Data Analytics for Cybersecurity

As the world becomes increasingly connected, it is also more exposed to a myriad of

cyber threats. We need to use multiple types of tools and techniques to learn and
understand the evolving threat landscape. Data are a common thread linking various
types of devices and end users. Analyzing data across different segments of
cybersecurity domains, particularly data generated during cyber–attacks, can help us
understand threats better, prevent future cyber–attacks, and provide insights into the
evolving cyber threat landscape. This book takes a data oriented approach to studying
cyber threats, showing in depth how traditional methods such as anomaly detection can
be extended using data analytics, and also applies data analytics to non–traditional
views of cybersecurity, such as multi domain analysis, time series and spatial data
analysis, and human-centered cybersecurity.

v a n d a n a p . j a n e j a is Professor and Chair of the Information Systems department

at the University of Maryland, Baltimore County. Most recently, she also served as an
expert at the National Science Foundation supporting data science activities in the
Directorate for Computer and Information Science and Engineering (CISE)
(2018–2021). Her research interests include discovering knowledge in presence of data
heterogeneity. Her research projects include anomaly detection in network
communication data, human behavior analytics in heterogeneous device environments,
geospatial context for IP reputation scoring, spatiotemporal analysis across
heterogeneous data, and ethical thinking in data science. She has been funded through
state, federal, and private organizations.

Published online by Cambridge University Press

Published online by Cambridge University Press
 Data Analytics for Cybersecurity

VANDANA P. JANEJA
University of Maryland, Baltimore County (UMBC)

Published online by Cambridge University Press

 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
103 Penang Road, #05-06/07, Visioncrest Commercial, Singapore 238467

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781108415279
DOI: 10.1017/9781108231954
© Vandana P. Janeja 2022
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2022
A catalogue record for this publication is available from the British Library.
ISBN 978-1-108-41527-9 Hardback

Cambridge University Press has no responsibility for the persistence or accuracy

of URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.

Published online by Cambridge University Press

 Contents

Preface page ix
Acknowledgment xii

1 Introduction: Data Analytics for Cybersecurity 1

1.1 What Is Cybersecurity? 1
1.1.1 Assets Affected 2
1.1.2 Motivation, Risks, and Attaining Security 4
1.2 Handling Cyberattacks 6
1.2.1 Subareas of Cybersecurity 6
1.3 Data Analytics 8
1.3.1 Why Is Data Analytics Important
for Cybersecurity? A Case Study of Understanding
the Anatomy of an Attack 9
1.3.2 How Can Data Analytics Help? 12
2 Understanding Sources of Cybersecurity Data 14
2.1 End-to-End Opportunities for Data Collection 14
2.2 Sources of Cybersecurity Data 16
2.2.1 Log Data 17
2.2.2 Raw Payload Data 20
2.2.3 Network Topology Data 21
2.2.4 User System Data 23
2.2.5 Other Datasets 24
2.3 Integrated Use of Multiple Datasets 26
2.4 Summary of Sources of Cybersecurity Data 27

Published online by Cambridge University Press

 vi Contents

3 Introduction to Data Mining: Clustering, Classification,

and Association Rule Mining 29
3.1 Knowledge Discovery and Data Mining Process Models 29
3.2 Data Preprocessing 31
3.2.1 Data Cleaning 33
3.2.2 Data Transformation and Integration 34
3.2.3 Data Reduction 35
3.3 Data Mining 36
3.3.1 Measures of Similarity 37
3.3.2 Measures of Evaluation 40
3.3.3 Clustering Algorithms 42
3.3.4 Classification 49
3.3.5 Pattern Mining: Association Rule Mining 55
4 Big Data Analytics and Its Need
for Cybersecurity: Advanced DM and Complex Data Types
from Cybersecurity Perspective 60
4.1 What Is Big Data? 60
4.2 Big Data in Cybersecurity 61
4.3 Landscape of Big Data Technologies 63
4.4 Mahout and Spark Comparative Analysis 64
4.5 Complex Nature of Data 66
4.5.1 Nature of Data: Spatial Data 67
4.5.2 Nature of Data: Graph Data 69
4.5.3 Nature of Data: Other Types of Complex Data 69
4.6 Where Does Analytics Fit in for Cybersecurity? 70
4.6.1 Change Detection in Massive Traffic Datasets 70
4.6.2 Multipronged Attacks 73
4.7 Privacy and Security Issues in Big Data 75
5 Types of Cyberattacks 78
5.1 Types of Attacks 78
5.2 Example: Social Engineering 80
5.2.1 Computational Data Model for Social
Engineering 81
5.3 Example: Advanced Persistent Threat 84
5.3.1 Attributes of APTs 85
5.3.2 Data Analytics Methods for Persistent Threats 88
6 Anomaly Detection for Cybersecurity 91
6.1 What Are Anomalies? 91
6.2 Context 93

Published online by Cambridge University Press

 Contents vii

6.3 Motivating Example: BGP Hijacking 95

6.4 Challenges in Understanding Anomalies 96
6.5 Interpretation 98
6.6 Treating Anomalies 99
7 Anomaly Detection Methods 101
7.1 Statistical Outlier Detection Tests 102
7.1.1 Discordancy Tests 103
7.1.2 IQR-Based Outlier Detection 104
7.2 Density-Based Outlier Detection:
OPTICS-OF Identifying Local
Outliers 105
7.3 Distance-Based Outlier Detection 108
7.4 Outlier Detection through Clustering 110
8 Cybersecurity through Time Series and
Spatial Data 112
8.1 Spatial and Temporal Data 112
8.1.1 Spatial Data 112
8.1.2 Spatial Autocorrelation 113
8.1.3 Spatial Heterogeneity 113
8.1.4 Temporal Data 115
8.1.5 Temporal Autocorrelation 115
8.1.6 Temporal Heterogeneity 116
8.2 Some Key Methods for Anomaly Detection in Spatial
and Temporal Data 116
8.2.1 Spatial Anomaly Detection 117
8.2.2 Spatial Outlier Detection 117
8.2.3 Spatial Neighborhood 119
8.2.4 Temporal Anomaly Detection 121
8.2.5 Temporal Neighborhoods 122
8.3 Cybersecurity through Spatiotemporal
Analysis 123
8.3.1 Spatial Anomalies in Cybersecurity 123
8.3.2 Neighborhood-Based Anomalies 124
8.3.3 Spatio temporally Clustered
Anomalies 125
8.3.4 Insights through Geo-Origination
of IP Addresses 126
8.4 Temporal Behaviors in Evolving Networks 127

Published online by Cambridge University Press

 viii Contents

9 Cybersecurity through Network and Graph Data 128

9.1 Graph Properties 130
9.2 Graphs for Cybersecurity: Understanding Evolving
Network Communication 133
9.3 Graphs for Cybersecurity: Similarity-Based
Redundancy in Router Connectivity 134
10 Human-Centered Data Analytics for Cybersecurity 137
10.1 Human Perspective to Cybersecurity: Phishing 138
10.2 Human Perspective to Cybersecurity: Insider Threat 140
10.3 User/Employee Viewpoint 142
10.4 Attacker Viewpoint: Anomaly Detection Methods 144
11 Future Directions in Data Analytics for Cybersecurity 147
11.1 Data Analytics in Cyberphysical Systems 147
11.1.1 Cyberphysical Systems 147
11.1.2 Internet-of-Things (IoT) 149
11.2 Multidomain Mining 151
11.2.1 Integrating Multiple Heterogeneous Data 151
11.2.2 Integrated Alerts from Multiple Sources 154
11.3 Advanced Machine Learning Models 157
11.3.1 Deep Learning 157
11.3.2 Generative Adversarial Networks 159
11.3.3 Model Reuse 161
11.4 Ethical Thinking in the Data Analytics Process 163

References 165
Index 189

Published online by Cambridge University Press

 Preface

Cybersecurity is a pervasive need in our connected world to counter threats

affecting individuals, organizations, and governments. The acceptance and
adoption of technology on multiple types of nontraditional devices force
cybersecurity solutions to address challenges emerging in the areas of not only
computer networks but also sensor networks, industrial control systems, and
user devices. Data are the common thread across all these types of devices and
end users, especially data generated during cyberattacks. Increasingly, the
focus of cybersecurity is shifting to analyzing data in not only a retrospective
manner but also a prospective manner across different cybersecurity domains.
This data-driven understanding of attacks can potentially prevent future cyber-
attacks and provide insights into the evolving cyber threats.
Data analytics pushes beyond the traditional themes of security to seam-
lessly weave the analysis of threats across several applications. This book
applies data analytics concepts and techniques to the domain of cybersecurity,
discusses methods to evaluate data sources in cyberattacks, and provides
insights into data mining methods that can be utilized for cybersecurity.
Finally, this book also looks at nontraditional views of data analytics for
cybersecurity in time series and spatial data, discussing the need and applica-
tion of big data analytics and offering a human-centered analytics perspective
to cybersecurity.
Although there are several books on network security, information assur-
ance, and forensics and other books on data mining, there is a need to address
both cybersecurity and data analytics in a synergistic manner. In addition,
I have developed a graduate and undergraduate course on data analytics for
cybersecurity. I taught this course for several semesters without a textbook.
While developing this course, I reviewed several very good books. However,
there was a gap in the material available for the data analytics perspective of
cybersecurity in depth. This motivated me to write a book that looks at the

ix

https://doi.org/10.1017/9781108231954.001 Published online by Cambridge University Press

 x Preface

different domains affected by cyberattacks, spanning computer networks,

industrial control systems, sensor networks, drones, and other connected
devices. Data analytics provides a window of learning into such systems by
looking at the massive amounts of data being generated that may go untapped.
For instance, this book addresses the human-centered perspective to cyberat-
tacks, and it addresses the multiple facets of data analytics for cybersecurity,
such as anomaly detection across spatial and temporal data.
I expect the book will be primarily useful for teaching graduate and under-
graduate cybersecurity courses that take a data analytics perspective. It should
also be relevant for the industry and government as it discusses the potential
avenues of understanding and discovering cyberattacks and additional know-
ledge about them to better inform future decision-making.
The need for data analytics is also evident from Information Assurance
requirements such as those stated in federal frameworks1,2 and security direct-
ives for Information Assurance Training, Certification, and Workforce
Management,3 where various technical cybersecurity positions and their func-
tions are outlined in detail. For instance, the Computer Network Defense
(CND) Service Provider Incident Responder (IR) position has functions such
as (CND-IR.2) – collect, analyze, intrusion artifacts – and (CND-IR.7) –
correlate incident data and perform CND trend analysis. Similarly, the CND-
Analyst (A) position has functions such as (CND-A.4) – perform analysis of
log files – (CND-A.5) – characterize and analyze network traffic to identify
anomalous activity and potential threats to network resources – and (CND-
A.8.) – perform event correlation. These and many other security functions,
such as forensics, threat hunting, and such clearly indicate the need to incorp-
orate an analytics perspective to cybersecurity.

