments. By using a two-way pattern-matching strategy, results are calculated
in less time complexity. Two-way pattern marching compares the signature of every incom-
ing packet for possible attacks. Furthermore, DPI, a packet filtering technique, is utilized to
scan the payload and header content of IP packets that are passed from public networks to
private networks.
Figure 10 represents Reinforcement based and pattern-matching RLPM architecture for
malicious packet filtration by the firewall. In RLPM, the input data is first passed to the
firewall, and firewall log then further classifies the data based on log information to gener-
ate new rules. The new rules are further compared with performance log data to generate
optimized rules. These updated rules are further fed into the firewall rule database for clas-
sifying the data packet as malicious or non-malicious. The process is repeated for all the
incoming packets passing through the network and their signatures are matched with rel-
evant packet signatures available in the database. If a signature is not matched, the packet is
blocked to enter the network. ECDSA is used to generate the dataset.
In 2021, (Tariq et al. 2021) created a hybrid solution for combating XSS in web
applications. They claim to be the first to combine a machine learning approach with a
metaheuristic algorithm, such as a Genetic Algorithm. They used a threat intelligence
model, reinforcement learning, GA, statistical inference, and a combination of GA and sta-
tistical inference to guard against any XSS assault. Their findings show that using only
genetic algorithms to identify XSS attacks leads to a large number of false positives. XSS

13
12748 J. Kaur et al.

Fig. 10  RLPM based firewall Architecture

vulnerabilities can be accessed via static analysis, dynamic analysis, or hybrid analysis.
However, one limitation of this approach is that only limited raw inputs must be fed into
the system to reduce the training time. Although there are innumerable techniques avail-
able for the detection of XSS attacks such as static analysis, dynamic analysis, hybrid
analysis (Mohammad & Reza 2017), input filtering methods, output escaping techniques,
and signature-based detection. However, these techniques have proved to be ineffective and
time-consuming due to numerous reasons. First and foremost, cybercriminals have devel-
oped stealth techniques to bypass the detection models. To explain, attackers use evasion
techniques (Jiang et al., 2020) (Li, 2021) and HTML/ URL/ Hex/ Base64/UTF-7 encod-
ing (Kascheev and Olenchikova, 2020) (Conti et al., 2018) to bypass static filters, escape
sequences, and comment symbols. In addition, although the static analysis method can
detect vulnerability before the code is exploited, it has high error and false-positive rates
(Katsikeas et al., 2021). Moreover, signature-based detection, heuristics, firewalls, and
dynamic analysis fail to detect new types of modern attacks. Antivirus software whether
online or offline implements signature-based attack detection models and hence, requires
subsequent updates of the attack vector definitions. Signature-based detection tools fail to
keep an eye on new types of attack vectors as the definition of a new attack is unavailable
in their detection mechanism. In signature-based detection methods, the signature of the
attack is captured and its pattern is then stored in its database to protect the device against
the same type of attack in the future. Owing to this, machine learning techniques have
gained attention for the past decade (Kaur et al., 2018) and researchers are enlisting ML
models in the detection of malware. They can even prove to be fruitful in the detection of
threats that are highly dynamic (Jiang et al., 2020).
The following evaluation parameters are used in machine learning XSS attack detection
models:
Confusion matrix:
Another important term in machine learning prediction models is a confusion matrix.
It’s an N-by-N matrix deployed for assessing the efficiency of prediction algorithms where
columns represent actual values, and rows represent predicted values, as depicted in
Table 6.

(i) Accuracy
The number of observations accurately recognized from the total number of predictions
is called accuracy. Accuracy is calculated using the following Eq. (4):

13
Detection of cross‑site scripting (XSS) attacks using machine… 12749

TP + TN
Accuracy = (4)
TP + TN + FP + FN
Even if the model’s consequent value is 99%, accuracy alone is insufficient to judge its
success. The reasons are two-fold. Firstly, accuracy will always result in a higher score if
the dataset is imbalanced as results will be higher for the class with the majority of sam-
ples (Onan 2019a). For instance, if the total number of benign links is 90 and malicious
links are only 10, you will consistently achieve accuracy greater than 90%. That’s what an
imbalanced dataset means. Therefore, in this case, accuracy alone is inefficient. Secondly,
if multiple classes exist in the dataset, predicting accuracy for each category from the over-
all accuracy result becomes ambiguous. Hence, it is always suggested to include preci-
sion, recall, and F1-score to verify the model’s performance. (Note: FN refers to missed
vulnerabilities)

(ii) Precision
Precision is the overall number of positive values divided by the total number of antici-
pated positive values. Precision is calculated via the following Eq. (5):
TP
Precision = (5)
TP + FP

(iii) Detection rate

The detection rate is the percentage of true positive values that are accurately antici-
pated. This is used to detect the malicious samples. Finding a balance between accuracy
and detection rate is critical. It can be calculated through the following Eq. (6):
TP
Detection rate = (6)
TP + FN

(iv) Escape rate

The escape rate is the percentage of malicious attack vectors incorrectly identified as
benign payloads. The following Eq. (7) is used to determine the escape rate:
Escape rate = 1 − Detection rate (7)

(v) F1-Score
The F1-Score is used to achieve a balance of accuracy and recall levels. This statistic
accounts for both false positives and false negatives without altering the prediction model’s
accuracy score. The following Eq. (8) is used to determine the F1-score:
Precision × Recall
F1 Score = 2 × (8)
Precision + Recall

(vi) FP-Rate
The false positive rate is the percentage of incorrectly identified attack results. The fol-
lowing Eq. (9) is used to determine the FP- rate:

13
 Table 6  Comparative analysis of existing web-attack and XSS attack detection techniques as per the literature survey
12750

ML solution Reference Data set Platform XSS included Accuracy

13
Supervised learning (Kaur et al. 2018) Linear separable dataset Vulnerable website Mutilli- ✔ 95.4%, 95.1% recall, false posi-
collected from GitHub, dae tive rate 0.111
GBHackers, XSS-payloads
(Syarif & Gata 2018) KDD CUP 1999 (KDD Cup MATLAB ✘ 99.9% accuracy for DoS
1999 Data, n.d.)
(Dada 2017) KD99 MATLAB ✘ 98.55% classification accuracy
(Olalere et al. 2016) 23582 malicious and benign Weka data mining tool ✘ 96.95%
URLs collected manually
(Kokila et al. 2015) 2000 DARPA (2000 DARPA Python ✘ 95.11%
Intrusion Detection Sce-
nario Specific Datasets |
MIT Lincoln Laboratory,
n.d.)
(Kascheev & Olenchikova 40,000 malicious samples Python ✔ 93.70% accuracy for Decision
2020) ((XSS Archive), 200,000 Tree
benign samples (Kaggle)
(Machine Learning Based CSIC HTTP dataset 2010: Python, Weka ✔ 94.7% Precision, 94.5% Recall,
Intrusion Detection System 36000 normal samples and 93.4% F1 Score with J48 ML
for Web-Based Attacks, more than 25000 malicious Classifier using Weka tool
2020) sample
(Munonye & Péter 2021) Set I-manual dataset gener- DEKANT tool ✔ Gradient boosting classifier
ated using fuzzing tech- generated 99.4% accuracy,
nique, set II- PHP security 80.34% precision, 74.70%
vulnerability dataset recall, 71.71% F1-score
J. Kaur et al.
 Table 6  (continued)
ML solution Reference Data set Platform XSS included Accuracy
(Rathore et al. 2017) XSSed (XSSed | cross site Weka ✔ Accuracy 97.2%, false positive
scripting (XSS) attacks rate 0.87
information and archive,
n.d.) (benign), Alexa
(Alexa—Top Sites, n.d.)
(malicious), Elgg (introduc-
ing a powerful open source
social networking engine,
n.d.) (SNSs infected by
XSS worm)
(Khan et al. 2017) 409 malicious samples, 1515 Weka ✔ Accuracy 99.22% with J48,
benign samples 97.14% with KNN
Decision- Tree Based Cyber (Vuong et al. 2015) Manual R programming ✘ 86.9%
Security
(Moon et al. 2015) Manual Weka ✘ 84.7%
Neural Networks based (Gao et al. 2010) DOS MSU SCADA security ✘ 97.1%
Detection of cross‑site scripting (XSS) attacks using machine…

solution laboratory
(Vollmer & Manic 2009) snort.org, emerging threats Nemesis ✘ 60% (DoS, misc activity,
attempted-recon)
Unsupervised Machine (Laghrissi et al. 2021) KD99 Google Colab ✔ Model based on PCA gener-
Learning Solutions ated best accuracy 99.49%,
99.15% Recall
(R. Wang et al. 2015) DMOZ and XSSed dataset Python ✔ 90% Precision, 96.5% Recall
(Shahid et al. 2012) Not mentioned MATLAB ✘ 99% fault-classification, 100%
fault-detection
(Hoang 2020) The HTTP param dataset Python ✔ 98.56% Accuracy
(GitHub—Morzeux/Http-
ParamsDataset, n.d.)
Reinforcement Machine (Fang et al. 2019) XSSed (33,426 samples) Python3, Keras-rl, and Opa- ✔ 99.5% Accuracy, 97.9% Recall,
Learning Solutions nAIGym 98.7% F1-score
12751

