Access to this full-text is provided by Taylor & Francis.
Content available from Applied Artificial Intelligence
This content is subject to copyright. Terms and conditions apply.
The Emerging Threat of Ai-driven Cyber Attacks: A Review
Blessing Guembe
a
, Ambrose Azeta
b
, Sanjay Misra
c
, Victor Chukwudi Osamor
a
,
Luis Fernandez-Sanz
d
, and Vera Pospelova
d
a
Department of Computer and Information Sciences, Covenant University, Ota, Ogun State, Nigeria;
b
Department of Computer Science, Namibia University of Science and Technology, NAMIBIA;
c
Department of Computer Science and Communication, Ostfold University College, Halden, Norway;
d
Department of Computer Science, University of Alcala, Madrid, Spain
ABSTRACT
Cyberattacks are becoming more sophisticated and ubiquitous.
Cybercriminals are inevitably adopting Articial Intelligence (AI)
techniques to evade the cyberspace and cause greater damages
without being noticed. Researchers in cybersecurity domain have
not researched the concept behind AI-powered cyberattacks
enough to understand the level of sophistication this type of
attack possesses. This paper aims to investigate the emerging
threat of AI-powered cyberattacks and provide insights into mal-
icious used of AI in cyberattacks. The study was performed
through a three-step process by selecting only articles based on
quality, exclusion, and inclusion criteria that focus on AI-driven
cyberattacks. Searches in ACM, arXiv Blackhat, Scopus, Springer,
MDPI, IEEE Xplore and other sources were executed to retrieve
relevant articles. Out of the 936 papers that met our search
criteria, a total of 46 articles were nally selected for this study.
The result shows that 56% of the AI-Driven cyberattack technique
identied was demonstrated in the access and penetration phase,
12% was demonstrated in exploitation, and command and con-
trol phase, respectively; 11% was demonstrated in the reconnais-
sance phase; 9% was demonstrated in the delivery phase of the
cybersecurity kill chain. The ndings in this study shows that
existing cyber defence infrastructures will become inadequate
to address the increasing speed, and complex decision logic of
AI-driven attacks. Hence, organizations need to invest in AI cyber-
security infrastructures to combat these emerging threats.
ARTICLE HISTORY
Received 5 November 2021
Revised 26 January 2022
Accepted 28 January 2022
Introduction
Cyberattacks are pervasive and are often regarded as one of the most tactically
significant risks confronting the world today (Dixon and Eagan 2019).
Cybercrimes can engender disastrous financial losses and affect individuals
and organizations as well. It is estimated that a data breach costs the United
States around 8.19 million Dollars and 3.9 Million Dollars on average, and the
annual effect on the global economy from cyberattack is approximately 400
CONTACT Sanjay Misra sanjay.misra@hiof.no Department of Computer Science and Communication,
Ostfold University College, Norway
This article has been republished with minor changes. These changes do not impact the academic content of the
article.
APPLIED ARTIFICIAL INTELLIGENCE
2022, VOL. 36, NO. 1, e2037254 (2409 pages)
https://doi.org/10.1080/08839514.2022.2037254
© 2022 The Author(s). Published with license by Taylor & Francis Group, LLC.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/
licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
Billion Dollars (Fischer 2016; Kirat, Jang, and Stoecklin 2018). A Cyberattack
is the intentional exploitation of computer systems, networks, and businesses.
With increasingly sophisticated cybersecurity attacks, cybersecurity specialists
are becoming incapable of addressing what has become the most significant
threat climate ever before (Chakkaravarthy et al., 2018).
The sophistication of cyberattack techniques poses an existential danger to
enterprises, essential services, and organization infrastructures, with the power to
interrupt corporate operations, wipe away critical data, and create reputational
damage. Today’s current wave of attacks outwits and outpaces humans and even
includes Artificial Intelligence (AI). Cybercriminals will be able to direct targeted
attacks at unprecedented speed and scale while avoiding traditional, rule-based
detection measures thanks to what’s known as “offensive AI” (DarkTrace, 2021). A
new generation of cybercriminals has emerged, one that is both subtle and
secretive, which will influence the future of cybersecurity. The new generation of
cyber threats will be smarter and capable of acting independently with the help of
AI. Future cyberattack methods will be able to be aware of their surroundings and
make informed decisions based on the target environment. The potential of AI to
learn and adapt will usher in a new era of scalable, custom-made, and human-like
assaults (Thanh and Zelinka 2019).
Cybercriminals today have many sophisticated AI-driven cyberattacks
methods by which they can create problems for the government, organiza-
tions, businesses, and individuals. Existing cybersecurity tools are no longer
viable against this advanced cyber-weaponry (Thanh and Zelinka 2019). The
negative use of AI to compromise digital security is known as AI-driven
cyberattack, in which cybercriminals can train robots to socially engineer
targets at human or superhuman levels of performance (Brundage et al.
2018). AI-assisted attacks will be able to adapt to the environment it infects.
Learning from contextual data to emulate trustworthy features of cyberspace
or target weak points it discovers (DarkTrace, 2020). AI-driven cyberattacks
are not a far-fetched future threat. The necessary tools and building blocks for
launching an offensive AI-driven cyberattack already exist (Dixon and Eagan
2019). Recent advancements in AI have influenced the tremendous growth in
automation and innovation. Although these AI systems have many benefits,
they can also be used maliciously. AI-driven cyberattacks have been on the rise
in recent years. Even as many organizations transit to more secure cyberspace,
such as cloud storage services, their resources remain vulnerable to cyber
criminals (Bocetta 2020). In general, AI-driven cyberattacks will only worsen,
then it will be almost impossible for traditional cybersecurity tools to detect
them. It is simply a question of machine efficiency vs. human effort. The
complexity and size of this growing trend are submerging cybersecurity
teams. In contrast, the advanced and qualified cybersecurity specialists neces-
sary to counter this threat successfully are increasingly expensive and difficult
to find. The consequences of these emerging AI-driven attack techniques
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2377
could be life-threatening and highly destructive. These subtle attacks under-
mine trust in organizations by undermining data security and integrity,
potentially resulting in systemic failures (Cabaj, Kotulski, Księżopolski &
Mazurczyk, 2018; Hamadah & Aqel, 2020). An AI-driven cyberattack will
harness a multitude of cyberspace and computer resources well beyond what
a human could enlist, resulting in an attack that is faster, more unpredictable,
more sophisticated than even the strongest cybersecurity team can respond
against. (Bocetta 2020). As AI becomes a more powerful tool in the hands of
malicious actors, cybersecurity researchers, practitioners, and governments
will need to respond with more inventive solutions to successfully safeguard
cyberspace from malicious actors (Hamadah & Aqel, 2020). AI-driven attacks
employ sophisticated obfuscating algorithms and frequently change identifi-
able characteristics that could reach a level of adaptability that renders it
virtually undetectable to both behavioral and signature-based antivirus tools
(Babuta, Oswald & Janjeva, 2020). High-security risks and elusive attacks in
benign carrier applications, such as DeepLocker, have demonstrated the inten-
tional use of AI for negative objectives. Attackers are constantly improving
and changing their attack strategy, with a particular focus on the use of AI-
based techniques in the attack process, known as AI-driven cyberattacks,
which can be used in collaboration with traditional cyberattack techniques
to cause more significant damage in cyberspace while remaining undetected
(Kaloudi and Li 2020; Thanh and Zelinka 2019; Usman et al. 2020).
AI-driven cyberattack techniques will have the capacity to adapt to the sur-
roundings where it executes. They can exploit the vulnerabilities or masquerade as
trusted system attributes by learning from contextual data or information. The
longer the attack exists in the host, the more they integrate and become indepen-
dent of its targets, environments, and countermeasures against cybersecurity
defense infrastructures (Thanh and Zelinka 2019). The consequences of these
emerging AI-driven attack techniques could be life-threatening and highly
destructive. Hence, this study investigates the emerging threat of AI-driven attacks
and reviews the negative impacts of this sophisticated cyber weaponry in
cyberspace.
The paper is divided into five parts. The mechanism for offering the review
process is presented in the next section. Section 3 contains the results. In
section 4, the findings are presented, and in section 5, a conclusion is formed.
Research Questions and Review Process
This study investigates the emerging threat of AI-driven cyberattacks and the
techniques utilized by cybercriminals to carry out AI-driven cyberattacks,
focusing on literature that addresses the research questions. Table 1 illustrates
the three research questions as well as the justification, which also establishes a
foundation for identifying the studies and formulation of our search criteria.
e2037254-2378 B. GUEMBE ET AL.