What Does the Book Cover?

The book spans from introductory concepts of cybersecurity, foundations of
data analytics, and applications of data analytics concepts to cybersecurity
applications.

1
Workforce Framework for Cybersecurity(NICE Framework), NIST Special Publication 800-
181 Revision 1, https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-181r1.pdf,
Last Accessed May 2021
2
Workforce Framework for Cybersecurity (NICE Framework), https://niccs.cisa.gov/
workforce-development/cyber-security-workforce-framework , Last Accessed May 2021
3
Information Assurance Workforce Improvement Program, DoD 8570.01-M www.esd.whs.mil/
Portals/54/Documents/DD/issuances/dodm/857001m.pdf; 8140.01 reissues and renumbers
DoDD 8570.01, Last Accessed May 2021

https://doi.org/10.1017/9781108231954.001 Published online by Cambridge University Press

 Preface xi

Chapter 1 introduces the basic concepts of cybersecurity and the data

analytics perspective to cybersecurity. It lays out the areas of study and how
data analytics should be a key part of the spectrum of cybersecurity solutions.
Chapter 2 focuses on understanding sources of cybersecurity data and the
end-to-end opportunities for data collection. It goes onto discuss the sources of
cybersecurity data and how multiple datasets can be leveraged in understand-
ing cyber threats.
Chapter 3 gets into the techniques of data analytics focusing on the three
pillars of data mining, namely clustering, classification, and association rule
mining, and how each can be used for cybersecurity. This chapter can be seen
as a crash course in data mining. It begins with an understanding of the overall
knowledge discovery and data mining process models and follows the elem-
ents of the data life cycle. This chapter outlines foundational elements such as
measures of similarity and of evaluation. It outlines the landscape of various
algorithms in clustering and classification and frequent and rare patterns.
Chapter 4 focuses on the big data elements of cybersecurity, looking at the
landscape of the big data technologies and the complexities of the different
types of data, including spatial and graph data. It outlines examples in these
complex data types and how they can be evaluated using data analytics.
Chapter 5 highlights the various types of cyberattacks and how data analytics
methods can potentially be used to analyze these attacks.
Chapter 6 and 7 holistically focus on anomaly detection. Chapter 6 focuses
on what anomalies are and more specifically what anomalies are in the
cybersecurity domain, and what some of the features of anomalies are.
Chapter 7, on the other hand, focuses on techniques of detecting anomalies
starting with some of the basic statistical techniques, going into data
analytics techniques.
While Chapter 4 introduces the complex types of data, Chapter 8 delves into
the specifics of spatial and temporal analytics with topics such as spatial
neighborhood and temporal evolution of large amounts of network traffic data.
Chapter 9 extends the ideas of complex data by looking into cybersecurity
through network and graph data. Chapter 10 brings in the human-centered data
analytics perspective to cybersecurity. Finally, Chapter 11 discusses several
key directions, such as data analytics in cyberphysical systems, multidomain
mining, machine Learning concepts such as deep Learning, generative adver-
sarial networks, and challenges of model reuse. Last but not the least, the
chapter closes with thoughts on ethical thinking in the data analytics process.

https://doi.org/10.1017/9781108231954.001 Published online by Cambridge University Press

 Acknowledgments

The views expressed in this book are my own and do not reflect those of the
organizations I am affiliated with.
I want to thank Carole Sargent, who planted the seed for this book and
helped me learn the ropes of the process. Many thanks to my editor at
Cambridge University Press, Lauren Cowles, and the Cambridge University
Press team, Amy He, Adam Kratoska, Johnathan Fuentes, Rebecca Grainger,
Andy Saff, and Neena S. Maheen, who were patient with the many deadlines
and generous in their time reviewing the book and its many facets. Many
thanks to the reviewers of this book who helped improve it with their
detailed feedback.
The work in this book is built on the foundation of my work with many of
my students, and the many conversations with them have inspired the writing
of this book. I would like to thank my PhD advisees, Josephine Namayanja,
Chhaya Kulkarni, Sara Khanjani, Ira Winkler, Faisal Quader, Mohammad
Alodadi, Ali Azari , Yanan Sun, Lei Shi, and Mike McGuire; my master’s
advisees, Henanksha Sainani, Prathamesh Walkikar, Suraksha Shukla,
Vasundhara Misal, Anuja Kench, Akshay Grover, Jay Gholap, Javad Zabihi,
Kundala Das, Prerna Mohod, Abdulrahman Alothaim, and Revathi
Palanisamy; my graduate students who worked on research projects with me,
Yuanyuan Feng, Ruman Tambe, Song Chen, Sandipan Dey, Tania Lobo,
Monish Advani, Ahmed Aleroud, Abu Zaher Md Faridee, Sai Pallaprolu,
and Justin Stauffer; and my undergraduate research students, Gabrielle
Watson, Olisaemeka Okolo, Aminat Alabi, Mackenzie Harwood, Adrian
Reyes, Jarrett Early, Alfonso Delayney, David Lewis, and Brian Lewis.
My thought process and contributions are also framed by work with
advisors, collaborators, colleagues, and industry affiliates who have contrib-
uted to my research and thinking in the area of security and data analytics,
including Lucy Erickson, Carolyn Seaman, Aryya Gangopadhyay, Anupam

xii

Published online by Cambridge University Press

 Acknowledgments xiii

Joshi, Susan Sterrett, Krishnan Chellakarai, Yatish Joshi, and Raghu

Chintalapati. I also want to thank my advisors and mentors: Vijay Atluri,
Nabil Adam, Jaideep Vaidya, Erwin Gianchandani, Keith Bowman, Philip
Rous, Patrice McDermott, Erin Lavik, and Chaitan Baru. This work is also
built through the support of Dr. Freeman A. Hrabowski III, who has always
been generous with his time as my mentor.
Last but not the least, my heartfelt thanks go to my biggest allies and pillars
of support: Neelu, Suhasini, Harish, Deiptii, Nisha, Viraj, and Mukul, and my
single most avid cheerleader, Vihan, who kept encouraging me with timeline
prompts to complete this book.

Published online by Cambridge University Press

Published online by Cambridge University Press
 1
Introduction
Data Analytics for Cybersecurity

1.1 What Is Cybersecurity?

Cybersecurity refers to securing valuable electronic assets and physical assets,
which have electronic access, against unauthorized access. These assets may
include personal devices, networked devices, information assets, and infra-
structural assets, among others.
Cybersecurity deals with security against threats also referred to as cyber
threats or cyberattacks. Cyberattacks are the mechanism by which security is
breached to gain access to assets of value.
The first aim of cybersecurity is prevention of cyberattacks against critical
assets. The second aim of cybersecurity is detection of threats. The third aim is
to respond to threats in the event that they penetrate access to critical assets,
and, finally, the fourth aim is to recover and restore the normal state of the
system in the event that an attack is successful. Cybersecurity is achieved by
addressing each of these three aspects to prevent, detect, and respond to threats
against critical assets. Essentially, it deals with securing everything that is in
cyberspace so these assets of value are not tampered with.
What really are these assets? Are they the data on a hard drive, a contact
list on a cell phone, an Excel sheet with sales numbers, or a program to switch
on a device?
It is all of these and more. With the ever-increasing use of electronics for
every possible function of life, cybersecurity is becoming a pervasive problem
permeating every individual’s life. If you look around yourself, there are
several electronic devices that you may see or use, and each electronic device
may in turn be connected to another device physically or virtually. This
connectivity enables us to access additional functionalities on the device.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 2 Introduction: Data Analytics for Cybersecurity

However, this also makes the device vulnerable since it has external connect-
ivity and access is fragmented through multiple external applications.
Now the obvious question is, does cybersecurity apply to connected devices
only? Although a big part of cybersecurity is a result of the high level of
connectivity, it also includes threats resulting from compromised physical
security. For instance, a highly secure desktop in a secure room, which is
not accessible via the traditional internet but only through a biometric authen-
tication or a card swipe, is also at risk. An unauthorized access to the secure
room poses a cybersecurity threat due to the electronic assets at risk and
potential risk to the systems providing access to the room. This is primarily
because some form of connectivity is always possible in this highly connected
world. Another example is that of an insider threat where an authorized user
accesses materials with a malicious intent.

1.1.1 Assets Affected

Cybersecurity is a major challenge due to the potential of damage. An empty
hard drive that is stolen is a theft, but a hard drive with data accessed in an
unauthorized manner poses a much bigger threat due to the value of the
information stored on it. A sensor controlling a chemical flow into a vat that
breaks down accidentally may lead to a hazmat accident, but an authorized
access leading to tampering of the program controlling the sensor is a cyber-
security risk. This is because the intent and extent of the tampering are
uncontrolled and may be much more catastrophic than a sensor breakdown.
Thus, cybersecurity aims to prevent unauthorized access to electronic assets
and physical assets with electronic access.
Let us look at some types of assets – personal, public, and corporate – that
can be impacted by cybersecurity risks, summarized in Figure 1.1. We are now
referring to the electronic access to physical assets or electronic assets. This list
is a small sample of such assets in the highly connected world that we live in.
These include the following.
• Personal assets (mostly used for personal needs): Phones (home and
mobile), tablets, personal computers (desktop and laptops), external physical
hard drive, cloud drive, email accounts, fitness trackers, smart watches,
smart glasses, media devices (TiVo, Apple TV, cable box), bank accounts,
credit cards, personal gaming systems, blogs, vlogs, photos, and videos.
• Public assets (for managing public utilities and services): Smart meters, power
grid, sewage controls, nuclear power plant, rail lines, air traffic, traffic lights,
citizen databases, websites (county, state and federal), space travel programs.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 1.1 What Is Cybersecurity? 3

Figure 1.1 Sample assets at risk due to cyberattacks.