13
 Table 6  (continued)
12752

ML solution Reference Data set Platform XSS included Accuracy

(Tariq et al. 2021) Deurgan dataset NetBeansIDE 8.0.2 (Java) ✔ Accuracy 99.89%

13
(Praise et al. 2020) Manually constructed Web ✔ 96.7% Efficiency
Deep Learning Solutions for (Dali Zhu et al. 2017) Automatic using flowdroid Android application ✘ 96.76%
defending against cyber-
space attacks
(Zhao et al. 2017) KDD CUP 1999 MATLAB ✘ 99.14%
(Vinayakumar et al. 2018) AndroidManifest GPU enables TensorFlow in ✘ 93.9% on dynamic analysis,
Ubuntu 14.04 97.5% on static analysis
(Feng et al. 2019) WAF logs and KDD Cup Python, NS-2 ✔ 98%
1999
(Abaimov & Bianchi 2019) SQL payload dataset, XSS Python ✔ 95% accuracy, 99% precision,
payload dataset (Github 92% recall
repository)
(Pavan Kumar et al. 2021) Kaggle benchmark dataset Python ✘ 94.8% accuracy with loss of
(phishing.csv) 0.2024%
(Yan et al. 2018) 1336 normal apps, 422 vul- Python with Keras on Ubuntu ✔ 97.55%
nerable apps 14.04 OS
(Pan et al. 2019) XSSed Weka ✔ F1-score 91% (No domain
knowledge)
(Mokbal et al. 2019) Manually collected using Python 3.6.7 + Keras ✔ Accuracy 99.32%, detection
web crawlers (148,157 2.2.4 + TensorFlow 1.12.0 probability98.35%, False
records) + XSSed + Open positive rate 0.3%, AUC-
Bug Bounty (38,569 mali- ROC score 99.02%
cious samples)
(Abdullah Alqarni et al. figshare.com: 1000 mali- Weka ✔ Accuracy: 99.96%
2022) cious samples and 100,000
benign samples
J. Kaur et al.
 Table 6  (continued)
ML solution Reference Data set Platform XSS included Accuracy
(Luo et al. 2021) HTTP CSIC 2010, CSIC DVWA, sqlmap, burpsuite, ✔ Accuracy: 0.9947, TPR:
2020 XSStrike 0.9929, FPR: 0.0033, Preci-
sion:0.9970
(Tian et al. 2020) HTTP CSIC 2010, FWAF, Python, tensorflow library ✔ Accuracy: 0.9941, TPR:
HttpParams 0.9891, DRN: 0.9955
(S. Kaur & Singh 2019) NSL-KDD and CICIDS 2017 Not mentioned ✔ Accuracy: 0.991
(Jiang et al. 2020) NSL-KDD Tensorflow (v1.2) ✘ Accuracy: 0.9894
Detection of cross‑site scripting (XSS) attacks using machine…
12753

13
12754 J. Kaur et al.

FP
FP Rate = (9)
FP + TN

6 Recent developments in XSS attack detection

This section discusses the recent developments in XSS attack detection techniques. Several
researchers have utilized Genetic algorithms, hybrid approaches, neural network models,
Deep Learning, and dynamic analysis-based approaches to optimize the accuracy of XSS
attack detection. Machine learning algorithms have been increasingly becoming popular
for detecting and preventing XSS assaults, although they are ineffective in innovative and
diverse XSS attacks (Tariq et al. 2021)(Zhou & Wang 2019)(S. Kaur & Singh 2019). Deep
Learning models have proved their efficiency by improving accuracy, reliability, and scala-
bility in cybersecurity applications (Dixit & Silakari 2021) (G. Zhang et al. 2022) (Onan &
Tocoglu 2021) (Onan 2019b). Therefore, security researchers are exploring novel methods
of web attack detection using Deep Learning (Luo et al. 2021). In Deep Learning, the word
‘deep’ refers to multiple hidden layers (up to 150) inside neural networks (Geetha & Thil-
agam 2020). The learning task is more difficult since data is handled at each layer, and thus
learning task becomes more complicated. The authors provided a study on Deep Learning
approaches for cybersecurity (Dixit & Silakari 2021). Deep Learning detection techniques
are separated into three categories: supervised, unsupervised, and reinforced.
The authors (Pavan Kumar et al. 2021), leverage Deep Learning techniques and Swarm
based Intelligence to detect website phishing attacks. The primary purpose of the study is
to use the Binary Bat Algorithm to distinguish fraudulent URLs from valid ones and to
enhance speed using the Adam optimizer. After pre-processing, 30 features are extracted
for the classification, such as URL length, IP address, user ID, iframe, port, redirect, mou-
seover, etc. SI-BBA model is evaluated using loss and accuracy parameters where loss is
0.2%, and accuracy is 94.8%. F. Jiang et al. (Jiang et al. 2020) presented a scheme that
detects DoS, U2R, Probe, and R2L attacks using multiple channels. This is the only paper
that has implemented multi-channel feature extraction and training for attack detection.
Their model also considers new variants of threat vectors. They used the NSL KDD train-
ing dataset to evaluate the performance of their model. The data is preprocessed first dur-
ing the training stage and second during the testing stage. The multi-channel training is
leveraged to generate classifiers based on LSTM-RNN neural networks. Their model is
evaluated on three parameters: Detection rate, False alarm rate, and accuracy. F. Feng et al.
in 2019 (Feng et al. 2019) implemented a plug-and-play ready-to-use device to detect DOS
attacks, XSS attacks, and SQL attacks using DNN, CNN, and LSTM models. If a mali-
cious payload is detected, an alarm is produced that could stop the spread of the attack to
other sites. LSTM performed best by generating 98% accuracy however took more time in
training. The authors experimented with KDD Cup 99 dataset and WAF (Web Application
Firewall) logs for attack detection.
S. Kaur et al. (S. Kaur & Singh 2019) leveraged a hybrid model based on a deep recur-
rent neural network that not only searches for malicious signatures but also performs
anomaly detection for web attacks and named their approach as D-Sign. Their hybrid
model consists of three sub-models: (1) Misuse Detection Engine (MDE), where the main
job is to differentiate the common attacks from the network traffic, (2) Anomaly Detection
Engine (ADE) identifies the change in the normal pattern that could occur due to presence
of anomaly and (3) signature generation for malicious vector through Signature Generation