Research Methodology and Eligibility Criteria
This study’s systematic literature review methodology was based on
Prisma International Standards (Moher, Liberati, Tetzlaff & Altman,
2010). This research aimed to find aset of studies on the topic of AI-
driven cyberattacks that were relevant. To do so, the authors outlined the
research questions offered in this study and explained the reasoning
behind each one. The authors also discuss the criteria for selecting the
required literature and the search method for retrieving relevant data and
published publications.
Search Criteria and Identication of Studies
The following search parameters were used to find relevant literature for this
study:
(1) Make a list of keywords from the research questions.
(2) Identify keywords in relevant literature;
(3) Recognize distinct keyword synonyms and spellings;
(4) To relate primary keywords and concepts using the Boolean operators
“AND” and “OR.”
The search keywords are developed from the research questions in Table 1.
The output of the search string used for searching relevant literature is as
follows: (“AI-driven cyberattacks” OR “AI-driven attack techniques” OR
“malicious use of AI”) AND (“AI-powered cyberattacks” OR “AI-based cyber-
attack” OR “Artificial intelligence in cyberattack” OR “Impact of AI-Driven
attack”).
Three factors were used to apply the eligibility criteria: inclusion, exclusiv-
ity, and quality criteria. These criteria were used to extract the literature from
the search results.
Table 1. Research Questions.
Research questions Rationale
RQ1: What are the current and emerging AI-driven
techniques malicious attackers utilize to carry out
cybercrime?
To identify the existing AI techniques malicious actors
utilize to cause more significant damage in
cyberspace without being noticed.
RQ2: What is the difference between traditional
targeted cyberattacks and AI-driven cyberattacks?
Identify the difference between traditional targeted
cyberattacks and AI-driven attacks and identify their
various logical components.
RQ3: What are the impacts of AI-driven cyberattacks? To identify the effects of AI-driven attacks and how
these techniques can be enhanced in the future to
conceal sophisticated cyber threats.
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2379
Exclusive Criteria
The following exclusive criteria were used to evaluate the retrieved literature:
EC1:Non-discussion of the research questions in the literature.
EC2:Articles on the same subject.
EC3:The same articles from different databases.
EC4:Articles that do not discuss AI-driven cyberattacks.
Inclusion Criteria
The retrieved literature was evaluated with the following inclusive criteria:
IC1:Literature that is relevant to AI-driven cyberattacks.
IC2:Methodologies, Journals, Conference and White Papers that addressed
AI-driven attacks.
IC3:Papers address AI-driven attacks techniques such as deep learning, bio-
inspired swarm intelligence, etc.
Quality Criteria
The retrieved papers were screened based on the following quality criteria:
QC1:Do the papers answer the majority of the research questions?
QC2:Is it possible to describe the AI techniques used for executing AI-
Driven attacks?
QC3:Do the studies provide answers to any of the research questions?
QC4:Do the papers adequately disclose the research methodology?
Selecting Procedure
The key search criteria for this study were ACM, arXiv Blackhat, Scopus,
Springer, MDPI, and IEEE research databases. The PRISMA flowchart,
depicted in Figure 1, depicts the systematic review process and selection
of relevant papers at various phases. Following the presentation of the
sources, criteria, and methodology for selecting relevant publications, a
quantitative evaluation was demonstrated to discover new methodologies,
measures, or contributions offered by researchers in the study domain.
(1) Stage 1 (Extracting Information): This is based on information extrac-
tion; a comprehensive search was conducted on nine electronic data-
bases, yielding 936 article outlines, which served as a pool of potential
articles for subsequent selection, as shown in Table 2.
(2) Stage 2 (Screening): A total of 936 potential articles were found based on
Table 2. There are 417 duplicate papers among the 936 research articles
because publications were found in more than one online resource. The
screening process was then undertaken based on the title of the articles’
irrelevancy, and 309 articles were deemed unsuitable for this study.
e2037254-2380 B. GUEMBE ET AL.
(3) Stage 3 (Eligibility): An article’s relevance and quality cannot be
determined solely by its abstract and title. As a result, complete
text-based selection criteria were used to extract relevant articles
for the study. A total of 210 papers were reviewed for eligibility,
with 164 being eliminated due to imprecise technique or
ambiguity.
Figure 1. PRISMA flowchart illustrating the systematic review process and article selection at
various stages.
Table 2. The number of publications found
in online databases.
S/N Database No. of articles
1 ACM 120
2 arXiv 65
3 Blackhat 12
4 Others 319
5 Scopus 153
6 Springer 49
7 MDPI 13
8 IEEE 205
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2381
(4) Stage 4 (Inclusion): The authors conducted a quality assessment based
on the research of the above topics for the remaining papers. After
resolving issues with article selection, a total of 46 primary papers were
examined for this investigation.
Search Strategy
This section outlines and examines the various search strategies used to
identify the 46 papers for this study. The forty-six articles were published
as follows: fifteen papers for journals, fourteen papers from conference
proceedings, eight articles in workshops, six from symposiums, two papers
from white papers, and one article from scholarly work, as shown in
Figure 2. The search approach was divided into four classes in this
study to explore all contributions made by past researchers in this
research domain. Techniques, status, source, and attack strategy are the
four search techniques used in this study.
Sources
In this study, data was extracted from seven digital databases. ACM, arXiv
Blackhat, Scopus, Springer, MDPI, and IEEE Xplore are available digital
libraries. Published conference proceedings, journal papers, workshops, sym-
posiums, and scholarly work were searched using titles, abstracts, and
keywords.
0
2
4
6
8
10
12
14
16
Conference Journal Scholarly Work Symposium White Paper Workshop
Figure 2. Number of collated studies.
e2037254-2382 B. GUEMBE ET AL.
Attack Strategy
In this paper, the AI-driven attack strategies identified in the forty-six selected
papers are deep learning, bio-inspired computation and swarm intelligence,
and fuzzy model. A large proportion of the selected papers are based on deep
learning strategy, as described in Section 4.
Types of Attacks
On the basis of the 46 selected papers, this section identifies nineteen uses
cases of offensive AI in six stages of the cybersecurity kill chain, as shown
in Figure 3. In the access and penetration phase (AI-aided attack), six
types of AI-driven attacks were identified, four types of AI-driven attacks
were identified in the access reconnaissance stage (AI-targeted attack),
four types of AI, three types of AI-driven attacks were identified in the
exploitation stage (AI-automated attack), two types of AI-driven attacks
were also identified in the delivery stage (AI-concealment attack) and C2
stage (AI-multi-layered attack) respectively. In contrast, one type of AI-
driven attack was identified in action on objectives stage (AI-malware
attack), as shown in Figure 3.
Figure 4 shows that the access and penetration stage has the most publica-
tions (6), followed by the reconnaissance stage (4), the exploitation stage has
three publications, and the delivery and C2 stages have two. In contrast, the
action on objectives stage has the least publication (1).
Figure 3. Modified Cybersecurity Kill Chain for AI-Driven Attack (Kaloudi and Li 2020).
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2383
Techniques
The selected studies demonstrated how malicious actors could utilize AI
techniques to execute vulnerability prediction, End-to-End (E2E) spear-phish-
ing, and intelligent target profiling/ intelligence collection in the cybersecurity
kill chain reconnaissance phase, as discussed in the cybersecurity kill chain
subsection 3.21. The selected studies also identified nineteen (19) AI techni-
ques that malicious actors can utilize to execute attacks in the access and
penetration phase of the cybersecurity kill chain. CNN has the highest appear-
ance (five) AI techniques utilized by the authors to demonstrate access and
penetration attacks. GAN and RNN were used two times respectively to
demonstrate access and penetration attacks. In contrast, LSTM, SVC, SVM,
cycle-GAN, TOD+CNN, RF, MP, GBRT, and KNN were demonstrated one
time in malicious attacks as discussed in subsection 3.2.2. The selected articles
identified three AI techniques that malicious actors can utilize to execute
attacks in the delivery stage of the cybersecurity kill chain. Two of the studies
identified GAN for intelligent concealment to generate adversarial malware
and undetectable malware URLs. One of the studies utilized LSTM to generate
automated malicious evasive payloads. At the same time, one study demon-
strated how Malicious actors could utilize DNNs to conceal malicious intent
and activate it when it gets to its specific target (Kirat, Jang, and Stoecklin
2018), as discussed in subsection 3.2.3. From the selected articles, three (3) AI
techniques were utilized to demonstrate behavioral analysis to find new ways
to exploit targeted infrastructures and automated disinformation generation
in the delivery phase of the cybersecurity kill chain, as discussed in subsection
Figure 4. Offensive AI Techniques Cyberattacks in the Modified Cybersecurity Kill Chain.
e2037254-2384 B. GUEMBE ET AL.