• Corporate assets (for managing business needs): Customer database,

websites, business applications, business network, emails, software,
intellectual property.
Some overlap may occur as corporate assets may link to public assets or
personal assets may be connected to corporate assets or public assets. As such,
these assets should be clearly separated through well-defined security policies.
However, with the level of connectivity through multiple devices, it is becom-
ing more and more difficult to keep assets completely discrete
and disconnected.
Thus, cybersecurity is securing everything of value that exists in cyberspace,
which includes computer networks, mobile devices, cloud data storage, sensor
networks, industrial control systems, emergency devices, railway lines, and air
traffic controls, to name a few. It is not simply limited to our own private data
or banking information but goes beyond to critical infrastructure and corporate
assets.
Certain geological regions are more prone to attacks than others due to the
availability and richness of such assets. Many tools are available to depict
attack maps of popular cyberattacks. Examples include ThreatCloud showing
real-time attacks based on monitoring sensors across the world (Check Point

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 4 Introduction: Data Analytics for Cybersecurity

2020), Digital Attack Map depicting top distributed denial of service (DDoS)
attacks worldwide (Netscout 2020), and Kaspersky Lab’s 3D interactive map
of cyberattacks (Kaspersky 2020). These are helpful in visualizing the spread
of types of attacks and level of activity in a region.

1.1.2 Motivation, Risks, and Attaining Security

What is the motivation behind such attacks? An unauthorized access to secured
resources may be due to various factors. Some such factors are summarized in
Figure 1.2 and listed as follows:
• Stealing intellectual property: The majority of corporate product develop-
ment strategies are stored on highly secure infrastructure. Small companies
and large corporations alike devise novel products, and their intellectual
property holds the key to revenue in the future. Loss of such information
could be devastating for the sales and profits for an organization.
• Gaining access to customer data: Banks and retailers are constantly at the
receiving end of cyberattacks due to valuable customer data, which can then
be sold. Recent notable examples include Target (Washington Post 2013).

Risks
Motivation
lnternet protocol that is inherently not
To steal Intellectual property
secure
To damage reputation
Applications with several dependencies
To gain access to data, which can
Logical errors in software code (ex.
then be sold
Heartbleed)
To gain access to information, which
Organizational risks (multiple partners ex.
is not generally available
Target, PNNL)
To make a political statement
Lack of user awareness of cybersecurity risks
To impede access to critical data and
(ex. social engineering, phishing)
applications
Personality traits of individuals using the
To make a splash/ for fun
systems

Attaining security

Protecting resources
Hardening defenses
Capturing data logs
Monitoring systems
Tracing the attacks
Predicting risks
Predicting attacks
Identifying vulnerabilities

Figure 1.2 Unauthorized access motivations, risks and attaining security.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 1.1 What Is Cybersecurity? 5

• Making a political statement: Hacktivism, or hacking for activism (Samuel

2004), is an act carried out primarily to send a political message. An example
included the cyber war in Estonia in 2007 (CSMonitor–Estonia 2007), where
a series of DDoS attacks were directed toward official websites and even
mobile phone networks. Estonia has several e-government initiatives, which
had to be shut down for access from foreign locations due to the massive
scale of the attacks. The Estonian foreign minister has also been quoted
blaming the Russian government for facilitating these attacks.
• Performing cyber espionage: These attacks originate from government-
supported cells, which target state secrets. Unlike theft of intellectual prop-
erty, cyber espionage is geared toward gaining access to sensitive infor-
mation. An example is the Titan Rain attack (Lewis 2005, Chi 2014), a type
of advanced persistent threat (APT), believed to have been initiated by
government-coordinated cells in China targeting US defense networks and
contractors such as Lockheed Martin, Sandia National Laboratories, and
NASA, among others. The SolarWinds hack led to a massive impact for
federal systems believed to be directed by the Russian government (Stubbs
et al. 2020). In some cases, this becomes a global supply chain attack, where
the hack from a trusted vendor leads to a major attack, such as in the case of
FireEye, which was impacted by the SolarWinds attack (Geenans 2020)
• In addition to the aforementioned motivations, cyberattacks may be aimed at
damaging reputation (Townsend et al. 2014), which can lead to substantial
losses; making a splash for fun (Gesenhues 2014), such as a traditional
search example using Google dorking; and impeding access to data and
applications (Risk Based Security 2014). Most of these scenarios are carried
out to gain unauthorized access to information.

1.1.2.1 Why Do We Have Security Risks?

Security risks arise due to various factors, which include the following:
applications with several dependencies, logical errors in software code (such
as Heartbleed), organizational risks (multiple partners, such as in cyber-attacks
at Target and the Pacific Northwest National Laboratories [PNNL]), lack of
user awareness of cybersecurity risks (such as in social engineering and
phishing), personality traits of individuals using the systems (phishing), and
inherent issues in the Internet protocol being used.

1.1.2.2 What Is the Level of Damage That Can Occur?

According to a McAfee report, the monetary loss resulting from cybercrime
costs about $600 billion, which is about 0.8% of the world Gross Domestic
Product (GDP) (McAfee–Cybercrime Impact 2018), with malicious actors

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 6 Introduction: Data Analytics for Cybersecurity

becoming more and more sophisticated. Such large numbers are difficult to
estimate, and the report outlines a sound strategy for reaching the
estimated loss.
The loss due to cyberattacks is not simply based on direct financial loss but
also based on several indirect factors that may lead to a major financial impact.
As an example, let us consider the Target cyberattack. According to a Reuters
news article (Skariachan and Finkle 2014), Target reported $61 million in
expenses related to the cyberattack out of which $44 million was covered by
insurance. Thus, the direct financial impact to Target was $17 million. Now let
us consider some other factors that led to the indirect loss for Target: there was
a 46% drop in net profit in the holiday quarter and a 5.5% drop in transactions
during the quarter, share price fluctuations led to further losses, cards had to be
reissued to several customers, and Target had to offer identity protection to
affected customers. All these losses amount to much more than the total $61
million loss. In addition, the trust of the customers was lost, which is not a
quantifiable loss and has long-term impacts.

1.2 Handling Cyberattacks

Now the key question is how one secures the resources against unauthorized
accesses. Understanding and preventing such risks can be done in several
ways, including protecting resources, hardening defenses, capturing data logs,
monitoring systems, tracing the attacks, predicting risks, predicting attacks,
and identifying vulnerabilities.
Several subdisciplines of cybersecurity deal with these aspects.

1.2.1 Subareas of Cybersecurity

Various areas of cybersecurity have developed techniques to prevent and
respond to cyberattacks. There are several overlaps in these areas of cyberse-
curity. These include the following (shown in Figure 1.3):
Application security: This area deals with incorporating security in the
software development process. This includes following best practices
for code development, designing strong and comprehensive test cases,
and following rigorous maintenance practices.
Data and information security: Increasingly with the advent of cloud data
storage and remote access to data warehouses, data and information are
at a risk of unauthorized access and misuse. Data and information
security deals with securing data from such threats.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 1.2 Handling Cyberattacks 7

Figure 1.3 Areas of cybersecurity.

Network security: This deals with the challenges faced in securing the
traditional computer networks and security measures adopted to secure,
prevent unauthorized access and misuse of either the public or the
private network.
Cyberphysical security: This focuses on the emerging challenges due to the
coupling of the cyber systems with the physical systems. For example,
the power plants being controlled by a cyber system present new security
challenges that can arise due to the risk of disruption of the cyber
component or risk of unauthorized control of the cyber system, thus
gaining control of the physical systems.
Data analytics: This crosscutting theme can apply to each of these areas to
learn from existing threats and develop solutions for novel and unknown
threats toward networks, infrastructure, data, and information. Threat
hunting (Sadhwani 2020), which proactively looks for malicious players
across the myriad data sources in an organization, is a direct application
of using data analytics on the various types of security data produced
from the various types of security streams. However, this does not
necessarily have to be a completely machine-driven process and should
account for user behaviors as well (Shashanka et al. 2016), looking at the
operational context. Data analytics can provide security analysts a much
focused field of vision to zero in on solutions for potential threats.
Cybersecurity affects many different types of devices, networks, and organ-
izations. Each poses different types of challenges to secure. It is important to
understand and differentiate between these challenges due to the changing
hardware and software landscape across each type of cybersecurity domain.
While hardware configurations and types of connectivity are out of scope for
this book, it is important to understand some of the fundamental challenges to
study their impact and how they can be addressed using techniques such as

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 8 Introduction: Data Analytics for Cybersecurity

data analytics, which is crosscutting across the many different types of con-
nectivity since data are ubiquitous across all these domains. In the current
connected environment, multiple types of networks and devices are used, such
as computer networks, cyberphysical systems (CPS), Internet of Things (IoT),
sensor networks, smart grids, and wired or wireless networks. To some extent,
IoT and sensor networks can be seen as specialized cases of CPS systems. We
can consider these systems to study how data analytics can contribute to
cybersecurity since any of these types of systems will generate data, which
can be evaluated to understand their functioning and any potential benign or
nonbenign malfunctions.
Computer networks are the most traditional type of networks where groups
of computers are connected in prespecified configurations. These configur-
ations can be designed using security policy deciding who has access to what
areas of networks. Another way networks form is by determining patterns of
use over a period of time. In both cases, zones can be created for access and
connectivity where each computer in the network and subnetworks can
be monitored.
Cyberphysical systems are an amalgamation of two interacting subsystems,
cyber and physical, that are used to monitor a function. cyberphysical systems
are used to monitor and perform the day-to-day functions of the many auto-
mated systems that we rely on, including power stations, chemical factories,
and nuclear power plants, to name a few.
With the ubiquitous advent of connected technology, many “smart” things
are being introduced into our connected environment. A connected network of
such smart things has led to the evolution of Internet of Things. IoT has
become excessively pervasive and prevalent from our smart homes to hos-
pitals. These new types of connected systems bring about new challenges in
securing them from attacks with malicious intent to disrupt their day-to-
day functioning.
Throughout this book, we will use examples from various types of such
connected systems to illustrate how data analytics can facilitate cybersecurity.