13
Detection of cross‑site scripting (XSS) attacks using machine… 12755

Fig. 11  Overview of D-Sign

Engine (SGE). Figure 12 illustrates the execution flow of their proposed XSS attack detec-
tion model.
In Fig. 11, emulated honeypot server acts as an emulated vulnerable server, attracting
hackers to the server for executing attack vectors. The NSL-KDD and CICIDS 2017 data-
sets were utilized in the experiments. The honeypot server catches any suspicious behavior
on the network and sends it to MDE as quickly as possible as the sensor detects it. MDE
further analyses the captured network packets using the rule-based approach for any suspi-
cious activity and, if found generates an alarming signal. If no suspicious packet is found,
traffic is forwarded to ADE for further analysis. ADE leverages a Deep Learning-based
model. i.e., LSTM to classify normal and malicious packets in the incoming traffic.
In 2020, Z. Tian et al. (Tian et al. 2020) proposed another approach for web attack
detection that can be applied to multiple CNN models on cloud-based edge devices. Their
technique involves preparing data to be further fed to the data discrimination model, and
finally, action to be taken for the normal or malicious requests. URLs are converted into
vector form in the data preparation stage using the CBOW word2vec model. For the opti-
mization of the CNN model, the ReLU function, dropout layers, and pooling layers are
leveraged in CNN architecture. Some changes are also implemented in Residual Network,
while ResNet is used to detect malicious URLs. For data discrimination, the FastText
Deep Learning neural model has been used with some modifications. The model is trained
over three datasets: HTTP CSIC 2010, FWAF, and HttpParams dataset. W. Melicher et al.
(Melicher et al. 2021) focus on DOM XSS vulnerabilities in the browser, which predomi-
nantly executes dynamic JavaScript code using Deep Learning. The authors presented a
strategy of combining the taint tracking method with DNN to detect DOM-based XSS
attacks. First, taint tracking is applied to identify vulnerable functions, and then, a heuristic
is applied to those functions to confirm if they are vulnerable. They found that a relatively
small deep neural network with four layers can be an effective pre-filter for taint tracking.
However, on the negative side, the algorithm captures only 50% of confirmed vulnerabili-
ties when used alone. The observed vulnerabilities may not be applicable to browsers other
than chromium v57. The restrictions of the dynamic analysis utilized for labeling also limit
the data. Moreover, the paper does not consider attacks on machine learning algorithms,
such as evasion and poisoning attacks. Moreover, there is a massive class imbalance across
labels in training data. A study by Abdullah Alqarni A. et al. (Abdullah Alqarni et al. 2022)
proposed a modular neural network approach based on neural networks. Using a deep neu-
ral network eliminates the need for security domain-related knowledge and expertise, as
illustrated in Fig. 12.
First, the pre-processing is done to remove noisy data and to transform the decoded
code into the original format. The first stage of pre-processing consists of the execution

13
12756 J. Kaur et al.

Fig. 12  MNN-XSS Architecture

of sub-three steps, namely, data decoding, generalization, and performing tokenization

to reduce the data length. After pre-processing, Word2Vec and CBOW are deployed for
the numerical representation of code in simple plain text. Apart from word2vec and bag
of words, improved text analysis schemes could have been utilized that cover using Deep
Learning, unsupervised learning algorithms (Onan 2019c), ensembled approach, genetic
algorithms (Onan & KorukoGlu 2017), and swarm optimization (Onan 2018a) to get
improved results in feature extraction. Feature selection is performed through Pearson’s
Correlation method. Equation 8 depicts Pearson’s Correlation method.
�� �� ��
� x−x y−y
r= √∑ ∑ (10)
(x − x)2 (y − y)2

Finally, MNN is applied to break down the features into further sub-features. For this,
MNN implements the divide and conquer method. MNN (Shukla et al. 2010) can merge
different types of independent neural networks and produce the desired output in the form
of a single modular network, as illustrated in Fig. 13.
Since the sub-network acts as a sub-module to build a vast network, MNN is named a
modular neural network. The challenging issue of processing a significant amount of train-
ing data by ANN is simplified by MNN. If the MNN network generates an output value
greater than 0.5, then the code is marked as malicious, else benign. MATLAB is used for
the implementation of MNN, while Weka is used for feature selection. The study achieved
exceptional results where the accuracy score is 0.9996, Precision score is 0.9995, recall
score is 0.9991, and F1-score is 0.9992 for malicious code. However, the dataset used in
this study is unbalanced, i.e., 10000 benign samples, while malicious samples are only
1000.
L. Lei et al. (Lei et al. 2020) added an attention mechanism to the hidden layer of the
already developed LSTM model to improve the attack detection rates. LSTM (Yu et al.

13
Detection of cross‑site scripting (XSS) attacks using machine… 12757

Fig. 13  Modular Neural Network (MNN) Architecture

2019) is an advanced Recurrent Neural Network with four hidden layers capable of remem-
bering the state values for long distances via cell states. A neural network is a model of
computation biologically inspired by the human brain and consists of artificial neurons
with a set of edges between each neuron (Lipton et al. 2015). Figure 14 represents RNN
architecture where it is the input sequence, and ­ht is the output sequence.
LSTM includes a hidden layer with the complex structure of gates as a permanent solu-
tion to the vanishing gradient problem faced by RNNs. Although RNNs also have a single
hidden layer, their design is simple. As RNNs have short-term memory, it is easier to for-
get the old information while covering the long-distance and may miss important details.
Unlike RNNs, LSTM, an artificial RNN, can regulate the flow of information via their
internal mechanism known as gates, as depicted in Fig. 15.
The attention mechanism, which V. Mnih introduced in 2014 (Mnih et al. 2014), focused
on the part of a sentence while performing the translation. The attention model was first
used in machine translation and can be applied to other applications. Firstly, the data is pre-
processed and cleaned using a decoder. Further, tokenization is applied to reduce the data
dimension as a part of preprocessing. Secondly, word2vec is applied for the feature extrac-
tion process and maps them to feature vectors for vectorization. Finally, the LSTM model
is trained and tested. For evaluation, the LSTM-Attention model is compared with three
classifiers: ADTree, AdaBoost, and SVM (Support Vector Machine), where their model
achieved higher Precision, Recall, and F1 as 99.3%, 98.2%, and 98.5%, respectively. The
model also performed better than other Deep Learning models, such as Recurrent Neu-
ral Networks, Gated Recurrent Unit, and LSTM. The layout of the LSTM attention model
is illustrated in Fig. 16. Comparative analysis of existing literature has been covered in
Table 6.
A study by Yong Fang et al. (Fang et al. 2018) used a decoder and word2vec for data
cleaning and feature extraction, respectively. The data is then fed into an LSTM neural
network model. Finally, the suggested method’s performance is compared to that of the
ADTree and AdaBoost algorithms using a tenfold cross-validation methodology. Using
DeepXSS, they were able to get results with 99.5% accuracy. However, the limitation of
their approach is the high time complexity. Rubio Yan et al. (Yan et al. 2018) proposed
hybrid Deep Learning based on the neural model HDLN to detect XSS attacks in hybrid
mobile applications. First, n-gram features are generated from AST, followed by key

13
12758 J. Kaur et al.

Fig. 14  Recurrent Neural Net-

work Architecture

Fig. 15  LSTM Architecture

feature selection. Furthermore, a fivefold cross-validation technique is employed to vali-

date the classification model. While the best accuracy is achieved with eight hidden layers,
the precision of malicious data is 99.24%. Results improve as the number of hidden lay-
ers is increased. According to Yao Pan et al. (Pan et al. 2019), traditional intrusion detec-
tion systems differentiate between normal and malicious traffic depending on supervised
machine learning. This necessitates a massive amount of data required for training the ML
model, which is expensive, inaccurate, and time-consuming. Furthermore, intrusion detec-
tion technologies used to protect against online threats frequently need a thorough under-
standing of the security domain. Furthermore, the current unsupervised machine learning
method necessitates manual feature selection and has a high probability of false positives.
The researchers claim that Deep Learning can automatically make categorization and
feature selection. Their method uses RSMT to detect many sorts of assaults with distinct
features.
(Mokbal et al. 2019) have proposed MLPXX: a multilayer perceptron scheme. Their
proposed approach is integrated with the dynamic feature extractor and Back-Propagation
algorithm for XSS attack detection. Three models were developed that work together to
detect XSS attacks. The researchers constructed a large dataset by selecting the optimal
approach of random walk and random jumping method for the training and testing of the
samples. Apart from this, a novel technique for dynamic feature extraction and a Multi-
layer Perceptron for the detection has also been proposed in this paper. Their model has
achieved promising results with a detection probability of 98.35%. Data is collected manu-
ally using python script scraping framework-based web crawler; extracting the raw data
from websites and deleting duplicate samples. The authors extracted the raw data using
their web crawler from XSSed and Bug Bounty websites for malicious data. The authors
also suggested avoiding using word vectors for text sample representation as XSS code is
a more generalized JavaScript code, and hence, word vector often is time-consuming for
XSS. The BeautifulSoup library and HTML5lib parser are integrated with the model to
extract the HTML raw files; for tokenization and extraction of abstract syntax tree from the
JS code, Esprisma JavaScript parser is utilized. The model developed is composed of three
sub-model having different feature sets such as HTML features (tags, events, attributes) for
developing attack vectors, JavaScript features, and URL features. Their results improved
on adding the additional hidden layer where mean accuracy increased to 0.991195 and