3.2.4. One of the selected studies demonstrated how NNs and Reinforcement
Learning (RL) could be utilized to execute behavioral analysis to exploit
vulnerabilities in web-based applications in other to bypass web-based appli-
cation authentication systems. K-means clustering was also utilized to demon-
strate how AI-driven self-learning malware can successfully exploit
vulnerabilities in security detection systems and act as though they were
unintentional failures on computer applications by exploiting and compro-
mising sensitive environmental control infrastructures in the exploitation
phase of the cybersecurity kill chain. Markov chains and LTSM were utilized
to execute automated machine-generated content disinformation by imple-
menting an end-to-end spear-phishing technique to generate personalized
content for high target users on Twitter. Two types of AI-driven cyberattacks
were identified: intelligent self-learning malware and automated domain gen-
eration, as discussed in subsection 3.2.5. Four types of AI techniques identified
malicious actors could utilize to execute AI-driven self-learning malware and
automated domain generation attack in the command and control of the
cybersecurity kill chain, as discussed in subsection 3.2.5. Two of the selected
studies utilized K-means clustering, Gaussian distribution, and DNNs to
demonstrate intelligent self-learning malware attacks, as discussed in subsec-
tion 3.2.5. 56% of the AI-Driven cyberattack technique were identified in the
access and penetration phase of the cybersecurity kill chain, 12% AI techniques
were identified in the exploitation and command and control phase, respec-
tively, 11% AI techniques were identified in the reconnaissance phase, 9% AI
Figure 5. Identified AI-Driven Cyberattack Techniques.
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2385
technique were identified in the delivery phase, and no AI technique was
demonstrated in action on the objective stage of the cybersecurity kill chain
to execute offensive AI attacks as shown in Figure 5.
Status
The literature chosen is grouped into three categories. This is accomplished as
illustrated in Table 3.
Table 3. Status of existing literature with respective narration.
S/N State Narration
1 Implemented This refers to the category of studies that designed a proof of concept to
demonstrate AI-driven attacks.
2 Proposed This category of studies is based on new techniques or methods without any proof
of concept and evaluation.
3 Implemented and
evaluated
This category of studies refers to those that designed a proof of concept
demonstrating AI-driven cyberattacks and evaluated the proof of concept
based on performance metrics.
Figure 6. Generated keywords from titles of selected articles.
e2037254-2386 B. GUEMBE ET AL.
Result Obtained
The results of the four search methodologies were analyzed in detail in this
section, which discussed the study findings. In several subsections, the study
presented a full discussion of the findings in relation to the outline study
topics, along with a concise interpretation of our findings. The results of a
word cloud analysis employing titles of selected articles on the orange machine
learning integrated environment are shown in Figure 6, with ‘Learning’ being
the most common occurrence, followed by ‘Artificial,’ ‘Machine,’ ‘attack,’
‘attacks,’ and ‘cybersecurity.’
Figure 7. Average numbers of relevant articles.
Figure 8. Publication Year.
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2387
Search Strategy 1: Source
The initial search exercise was carried out using a hierarchical search
method to find related articles on AI-driven cyberattacks using the article
title and keywords before designing a final search strategy.
The following databases were utilized to find related literature for publica-
tions published between 2014 and 2021: ACM, arXiv Blackhat, MDPI, Scopus,
Springer, and IEEE Xplore. The retrieved findings for relevant article sources
are shown in Figure 7, while the number of relevant publications released
during the research year is shown in Figure 8.
Figure 5 shows that of the final 46 papers, ACM had the most relevant
publications with eleven (11), followed by arXiv with five (5), Blackhat and
IEEE have four (4) publications respectively, Scopus and Springer have two
(publications) respectively, while other fifteen (15) relevant publications were
retrieved from other sources.
Figure 6 demonstrates that 2018 had the most relevant papers (14), whereas
2014 had the least publications (2) respectively. The threat of AI-driven
cyberattacks grows despite ongoing research attempts to understand and
combat these advanced cyber weapons.
Figure 9 depicts the AI techniques used by the selected studies to
demonstrate the malicious use of AI in cyberattacks in the access and
penetration stage of the modified cybersecurity kill chain.
Figure 10 depicts the AI techniques used by the selected authors to
demonstrate the malicious use of AI in the delivery stage of the modified
cybersecurity kill chain. The results indicate that GAN has the most
publications (2), while DNNs and LSTM have one publication,
respectively.
5
1
2
11111111
2
1
0
1
2
3
4
5
6
Figure 9. AI Techniques in the Access and Penetration Attack Stage.
e2037254-2388 B. GUEMBE ET AL.
Search Strategy 2: Techniques
The outcome of this segment is broken down into six stages of the cyberse-
curity kill chain, which include reconnaissance (AI-targeted attack), access
and penetration (AI-aided attack), delivery (AI-concealment attack), exploita-
tion (AI-automated malware), command on control (AI-multi-layered attack),
action on objectives (AI-massive attack).
Reconnaissance Stage
Three types of AI techniques were identified in the reconnaissance stage
of the cybersecurity kill chain. The selected studies demonstrated how
malicious actors could utilize Markov chains/LTSM, NNs, DNNs to exe-
cute vulnerability prediction, End-to-End (E2E) spear-phishing, and intel-
ligent target profiling/intelligent collection, respectively. Table 4
summarizes the AI techniques utilized to execute AI attacks in the recon-
naissance stage.
Access and Penetration Stage
This study identified six (6) AI-driven attacks in the access and penetra-
tion phase. They include; password guessing/password cracking (brute-
force attack), intelligent captcha/manipulation, smart abnormal behavioral
generation, AI model manipulation, and smart fake reviews generation.
The study also identified nineteen (19) AI techniques that can be utilized
0
0.5
1
1.5
2
2.5
DNNs GAN LSTM
Figure 10. AI Techniques in the Delivery Stage.
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2389
Table 4. Offensive Utilization of AI Techniques in the Reconnaissance Stage of Cybersecurity Kill
Chain.
Authors Attack Class of Attack Attack Goal Technique
Seymour and
Tully
(2016)
E2E spear
phishing
Intelligent Malware An automated end-to-end spear-phishing
strategy that includes identifying high-value
targets and propagating personalized
machine-generated information on Twitter.
Markov
chains
and
LTSM
Dheap (2017) Vulnerability
Prediction
Outcome Prediction Boost malicious actors’ confidence to seek
risker and high-value outcomes to
overpower state-of-the-art cybersecurity
infrastructures.
NNs
Kirat, Jang,
and
Stoecklin
(2018).
DeepLocker Intelligent Target
Profiling/
Intelligent
Collection
Hide payload without being detected in video
conferencing application.
DNNs
Table 5. Offensive Utilization of AI Techniques in the Access and Penetration Stage of
Cybersecurity Kill Chain.
Authors Attack Class of
Attack
Attack Goal Technique
Bahnsen
et al.
(2018).
DeepPhish:
Simulating
Malicious AI.
Automated
Payload/
Phishing.
To prevent AI cyberattack detection model and
execute more effective phishing attacks.
LSTM.
Hitaj et al.
(2019).
PassGAN. Password
Guessing.
Guess password based on learning the
distribution of actual password leaks.
GAN.
Trieu and
Yang
(2018).
Intelligent
password brute
force attack.
Password
Cracking
To obtain previous password sequences, in
order to create new passwords by guessing
one character at a time.
RNN
Lee and
Yim
(2020).
Offensive password
authentication
technique.
Password
Guessing.
Predict and steal users’ actual passwords based
on keyboard strokes.
LR, SVM, SVC,
RF, KNN,
GBRT, MP
Bursztein
et al.
(2014).
Single Step
Captcha Solver.
Intelligent
Captcha
Attack.
To Combine segmentation and recognition
issues to attack captcha in a single phase.
CNN.
Gao et al.
(2017)
Breaking text-
based Captchas.
Intelligent
Captcha
Attack.
Four deep learning models with 2, 3, 5, and 6
convolutional layers as recognition engines
to attack text-based captchas.
CNN.
Ye et al.
(2018)
Text-based
Captcha solver.
Intelligent
Captcha
Attack.
To generate synthetic captchas and then fine-
tune the base solver on a limited selection of
real Captchas using transfer learning.
GAN
Gao et al.
(2017).
Attacking two-layer
captcha.
Intelligent
Captcha
Attack.
To break two-layer captchas with an enhanced
LeNet-5 and a radical CNN model as a
recognition engine.
CNN.
Yu and
Darling
(2019).
AI-based Captcha
solver.
Intelligent
Captcha
Attack.