1.3 Data Analytics

Data analytics deals with analyzing large amounts of data from disparate
sources to discover actionable information leading to gains for an organization.
Data analytics includes techniques from data mining, statistics, and business
management, among other fields.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 1.3 Data Analytics 9

The large amount of data collected has led to the “big data” revolution. This
is also the case in the domain of cybersecurity. Big data (Manyika et al. 2011,
Chen et al. 2014) refers to not only massive datasets (volume) but also data that
are generated at a rapid rate (velocity) and have a heterogeneous nature
(variety), and that can provide valid findings or patterns in this complex
environment (veracity). These data can also change by location (venue).
Thus, big data encompasses the truly complex nature of data particularly in
the domain of cybersecurity.
Every device, action, transaction, and event generates data. Cyber threats
leave a series of such data pieces in different environments and domains.
Sifting through these data can lead to novel insight into why a certain event
occurred and potentially allow the identification of the responsible parties and
lead to knowledge for preventing such attacks in the future.
Let us next understand how data analytics plays a key role in
understanding cyberattacks.

1.3.1 Why Is Data Analytics Important for Cybersecurity? A Case

Study of Understanding the Anatomy of an Attack
Cyberattacks evolve very fast. Let us consider a vulnerability in software code.
A software patch can be developed to address the vulnerability; however, a
slight modification may lead to a new vulnerability. There is a constant back
and forth between the conspirators and the legitimate users. In several cases,
the more serious attacks are not just a one-stop exploit that breaches a system
through a single vulnerability, but a multipronged attack that uses several
different channels to gain access to the secured systems.
To understand the intricacies of such a complex attack, let us consider the
anatomy of an attack based on a scenario motivated from a recent attack on a
federal lab (Dark Reading 2011). This example demonstrates the challenges
posed by multiple aspects of cyberattacks coming from multiple sources and
spread over time. Similar patterns of threat propagation have been seen in other
attacks (such as Skariachan and Finkle 2014, FireEye 2020).
Figure 1.4 shows the flow of a multipronged attack. Despite the lab’s well-
protected information technology (IT) security perimeter, the attacks made it
through in a very coordinated and prolonged process. Similar to a global
supply chain attack (FireEye 2020), in the PNNL case, first there is an attack
on the organization, and second, there is an attack on a partner that shares key
resources. In the first part of the attack, intruders take advantage of vulnerabil-
ities in public-facing web servers. In addition, hackers secretively scout the

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 10 Introduction: Data Analytics for Cybersecurity

Figure 1.4 Anatomy of a multilevel attack.

network from compromised workstations that have already been targeted

beforehand as part of a coordinated prolonged attack. The second part of the
attack starts with spear phishing. Instead of casting out thousands of emails
randomly, spear phishers target select groups of people with something in
common, such as a common employer, a similar banking or financial institution,
the same college, etc. So potentially a second group of hackers institutes a spear-
phishing attack on the organization’s major business partners with which it shares
network resources. The hackers are able to obtain a privileged account and
compromise a root domain controller that is shared by the organization and its
partner. When the intruders try to recreate and assign privileges, it triggers an
alarm, alerting the organization’s cybersecurity team.
This scenario clearly demonstrates that simply looking at one dimension of
the attack is not enough in such prolonged attacks. For such multipronged
attacks, a multifaceted approach is required. Events of interest can be identified
using a combination of factors such as proximity of events in time, in terms of
series of communications, and even in terms of the geographic origin or
destination of the communication, as shown in Figure 1.4.
For this specific example, we can evaluate three important aspects of
analysis of the cyber traffic data, which can potentially generate new insights

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 1.3 Data Analytics 11

in detecting unusual events of interest. The three aspects are temporal, spatial,
and data-driven understanding of human behavioral aspects (particularly of
attackers):
• Firstly, computer networks evolve over time, and communication patterns
change over time. Can we identify these key changes, which deviate from
the normal changes in a communication pattern, and associate them with
anomalies in the network traffic?
• Secondly, attacks may have a spatial pattern. Sources and destinations in
certain key geolocations are more important for monitoring and preventing
an attack. Can key geolocations, which are sources or destinations of attacks,
be identified?
• Thirdly, any type of an attack has common underpinnings of how it is carried
out; this has not changed from physical security breaches to computer
security breaches. Can this knowledge be leveraged to identify anomalies
in the data where we can see certain patterns of misuse?
Utilizing the temporal, spatial, and human behavioral aspects of learning new
knowledge from the vast amount of cyber data can lead to new insights of
understanding the challenges faced in this important domain of cybersecurity.
Thus, simply looking at one dimension of the data is not enough in such
prolonged attack scenarios. For such a multipronged attacks, we need a
multilevel framework that brings together data from several different data-
bases. Events of interest can be identified using a combination of factors such
as proximity of events in time, in terms of series of communications and even
in terms of the geographic origin or destination of the communication, as
shown in Figure 1.5. Some example tasks that can be performed to glean
actionable information are the following:
(a) Clustering based on feature combinations: One important piece of data
collected in most organizations is Intruder Detection System (IDS) logs
such as SNORT. These can be leveraged, and a keyword matrix and a
word frequency matrix can be extracted to use for various analytical tasks.
For example, the keyword matrix can be used to perform alarm clustering
and alarm data fusion to identify critical alerts that may have been missed.
Instead of clustering the entire set of features seen in a snort alarm, we can
perform clustering based on a combination of features.
(b) Collusions and associations: Using the keyword matrix, we also extract
associations to identify potentially repeated or targeted communications. This
information in conjunction with network mapping can also be used to deter-
mine which attacks are consistently targeted to specific types of machines.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 12 Introduction: Data Analytics for Cybersecurity

Š Events become relevant when they occur

together
ŠThese events become relevant with proximities
rather than causation
Š The two events are In close proximity , based on
• Source proximity
• Destination proximity
• Temporal proximity or delay
Spatial distance

N1

N12
N2
N8

N5 N4
N3

N6
N4 N9 N10 N11 N3

0 4 8 12 16
Time
Goal : to identify potential “collusions” among the entities responsible for related events

Figure 1.5 Multidimensional view of threats.

(c) Time proximity and network evolution: Clustering can be performed

after creating time intervals that accounts for time proximity. This not only
allows mining the data in proximity of time but also evaluating how the
networks evolve over time and which time interval may be critical. For
instance, we might want to identify if there are repeated events of interest
in certain time periods. Lastly, looking at the clustering in different
segments of time, we can look at mining for possible attack paths based
on variations in cluster content and cluster cohesion.

1.3.2 How Can Data Analytics Help?

The attack case study shows the potential of data analytics and how data from
multiple sources can be used to glean novel information. Data analytics can
help in several ways to support the defense of cyber systems: mining system
logs, monitoring systems, tracing the attacks, predicting risks by identifying
critical systems in a network flow, predicting attacks based on prior or similar
attacks, identifying vulnerabilities by mining software code, understanding
user behavior by mining network logs, and creating robust access control rules
by evaluating prior usage and security policies.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 1.3 Data Analytics 13

What this book is not about: This book does not address the traditional
views of security configurations and shoring up defenses, including setting up
computer networks, setting up firewalls, web server management, and patching
of vulnerabilities.
What this book is about: This book addresses the challenges in cybersecurity
that data analytics can help address, including analytics for threat hunting or
threat detection, discovering knowledge for attack prevention or mitigation,
discovering knowledge about vulnerabilities, and performing retrospective and
prospective analysis for understanding the mechanics of attacks to help prevent
them in the future.
This book will provide multiple examples across different types of networks
and connected systems where data can be captured and analyzed to derive
actionable knowledge fulfilling one of the several aims of cybersecurity,
namely prevention, detection, response, and recovery.

https://doi.org/10.1017/9781108231954.002 Published online by Cambridge University Press

 2
Understanding Sources of Cybersecurity Data

Cyber threats often lead to loss of assets. This chapter discusses the multitude
of datasets that can be harvested and used to track these losses and origins of
the attack. This chapter is not about the data lost during cyberattacks but the
data that organizations can scour from their networks to understand threats
better so that they can potentially prevent or even predict future attacks.

2.1 End-to-End Opportunities for Data Collection

The information systems used to perform business functions have a well-
defined process spanning over connected systems. In a typical client server
scenario, as shown in Figure 2.1, a user connects to a system via an internet
pipeline. The system has built-in application functionality important to run the
business function. A return pipeline sends a response back to the user. The
functionality of the system allows the delivery of the information commodity
requested by the user.
As the example layout in Figure 2.1a shows, the logical view of the user
requesting access to a business application can appear to be fairly straightfor-
ward. However, within this pipeline there could be several points through
which the request and response pass, as shown in Figure 2.1b, leading to
several opportunities in the end-to-end process for data collection to help
understand when a cyber threat may occur in this process.
As we can see in Figure 2.1b, when the user requests a resource, it has to go
through a complex networking pipeline. The user may have a firewall on their
own system and the router through which they send out the request. This
request can be filtered through the internet service provider, lookups can be
performed in the domain name system (DNS) and the data can be routed
through multiple paths of routers, which are linked through the routing table.

14

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 2.1 End-to-End Opportunities for Data Collection 15

Figure 2.1 Logical and physical view of user request and response in a network-
based environment.

The request on the other side may again have to pass through the routers and
firewalls at multiple points in the system being accessed by the user. There
may be multiple intrusion detection systems (IDS) posted throughout the
systems to monitor the network flow for malicious activity. This is just one
example scenario; different network layouts will result in different types of
intermediate steps in this process of request and response, particularly based on
the type of response, the type of network being used, the type of organization
of business applications, the cloud infrastructure being used, to name a few
factors. However, certain key components are always present that allow for
multiple opportunities to glean and scour for data related to potential cyber
threats.
There can be several opportunities to collect data to understand potential
threats. Data collection can begin at a user access point, system functionality
level, and commodity level (particularly if the data is being delivered). For
example, at the user level, we can utilize data such as the following: (a) Who is
the user? The psychology of the user, personality types, etc., can influence

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 16 Understanding Sources of Cybersecurity Data

Figure 2.2 Common types of cybersecurity data.

whether a user will click on a link or give access to information to others. (b)
What type of interface is being used by the user? Is there clear information about
what is acceptable or not acceptable in the interface? (c) What type of access
system is being used? Is there access control for users? (d) What data are available
about the access pipeline, such as the type of network or cloud being used.
Several common types of datasets can be collected and evaluated, as shown
in Figure 2.2, including various types of log data such as key stroke logs, web
server logs, and intrusion detection logs, to name a few. We next discuss
several types of such datasets.