13
Detection of cross‑site scripting (XSS) attacks using machine… 12759

Fig. 16  LSTM Attention model

for XSS attack detection

decreased variance value to 0.000759. The experiment was executed on Python framework
3.6.7 and Keras (2.2.4) with Tensorflow (1.12.0). For Non-XSS class, precision is 0.9924,
detection rate (True Positive Rate) is 0.9969, F-score is 0.9953. While for XSS class, preci-
sion is 0.9921, detection rate is 0.9835, F-score is 0.9877, and false-positive rate is 0.0031.
Abaimov & Bianchi (2019) proposed a supervised approach based on Deep Learning
for code injection attacks such as SQLi and XSS. Their model achieved an overall accuracy
of 95% with 99% precision and 92% recall value with a preprocessing module that can be
integrated with the detection model before the training phase or in the online query injec-
tion phase. The concept of the pre-processing module, based on the type of keyword or
symbol, removes noisy data from input queries so that ambiguous data value pairs from the
dataset can be removed. This directly impacts the neural network’s performance, thereby
increasing the detection accuracy from 75 to 95%. Furthermore, a local search optimiza-
tion algorithm is used in CODDLE to automate the process of parameter selection and of
effective neural network shape that otherwise requires considerable time for trial and error.
Figure 17 represents the CODDLE architecture in brief.
C. Luo et al. in (Luo et al. 2021) proposed a Deep Learning and ensemble learning
based web attack detection model for IoT (Internet of Things) networks. A combination of
three Deep Learning-based models i.e., MRN, CNN, and LSTM is utilized to analyze the
URL requests for anomalies. CBOW and TF-IDF text analysis algorithms are employed

13
12760 J. Kaur et al.

to transform URL requests into vectors. Finally, the results produced by Deep Learning
models are fed into the MLP ensemble classifier to improve the predictive model results.
The experiments were performed on DVWA vulnerable website and used HTTP CSIC
2010 dataset and another dataset provided by the security company. Their model achieved
99.47% of accuracy and 99.70% of precision. However, the performance score was reduced
when CSIC 202 dataset was applied. Moreover, the model does not consider attacks on ML
models.

7 Discussions, challenges and opportunities

Our study was solely based upon the clarification of the following research questions.

RQ1: What new advances have been made in the realm of XSS attack detection?
RQ2: Why are traditional techniques not effective at combating XSS attacks?
RQ3: How can XSS attacks be detected and prevented using Machine Learning
techniques?
RQ4: What are the most popular and effective cyber-attack mitigation techniques used
by active researchers?
RQ5: What are the challenges and opportunities for young researchers in the field of
cyber security?

During our survey, extensive analysis of recent advancements is observed that suggests
that for RQ1 and RQ4, several researchers are using Deep Learning in the web security
field, where hybrid neural techniques perform better than other techniques. Apart from
this, some researchers are focusing on merging metaheuristic algorithms with Deep Learn-
ing algorithms for cyber-attack detection as metaheuristic algorithms can optimize Deep
Learning-based solutions (Akay et al. 2021). With the ever-increasing usage of web access
and advancements in the digital world, hackers have started implementing new techniques
and threats to deceive the victim. The adoption of new networking terminologies such as
SDNs, OSNs, IoT, and Cloud computing induces the need for advanced intrusion-detec-
tion systems based on machine learning techniques. One of the most critical challenges
is to defend against cyberattacks by searching for vulnerabilities before the attacker can
exploit them (Conti et al. 2018). Furthermore, the heterogeneous nature of new attacks
and anti-forensics-evasion techniques used by hackers makes it challenging to detect and
mitigate cyberattacks. Most XSS attack detection models based on machine learning algo-
rithms always assume that the ML attack detection environment is benign. However, this
assumption is not always valid (Fang et al. 2019), and the adversary can take advantage of
this assumption by targeting the training data (poisoning attacks) or testing data (evasion
attacks) (Pitropakis et al. 2019) as illustrated in Fig. 18.
Moreover, according to X. Zhang et al. in (X. Zhang et al. 2020), many Deep Learn-
ing algorithms are vulnerable to adversarial attacks. There is limited research on defense
mechanisms against adversarial attacks on machine learning models. While doing this sur-
vey, we found only five articles (Chen et al. 2019; Fang et al. 2019; Zhang et al. 2020),(Q.
Wang et al.2022a 2022b), (Aldahdooh et al. 2022) focusing on adversarial attacks in web
security based upon machine learning algorithms in the context of neural networks. This
answers our RQ1 and RQ2. For RQ3, it has been observed that to prevent XSS attacks,

13
Detection of cross‑site scripting (XSS) attacks using machine… 12761

Fig. 17  CODDLE XSS attack

detection flow diagram

proper input sanitization must be done at the backend of a website, password strength must
be kept as strong, the web developer must enforce high encryption on the network traffic,
and regular updates in security patch must be performed by the web application owner.
Moreover, web users must be warned about online threats via emails, or text messages as
most often attack is executed by sending a malicious link to the user either on their email
accounts or via text messages.
Adversarial attacks have affected the performance of several ML-based applications,
such as intrusion detection systems, credit card fraud detection, biometric recognition sys-
tem, and recommendation systems. To defend against these attacks, Adversarial Learning
has also become popular among the cyber research communities over the past two dec-
ades [80]. In addition, innumerable dynamic web applications leverage the use of their-
party tools and components for user comfort such as AJAX, JavaScript, HTML /HTML5,
XHTML, RDF, OWL, etc. (Tekli 2021). However, these trends give birth to new hidden
vulnerabilities that require extensive research and analysis to get exploited. Develop-
ing detection algorithms with few false positives, false negatives, and 100% accuracy is
still complex and requires more research. Advances in communication technology have
made the attack detection task complicated because of the differences in syntax imple-
mentations of underlying platforms deployed by multiple third-party organizations (Abai-
mov & Bianchi 2019). For example, MySQL differs greatly in the implementation from
Oracle, MSSQL, Postgre SQL, etc. The majority of the XSS attack detection models are

13
12762 J. Kaur et al.

Fig. 18  Adversarial attack on Machine Learning-based system

weak at defending against zero-day attacks. These challenges open the door for innumer-
able research opportunities for young researchers. On the one hand, artificial intelligence,
machine learning, and Deep Learning have gained the attention of researchers in resolving
cybersecurity-related threats and risks. On the other hand, Deep Learning suffers from the
semantic gap, training of the neural network, complex result evaluation strategies, and high
cost of misclassification. For RQ5, we have observed that Deep Learning models aimed
especially at cyber security applications must not focus on a single attack vector and should
consider multiple attack vectors. Most ML and DL algorithms focus solely on TPR and
FPR for performance evaluation and ignore training speed, storage consumption, and sus-
pension of data poisoning parameters (Dixit & Silakari 2021). In addition, the usage of
genetic algorithms and fuzzy algorithms can be better utilized to improve accuracy. Finally,
metaheuristics algorithms can improve the efficiency of Deep Learning-based XSS attack
detection techniques.

8 Conclusion and future directions

Cyber security threats have consistently risen over the past few decades, and several
organizations are at potential risk. Every year cyber minds come up with sophisticated
cyber-attacks that are challenging for traditional attack detection tools. Organizations are
facing many challenges and losses due to XSS attacks on their web applications. There
is an urgent need for a survey that can cover the limitations of traditional attack detec-
tion techniques and the scope of improvements required in these techniques to tackle novel
cyber-attacks. Our survey provides the direction to researchers and security analysts with a
detailed description of sophisticated vulnerability analysis techniques as compared to tra-
ditional attack detection techniques. Machine learning techniques provide better accuracy
in detecting XSS attacks as compared to state-of-the-art approaches. In this survey, super-
vised, unsupervised and semi-supervised techniques are discussed along with advantages
and limitations. We have highlighted five Research Questions for which we have tried to
find answers from recent research works. After doing this survey, all RQs are answered
which will be beneficial for researchers.
After doing this survey, it can be concluded that the ML techniques have proved to scan
subtler bugs and reduce false alarms. However, there are a few challenges that are observed
while doing a survey of existing techniques. The applications must be tested for XSS attack
vulnerability at regular intervals. Leaving web applications for years without any security
assessment is one of the biggest mistakes that many developers and application owners do.

13
Detection of cross‑site scripting (XSS) attacks using machine… 12763

They must have sound knowledge of security risks related to XSS. Moreover, most existing
approaches do not consider new XSS threat vectors and payloads that must be identified
through an appropriate machine-learning model. It should be noted that XSS attack detec-
tion models can also be compromised by hackers. Therefore, it is significant to retrain the
attack detection model further through effective machine learning models and training. To
obtain high accuracy, a feature subset selection mechanism must be implemented to extract
high-profile features from overall generated features that will significantly impact the detec-
tion model’s performance. To eliminate XSS attacks from their roots, Deep Learning tech-
niques act as an alternative to inefficient traditional techniques of machine learning. Social
networking sites are more affected by XSS attacks as the attack is easier to pass on to other
users through a single infected victim via social connections. Deep Learning models such
as LSTM, AutoEncoder, and MLP are applied by researchers to detect XSS attacks more
effectively. In our survey, these techniques are covered for a better understanding of the
reader. It is concluded that more sophisticated Deep Learning models should be proposed
to achieve effective results.