To determine which character was contained in
a segmented sample in order to crack
captcha.
TOD+CNN
Noury and
Rezaei
(2020).
Captcha solver for
vulnerability
assessment
Intelligent
Captcha
Attack.
To explore the flaws and weaknesses of existing
Captcha generator systems.
CNN.
Chen et al.
(2018).
Hollow Captcha
solver.
Intelligent
Captcha
Attack.
To improve attack accuracy and reduce the
attack time, use precise filling and
nonredundant merging.
CNN.
Li et al.
(2021).
End-to-end attack
on text-based
Captchas.
Intelligent
Captcha
Attack.
Using Captcha synthesizers based on the cycle-
GAN, create some false samples.
cycle-GAN
Yao et al.
(2017).
Smart Fake
Review
Generation.
To generate fake smart reviews. RNN
e2037254-2390 B. GUEMBE ET AL.
by malicious actors to execute access and penetration attacks. Table 5
summarizes the AI techniques utilized to execute AI attacks in the access
and penetration stage.
Delivery Stage
In the delivery phase, two types of AI-driven cyberattacks were identified:
intelligent concealment and evasive malware. The selected studies identi-
fied three types of AI techniques; malicious actors could utilize that to
execute AI-driven concealment and evasive attack, as illustrated in Table 6
and Figure 8. The study identified three AI techniques that malicious
actors can utilize to execute attacks in the delivery stage. Table 6 sum-
marizes the AI techniques utilized to execute AI attacks in the delivery
stage.
Exploitation Stage
The exploitation phase involves gaining authorized access to computer appli-
cations and resources, and after gaining access to the target, malicious actors
can utilize AI techniques to execute complex attacks that are difficult to detect
using NNs and DNNs. Table 7 summarizes the AI techniques utilized to
execute AI attacks in the exploitation stage.
Command and Control Stage
In the command and control (C2) phase, two types of AI-driven cyberattacks
were identified: intelligent self-learning malware and automated domain gen-
eration. Table 8 summarizes the AI techniques utilized to execute AI attacks in
the C2 stage.
Table 6. Offensive Utilization of AI Techniques in the Delivery Stage of Cybersecurity Kill Chain.
Authors Attack Class of Attack Attack Goal Technique
Bahnsen et al.
(2018).
Malicious
Payload.
Intelligent
Concealment.
Automated generation of undetected phishing
URLs.
LSTM
Hu and Tan
(2021).
Adversarial
malware
generation.
Intelligent
Concealment.
Generating undetectable adversarial malware to
bypass machine learning black-box cyber
threat detection systems
GAN
Anderson,
Woodbridge,
and Filar
(2016).
Undetectable
malware
URL.
Intelligent
Concealment.
GAN-based automatic generation of
undetectable malware URL that learns to
bypass DNNs-based malware detection
system.
GAN
Kirat, Jang, and
Stoecklin
(2018).
DeepLocker Evasive
Malware
Conceal its attack and only activate it for specific
targets.
DNNs
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2391
Table 7. Offensive Utilization of AI Techniques in the Exploitation Stage of Cybersecurity Kill Chain.
Authors Attack Class of Attack Attack Goal Technique
Petro and Morris
(2017).
DeepHack: Open
source Hacking AI
Behavioral analysis to
exploit vulnerabilities.
Breaking into web applications. NNs, Reinforcement
Learning (RL).
Chung, Kalbarczyk,
and Iyer (2019)
Self-Learning
Malware.
Behavioral analysis to
exploit vulnerabilities.
A malicious attack that exploits and compromises environmental control systems while
masquerading as an unintentional failure on computer infrastructures.
K-means clustering
Seymour and Tully
(2016)
Machine generated
spear-phishing.
Automated
disinformation
generation
To generate personalized content for high-value targets on Twitter. Markov chains and
LTSM
e2037254-2392 B. GUEMBE ET AL.
Table 8. Offensive Utilization of AI Techniques in the C2 of Cybersecurity Kill Chain.
Authors Attack Class of Attack Attack Goal Technique
Chung, Kalbarczyk, and
Iyer (2019)
To execute attack
planning phase.
Intelligent self-
learning malware
To attack computers at a supercomputer facility without the attacker’s knowledge
by interfering with the cyber-physical systems.
K-means clustering, and
Gaussian distribution
Anderson, Woodbridge,
and Filar (2016)
DGA classifier to analyze
infected hosts.
Automated domain
Generation.
To assess successful DNS queries made by infected hosts and assign scores based
on values derived from training datasets.
GAN
Kirat, Jang, and Stoecklin
(2018)
DeepLocker for self-
learning malware.
Intelligent self-
learning malware
To establish a self-learning attack in the C2 channel. DNN
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2393
Search Strategy 3: Status
The current state of the AI-driven attack tool was assessed using three cate-
gories, and the results of the analysis for the selected publications are shown in
Figure 11. Based on existing AI-driven cyberattack technologies, the analysis
result for the 46 articles evaluated in this study shows that 63% of the existing
AI-driven cyberattack tools are implemented and evaluated, 25% are proposed
only, and 12% of the existing AI-driven cyberattack tools are implemented
without evaluation. As a result of the current state of the study, the majority of
existing AI-driven cyberattack tools are based on implementation and
evaluation.
Discussion on Research Questions
This section reviews the core concepts that make up this study, such as the
current and emerging AI-Driven cyberattack techniques, weaponization of
machine learning and deep learning techniques in cyberattacks, and types of
AI-Driven attacks in the cybersecurity kill chain and existing AI-Driven
attacks. Also discussed in the section are the result and findings of this study.
RQ1:Current and Emerging AI-Driven Cyberattack Techniques
The advancement of cyberattack tools and recent techniques are shaping and
expanding the cyberattack domain, which opens up cyberspace to a wide range
of sophisticated cyber weaponry with many powerful negative effects (Kaloudi
and Li 2020). Brundage et al. (2018) established a scenario that notifies
cybersecurity researchers and industry about the malevolent utilization of AI
by embedding some hypothetical concepts within digital, physical and political
security domains. Researchers have established a few concepts that showed the
potential of an automatic exploit generation in state-of-the-art applications.
12%
25%
63%
Implemented Proposed Implemented and evaluated
Figure 11. Status of existing AI-driven cyberattack tools.
e2037254-2394 B. GUEMBE ET AL.
Malicious actors are utilizing fuzzy models to develop a next-generation
malware capable of learning from its environment, continuously updating
itself with new variants, and infecting vulnerable and sensitive computer
infrastructures without being noticed (Kaloudi and Li 2020). Malicious actors
can utilize these concepts to deploy a new type of sophisticated and stealthy
cyber weaponries.
AI-Driven Attacks in the Reconnaissance Stage of the Cybersecurity Kill Chain
Malicious actors can use AI techniques to improve reconnaissance to study
normal behavior and operations about a cybersecurity defense mechanism,
computer infrastructures, and devices (Kaloudi and Li 2020). In this case, a
malicious actor can obtain structural, operational, and topological data about
the user’s devices, network flows, and network infrastructure to identify a
critical relationship with the intended targets. Malicious actors may be able
to use AI technology to detect patterns of targeted attacks in massive
volumes of data. A reconnaissance attack, also classified as AI-targeted,
depends on a well-prepared planning phase to execute its attack. AI’s cap-
ability to interpret, discover and comprehend patterns in large amounts of
the dataset can be utilized to provide in-depth analysis and to develop
targeted exploration processes by overcoming human limitations (Kaloudi
and Li 2020). The authors identified four AI-driven threat use cases in the
reconnaissance phase of AI-targeted attacks; they include intelligent target
profiling, clever vulnerability detection/intelligent malware, intelligent col-
lection/automated learn behavior, and intelligent vulnerability/outcome
prediction.
Intelligent Target Profiling. AI has already been shown to have an impact on
the ability to profile the use of information and communication technology.
Bilal et al. (2019) presented a taxonomy of profiling approaches as well as the
AI algorithms that enable them. The authors pointed out that there are two
forms of profiling: individual and group profiling, and that fuzzy logic ontol-
ogy, machine learning, and convolutional neural networks are the most
commonly used AI methodologies. With the advancement of AI techniques,
cyberattack targets can be profiled based on their social media activity and
public social media profiles. To maximize persuasive potential, AI systems
may allow groups to target precisely the correct message at precisely the right
time (Brundage et al. 2018). Malicious actors can utilize AI techniques to
improve the chances of profiling their targets (Kirat, Jang, and Stoecklin 2018).