2.2 Sources of Cybersecurity Data

Cybersecurity-related data collection will vary across the type of networks,
including computer networks, sensor networks, or cyberphysical systems. The
method and level of data collection will also vary based on the application
domains for which the networks are being used and the important assets being
protected. For example: (a) social media businesses, such Facebook, are
primarily user data driven, where the revenue is based on providing access
to user data and monitoring usage data; (b) e-commerce businesses, such as
Amazon, are usage and product delivery based; (c) portals, such as Yahoo, are
again user data driven but more heavily reliant on advertisements, which can
target users based on what they see and use most often; (d) cyberphysical
systems, such as systems for monitoring and managing power grids, are based
on accurate functioning of physical systems and delivery of services to users
over these physical infrastructural elements.

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 2.2 Sources of Cybersecurity Data 17

In each of these types of systems, the underlying infrastructure has to be

monitored to ensure accurate functioning and prevention, detection, and recov-
ery from cyber threats. The level of monitoring and management of such data
will vary with the level of prevention, detection, or recovery expected in the
domain. Some domains have a high emphasis on prevention; others may have
a high level of emphasis on detection or recovery. In all such cases, multiple
types of datasets can be collected to provide intelligence on the cyber threats,
and user behaviors can be evaluated to prevent future threats or even identify
an insider propagating the threats.
In the following discussion for each dataset, we examine the following: (a)
What is the data? (b) What is an example of its use in literature? (c) What type
of detection can it be used for? In the chapters throughout this book, we will
discuss how some of these datasets can be leveraged to discover anomalies
identifying potential threats using data analytics methods.

2.2.1 Log Data

The nature of electronic communication and activities allows for several types
of datasets to be logged. Some examples include the following: (1) intrusion
detection system (IDS) logs including alarms raised by IDS; (2) key stroke
logs; (3) router connectivity data/ router logs; (4) web server logs; and (5)
firewall logs. This is not an exhaustive list but includes some of the major
types of logs that can be collected.

2.2.1.1 Keystroke Logs

Keystroke logging or key logging is a mechanism to capture every key being
pressed on a keyboard, but can also go beyond key presses to actions such as
copying materials to the clipboard or other interactions with the user system.
Key logging has been extensively studied for many applications, from writing
to cognitive analysis to security threats. A survey on key logging (Heron 2007)
outlines mechanisms, including hardware installation, kernel-level, system
hook, and function-based methods, for key logging.
Key logging has also been studied for smart phones (Gupta et al. 2016, Cai and
Chen 2011). A recent survey (Hussain et al. 2016) extensively outlines motion-
based key logging and inference attacks that can result from smart phone key
logging. This survey classifies key logging as in-band logging through the main
channels of the keystrokes and out-of-band logging using side channels such as
acoustics, power consumption, etc. Thus, key logging is not necessarily limited to
keyboard-based data collection but can get quite sophisticated.
This type of data collection allows studying user behaviors but may also
be used to maliciously detect user credentials, user preferences, or other

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 18 Understanding Sources of Cybersecurity Data

sensitive information. Thus, it is also essential to understand the capabilities

of key loggers to create any type of defense against threats utilizing
key loggers.

2.2.1.2 Intrusion Detection System Logs

Intrusion detection system (IDS) log data (e.g., from Snort) provide data about
alerts that are raised by matching any known signatures of malicious activities
in the header and payload data. Generally, IDS will also provide an alert level
of low, medium, or high. IDS logs analyze the packets based on malicious
signatures and provide information on time stamp, service used, protocol,
source, and destination. IDS can be placed at various points in a network,
and multiple such datasets can be collected and correlated (Deokar and
Hazarnis 2012). IDS logs are also commonly used for anomaly detection
methods, which are utilized to detect threats beyond signature matching.
Here anomalous packets indicate an unusual behavior with respect to the
normal, where the normal can be discovered and predefined through various
analytics methods.
IDS log data lends itself well to secondary analysis such as through data
mining methods including association rule mining (such as Vaarandi and
Podiņš 2010 and Quader et al. 2015), human behavior modeling (such as
Quader and Janeja 2014 and Chen et al. 2014), and prediction of attacks, to
name a few examples. Multiple IDS and other types of logs are also correlated
to detect significant anomalies, which are not otherwise detectable (such as
illustrated in Janeja et al. 2014 and Abad et al. 2003). Visualization of logs
(such as in Koike and Ohno 2004) has been explored to facilitate the analysis
of the logs by looking at the information selectively, slicing and dicing the data
by certain features, such as by time or by event.

2.2.1.3 Router Connectivity and Log Data

The internet is a network of networks or subnetworks. The networks at each
level are connected by routers. A router connects computer networks and
forwarding data across computer networks. Each of these routers is connected
for data transmission. This can range from a simple home router to corporate
routers that connect to the internet backbone. A routing table stores infor-
mation about the paths to take for forwarding and transmitting the data. The
routing table stores the routes of all reachable destinations, including routers,
from it. Various algorithms devise an efficient path through these connected
routers (such as Sklower 1991 and Tsuchiya 1988).
A router provides not only route information but also all the raw IP
addresses that pass through the router. These IP addresses can be mapped to

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 2.2 Sources of Cybersecurity Data 19

identify possible malware activity when data are sent to suspicious

geolocations in an unauthorized manner (Geocoding-Infosec 2013).
However, care must be taken in using the IP addresses in isolation as they
can be subject to IP spoofing, which hides the identity of the sender. Router
data can also be utilized to study and possibly identify traffic hijacking (Kim
Zetter Security 2013) and bogus routes by looking at historic route data stored
in a knowledge base (Qiu et al. 2007).

2.2.1.4 Firewall Log Data

Firewalls act as a first line of defense that can stop certain types of traffic
based on firewall security policies. In addition, these policies also have to be
maintained to stay up to date with the changing landscape of the network
usage. Essentially every access entry can be logged as it has to pass through
the firewall. Some threats can be directly identified and blocked based on a
clearly defined firewall policy or rule. For instance, if there is a clearly
unauthorized access to an internal server, a well-configured firewall can
block it to prevent access to the system. Major threat-related activities such
as port scans, malware, and unauthorized access can easily be filtered
through robust firewall rules. It essentially filters traffic based on the config-
uration of access to the systems protected by the firewall. Firewalls are
typically designed to look at the header information in the data packets to
match against prespecified rule sets. Firewalls can be host based or network
based depending on whether they are deployed at an individual user’s system
or at a network interface.
Firewalls differ from IDS since they are generally limited to header infor-
mation screening, whereas IDS can look at the payload data as well and block
connections with malicious signatures. However, there has been a convergence
in these functionalities in more recent times.
Firewall policy rules are one area where data mining may benefit by
allowing the creation of a dynamic set of rules based on the traffic passing
through the firewall. Analysis of policy rules and network traffic is used
(Golnabi et al. 2006) to generate efficient rule sets based on the network traffic
trends and potentially identify misconfigurations in the policy rules. This
particular work uses association rule mining (ARM) and simple frequency
counting of rules to generate firewall policy rules. In addition, it also identifies
different types of policy anomalies, including blocking of legitimate traffic,
allowing traffic to nonexisting services or redundant policy anomalies.
Similarly, Abedin et al. (2010) regenerates firewall policy rules and com-
pares them with existing policies to discover anomalies.

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 20 Understanding Sources of Cybersecurity Data

2.2.2 Raw Payload Data

Any data sent over the network are divided into multiple parts. Two key parts
include (a) the header information, which stores data about source and destin-
ation among other things; and (b) the actual content being transmitted, referred
to as the payload. There are several privacy concerns in accessing these
payload data since these data are the actual content that is being sent, which
may be under strict access control. Such payload data can be accessed only
where legally allowed and users have provided permissions to access the data.
Additionally, the data may be encrypted, so its usefulness as raw data to be
mined is limited.
Payload data are accessible through packet sniffers such as Wireshark,1
where the data dump of the traffic can be retrieved. Payload data can be
massive even for a few minutes of data capture. Thus, it provides a strong
motivation for using big data technologies to collect and mine such data where
permissible. In addition, for web-based traffic the browser cache is another
way to access the payload data from the client or end user’s side.
Payload data have been shown (Wang and Stolfo 2004, Kim et al. 2014,
Limmer and Dressler 2010) to be effective in identifying anomalous threats in
network intrusion detection systems. For example, one recent study (Limmer
and Dressler 2010) selectively analyzed parts of the payload, thus reducing the
challenges in high-speed network intrusion detection systems. Parekh et al.
(2006) utilize suspicious payload sharing in a privacy-preserving manner to
identify threats across multiple sites.
Payload data can be used in multiple ways, such as to discover an individual
user’s behavior, the presence of malwares in the payloads, and other security
threats that can be detected based on the actual content of the payload. One
common use of payload data is to identify threats based on signatures of
malware that may be present in the payloads. For example, if a virus is
embedded in a packet and this virus has a known signature, then this can be
captured by traditional intrusion detection system rules. One such open-source
network intrusion detection system is Snort, which provides Snort rules (Snort
2020). Snort can also be used as a packet sniffer, like Wireshark, but can also
be used as an IDS. Packets with malware embedded in them can be detected
using multiple mechanisms such as simple keyword searches or complex
regular expression matches and flagged. The traffic can be blocked or marked
for further analysis, such as using Snort alarms or Wireshark coloring rules
(Cheok 2014).

1
www.wireshark.org.

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 2.2 Sources of Cybersecurity Data 21

Figure 2.3 Example extraction of a communication graph from network traffic.

2.2.3 Network Topology Data

A computer network can be represented as a graph in terms of the structure of
the network and in terms of the communication taking place over the network.
Network traffic data dump can be used to generate the communication graph of
all exchanges taking place over the network. As shown in Figure 2.3, for
example, header data collected from a traffic dump file through Wireshark can
be utilized to plot the communication between the source and destination IP
addresses, which become the vertices, and the exchange between the two
vertices forms the edge in the graph. In this example, NodeXL2 is used to plot
the graph data.
Once the communication data are in the graph form, graph metrics (for
example, as discussed in Nicosia et al. 2013) can be computed, such as node-
level metrics, including centrality, page rank, etc.; and network-level metrics,
such as diameter, density, etc. In addition, based on the network properties,
predictions can also be made about future network evolution. The example in
Figure 2.4 illustrates one such task in a sample traffic data.