References
2000 DARPA Intrusion Detection Scenario Specific Datasets | MIT Lincoln Laboratory. (n.d.). https://​www.​
ll.​mit.​edu/r-​d/​datas​ets/​2000-​darpa-​intru​sion-​detec​tion-​scena​rio-​speci​fic-​datas​ets. Accessed 2 Nov
2021
Abaimov S, Bianchi G (2019) CODDLE: Code-injection detection with deep learning. IEEE Access
7:128617–128627. https://​doi.​org/​10.​1109/​ACCESS.​2019.​29398​70
Abdullah Alqarni A, Alsharif N, Ahmad Khan N, Georgieva L, Pardade E, Alzahrani YM (2022) MNN-
XSS: modular neural network based approach for XSS attack detection. Comput Mater Continua
70(2):4075–4085. https://​doi.​org/​10.​32604/​CMC.​2022.​020389
Akay B, Karaboga D, Akay R (2021) A comprehensive survey on optimizing deep learning models by
metaheuristics. Artif Intell Rev 55(2):829–894. https://​doi.​org/​10.​1007/​S10462-​021-​09992-0
Akrout R, Alata E, Kaaniche M, Nicomette V (2014) An automated black box approach for web vulnerabil-
ity identification and attack scenario generation. J Brazilian Comput Soc 20(1):1–16. https://​doi.​org/​
10.​1186/​1678-​4804-​20-4
Alam F, Pachauri S (2017) Comparative study of J48, naive bayes and one-R classification technique for
credit card fraud detection using WEKA. Adv Comput Sci Technol 10(6):1731–1743
Aldahdooh A, Hamidouche W, Fezza SA, Déforges O (2022) Adversarial example detection for DNN
models: a review and experimental comparison. Artif Intell Rev 2022:1–60. https://​doi.​org/​10.​1007/​
S10462-​021-​10125-W
Alexa - Top sites. (n.d.). https://​www.​alexa.​com/​topsi​tes. Accessed 2 Nov 2021
Banerjee, R., Baksi, A., Singh, N., & Bishnu, S. K. (2020, October 2). Detection of XSS in web applica-
tions using Machine Learning Classifiers. 2020 4th International Conference on Electronics, Materi-
als Engineering & Nano-Technology (IEMENTech). https://​doi.​org/​10.​1109/​IEMEN​Tech5​1367.​2020.​
92700​52
Brewer R (2016) Ransomware attacks: detection, prevention and cure. Netw Secur 2016(9):5–9. https://​doi.​
org/​10.​1016/​S1353-​4858(16)​30086-1
Chauhan S, Vig L, Filippo De Grazia M, Corbetta M, Ahmad S, Zorzi M (2019) A Comparison of shallow
and deep learning methods for predicting cognitive performance of stroke patients from mri lesion
images. Front Neuroinform. https://​doi.​org/​10.​3389/​fninf.​2019.​00053
Chen T, Liu J, Xiang Y, Niu W, Tong E, Han Z (2019) Adversarial attack and defense in reinforcement
learning-from AI security view. Cybersecurity 2:1. https://​doi.​org/​10.​1186/​s42400-​019-​0027-x
Cimpanu Catalin. (2018). British Airways breach caused by the same group that hit Ticketmaster | ZDNet.
ZDNET, A RED VENTURES COMPANY. https://​www.​zdnet.​com/​artic​le/​briti​sh-​airwa​ys-​breach-​
caused-​by-​the-​same-​group-​that-​hit-​ticke​tmast​er/
CISCO. (2021). Defending Against Critical Threats. https://​www.​cisco.​com/c/​en/​us/​produ​cts/​secur​ity/​defen​
ding-​again​st-​criti​cal-​threa​ts.​html?​CCID=​cc000​160&​DTID=​odicd​c0000​16&​OID=​rptsc​024689

13
12764 J. Kaur et al.

Conti, M., Dargahi, T., & Dehghantanha, A. (2018). Cyber threat intelligence: Challenges and opportuni-
ties. In Advances in Information Security (Vol. 70, pp. 1–6). Springer New York LLC. https://​doi.​org/​
10.​1007/​978-3-​319-​73951-9_1
Dada, E. G. (2017). A Hybridized SVM-kNN-pdAPSO Approach to Intrusion Detection System. In Univer-
sity of Maiduguri Faculty of Engineering Seminar Series (Vol. 8). https://​www.​resea​rchga​te.​net/​publi​
cation/​31614​5216_A_​Hybri​dized_​SVM-​kNN-​pdAPSO_​Appro​ach_​to_​Intru​sion_​Detec​tion_​System
Zhu Dali, Jin Hao, Ying Yang Wu, D., & Weiyi Chen. (2017) DeepFlow: deep learning-based malware
detection by mining android application for abnormal usage of sensitive data. 2017 IEEE Symposium
Comput Commun (ISCC). https://​doi.​org/​10.​1109/​ISCC.​2017.​80245​68
Dixit P, Silakari S (2021) Deep learning algorithms for cybersecurity applications: a technological and sta-
tus review. Comput Sci Rev 39:100317. https://​doi.​org/​10.​1016/J.​COSREV.​2020.​100317
Dora JR, Nemoga K (2021) Ontology for cross-site-scripting (XSS) attack in cybersecurity. J Cybersecur
Privacy 1(2):319–339. https://​doi.​org/​10.​3390/​jcp10​20018
EOIN KEARY. (n.d.). 2019 VULNERABILITY STATISTICS REPORT. https://​www.​edges​can.​com/​wp-​conte​
nt/​uploa​ds/​2019/​02/​edges​can-​Vulne​rabil​ity-​Stats-​Report-​2019.​pdf. Accessed 2 Sep 2021
Fang Y, Huang C, Xu Y, Li Y (2019) RLXSS: Optimizing XSS detection model to defend against adversar-
ial attacks based on reinforcement learning. Future Internet 11:8. https://​doi.​org/​10.​3390/​FI110​80177
Fang, Y., Li, Y., Liu, L., & Huang, C. (2018). DeepXSS: Cross site scripting detection based on deep learn-
ing. ACM International Conference Proceeding Series. https://​doi.​org/​10.​1145/​31944​52.​31944​69
Feng F, Liu X, Yong B, Zhou R, Zhou Q (2019) Anomaly detection in ad-hoc networks based on deep
learning model: a plug and play device. Ad Hoc Netw 84:82–89. https://​doi.​org/​10.​1016/J.​ADHOC.​
2018.​09.​014
Furnell S, Emm D (2017) The ABC of ransomware protection. Comput Fraud Secur 2017(10):5–11. https://​
doi.​org/​10.​1016/​S1361-​3723(17)​30089-1
Gao W, Morris T, Reaves B, Richey D (2010) On SCADA control system command and response injection
and intrusion detection. General Mem Meet ECrime Res Summit ECrime. https://​doi.​org/​10.​1109/​
ECRIME.​2010.​57066​99
Guo Y, Pan Y, Zhang Z, Li L, Jamshed MA, Moon Y, Kim D, Han D, Park K, Jamshed M A, Berger DS,
Sitaraman RK, Harchol-Balter M, Pfaff B, Pettit J, Koponen T, Jackson E, Zhou A, Rajahalme J, …
Security I. T (2017) Same-origin policy: Evaluation in modern browsers. In Proceedings of the Same-
Origin Policy: Evaluation in Modern Browsers. Nsdi, 40(4): 97–112
Geetha R, Thilagam T (2020) A review on the effectiveness of machine learning and deep learning algo-
rithms for cyber security. Arch Comput Methods Eng 28(4):2861–2879. https://​doi.​org/​10.​1007/​
S11831-​020-​09478-2
GitHub - Morzeux/HttpParamsDataset. (n.d.). https://​github.​com/​Morze​ux/​HttpP​arams​Datas​et. Accessed 2
Nov 2021
Gkioulos V, Chowdhury N (2021) Cyber security training for critical infrastructure protection: a literature
review. Comput Sci Rev 40:100361. https://​doi.​org/​10.​1016/J.​COSREV.​2021.​100361
Gupta B.B., & Chaudhary Pooja. (2020). Cross-Site Scripting Attacks: Classification, Attack, and Counter-
measures. (First). CRC Press. https://​www.​google.​co.​in/​books/​editi​on/​Cross_​Site_​Scrip​ting_​Attac​ks/​
697SD​wAAQB​AJ?​hl=​en&​gbpv=​0&​kptab=​overv​iew
Gupta S, Gupta BB (2015) Cross-Site Scripting (XSS) attacks and defense mechanisms: classifica-
tion and state-of-the-art. Int J Syst Assurance Eng Manag 8(1):512–530. https://​doi.​org/​10.​1007/​
S13198-​015-​0376-0
Hassan MdM, Nipa SS, Akter M, Haque R, Deepa FN, Rahman MM, Siddiqui Md, Sharif MdH (2018)
Broken authentication and session management vulnerability: a case study of web application. Int
J Simul. https://​doi.​org/​10.​5013/​ijssst.​a.​19.​02.​06
Heiderich, M., Schwenk, J., Frosch, T., Magazinius, J., & Yang, E. Z. (2013). mXSS attacks: Attack-
ing well-secured web-applications by using innerHTML mutations. Proceedings of the ACM Con-
ference on Computer and Communications Security, 777–788. https://​doi.​org/​10.​1145/​25088​59.​
25167​23
Hickling J (2021) What is DOM XSS and why should you care? Comput Fraud Secur 2021(4):6–10.
https://​doi.​org/​10.​1016/​S1361-​3723(21)​00040-3
Hoang XD (2020) Detecting common web attacks based on machine learning using web log. Lecture
Notes Networks and Syst 178:311–318. https://​doi.​org/​10.​1007/​978-3-​030-​64719-3_​35
Introducing a powerful open source social networking engine. (n.d.). https://​elgg.​org/. Accessed 2 Nov
2021
Jagajeevan Rao L, Nazeer Basha SK, Rama Krishna V (2021) Prevention and analysing on cross site
scripting. Adv Intell Syst Comput 1171:731–739. https://​doi.​org/​10.​1007/​978-​981-​15-​5400-1_​69