Malicious actors can utilize DNNs and NNs to classify and profile targets
(Dheap 2017). It is possible to draw conclusions from research prototypes like
SNAP_R that both the technology readiness level and the chance of malicious
end applications for executing intelligent profiling are high (Seymour and
Tully 2016)
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2395
Intelligent Collection. Intelligence collection is a type of reconnaissance that
aids in the planning and formulation of cyberattack policies. The AI techni-
ques utilized to execute this type of attack include NLP and DNNs. These
analyses should be classified as dual-use because they automate the collection
of generic information on specific types of attacks and specific features that
affect risk (both defensive and offensive) (Dheap 2017; Kirat, Jang, and
Stoecklin 2018).
Intelligent Malware. By infiltrating the environmental control systems, intel-
ligent malware §can initiate indirect cyber weaponries that pretend to be
unintentional failures on computing infrastructure (Chung, Kalbarczyk, and
Iyer 2019). Malicious actors can use a highly automated end-to-end spear-
phishing technique that involves identifying high-priority targets and auto-
matically disseminating personalized machine-generated information
(Seymour and Tully 2016).
Outcome Prediction. AI techniques can examine current and previous occur-
rences in order to predict the outcomes of planned activities in the future.
Cyber-related evaluation and simulation development techniques could be
critical steps toward more advanced AI prediction models. For cybercriminals,
offensive AI could increase their confidence in pursuing high-risk, high-value
outcomes in order to defeat state-of-the-art cybersecurity systems (Dheap
2017).
AI-Driven Attacks in the Access and Penetration Stage of the Cybersecurity Kill
Chain
This phase of cyberattack is also known as an AI-aided attack. This study
identified six (6) AI-driven attacks in the access and penetration phase. They
include; password guessing/password cracking (brute-force attack), intelligent
captcha/manipulation, smart abnormal behavioral generation, AI model
manipulation, and smart fake reviews generation.
Automated Payload Generation/Phishing. Malicious actors are capable of
weaponizing machine learning algorithms to improve phishing attacks and
make them invisible by cybersecurity detection systems, as demonstrated by
Bahnsen et al. (2018) in DeepPhish: Simulating Malicious AI. DeepPhish is an
AI algorithm that learns patterns from the most effective phishing URLs in the
past to generate new synthetic phishing URLs. The objective is to create more
effective phishing URLs to avoid AI detection and conduct more effective
phishing attacks. To create phishing URLs in the past, attackers employed
randomly generated segments. By utilizing an LSTM model to create a phish-
ing URL classifier and produce new effective synthetic phishing URLs, the
authors demonstrated the effectiveness of phishing attacks’ by improving the
e2037254-2396 B. GUEMBE ET AL.
efficiency and success rate. The authors claimed that by training DeepPhish on
two different threat actors raised the attack’s effective rate from 0.69% to 20.9%
and from 4.91% to 36.28%.
AI-Driven Password Guessing/Password Cracking. Three types of AI-driven
password attacks were identified; password brute-force attack, password gues-
sing, and password stealing. A deep learning model for password guessing was
proposed by Hitaj et al. (2019). The authors evolved an automated password
guessing technique based on GAN by learning the distribution from actual
password breaches. Brute force, which entails testing all possible character
combinations exhaustively; a dictionary, which entails using a list of likely
words and previous password leaks in the hopes of correctly guessing; and
rule-based approaches, which entail defining generation rules for possible
password transformations such as concatenation.
Hitaj et al. (2019) evolved a model for correctly training a GAN such that
tailored samples can be generated from the training set. GAN is used to
automatically create passwords in the following way: The GAN is made up
of a generating DNN (G) and a discriminative DNN (D) (D). There is also a
training dataset using actual password samples, which are a collection of
leaked passwords. A noise vector was used to train the generator G, which
represents a random probability distribution and generates a sequence of
vectors known as false password samples. The real and false samples are fed
into the discriminator D, which subsequently learns to tell the difference
between the two. When attempting to understand the original distribution
of true password leaks, G compel D to disclose data to G. The rockyou dataset,
which is an industry-standard password list, was used to train PassGAN,
which achieved an effective result of both guessing new unique passwords
and mimicking the distribution of rockyou dataset. PassGAN was able to
correctly match 10,478,322 (24.2%) out of the 43,354,871 unique passwords
from the LinkedIn data breach. GAN was never exposed to any of the
LinkedIn datasets, but it was nevertheless able to produce meaningful, unique
passwords based on the rockyou words. PassGAN, in combination with
HashCat, was able to predict between 51 and 73% of unique passwords more
accurately than HashCat alone.
Lee and Yim (2020) implemented a K-Nearest Neighbors (KNN), logistic
regression (LR), decision tree (DT), linear support vector classifier (SVC),
random forest (RF), support vector machine (SVM), gradient boosting regres-
sion tree, and multilayer perceptron models for data classification from key-
board strokes. The implemented model was 96.2% accurate when it came to
stealing keyboard data. This means that cybercriminals can steal users’ actual
keyboard data in the real world with AI techniques. Trieu and Yang (2018)
utilized Torch-rnn, an open-source machine learning technique to generate
new candidate passwords based on a pattern similar to prior passwords. As
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2397
demonstrated in Figure 11, attackers can build new passwords by guessing one
character at a time using the RNN, which is trained on previously obtained
password sequences. The RNN produces a prediction by updating its hidden
state at each timestamp by finding patterns over sequences. With his techni-
ques, attackers are capable of constructing new terms that will present incred-
ibly likely passwords. The comparisons were carried out for different
dictionary lengths (i.e. the dictionary’s total amount of words): 50, 100, 250,
500, 750, and 1000. Calculate the average of 100 trials for each trial. The result
shows that the success rate of AI-driven password Brute-force attack out-
performed the traditional algorithm Brute-force attack as shown in Figure 12.
Figure 13 also illustrates the concept of AI-driven password brute-force attack
(Trieu and Yang 2018), as illustrated in Figure 15.
0
20
40
60
0 200 400 600 800 1000 1200
Success Rates of AI- Based Algorithm
Success Rates of Traditional Algorithm
Figure 12. Success Rates of AI-Driven Password Brute-force Vs. Traditional Brute-force Attack.
Figure 13. Password Brute-Force Attacks Powered by AI.
e2037254-2398 B. GUEMBE ET AL.
Intelligent Captcha Attack and Manipulation. Yu and Darling (2019) utilized
an open-source Python Captcha package to boost recognition accuracy by
combining TensorFlow object detection (TOD) and a speech segmentation
method with CNN. The implemented model was able to determine which
character was contained in a segmented sample. The result shows that the
well-designed TOD+CNN model can crack open-source CAPTCHA libraries
like Python Captcha and external captcha like the Delta40 benchmark. It has
also been demonstrated that TOD+CNN can crack various types of
CAPTCHAs, such as HashKiller13. Bursztein et al. (2014), developed a novel
method for attacking captcha in a single step by combining segmentation and
recognition problems using machine learning techniques. When both actions
are done at the same time, the technique can take advantage of knowledge and
context that would not be available if they were done separately. At the same
time, it removes the need for any hand-crafted components, allowing this
method to be applied to new Captcha schemes that the previous method could
not. Without making any modifications to the algorithm or its settings, the
authors were able to solve all of the real-world Captcha schemes they inves-
tigated exactly enough to consider the scheme insecure in reality, including
Yahoo (5.33%) and ReCaptcha (33.34%). The success of this strategy against
the Baidu (38.68%) and CNN (51.09%) schemes, both of which use occluding
lines and character collapsing, implies that it can beat occluding lines in a
broad sense. Noury and Rezaei (2020) proposed a vulnerability assessment
Captcha solution based on deep learning. To explore the weaknesses and
vulnerabilities of existing Captcha generation systems, the authors used a
CNN model called DeepCAPTCHA. The numerical and alpha-numerical
test datasets have cracking accuracy rates of 0.9894 and 0.983, respectively.
That means more effort will be required to develop powerful Captchas that are
resistant to AI-driven Captcha attack models.
Chen et al. (2018) suggested a hollow captcha attack that uses exact filling
and nonredundant merging to improve attack accuracy and reduce attack
time. To begin, the character shapes were methodically fixed using a thinning
approach. Secondly, an inner-outer contour filling technique was developed
for obtaining solid characters, which only fills the vacant character compo-
nents rather than noise blocks. Finally, segmenting solid characters yields
many distinct characters but only a few character components. Fourth, to
obtain individual characters without duplication, a minimum-nearest neigh-
bor merging technique was proposed. Finally, to obtain the final recognition
results, (CNN) was used.