2
www.smrfoundation.org/nodexl/.

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 22 Understanding Sources of Cybersecurity Data

Figure 2.4 Exploratory analysis using Degree Centralities.

Data from network traffic is collected through packet sniffers such as

Wireshark. To find communication behaviors of IP addresses along with
anomalous fluctuations, exploratory analysis is performed on these data. The
data from the network traffic need to be preprocessed, and this preprocessing
will change with the task being performed. For instance, if in this example we
wish to perform analysis by day of the week, the traffic data are sorted by the
day of the week to get patterns by day, such as all Mondays or all Tuesdays.
We can then compute the degree (i.e., number of edges incident on a vertex) of
each node by day of the week. This can be performed for specific dates also;
however, in this particular example we are interested to see behavior on certain
days of the week by each of the IP addresses. We can sort the IP addresses by
their degrees across days of the week, and the top ones appear to be consist-
ently present in the traffic. Similarly, nodes with low degrees can also be
identified. In such a scenario, it would be interesting to find a node, which is
generally highly consistent as a high-degree node, to appear in the list of nodes
with a lower degree, indicating a shift in the traffic pattern. Now let us consider
the bar chart of the degrees for each IP address across each day of the week.
We can observe that some IP addresses are consistently higher degree across
all days of the week, which is further illustrated by the plot for IP1, IP2, and

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 2.2 Sources of Cybersecurity Data 23

IP3 across all days of the week. We can also see that the degrees of IP9 and IP7
seem to be higher on some days but lower on other days. This is further
clarified by the plot for IP9, which shows Wednesday as a day where IP9 has
inconsistent behavior.
Thus, through such exploratory analysis it is not only possible to identify
nodes that are inconsistent but also time points where the behavior is inconsist-
ent. Alternatively, this method can also be used to identify highly connected
nodes (such as nodes receiving higher than normal connections during a
breach) or least connected nodes (perhaps nodes that are impacted by a breach
and lose connectivity). This type of consistency and inconsistency can be
identified at the node level and at the graph level as discussed in Namayanja
and Janeja (2015 and 2017)
Another study (Massicotte et al. 2003) introduces a prototype network
mapping framework that uses freely available network scanners (nmap,
Xprobe) on built-in network protocols (ICMP, ARP, NetBIOS, DNS,
SNMP, etc.) to create a real-time network topology mapping with the help of
intelligence databases. It must be used in tandem with an intrusion detection
system. Studies discussed earlier for graph metrics can be applied to such
works as well after the topology is discovered.

2.2.4 User System Data

Figure 2.5 outlines several key features that can be extracted to monitor
unusual activities at the individual system level. Example features include
active process resident memory usage, which is available for all operating
systems (OS) and allows for building a profile on the normal memory usage
of a process over time. As an example, an abnormal spike in memory usage
can be attributed to processing a large volume of data. This might be useful
in detecting a potential insider threat, especially when integrated with other
user behavioral data from sensors monitoring user stress levels or integrat-
ing with other log datasets. Similarly, CPU time utilization can be used for
measuring system usage. Several OS-specific features, such as kernel
modules and changes in registry values, are also identified in Figure 2.5.
However, it is important to use multiple signatures over time from several
of the features to eliminate the regular spikes of day-to-day operations. This
is the key differentiator for a robust analysis where we do not simply rely
on one or two features but multiple features and their stable signatures (as
compared to historical data) to distinguish alerts. Tools such as OSQuery

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 24 Understanding Sources of Cybersecurity Data

Feature name OS specific

Active process name, active process filesystem path, active process All OS
ports and sockets, active process file access, active process
resident memory usage, active process CPU time utilization, active
process system calls, active process priority value, active process
owner and group information, loaded peripherals drivers, key-
store access patterns
Loaded kernel modules Linux/Unix
Loaded kernel extensions Mac OSX
Change in registry values Windows
File system journaling (metadata) information All major file systems (NTFS,
ext4, HFS+)
Network routing tables All major OS
Network firewall rules All major firewall
implementations
System-level sensors (current, voltage in different bus inside PC, Almost all peripherals
CPU/GPU fan speed, etc.)

Figure 2.5 OS-specific variables for CPU processing.

(OSQuery 2016) and Snare (SNARE 2016) can facilitate capture of

these features.
Stephens and Maloof (2014) provide a very general framework for insider
threat detection by gathering information from file read/write activities,
printing, emailing, and search queries, then building a probabilistic Bayesian
belief network from the sensor and context data, such as a user behavior profile
from past actions. Van Meigham (2016) focuses on macOS malware detection
using a kernel module to intercept system calls and generating a heat map
analysis on the results.

2.2.5 Other Datasets

In addition to the datasets discussed, there are additional datasets that can be
utilized to leverage knowledge about cyberattacks.
Access control data: These data can help better understand usage of the
assets that need to be protected. Role mining (Vaidya et al. 2007, Mitra et al.
2016) from access control data can help shape and create better and more
robust roles.
Eye tracker data: A user’s behavior can be judged by the interactions of
the user with the system being used. One such mode of input is the screen.
Data collected from the user’s eye gaze, captured through an eye tracker, can
help analyze the user’s level of engagement with a system and user preferences
or positioning important items on the screen (such as those discussed in

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 2.2 Sources of Cybersecurity Data 25

Darwish and Bataineh 2012) to evaluate browser security indicators. The data
collected through the eye tracker can be mined for patterns such as associations
between security cue locations on the screen and number of views or clicks.
Clustering can be performed on eye gaze data to identify presence or absence
of clusters around security cues. Associations can be analyzed between user’s
perception of security, backgrounds, and demographics to different zones of
eye gaze foci in a stratified manner. If users perceive disclosing important
information through emails as a low-risk activity, they are less likely to see the
security cues. Similarly, if they see the security cues, their perceived risk of
responding will be high. Studies have hypothesized that user education can
change user’s perception of security and help them to better see these security
cues, increasing the likelihood of threat detection or identifying threats through
visual cues such as in the case of phishing.
Vulnerability data: Software vulnerability is a defect in the system (such as
a software bug) that allows an attacker to exploit the system and potentially
pose a security threat. Vulnerabilities can be investigated, and trends can be
discovered in various operating systems to determine levels of strength or
defense against cyberattacks (Frei et al. 2006). Using the National
Vulnerability Database from the National Institute of Standards and
Technology (NIST) (NIST 2017), trends can be analyzed for several years
and across major releases for operating systems to reinforce knowledge of
choices for critical infrastructural or network projects.
NVD is built on the concept of Common Vulnerabilities and Exposures
(CVE),3 which is a dictionary of publicly known vulnerabilities and exposures.
CVEs allow the standardization of vulnerabilities across products around the
world. NVD scores every vulnerability using the Common Vulnerability
Scoring System (CVSS).4 CVSS is comprised of several submetrics, including
(a) base, (b) temporal, and (c) environmental metrics. Each of these metrics
quantifies some type of feature of a vulnerability. For example, base metrics
capture characteristics of a vulnerability constant across time and user environ-
ments, such as complexity, privilege required, etc. The environmental metrics,
on the other hand, are the modified base metrics reevaluated based on organ-
ization infrastructure. NVD allows searches based on subcomponents of these
metrics and also based on the basic security policies of confidentiality, integ-
rity, and availability. These searches can provide data for analysis to identify
trends and behaviors of vulnerabilities across operating systems or other
software for different types of industries.

3 4
http://cve.mitre.org/. https://nvd.nist.gov/cvss.cfm; www.first.org/cvss.

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 26 Understanding Sources of Cybersecurity Data

Windows Mac_OS_X Linux Android

14

12

10

8
Buffer XSS
6 50

45
4
40

35
2

Number
30

0 25
2006 2007 2008 2009 2010 2011 2012 20

15

10

0
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

Figure 2.6 Comparison of vulnerabilities over operating systems.

Let us consider cross-site scripting vulnerability.5 When data regarding the

number of vulnerabilities are pulled from NVD across 2006 to 2012, we can
see the trends of operating systems that are most impacted by this vulnerabil-
ity, as shown in Figure 2.6. In addition, we can also compare the occurrences
of different types of vulnerabilities such as cross-site scripting and buffer
overflow. While this is a straightforward plotting of number of vulnerabilities
across years, it provides insights into the robustness of operating systems
for different types of vulnerabilities and across different CVSS metrics.
Such analyses can be an important feed into decision making before choices
for adopting software are made from a security point of view in
organizational applications.

2.3 Integrated Use of Multiple Datasets

Let us consider a scenario where multiple datasets can be utilized to study
potential cyberattacks. Cyberattacks are rare compared to the day-to-day traffic
in a computer network; therefore, they appear in datasets as anomalies.
Anomalies are essentially data points or patterns that are unusual with respect
to the normal. It is clear that there needs to be a frame of reference that is

5
https://tools.cisco.com/security/center/viewAlert.x?alertId=35601, Adobe Flash Player Cross-
Site Scripting Vulnerability.

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 2.4 Summary of Sources of Cybersecurity Data 27

“normal” compared to which something is deemed an “anomaly.” A single

dataset such as any of the ones discussed so far can be used for anomaly
detection, but it is important to note that if multiple datasets result in similar
types of anomalies, then the credibility of labeling an anomaly is higher.
One such integrated evaluation would be to discover anomalies in
network traffic data with a temporal, spatial, and human behavioral per-
spective. Studying how network traffic changes over time, which locations
are the sources, where is it headed, and how are people generating this
traffic – all these aspects become very critical in distinguishing the normal
from the abnormal in the domain of cybersecurity. This requires shifting
gears to view cybersecurity as a holistic people problem rather than a
hardened defense problem. By utilizing some of the datasets discussed in
this chapter, we can answer the following important questions in studying
these aspects:
Firstly, computer networks evolve over time, and communication patterns
change over time. Can we identify these key changes that are deviant from the
normal changes in a communication pattern and associate them with anomalies
in the network traffic?
Secondly, as attacks may have a spatial pattern, sources and destinations in
certain geolocations can be more important for monitoring and preventing an
attack. Therefore, can key geolocations that are sources of attacks, or key
geolocations that are destinations of attacks, be identified? Moreover, can IP
spoofing be mitigated by looking at multiple data sources to supplement the
knowledge of a geospatial traffic pattern?
Thirdly, any type of an attack has common underpinnings of how it is
carried out; this has not changed from physical security breaches to computer
security breaches. Can this knowledge be leveraged to identify behavioral
models of anomalies where we can see patterns of misuse?
Recent work highlights some of these questions in discovering anomalies
utilizing network data to study human behavioral models such as Chen et al.)
2014) and Quader and Janeja (2014). These will be discussed further in
Chapter 10.