13
Detection of cross‑site scripting (XSS) attacks using machine… 12765

Jiang F, Fu Y, Gupta BB, Liang Y, Rho S, Lou F, Meng F, Tian Z (2020) Deep learning based multi-
channel intelligent attack detection for data security. IEEE Trans Sustain Comput 5(2):204–212.
https://​doi.​org/​10.​1109/​TSUSC.​2018.​27932​84
Jian-hua Li. (2021). Cyber Security Meets Machine Learning. In Cyber Security Meets Machine Learn-
ing. Springer Singapore. https://​doi.​org/​10.​1007/​978-​981-​33-​6726-5
Kascheev, S., & Olenchikova, T. (2020). Detecting Cross-Site Scripting (XSS) Using Machine Learn-
ing Methods. 2020 Global Smart Industry Conference (GloSIC). https://​doi.​org/​10.​1109/​GloSI​
C50886.​2020.​92678​66
Katsikeas S, Johnson P, Ekstedt M, Lagerström R (2021) Research communities in cyber security: a
comprehensive literature review. Comput Sci Rev. https://​doi.​org/​10.​1016/j.​cosrev.​2021.​100431
Kaur, G., Malik, Y., Samuel, H., & Jaafar, F. (2018). Detecting blind cross-site scripting attacks using
machine learning. ACM International Conference Proceeding Series, 22–25. https://​doi.​org/​10.​
1145/​32970​67.​32970​96
Kaur, J., & Garg, U. (2022). State-of-the-Art Survey on Web Vulnerabilities, Threat Vectors, and Coun-
termeasures. In: Dr. R. Aggarwal, Dr. J. He, Dr. E. Shubhakar Pilli, & Dr. S. Kumar (Eds) Cyber
Security in Intelligent Computing and Communications. Springer, Singapore. (pp. 3–17).
Kaur S, Singh M (2019) Hybrid intrusion detection and signature generation using deep recurrent neural
networks. Neural Comput Appl 32(12):7859–7877. https://​doi.​org/​10.​1007/​S00521-​019-​04187-9
KDD Cup 1999 Data. (n.d.). http://​kdd.​ics.​uci.​edu/​datab​ases/​kddcu​p99/​kddcu​p99.​html. Accessed 2 Nov
2021
Keary Eoin. (n.d.). 2021 VULNERABILITY STATISTICS REPORT EDGESCAN. In EdgeScan Report.
https://​info.​edges​can.​com/​hubfs/​Edges​can20​21Sta​tsRep​ort.​pdf?​hsCta​Track​ing=​9601b​027-​23d3-​
443f-​b438-​fcb67​1cfda​06%​7Cb22​2011c-​0b6d-​440b-​aed8-​64d37​dec66​e2. Accessed 2 Sep 2021
Khan N, Abdullah J, Khan AS (2017) Defending malicious script attacks using machine learning classi-
fiers. Wireless Commun Mobile Comput. https://​doi.​org/​10.​1155/​2017/​53604​72
Kharraz A, Robertson W, Balzarotti D, Bilge L, Kirda E (2015) Cutting the Gordian knot: a look under
the hood of ransomware attacks. Lecture Notes Comput Sci 9148:3–24. https://​doi.​org/​10.​1007/​
978-3-​319-​20550-2_1
Kokila, R. T., Thamarai Selvi, S., & Govindarajan, K. (2015). DDoS detection and analysis in SDN-
based environment using support vector machine classifier. 6th International Conference on
Advanced Computing, ICoAC 2014, 205–210. https://​doi.​org/​10.​1109/​ICOAC.​2014.​72297​11
Kumar R, Goyal R (2019) On cloud security requirements, threats, vulnerabilities and countermeasures:
a survey. Comput Sci Rev 33:1–48. https://​doi.​org/​10.​1016/J.​COSREV.​2019.​05.​002
Laghrissi FE, Douzi S, Douzi K, Hssina B (2021) Intrusion detection systems using long short-term
memory (LSTM). J Big Data 8:1. https://​doi.​org/​10.​1186/​s40537-​021-​00448-4
Lei, L., Chen, M., He, C., & Li, D. (2020). XSS Detection Technology Based on LSTM-Attention. 2020
5th International Conference on Control, Robotics and Cybernetics, CRC 2020, 175–180. https://​
doi.​org/​10.​1109/​CRC51​253.​2020.​92534​84
Leyden, J. (2020). XSS vulnerability in ‘Login with Facebook’ button earns $20,000 bug bounty.
PortSwigger. https://​ports​wigger.​net/​daily-​swig/​xss-​vulne​rabil​ity-​in-​login-​with-​faceb​
ook-​button-​earns-​20-​000-​bug-​bounty
Li J, Hua (2018) Cyber security meets artificial intelligence: a survey. Front Inform Technol Electron
Eng 19(12):1462–1474. https://​doi.​org/​10.​1631/​FITEE.​18005​73
Lipton, Z. C., Berkowitz, J., & Elkan, C. (2015). A Critical Review of Recurrent Neural Networks for
Sequence Learning. http://​arxiv.​org/​abs/​1506.​00019
Luo C, Tan Z, Min G, Gan J, Shi W, Tian Z (2021) A novel web attack detection system for internet of
things via ensemble classification. IEEE Trans Industr Inf 17(8):5810–5818. https://​doi.​org/​10.​1109/​
TII.​2020.​30387​61
Matt, F. (2021). Application Model & Same-Origin Policy. In Lecture Notes on Web Security: Application
Model & Same-Origin Policy.
Melicher, W., Fung, C., Bauer, L., & Jia, L. (2021). Towards a lightweight, hybrid approach for detect-
ing DOM XSS vulnerabilities with machine learning. The Web Conference 2021—Proceedings of
the World Wide Web Conference, WWW 2021, 2684–2695. https://​doi.​org/​10.​1145/​34423​81.​34500​62
C.C. Michael, & Steven Lavenhar. (2013). Source Code Analysis Tools - Overview | CISA. Cybersecurity &
Infrastructure Security Agency. https://​us-​cert.​cisa.​gov/​bsi/​artic​les/​tools/​source-​code-​analy​sis/​source-​
code-​analy​sis-​tools---​overv​iew
Microsoft Teams security vulnerability left users open to XSS via flawed stickers feature | The Daily Swig.
(n.d.). https://​ports​wigger.​net/​daily-​swig/​micro​soft-​teams-​secur​ity-​vulne​rabil​ity-​left-​users-​open-​to-​
xss-​via-​flawed-​stick​ers-​featu​re. Accessed 10 Oct 2022

13
12766 J. Kaur et al.