On text-based Captchas, Li et al.((2021) utilized cycle-GAN to train
Captcha synthesizers to make several fake samples. Basic recognizers based
on convolutional recurrent neural networks were trained using the fake
dataset. After that, an active transfer learning mechanism optimizes the basic
recognizer using small numbers of labeled real-world Captcha samples. This
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2399
method successfully solved the Captcha techniques used by ten (10) popular
websites such as Amazon (88.4% success rate), Apple (0.877 success rate), Sina
(0.85 success rate), Baidu (0.807 success rate), Weibo (0.798 success rate), eBay
(0.743 success rate), Sogou (0.717 success rate), and Microsoft’s two-layer
approach had a success rate of 0.224, indicating that the attack is likely wide-
spread. The findings demonstrate that combining multiple anti-recognition
measures can increase captcha security, but only to a limited extent.
Gao et al. (2017) utilized deep learning to crack text-based Captchas and
create image-based Captchas. As recognition engines, the authors used four
CNN models with 2, 3, 5, and 6 convolutional layers. With success rates
ranging from 0.10 to 0.90, the authors were able to defeat the Roman char-
acter-based Captchas used by the fifty most popular websites in the world, as
well as three Chinese Captchas that use a broader character set. The average
pace of this attack is substantially faster than prior attacks. Because these
focused tactics cover almost all known resistance mechanisms, this offensive
AI technique can breach other existing Captchas. Ye et al. (2018) presented a
GAN-based approach for text-based Captcha solver. This was accomplished
by first employing a Captcha synthesizer that generates synthetic captchas
automatically so that a base solver may be learned, and then using transfer
learning to fine-tune the basic solver on a limited set of real Captchas. The
authors evaluated the implemented model on 33 different Captcha schemes,
including 11 that are presently employed by 32 of the top 50 most visited
websites. The authors demonstrated that their method is incredibly effective,
cracking state-of-the-art Captchas in under 0.05 seconds. Gao et al. (2017)
proposed a simple yet effective AI-driven solution for defeating Microsoft’s
two-layer captcha. The authors created an improved LeNet-5, a radical CNN
model, as the recognition engine. The implemented model had a success rate
of 44.6% and an average speed of 9.05 seconds on a standard desktop com-
puter with a 3.3 GHz Intel Core i3 CPU.
Smart Fake Review Generation. Yao et al. (2017) presented a two-phase review
generation and customization attack that can generate reviews that are unrec-
ognizable by statistical detectors. The authors implemented an RNN-based
fake review generation that is capable of generating misleading but realistic-
looking reviews aimed at restaurants on the Yelp App. The result showed that
the difficulty in detecting this form of attack is evidenced by the high quality of
reviews generated.
AI-Model Manipulation. Malicious actors can purposely manipulate the data
of machine learning models with adversary techniques to undermine the
model. In several instances, malicious actors can insert a fake input set or
manipulate the text of spam e-mails to bypass the spam filters, classification
model. Cybercriminals can utilize a Naïve Bayes (NB) model that is used for
e2037254-2400 B. GUEMBE ET AL.
spam mail filtering by altering the input and training data to bypass the spam
filter (Dheap 2017; Truong et al. 2020). Zhou et al. (2021) conducted a
thorough investigation on deep model poisoning attacks on federated learn-
ing. The authors utilized the regularization term in the objective function to
inject malicious neurons in the redundant network space to improve poison-
ing attacks in persistence, effectiveness, and robustness. DNNs models are
subject to purposefully manipulated samples known as adversarial instances.
These adversarial examples are created with little changes, yet they can cause
DNN models to make incorrect predictions.
AI-Driven Attacks in the Delivery Stage of the Cybersecurity Kill Chain
From the selected studies, AI-driven concealment and AI-driven evasive
attacks were identified as discussed below.
Intelligent Concealment and Evasive Malware. Bahnsen et al. (2018) utilized
LSTM to generate sophisticated phishing URLs that are sufficient enough to be
undetected by state-of-the-art cybersecurity detection infrastructures. Hu and
Tan (2021) proposed a GAN technique that is capable of generating undetect-
able adversarial malware to bypass machine learning black-box detection sys-
tems. Anderson, Woodbridge, and Filar (2016) proposed a GAN-based
automatic generation of undetectable malware URL that learns to bypass
DNNs-based detection systems. The result shows that domains generated
from the implemented GAN model bypass the DNNs, and GAN malware
detection systems. Also, a random forest classifier that relies on hand-crafted
features was easily bypassed. Kirat, Jang, and Stoecklin (2018) proposed a
sophisticated evasive malware that is capable of hiding its malicious payload
attack in video conferencing applications without being detected. The authors
utilized DNNs to conceal its nefarious aim and only enable them for selected
targets.
AI-Driven Attacks in the Exploitation Stage of the Cybersecurity Kill Chain
The selected studies identified AI-driven exploitation attacks, also known as
AI-automated malware, as discussed below.
Behavioral Analysis to Find New Ways to Exploit. Petro and Morris (2017)
evolved a Machine Learning model called DeepHack. The authors demon-
strated how the implemented model could be utilized to break and bypass
web-based applications using NNs and reinforcement learning (RL). Chung,
Kalbarczyk, and Iyer (2019) demonstrated how k-means clustering could be
utilized to determine attack effects on the target system, the author’s utilized
logical control data from the targeted system and a Gaussian distribution. The
idea behind this technique was that once malware had gained access to a
system, it needed to know how to operate and exploit weaknesses without the
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2401
need for further assistance from the attacker. The model effectively illustrated
how AI-driven malware might launch their own attacks using a self-learning
algorithm, reducing the level of knowledge needed by the attacker to success-
fully influence and exploit the target system.
Automated Disinformation Generation. Seymour and Tully (2016) demon-
strated the automation of malicious payload in the phishing process by utiliz-
ing data science techniques to the target audience with personalized phishing
messages. As a result, this malicious technique learned how to post phishing
messages aimed just at high-value users, resulting in an automated targeted
spear-phishing campaign. To find the high-value targets, the authors used the
k-means clustering technique to cluster a collection of Twitter accounts into
groups based on public profiles and social engagement indicators such as
retweets, likes, and a number of followers. The attack disseminates custo-
mized, computer-generated posts with a truncated URL inserted in them once
targets have been identified and established. NLP was used to determine which
topics the target is interested in. As a result, it uses both Markov models and
LSTMs to construct the content of the postings and also learns to guess the
next word by analyzing the preceding context in the target’s posting history.
AI-Driven Attacks in the Command and Control Stage of the Cybersecurity Kill
Chain
Malicious actors commonly try to establish channels for the further commu-
nications link between it and the target with the objective of exerting influence
over the compromised computer infrastructure and other systems on its
internal network infrastructures. However, by utilizing AI techniques, cyber-
criminals do not require a C2 channel to execute their attacks (Kirat, Jang, and
Stoecklin 2018). Based on the existing target attributes, the AI-Driven C2
malware automatically predicts when it will be unlocked across various sorts
of nodes. As a result, a multi-layered AI-driven attack is capable of providing
access to other computer infrastructure components remotely and automati-
cally. Depending on the intents of the attacker, a successful AI-driven C2
attack can be used to disseminate the virus to other computers on the network,
prompting the target to establish botnets, and downloading and installing
remote access trojans (Zouave et al. 2020).
Intelligent Self-Learning Malware. Self-learning malware could be used to
infiltrate cybersecurity defense systems in a supercomputer facility indirectly
by interfering with the cyber-physical systems (CPS) automation system of the
building (Chung, Kalbarczyk, and Iyer 2019). To classify the target system’s
logical control data and determine attack effects on the target system, the
authors employed k-means clustering and Gaussian distribution. The goal of
the simulation was to teach malware how to behave without the help of the
e2037254-2402 B. GUEMBE ET AL.
attacker once it had gained access to the system. The self-learning malware can
use a self-learning algorithm to carry out its own attack plans, reducing the
amount of information necessary by malicious actors to successfully manip-
ulate the target system.
Automated Domain Generation. DGA classifiers employ GAN to evaluate
and grade DNS queries executed by compromised hosts that were success-
ful based on the values generated from the different training sets
(Anderson, Woodbridge, and Filar 2016). DGAs are identified as queries
that fall below a specific threshold and are prohibited. Anderson,
Woodbridge, and Filar (2016) demonstrated how Cybercriminals could
utilize domain Generation Algorithms (DGAs) to carry out sophisticated
cyberattacks in the C2 phase of the cybersecurity kill chain. Malicious
actors can also utilize this technique to establish data exfiltration (Sood,
Zeadally, and Bansal 2017).
AI-Driven Attacks in the Action on Objective Stage of the Cybersecurity Kill Chain
AI-Driven DDoS Attack. There were automatic malware distribution and
vulnerability type changes via C&C servers, but there was still a limitation.