2.4 Summary of Sources of Cybersecurity Data

Through this chapter, multiple types of sources of cybersecurity data have
been discussed. Table 2.1 summarizes these data under the following: (a) data
source, (b) literature study examples, and (c) type of detection it can be
used for.

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 28 Understanding Sources of Cybersecurity Data

Table 2.1 Summary of sources of cybersecurity data

Source of
cybersecurity Type of detection it can be
data Literature study examples used for
Keystroke Heron 2007, Cai and Hao 2011, User behavior, malicious use to
logging Gupta et al. 2016, Hussain et al. detect user credentials
2016
IDS log data Abad et al. 2003, Koike and Association rule mining, human
Ohno 2004, Vaarandi and Podiņš behavior modeling, log
2010, Deokar and Hazarnis 2012, visualization, temporal analysis,
Chen et al. 2014, Janeja et al. anomaly detection
2014, Quader and Janeja 2014,
2015
Router Tsuchiya 1988, Sklower 1991, Suspicious rerouting, traffic
connectivity Qiu 2007, Geocoding Infosec hijacking, bogus routes
and log data 2013, Kim Zetter Security 2013
Firewall log Golnabi et al. 2006, Abedin et al. Generate efficient rule sets,
data 2010 anomaly detection in policy
rules
Raw payload Wang and Stolfo 2004, Parekh Malware detection, embedded
data et al. 2006, Limmer and Dressler malware, user behavior
2010, Kim et al. 2014, Roy 2014
Network Massicotte et al. 2003, Nicosia Consistent and inconsistent
topology 2013, Namayanja and Janeja nodes, time points
2015, 2017 corresponding to anomalous
activity
User system Stephens and Maloof 2014, Van User profiles, user behavior
data Meigham 2016 data, insider threats
Access Vaidya et al. 2007, Mitra et al. Generate efficient access control
control data 2016 roles
Eye tracker Darwish and Bataineh 2012 Browser security indicators,
data security cues, user behavior
Vulnerability Frei et al. 2006 Vulnerability trend discovery
data

https://doi.org/10.1017/9781108231954.003 Published online by Cambridge University Press

 3
Introduction to Data Mining
Clustering, Classification, and Association Rule Mining

3.1 Knowledge Discovery and Data Mining

Process Models
Data mining (DM) is the discovery of hidden and nontrivial patterns in very
large datasets. The discovered patterns are potentially actionable and beneficial
to study the problem at hand.
There are several steps involved in this process as proposed initially by
Fayyad and Shapiro (1996). Figure 3.1 depicts the goals of knowledge discov-
ery (KD) and components of data mining methods they discussed. These goals
of KD include the discovery of new patterns or verifying a hypothesis that the
users generally accept or are interested in evaluating, such as prediction of
future events or explanation to the user in the form of understandable and
intuitive knowledge discovered. Increasingly, as data are becoming very large,
data mining fits into the core analytics of any big data solution. In such
scenarios, the core functionality of the mining algorithms does not change
drastically; however, the infrastructure managing the parallelization is a key
distinguishing factor in a big data environment. We further discuss the big data
perspective to data mining in Chapter 4.
A data mining method that is used for the knowledge discovery will
generally comprise multiple components, including the following:
(a) The model being used and its representation. This refers to the language
used to describe the patterns along with any representational assumptions.
(b) A method to evaluate the model. DM algorithms have an expected output,
and quality of results based on this can be quantified through an evaluation
metric (such as accuracy and error metrics).

29

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 30 Introduction to Data Mining

Goals of
knowledge
discovery

Discover new Verify user

patterns hypothesis
Components of
data mining
methods
Prediction
Model
Model evaluation Search
representation
Description
/presentation to
user

Figure 3.1 Goals of knowledge discovery and components of data

mining methods.

DM method Example evaluation criteria

Classification Accuracy, precision, recall,

F-measure

Clustering Sum of Squared Error,

silhouette coefficient

Association rule mining Confidence, support, lift

Figure 3.2 DM method example evaluation criteria.

(c) A search component. This involves a parameter search and model search.
This component allows for finding the optimal fit of both the model and
parameter to use based on maximizing the evaluation criteria.
Some example methods and their evaluation criteria are shown in Figure 3.2.
These will be discussed further in the following sections.
Figure 3.3 summarizes the steps for a standard knowledge discovery pro-
cess. The process starts with large and potentially heterogeneous data sources.
Through user input, data need to be identified and selected based on the task
being addressed. As discussed in some models (such as Cross Industry
Standard Process for DM [CRISP-DM]; see Chapman et al. 2000, Shearer
2000), it is important to have user input at this stage to clearly understand the
business or user requirements that can facilitate better and more targeted data
selection. Once the right data sources are identified, integrated data

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.2 Data Preprocessing 31

Heterogeneous data
sources

Data integration Data Data Pattern

and selection preprocessing mining evaluation

Actionable
knowledge

Figure 3.3 KDD process.

preprocessing is performed on these data sources. Data preprocessing tasks can

include filling missing values, semantic mapping, removing redundancy in the
data through correlation, noise removal, smoothing of the data, data volume
reduction, and feature selection, among others. A large amount of preproces-
sing can done while integrating the data to resolve data inconsistencies. Data
preprocessing generally takes up a majority of time, up to 80%, in any data
mining project. A good preprocessing leads to good data mining. Data that
have not been sufficiently preprocessed may lead to frivolous and random
results. Thus, sufficient time should be allocated to the preprocessing tasks.
Once the data have been processed, then the relevant data mining task can be
applied. Then the results need to be evaluated, and once this evaluation is
quantified any patterns found can become actionable knowledge. It is clearly
possible that any stage can send the process back to any of the preceding steps.
In addition, once actionable patterns have been tested or examined by the end
users, changes can be made to how and what data are collected for better
knowledge discovery. All these steps are further described in the following
sections. A detailed comparison of multiple DM process models is provided
by Kurgan and Musilek (2006), which discusses a step-by-step comparison of
over six such models.

3.2 Data Preprocessing

Data in the real world tend to be messy. This can be caused by device
malfunction, data entry errors, or data integration from heterogeneous sources.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 32 Introduction to Data Mining

Attribute data Example

type
Numeric Number of bytes sent
Categorical Type of attack
Binary Presence or absence of virus
Ordinal Severity of alert (low, medium, high)
Text Intrusion alarm text
Complex Types of Datasets Example

Unstructured Spam email

text Intrusion log data
Spatial data Geographical server locations
Locations of cyberphysical sensors
Locations of spam origination
Graph data IP communication graph
Routing paths for communication

Figure 3.4 Examples of some data types.

To get high-quality mining results that are actionable and beneficial, it is

important to clean up the messy data. Data can be very large in terms of size
and number of attributes, meaning that important patterns get hidden in the
subspaces. It is important to focus on the right portions of the data both in
terms of the volume and the attributes. Data preprocessing such as data
cleaning, data transformation, and data reduction, as well as careful data
integration, prepares the data for mining.
These tasks will vary based on the data types and the complexity of the data.
Data can be of different types, including text, numeric (discrete or continuous),
ordinal (order or ranked data, such as severity scores, which have an inherent
order to them), binary data (with only two possible outcomes), and categorical
data (with multiple categories). Various data types and types of complex
datasets are shown in Figure 3.4, along with brief descriptions and examples
of relevant datasets of these types. Some complex datasets may consists of
multiple types of attributes in them; for example, intrusion log data will consist
of text, categorical, and numeric data attributes.
The preprocessing tasks to apply are selected based on the complexity of the
data and the attribute types. For example, similarity in the data is defined
differently for numeric attributes versus binary valued attributes. Similarly, for
complexity of the data, for example, spatial data, would require additional
steps such as discovering spatial proximity, such that the mining task can be
separated by spatial neighborhoods.
Data preprocessing includes several tasks as discussed in the
following subsections.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.2 Data Preprocessing 33

3.2.1 Data Cleaning

Many times in network traffic data we might see missing data values
due to dropped packets, traffic fluctuations over different time periods,
and generally redundancies in network communication data. This step
addresses the general cleanup of messy data, including (a) missing data
values, (b) fluctuations in the data, and (c) inconsistencies and redundancies
in the data.
Missing data: To address the missing values, the mean of the attribute
(a) if numerical or mode if categorical, can be used to replace the missing
values (see Figure 3.5). However, if the dataset is large or has multiple
classes/groups, a fairly simple way to deal with missing values is to use
the mean of the group to replace the missing values in that particular
group (Luengo et al. 2012). This is a fairly intuitive and simple way to
deal with missing values. Other more involved methods can be used
to predict the missing values by training based on the records that have
the complete data.
How much missing data can be replaced? A rule of thumb is that if roughly
5% of the data are missing, then it is feasible to replace it. One could also

Figure 3.5 Example data preprocessing.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 34 Introduction to Data Mining

consult with a domain expert or utilize domain heuristics to decide if data

should be discarded or filled in to impute the missing data.
Noisy data: Fluctuations in the data, when taken together in similarity or
distance-based methods, can cause results to be messy. Thus, it is useful to
clean the noise by smoothing out the data to provide simpler and more refined
representations. If the data can be divided into small groups and the mean
behavior of the groups can be used to represent all the data in the group, then
the data will be much smoother.
Small groups can be created by creating small bins or buckets in the data,
and the mean or mode of the data can replace all the values in each bin.
Binning is essentially an unsupervised splitting-based discretization method of
the data. Several discretization techniques are discussed in by Liu et al. (2002)
and Garcia et al. (2013). Equal frequency binning roughly deposits an equal
number of points are across a fixed number of bins as shown in Figure 3.5b.
Similarly, equal distance binning distributes an equal distance across a fixed
number of bins (Figure 3.5b).
Clustering where similar data points are grouped together and the mean or
centroid of the cluster can be used to replace the other values can also create a
smooth dataset. In both of binning and clustering, a relatively large number of
groups should be used to preserve the data patterns.
Outliers: Outliers in the data can drastically alter data mining
output, especially in distance-based methods. It is important to decide before-
hand through discussions with end users whether outliers need to be accom-
modated or eliminated (Kauffman and Rousseeuw 1990). This will also be
determined by the types of patterns being discovered. If the main task of the
mining is to find anomalies, in that case outliers might be kept in the data;
however, if the mining task is finding the normal behavior, such as the baseline
network signatures, then it might be more practical to remove outliers. For
univariate data, several statistical methods such as interquartile range (IQR)–
based outlier detection, discordancy tests, and standard deviation–based tests
can be used. For multivariate data, distance-based outlier detection methods
can be utilized. Outlier detection as part of the anomaly detection tasks is
further discussed in Chapters 6 and 7.