Mnih, V., Heess, N., Graves, A., & Kavukcuoglu, K. (2014). Recurrent Models of Visual Attention.
Advances in Neural Information Processing Systems, 3(January), 2204–2212. https://​arxiv.​org/​abs/​
1406.​6247v1
Mohammad G, Reza S (2017) Software vulnerability analysis and discovery using machine-learning and
data-mining techniques. ACM Comput Surv (CSUR) 50:4. https://​doi.​org/​10.​1145/​30925​66
Mokbal FMM, Dan W, Imran A, Jiuchuan L, Akhtar F, Xiaoxi W (2019) MLPXSS: an Integrated XSS-
based attack detection scheme in web applications using multilayer perceptron technique. IEEE
Access 7:100567–100580. https://​doi.​org/​10.​1109/​ACCESS.​2019.​29274​17
Moon D, Im H, Kim I, Park JH (2015) DTB-IDS: an intrusion detection system based on decision tree
using behavior analysis for preventing APT attacks. J Supercomput 73(7):2881–2895. https://​doi.​org/​
10.​1007/​S11227-​015-​1604-8
Munonye K, Péter M (2021) Machine learning approach to vulnerability detection in OAuth 2.0 authentica-
tion and authorization flow. Int J Inf Secur 2021:1–15. https://​doi.​org/​10.​1007/​S10207-​021-​00551-W
Nidecki Tomasz Andrzej. (2019). Mutation XSS in Google Search. THE ACUNETIX BLOG. https://​www.​
acune​tix.​com/​blog/​web-​secur​ity-​zone/​mutat​ion-​xss-​in-​google-​search/
G. Nick. (2021). The Most Telling Cyber Security Statistics in 2021 [Infographic]. Cyber Security Stats—
Infographic. https://​techj​ury.​net/​blog/​cyber-​secur​ity-​stati​stics/
Novinson Michael. (2021). The 10 Biggest Data Breaches Of 2021 (So Far). CRN News. https://​www.​crn.​
com/​slide-​shows/​secur​ity/​the-​10-​bigge​st-​data-​breac​hes-​of-​2021-​so-​far-/2
Olalere, M., Abdullah, M. T., Mahmod, R., & Abdullah, A. (2016). Identification and Evaluation of Dis-
criminative Lexical Features of Malware URL for Real-Time Classification. 2016 International Con-
ference on Computer and Communication Engineering (ICCCE). https://​doi.​org/​10.​1109/​ICCCE.​
2016.​31
Onan A (2018) Biomedical text categorization based on ensemble pruning and optimized topic modelling.
Comput Math Methods Med. https://​doi.​org/​10.​1155/​2018/​24974​71
Onan A (2018b) An ensemble scheme based on language function analysis and feature engineering for text
genre classification. J Inf Sci 44(1):28–47. https://​doi.​org/​10.​1177/​01655​51516​677911
Onan A (2019) Consensus clustering-based undersampling approach to imbalanced learning. Sci Program.
https://​doi.​org/​10.​1155/​2019/​59010​87
Onan A (2019b) Topic-enriched word embeddings for sarcasm identification. Adv Intell Syst Comput
984:293–304. https://​doi.​org/​10.​1007/​978-3-​030-​19807-7_​29/​COVER
Onan A (2019c) Two-stage topic extraction model for bibliometric data analysis based on word embeddings
and clustering. IEEE Access 7:145614–145633. https://​doi.​org/​10.​1109/​ACCESS.​2019.​29459​11
Onan A (2020) Mining opinions from instructor evaluation reviews: a deep learning approach. Comput
Appl Eng Educ 28(1):117–138. https://​doi.​org/​10.​1002/​CAE.​22179
Onan A (2021a) Sentiment analysis on massive open online course evaluations: a text mining and deep
learning approach. Comput Appl Eng Educ 29(3):572–589. https://​doi.​org/​10.​1002/​CAE.​22253
Onan A (2021) Sentiment analysis on product reviews based on weighted word embeddings and deep neural
networks. Concurrency Comput 33(23):e5909. https://​doi.​org/​10.​1002/​CPE.​5909
Onan A (2022) Bidirectional convolutional recurrent neural network architecture with group-wise enhance-
ment mechanism for text sentiment classification. J King Saud University—Comput Information Sci
34(5):2098–2117. https://​doi.​org/​10.​1016/J.​JKSUCI.​2022.​02.​025
Onan A, KorukoGlu S (2017) A feature selection model based on genetic rank aggregation for text senti-
ment classification. J Inf Sci 43(1):25–38. https://​doi.​org/​10.​1177/​01655​51515​613226
Onan A, Korukoǧlu S, Bulut H (2016) Ensemble of keyword extraction methods and classifiers in text clas-
sification. Expert Syst Appl 57:232–247. https://​doi.​org/​10.​1016/J.​ESWA.​2016.​03.​045
Onan A, Korukoğlu S, Bulut H (2017) A hybrid ensemble pruning approach based on consensus clus-
tering and multi-objective evolutionary algorithm for sentiment classification. Inform Processing
Manag 53(4):814–833. https://​doi.​org/​10.​1016/J.​IPM.​2017.​02.​008
Onan A, Tocoglu MA (2021) A term weighted neural language model and stacked bidirectional LSTM
Based framework for sarcasm identification. IEEE Access 9:7701–7722. https://​doi.​org/​10.​1109/​
ACCESS.​2021.​30497​34
owasp. (2017). OWASP Top Ten. OWASP. https://​owasp.​org/
Pan Y, Sun F, Teng Z, White J, Schmidt DC, Staples J, Krause L (2019) Detecting web attacks
with end-to-end deep learning. J Internet Serv Appl 10(1):1–22. https://​doi.​org/​10.​1186/​
S13174-​019-​0115-X
Pavan Kumar P, Jaya T, Rajendran V (2021) SI-BBA—a novel phishing website detection based on swarm
intelligence with deep learning. Mater Today. https://​doi.​org/​10.​1016/J.​MATPR.​2021.​07.​178

13
Detection of cross‑site scripting (XSS) attacks using machine… 12767

Pitropakis N, Panaousis E, Giannetsos T, Anastasiadis E, Loukas G (2019) A taxonomy and survey of