The requirement for human intervention was this constraint. The rise of
AI in DDoS ushers is a new era of attack that does not necessitate the
presence of humans. AI-Driven DDoS attack eliminates the need for
human intervention entirely. Machines are now assaulting applications
and state-of-the-art cybersecurity defense infrastructures. They are com-
pletely automated, altering vulnerability types and attack vectors in
response to the defense’s response. If one attacking signature fails, the
machine can think for itself and switch to a different signature. All of this
is carried out automatically, without the need for human intervention
(Kaloudi and Li 2020; Kirat, Jang, and Stoecklin 2018).
RQ2: Traditional Targeted Cyberattacks and AI-Driven Cyberattacks
The traditional targeted cyberattack is a simplistic if-then conditional con-
struct where it asks this question; is this a target? And if the answer is “No” the
malicious program is going to end, and if the answer is “Yes” the malicious
program is going to execute its attack (Kirat, Jang, and Stoecklin 2018). Figure
14 illustrates the decision logic of the traditional targeted attack.
Since cybercriminals realized that cybersecurity experts are using sandboxes
to analyze and combat these traditional targeted attacks, they are now trans-
forming this simplistic form of the if-then conditional construct to a very
convoluted and complicated decision logic using Deep Neural Networks
(DNN). With the concept of DNN, malicious actors can decide whether to
attack or not. The problem for the defender is that it will be extremely difficult
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2403
to figure out what is the actual malicious code and what it is the right target
(Kirat, Jang, and Stoecklin 2018). Figure 15 illustrates a DNN cyberattack
decision logic.
Figure 14. Traditional Targeted Cyberattack Decision Logic.
Figure 15. DNN Cyberattack Decision Logic.
e2037254-2404 B. GUEMBE ET AL.
RQ3: Impact of AI-Driven Cyberattacks
The consequences of these emerging AI-driven cyberattacks could be life-threa-
tening and highly destructive. By undermining data confidentiality and integrity,
highly sophisticated and stealthy attacks will erode trust in organizations and
perhaps result in systemic failures. Consider a medical expert or physician giving
a diagnosis based on tampered medical data or an oil rig drilling for crude in the
wrong spot based on inaccurate geo-prospecting information. The potential of AI
to learn and adapt has ushered in a new era of scalable, highly targeted, and
human-like attacks. A smart and hostile offensive AI-driven attack will be able to
adapt as it learns from its surroundings, allowing it to easily infect systems with
little possibility of detection (Dixon and Eagan 2019). An AI-driven attack such as
PassGAN is capable of generating a large number of efficient password guesses
bypassing existing cybersecurity authentication infrastructures and causing greater
damages without being noticed (John and Philip 2018).
Discussion
As discussed in this study, malicious actors are beginning to utilize AI-advanced
data mining capabilities to execute more informed decisions. Learning from
contextual data will specifically mimic trusted features of cyberspace or target
weak points it discovers. This will enable AI-driven cyberattacks to avoid detection
and maximize the damage they inflict on cyberspace. AI-driven attacks will be able
to evolve as they learn from their surroundings, allowing them to effortlessly
compromise systems with little possibility of detection. In general, it’s apparent
that AI-driven cyberattacks will only worsen, then it will be almost impossible for
traditional cybersecurity tools to detect them It’s simply a question of machine
efficiency versus human labor. AI-driven threats will harness a multitude of
cyberspace and computer resources well beyond what a human could enlist,
resulting in an attack that is faster, more unpredictable, more sophisticated than
even the strongest cybersecurity team can respond against. However, by using AI
to fight AI, cybersecurity researchers, organizations, cybersecurity experts, and
Government institutions can begin to prepare more advanced and sophisticated
countermeasures to combat AI-driven attacks. The best method to prepare for this
right now is to harden cybersecurity defense infrastructures to the best of their
capacity, with the lowest possible number of false positives and negatives.
Conclusion
Cybercriminals are constantly changing and improving their attack efficiency,
emphasizing the use of AI-driven techniques in the attack process. This study
investigates the offensive capabilities of AI, allowing attackers to initiate
attacks on a larger scale, with a broader scope, and at a faster pace. This
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2405
study reviewed existing literature on AI-driven cyberattacks, the improper use
of AI in cyberspace, and the negative impact of AI-driven cyberattacks. The
findings show that 56% of the AI-Driven cyberattack techniques identified
were demonstrated in the access and penetration stage of the modified cyber-
security kill chain, 12% in the exploitation and C2 stage, 11% in the recon-
naissance, and 9% in the delivery stage. CNN has the most appearances (five)
among the AI techniques used by the selected authors to demonstrate access
and penetration attacks. This study determined the status of existing AI-driven
cyberattack research because 63% of current studies were based on implemen-
tation and evaluation, 25% on the proposed framework, and 12% on imple-
menting AI techniques to execute AI-driven attacks. The findings show that
traditional cybersecurity techniques’ inability to detect and mitigate AI-driven
attacks is directly related to their inability to cope with the speed, complex
decision logic, and multiple variant nature of AI-driven attacks. With the
emergence of these sophisticated attacks, organizations and security teams
must quickly reform their strategies, be prepared to defend their digital assets
with AI, and regain the advantage over this new wave of sophisticated attacks.
Finally, this study recommends that it is essential for the security research
community, government, and cybersecurity experts to prepare and invest in
advanced and sophisticated countermeasures to combat AI-driven cyberat-
tacks and utilize AI to fight offensive AI. A trustworthy AI framework will be
developed in the future to combat AI-Driven attacks while explaining essential
features that influence the detection logic.
Disclosure Statement
No potential conflict of interest was reported by the author(s)
ORCID
Sanjay Misra http://orcid.org/0000-0002-3556-9331
Luis Fernandez-Sanz http://orcid.org/0000-0003-0778-0073
Vera Pospelova http://orcid.org/0000-0001-5801-1923
References
DarkTrace. 2021. The Next Paradigm Shift AI-Driven Cyber-Attacks. DarkTrace Research
White Paper. https://www.oixio.ee/sites/default/files/the_next_paradigm_shift_-_ai_dri
ven_cyber_attacks.pdf (accessed June 9, 2021).
Anderson, H. S., J. Woodbridge, and B. Filar. 2016. Deepdga: Adversarially-tuned domain
generation and detection. In Proceedings of the ACM Workshop on Artificial Intelligence
and Security, Vienna, Austria, 13–21.
Babuta, A., M. Oswald, and A. Janjeva. 2020. Artificial Intelligence and UK National Security
Policy Considerations. Royal United Services Institute Occasional Paper.
e2037254-2406 B. GUEMBE ET AL.
Bahnsen, A. C., I. Torroledo, L. Camacho, and S. Villegas. 2018. DeepPhish: Simulating
malicious AI. In APWG Symposium on Electronic Crime Research, London, United
Kingdom, 1–8.
Bilal, M., A. Gani, M. Lali, M. Marjani, and N. Malik. 2019. Social profiling: A review,
taxonomy, and challenges. Cyberpsychology, Behavior and Social Networking 22
(7):433–50. doi:10.1089/cyber.2018.0670.
Bocetta, S. 2020. Has an AI cyberattack happened yet? https://www.infoq.com/articles/ai-
cyberattacks/ (accessed December 9,2020).
Brundage, M., S. Avin, J. Clark, H. Toner, P. Eckersley, B. Garfinkel, A. Dafoe, P. Scharre, T.
Zeitzoff, B. Filar, et al. 2018. The malicious use of artificial intelligence: forecasting, preven-
tion, and mitigation. Oxford: Future of Humanity Institute.
Bursztein, E., J. Aigrain, A. Moscicki, and J. C. Mitchell. 2014. The end is nigh: generic solving
of text-based CAPTCHAs. 8th Usenix workshop on Offensive Technologies WOOT ‘14, San
Diego, CA, USA.
Cabaj, K., Z. Kotulski, B. Księżopolski, and W. Mazurczyk. 2018. Cybersecurity: trends, issues,
and challenges. EURASIP Journal On Information Security. doi:10.1186/s13635-018-0080-0.
Cani, A., M. Gaudesi, E. Sanchez, G. Squillero, and A. Tonda (2014). Towards automated
malware creation. Proceedings of The 29Th Annual ACM Symposium On Applied
Computing, Gyeongju Republic of Korea, 157–60. doi: 10.1145/2554850.2555157.
Chakkaravarthy, S. S, D. Sangeetha, V. M. Rathnam, K. Srinithi, and V. Vaidehi. 2018.
Futuristic cyber-attacks. International Journal of Knowledge-Based and Intelligent
Engineering Systems 22 (3):195–204. doi: 10.3233/kes-180384.