attacks against machine learning. Comput Sci Rev 34:100199. https://​doi.​org/​10.​1016/J.​COSREV.​
2019.​100199
Praise JJ, Raj RJS, Benifa JVB (2020) Development of reinforcement learning and pattern matching
(RLPM) based firewall for secured cloud infrastructure. Wireless Personal Commun 115(2):993–
1018. https://​doi.​org/​10.​1007/​S11277-​020-​07608-4
Rathore S, Sharma PK, Park JH (2017) XSSClassifier: an efficient XSS attack detection approach based
on machine learning classifier on SNSs. J Inform Processing Syst 13(4):1014–1028. https://​doi.​
org/​10.​3745/​JIPS.​03.​0079
Rodríguez GE, Torres JG, Flores P, Benavides DE (2020) Cross-site scripting (XSS) attacks and mitiga-
tion: a survey. Computer Networks 166:106960. https://​doi.​org/​10.​1016/J.​COMNET.​2019.​106960
Sarmah U, Bhattacharyya DK, Kalita JK (2018) A survey of detection methods for XSS attacks. J Netw
Comput Appl 118:113–143. https://​doi.​org/​10.​1016/J.​JNCA.​2018.​06.​004
Screencastify Chrome extension flaws allow webcam hijacks. (n.d.). https://​www.​bleep​ingco​mputer.​com/​
news/​secur​ity/​scree​ncast​ify-​chrome-​exten​sion-​flaws-​allow-​webcam-​hijac​ks/. Accessed 10 Oct 2022
Shabut, A. M., Lwin, K. T., & Hossain, M. A. (2016). Cyber attacks, countermeasures, and protection
schemes—a state of the art survey. 2016 10th International Conference on Software, Knowledge,
Information Management & Applications (SKIMA). https://​doi.​org/​10.​1109/​SKIMA.​2016.​79161​94
Shahid N, Aleem SA, Naqvi IH, Zaffar N (2012) Support vector machine based fault detection & clas-
sification in smart grids. 2012 IEEE Globecom Workshops. GC Wkshps 2012:1526–1531. https://​
doi.​org/​10.​1109/​GLOCO​MW.​2012.​64778​12
Machine Learning based Intrusion Detection System for Web-Based Attacks, Proceedings - 2020 IEEE
6th Intl Conference on Big Data Security on Cloud, BigDataSecurity 2020, 2020 IEEE Intl Con-
ference on High Performance and Smart Computing, HPSC 2020 and 2020 IEEE Intl Conference
on Intelligent Data and Security, IDS 2020 227 (2020).
Shukla A, Tiwari R, Kala R (2010) Modular neural networks. Stud Comput Intell 307:307–335. https://​
doi.​org/​10.​1007/​978-3-​642-​14344-1_​14
Snehi M, Bhandari A (2021) Vulnerability retrospection of security solutions for software-defined
cyber-physical system against DDoS and IoT-DDoS attacks. Comput Sci Rev. https://​doi.​org/​10.​
1016/j.​cosrev.​2021.​100371
Syarif, A. R., & Gata, W. (2018). Intrusion detection system using hybrid binary PSO and K-nearest
neighborhood algorithm. Proceedings of the 11th International Conference on Information and
Communication Technology and System, ICTS 2017, 2018-January, 181–186. https://​doi.​org/​10.​
1109/​ICTS.​2017.​82656​67
Tariq I, Sindhu MA, Abbasi RA, Khattak AS, Maqbool O, Siddiqui GF (2021) Resolving cross-
site scripting attacks through genetic algorithm and reinforcement learning. Exp Syst Appl
168:114386. https://​doi.​org/​10.​1016/J.​ESWA.​2020.​114386
Tekli G (2021) A survey on semi-structured web data manipulations by non-expert users. Comput Sci
Rev 40:100367. https://​doi.​org/​10.​1016/J.​COSREV.​2021.​100367
Thakkar A, Lohiya R (2021) A survey on intrusion detection system: feature selection, model, perfor-
mance measures, application perspective, challenges, and future research directions. Artificial
Intell Rev 55(1):453–563. https://​doi.​org/​10.​1007/​S10462-​021-​10037-9
Tian Z, Luo C, Qiu J, Du X, Guizani M (2020) A distributed deep learning system for web attack detection
on edge devices. IEEE Trans Industr Inf 16(3):1963–1971. https://​doi.​org/​10.​1109/​TII.​2019.​29387​78
Tran NK, Sheng QZ, Babar MA, Yao L (2017) Searching the web of things: state of the art, challenges,
and solutions. ACM Comput Surv 50(4):1–34. https://​doi.​org/​10.​1145/​30926​95
Vinayakumar R, Soman KP, Poornachandran P, Sachin Kumar S (2018) Detecting android malware using
long short-term memory (LSTM). J Intell Fuzzy Syst 34:3. https://​doi.​org/​10.​3233/​JIFS-​169424
Vollmer, T., & Manic, M. (2009). Computationally efficient neural network intrusion security awareness.
Proceedings - ISRCS 2009—2nd International Symposium on Resilient Control Systems, 25–30.
https://​doi.​org/​10.​1109/​ISRCS.​2009.​52513​57
von Solms R, van Niekerk J (2013) From information security to cyber security. Comput Secur 38:97–102.
https://​doi.​org/​10.​1016/j.​cose.​2013.​04.​004
Vuong, T. P., Loukas, G., Gan, D., & Bezemskij, A. (2015). Decision tree-based detection of denial of service
and command injection attacks on robotic vehicles. 2015 IEEE International Workshop on Information
Forensics and Security, WIFS 2015 - Proceedings. https://​doi.​org/​10.​1109/​WIFS.​2015.​73685​59
Wang Q, Yang H, Wu G, Choo KKR, Zhang Z, Miao G, Ren Y (2022) Black-box adversarial attacks on XSS
attack detection model. Comput Secur 113:102554. https://​doi.​org/​10.​1016/J.​COSE.​2021.​102554

13
12768 J. Kaur et al.

Wang, R., Jia, X., Li, Q., & Zhang, D. (2015). Improved N-gram approach for cross-site scripting detection
in Online Social Network. Proceedings of the 2015 Science and Information Conference, SAI 2015,
1206–1212. https://​doi.​org/​10.​1109/​SAI.​2015.​72372​98
Wang, R., Jia, X., Li, Q., & Zhang, S. (2014). Machine Learning Based Cross-Site Scripting Detection in
Online Social Network. 2014 IEEE Intl Conf on High Performance Computing and Communications,
2014 IEEE 6th Intl Symp on Cyberspace Safety and Security, 2014 IEEE 11th Intl Conf on Embedded
Software and Syst (HPCC,CSS,ICESS), 823–826. https://​doi.​org/​10.​1109/​HPCC.​2014.​137
Wang Z, Fok KW, Thing VLL (2022) Machine learning for encrypted malicious traffic detection:
Approaches, datasets and comparative study. Comput Secur 113:102542. https://​doi.​org/​10.​1016/J.​
COSE.​2021.​102542
Wikipedia. (2021). Session ID. Online. https://​en.​wikip​edia.​org/​wiki/​Sessi​on_​ID
WordPress Stored XSS Vulnerability—Update Now. (n.d.). https://​www.​searc​hengi​nejou​rnal.​com/​wordp​
ress-​core-​vulne​rabil​ity-​2022/​44179​5/#​close. Accessed 10 Oct 2022
XSS in Gmail’s AMP For Email earns researcher $5,000 | The Daily Swig. (n.d.). https://​ports​wigger.​net/​
daily-​swig/​xss-​in-​gmails-​amp-​for-​email-​earns-​resea​rcher-5-​000. Accessed 8 Oct 2022
XSS vulnerabilities in Google Cloud, Google Play could lead to account hijacks | The Daily Swig. (n.d.).
https://​ports​wigger.​net/​daily-​swig/​xss-​vulne​rabil​ities-​in-​google-​cloud-​google-​play-​could-​lead-​to-​
accou​nt-​hijac​ks. Accessed 8 Oct 2022
XSSed | Cross Site Scripting (XSS) attacks information and archive. (n.d.). http://​xssed.​com/. Accessed 2
Nov 2021
Yan R, Xiao X, Hu G, Peng S, Jiang Y (2018) New deep learning method to detect code injection attacks on
hybrid applications. J Syst Softw 137:67–77. https://​doi.​org/​10.​1016/J.​JSS.​2017.​11.​001
Yu Y, Si X, Hu C, Zhang J (2019) A review of recurrent neural networks: Lstm cells and network architec-
tures. Neural Comput 31(7):1235–1270. https://​doi.​org/​10.​1162/​NECO_A_​01199
Zhang G, Liu B, Zhu T, Zhou A, Zhou W (2022) Visual privacy attacks and defenses in deep learning: a
survey. Artif Intell Rev 2021:1–55. https://​doi.​org/​10.​1007/​S10462-​021-​10123-Y
Zhang X, Zhou Y, Pei S, Zhuge J, Chen J (2020) Adversarial examples detection for XSS attacks based
on generative adversarial networks. IEEE Access 8:10989–10996. https://​doi.​org/​10.​1109/​ACCESS.​
2020.​29651​84
Zhang Z, Ning H, Shi F, Farha F, Xu Y, Xu J, Zhang F, Choo KKR (2021a) Artificial intelligence in cyber
security: research advances, challenges, and opportunities. Artif Intell Rev. https://​doi.​org/​10.​1007/​
s10462-​021-​09976-0
Zhang Z, Ning H, Shi F, Farha F, Xu Y, Xu J, Zhang F, Choo KKR (2021) Artificial intelligence in cyber
security: research advances, challenges, and opportunities. Artif Intell Rev 55(2):1029–1053. https://​
doi.​org/​10.​1007/​S10462-​021-​09976-0
Zhao, G., Zhang, C., & Zheng, L. (2017, July). Intrusion Detection Using Deep Belief Network and Proba-
bilistic Neural Network. 22017 IEEE International Conference on Computational Science and Engi-
neering (CSE) and IEEE International Conference on Embedded and Ubiquitous Computing (EUC).
https://​doi.​org/​10.​1109/​CSE-​EUC.​2017.​119
Zhou Y, Wang P (2019) An ensemble learning approach for XSS attack detection with domain knowledge
and threat intelligence. Comput Secur 82:261–269. https://​doi.​org/​10.​1016/J.​COSE.​2018.​12.​016

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.

13
Detection of cross‑site scripting (XSS) attacks using machine… 12769

Authors and Affiliations

Jasleen Kaur1 · Urvashi Garg1 · Gourav Bathla2

Jasleen Kaur
[email protected]
Urvashi Garg
[email protected]
1
Chandigarh University, Punjab, India
2
University of Petroleum and Energy Studies, Dehradun, India

13