Chen, J., X. Luo, J. Hu, D. Ye, and D. Gong. 2018. An Attack on Hollow CAPTCHA Using
Accurate Filling and Nonredundant Merging. IETE Technical Review, 35(sup1):106–118.
doi:10.1080/02564602.2018.1520152.
Chung, K., Z. T. Kalbarczyk, and R. K. Iyer. 2019. Availability attacks on computing
systems through alteration of environmental control: Smart malware approach.
Proceedings of the 10th ACM/IEEE International Conference on Cyber-Physical
Systems, Montreal Quebec, Canada, 1–12.
Dheap, V. 2017. AI in cybersecurity: A balancing force or a disruptor? https://www.rsaconfer
ence.com/industry-topics/presentation/ai-in-cybersecurity-a-balancing-force-or-a-disrup
tor (accessed February 13, 2020).
Dixon, W., and N. Eagan. 2019. AI will power a new set of tools and threats for the
cybercriminals of the future. https://www.weforum.org/agenda/2019/06/ai-is-powering-
a-new-generation-of-cyberattack-its-also-our-best-defence/ (accessed December 3, 2020).
Fischer, E. 2016. Cybersecurity issues and challenges: In brief. . https://fas.org/sgp/crs/misc/
R43831.pdf (accessed December 1, 2020).
Gao, H., M. Tang, Y. Liu, P. Zhang, and X. Liu. 2017. Research on the security of microsoft’s
two-layer captcha. IEEE Transactions On Information Forensics And Security 12 (7):1671–85.
doi:10.1109/tifs.2017.2682704.
Hamadah, S., and D. Aqel. 2020. Cybersecurity becomes smart using artificial intelligent and
machine learning approaches: An overview. ICIC Express Letters, Part B: Applications 11
(12):1115–1123. doi:10.24507/icicelb.11.12.1115.
Hitaj, B., P. Gasti, G. Ateniese, and F. Perez-Cruz. 2019. PassGAN: A deep learning approach
for password guessing. Applied Cryptography and Network Security 11464:217–37.
doi:10.1007/978-3-030-21568-2_11.
Hu, W., and Y. Tan. 2021. Generating adversarial malware examples for black-box attacks
based on GAN.https://arxiv.org/abs/1702.05983 (accessed August 12, 2021).
John, S., and T. Philip. 2018. Generative models for spear phishing posts on social media. NIPS
Workshop On Machine Deception, California, USA. arXiv:1802.05196
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2407
Kaloudi, N., and J. Li. 2020. The AI-based cyber threat landscape. ACM Computing Surveys 53
(1):1–34. doi:10.1145/3372823.
Kirat, D., J. Jang, and M. Stoecklin. 2018. DeepLocker concealing targeted attacks with AI
locksmithing. https://www.blackhat.com/us-18/briefings/schedule/index.html#dee
plocker—concealing-targeted-attacks-with-ailocksmithing-11549 (accessed December 4,
2020).
Lee, K., and K. Yim. 2020. Cybersecurity threats based on machine learning-based offensive
technique for password authentication. Applied Sciences 10 (4):1286. doi:10.3390/
app10041286.
Li, C., X. Chen, H. Wang, P. Wang, Y. Zhang, and W. Wang. 2021. End-to-end attack on
text-based CAPTCHAs based on cycle-consistent generative adversarial network.
Neurocomputing 433:223–36. doi:10.1016/j.neucom.2020.11.057.
Meng, G., Y. Xue, C. Mahinthan, A. Narayanan, Y. Liu, J. Zhang, and T. Chen. 2016. Mystique.
Proceedings of the 11Th ACM On Asia Conference On Computer and Communications
Security, Xi'an, China, 365–76. doi:10.1145/2897845.2897856.
Moher, D., A.Liberati, J.Tetzlaff, and D. G.Altman. 2010. Preferred reporting items for sys-
tematic reviews and meta-analyses: The PRISMA statement. International Journal of
Surgery 8(5): 336–341. doi:10.1016/j.ijsu.2010.02.007.
Ney, P., K. Koscher, L. Organick, L. Ceze, and T. Kohno. 2017. Computer security, privacy, and
dna sequencing: compromising computers with synthesized DNA, privacy leaks, and more.
USENIX Security Symposium, Vancouver, BC, Canada, 765–79.
Noury, Z., and M. Rezaei. 2020. Deep-CAPTCHA: A deep learning based CAPTCHA solver for
vulnerability assessment. ArXiv, abs/2006.08296.
Petro, D., and B. Morris. 2017. Weaponizing machine learning: Humanity was overrated
anyway. DEF CON.
Rigaki, M., and S. Garcia. 2018. Bringing a GAN to a knife-fight: adapting malware commu-
nication to avoid detection. IEEE Security and Privacy Workshops (SPW), San Francisco,
CA, USA. doi:10.1109/spw.2018.00019.
Seymour, J., and P. Tully. 2016. Weaponizing data science for social engineering: Automated
E2E spear phishing on Twitter. https://www.blackhat.com/docs/us-16/materials/us-16-
Seymour-Tully-Weaponizing-Data-Science-For-Social-EngineeringAutomated-E2E-Spear-
Phishing-On-Twitter-wp.pdf (accessed December 21, 2020).
Sood, A., S. Zeadally, and R. Bansal. 2017. Cybercrime at a Scale: A Practical Study of
Deployments of HTTP-Based Botnet Command and Control Panels. IEEE
Communications Magazine 55 (7):22–28. doi:10.1109/mcom.2017.1600969.
Tang, M., H. Gao, J. Yan, F. Cao, ZhangZ., L. Lei, M. Zhang, P. Zhou, X. Wang, X. Li, and L. X.
Jiawei. 2016. A simple generic attack on text captchas. Proceedings 2016 Network And
Distributed System Security Symposium, San Diego, California. doi:10.14722/
ndss.2016.23154.
Thanh, C., and I. Zelinka. 2019. A survey on artificial intelligence in malware as
next-generation threats. MENDEL 25 (2):27–34. doi:10.13164/mendel.2019.2.027.
Trieu, K., and Y. Yang. 2018. Artificial intelligence-based password brute force attacks.
Proceedings of Midwest Association for Information Systems Conference, St. Louis,
Missouri, USA, 13(39).
Truong, T., I. Zelinka, J. Plucar, M. Čandík, and V. Šulc. 2020. Artificial intelligence and
cybersecurity: past, presence, and future. Advances In Intelligent Systems And Computing
351–63. doi:10.1007/978-981-15-0199-9_30.
Usman, M., M. Jan, X. He, and J. Chen. 2020. A survey on representation learning efforts in
cybersecurity domain. ACM Computing Surveys 52 (6):1–28. doi:10.1145/3331174.
e2037254-2408 B. GUEMBE ET AL.
Xu, W., D. Evans, and Y. Qi. 2018. Feature squeezing: Detecting adversarial examples in deep
neural networks. Proceedings 2018 Network and Distributed System Security Symposium,
San Diego, California, USA. doi:10.14722/ndss.2018.23198.
Yao, Y., B. Viswanath, J. Cryan, H. Zheng, and B. Zhao. 2017. Automated crowdturfing
attacks and defenses in online review systems. Proceedings Of The 2017 ACM SIGSAC
Conference On Computer And Communications Security, Dallas Texas, USA.
doi:10.1145/3133956.3133990.
Ye, G., Z. Tang, D. Fang, Z. Zhu, Y. Feng, P. Xu, X. Chen, and Z. Wang. 2018. Yet another text
captcha solver. Proceedings of The 2018 ACM SIGSAC Conference On Computer And
Communications Security, Toronto, Canada. doi:10.1145/3243734.3243754.
Yu, N., and K. Darling. 2019. A low-cost approach to crack python CAPTCHAs using
AI-based chosen-plaintext attack. Applied Sciences 9 (10):2010. doi:10.3390/
app9102010.
Zhou, X., M. Xu, Y. Wu, and N. Zheng. 2021. Deep model poisoning attack on federated
learning. Future Internet 13 (3):73. doi:10.3390/fi13030073.
Zouave, E., M. Bruce, K. Colde, M. Jaitnee, I. Rodhe, and T. Gustafsson. 2020. Artificially
intelligent cyberattacks.https://www.statsvet.uu.se/digitalAssets/769/c_769530-l_3-k_rap
port-foi-vt20.pdf (accessed December, 21 2020).
APPLIED ARTIFICIAL INTELLIGENCE e2037254-2409
Content uploaded by Sanjay Misra
Author content
All content in this area was uploaded by Sanjay Misra on Mar 28, 2022
Content may be subject to copyright.