3.2.2 Data Transformation and Integration

When data are merged from multiple sources, the resolution of the data may
be different; for example, it could be years rather than days. In addition, the
attribute ranges may vary, and these ranges may be very large. In these and
other scenarios, the data will need to be transformed such that the

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.2 Data Preprocessing 35

traditional data mining algorithms can be uniformly applied. Other data

integration steps such as mapping the structure and content also need to
be performed.
Normalization: Many normalization methods have been presented in the
traditional statistical literature. One common method of normalization is the
min–max normalization (as shown in 3.1) or feature scaling such that d0 is the
transformed value and d is the original value in an attribute ‘D’. An example is
shown in Figure 3.5c:
d  minD
d0 ¼ ðnewmaxD  newminD Þ þ newminD (3.1)
maxD  minD

In this method, a new range, for example 0–1, is created for the data such
that the old min becomes 0 and the old max becomes 1 so that all the data are
within this range. This allows for a rescaling of the data to a new well-defined
range. Other measures such as z-score normalization work well when the data
are normally distributed and the mean is well defined. A comparison of
normalization methods, including min–max, z-score, and decimal scaling, is
provided by al Shalabi et al. (2006).

3.2.3 Data Reduction

Data mining is inherently useful when patterns need to be discovered in very
large amounts of data. For example, network traffic data from Wireshark can
become terabytes of data even within a few seconds depending on the network
size. Similarly, source and destination-based communication graph can be a
massive dataset. In such cases data reduction is an important strategy to help
reduce the volume and attribute set of the data.
One important step in performing data mining on such massive datasets is
reducing the dimensions that may not have sufficient data to begin with or may
not be representative of the patterns to be found. For example, in the Wireshark
data capture, we might mainly be interested in looking at just the source and
destination to generate a communication graph and annotate the graph with
packet size. In such scenarios, it is useful to reduce the volume and dimen-
sionality of the data. Moreover, in sparse datasets, data reduction helps identify
the key subspaces of the data. Data reduction, particularly volume reduction, is
also related to the smoothing techniques mentioned in the previous sections,
because the central measures of the smoothed data can be used instead of all
the data points.
Methods such as feature selection can be used to identify the relevant
features for performing data mining (Molina et al. 2002, Chandrashekar and

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 36 Introduction to Data Mining

Sahin 2014). Feature selection methods can be ranking-based methods that use
entropy measures to quantify how much data are useful in each of the
attributes. Methods such as singular-value decomposition and principal com-
ponent analysis (Fodor 2002 and Wall et al. 2003) work on the common
principal to transform the data such that a combination of the features can be
used as transformed features.

3.3 Data Mining

Data in security applications are heterogeneous, high-dimensional, and often
complex. In this section, we consider three key groups of methods, namely
clustering, classification, and association rule mining. The supervised methods
are also referred to as machine learning, where historic data help train the
algorithm to make predictions. Additional methods such as deep learning are
discussed in Chapter 11.
Each of these methods also feeds into the discovery of anomalies or unusual
patterns, which is particularly relevant to cybersecurity (see Figure 3.6). In
particular, anomaly detection deals with finding individual objects and rare

Figure 3.6 Data mining methods.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 37

combinations of objects that are unusual (anomalous) with respect to other

“normal” data points. Often the normal is a cluster or a set of clusters and the
anomalies are the outlier data points.
We start by introducing measures of similarity, which are essential for
clustering and various metrics for evaluation of the results of data mining.

3.3.1 Measures of Similarity

3.3.1.1 Distance Measures
If we want to identify a network traffic data point that looks deviant compared
to others, the fundamental way to do that is to see how similar or dissimilar it is
to other data points in the network traffic. We use these measures when we
identify clusters as discussed later in this chapter and also when we identify
distance-based outliers, discussed in Chapter 7.
The similarity or dissimilarity between two objects can be interpreted as the
distance between them. A nonnegative function d(x,y), measuring distance
between objects x and y is a metric if it satisfies the following:
• The identity property, d(x,x) = 0
• The symmetry property, d(x,y) = d(y,x)
• The triangle inequality, g(x,y) + g(y,z)  g(x,z)
Euclidean distance, the straight-line distance between two points in Euclidean
space in terms of the attributes of each data point, is one of the most commonly
used distance metrics.
Euclidean distance (ED) is computed given two n-dimensional data points
X(x1, x2, . . . , xn) and Y(y1, y2, . . . , yn), as shown in (3.2):
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
EDðX,YÞ ¼ ðx1  y1 Þ2 þ ðx2  y2 Þ2 þ . . . þ ðxn  yn Þ2 (3.2)

Manhattan distance (MD) is the “city block” distance or the distance measured
on each axis. MD is computed as shown in (3.3):
MDðX,YÞ ¼ ðjx1  y1 j þ jx2  y2 j þ . . . þ j xn  yn j Þ (3.3)
The Minkowski distance (MnD) is the generalizable distance from which
both Euclidean and Manhattan distance are derived. MnD is computed as
shown in (3.4):
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
MnDðX,YÞ ¼ ðx1  y1 Þp þ ðx2  y2 Þp þ . . . þ ðxn  yn Þp
p
(3.4)

For p = 2, the Minkowski distance generalizes to Euclidean distance.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 38 Introduction to Data Mining

Figure 3.7 Example distance metrics.

Hausdorff distance (HD) is the maximum distance of a set to the nearest

point in the other set. Hausdorff distance is a distance metric that finds the
distances between two points considering the contribution of each dimension.
It is defined as follows: Given two n-dimensional objects X(x1, x2, . . . , xn) and
Y(y1, y2, . . . , yn),

HDðX,YÞ ¼ Max fhðX,YÞ, hðY, XÞg,

where hðX,YÞ ¼ max ðmin ðxi  yi Þ
and hðY,XÞ ¼ max ðmin ðyi  xi Þ:
This distance takes each dimension into account while calculating the distance
between two points, thus bringing out the stark differences between the specific
dimensions. HD has been used in image processing literature to discover
difference between images. As deep fakes in video and imaging data become
prevalent, distance measures such as Hausdorff distance can be quite useful.
Some of the distance measures are summarized in Figure 3.7.

3.3.1.2 Similarity in Binary-Valued Variables

Consider two vulnerabilities where you are trying to identify vulnerabilities
that are similar to each other. We can extract keywords from all vulnerability
descriptions and produce a binary feature vector for each vulnerability with a
presence of a keyword (1) or absence of a keyword (0). We can use similarity
measures on these binary-valued feature vectors to find vulnerabilities that are
similar to each other.
When computing a similarity coefficient matrix for binary-valued attributes,
such as in the preceding example and as illustrated in Figure 3.8, we start with
the data matrix comprised of feature vectors for each object. The data matrix is
thus an n  m matrix consisting of n objects each with a feature vector of m
features. The first step is to determine the contingency matrix of this data
matrix, which essentially is the values of a, b, c, and d. Here, a is the number
of positive matches for each object pair where objects oi and oj both have a
value of 1, b is the number of mismatches where oi has a value of 1 and oj has a
value of 0, c is the number of mismatches such that oi has a value of 0 and oj

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 39

Figure 3.8 Process of similarity computation.

has a value of 1, and d equals the number of negative matches such that oi
and oj both have a value of 0. Once these values are computed, they can be
used to compute the similarity using different types of similarity coefficients
such as the Jaccard coefficient (a/a + b + c). The output is a n  n triangular
matrix with the computed similarity between each object pair.
The Jaccard coefficient is an asymmetric coefficient that ignores negative
matches. Asymmetric coefficients should be applied to data where absences
(0s) are thought to carry no information. For example, if we would like to
study attacks, then a system that did not have a set of attacks is not a
relevant question so features with a 0–0 match (i.e., the system has not had
an attack) is not relevant. Therefore, asymmetric coefficient is useful to
employ. Symmetric coefficients acknowledge negative matches and should
therefore be applied to data where absences are thought to carry information.
An example is the simple matching coefficient (SMC). For example, if we
are studying impact of a vulnerability patch on a software, then if in two
types of applications there was no impact of the patch, then this is pertinent
information we want to capture. Thus, in this scenario we would use a
symmetric coefficient. There are some other types of coefficients, called
hybrids, that include the 0–0 match in either the numerator or denominator
but not both. Several different types of similarity coefficients are proposed
and evaluated (Lewis and Janeja 2011), and some examples are shown in
Figure 3.9.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 40 Introduction to Data Mining

Asymmetric

Coefficient Expression

Anderberg (range 0–1)

+ 2( + )

Jaccard–Tanimoto (range 0–1)

+ +

Kulczynski (range 0–1) 1

+
2 + +
Ochiai–Cosine (range 0–1)
√( + )( + )
Sorensen–Dice (range 0–1) 2
2 + +
Symmetric

Baulieu (range –1 to 1) 4( − )
( + + + )2
Simple matching (range 0–1) +
+ + +
Russel–Rao (hybrid, range 0–1)
+ + +

Figure 3.9 Example types of similarity coefficients.

3.3.2 Measures of Evaluation

Every data mining method needs to have a way to evaluate the results to ensure
quality of results from an algorithmic perspective as well as usefulness, ground
truth evaluation, and business evaluation.
For classification approaches where the accuracy of predicted labels is
being evaluated, measures such as accuracy, precision, recall, and F-measure
are used. These are based on validating the results against real-world or known
class labels. We compare the experimental finding with the labeled ground
truth for class 1 and class 2, for example anomaly or not an anomaly, to
measure the accuracy of the approach. These measures are based on true
positive (TP), which indicates that the label is classified correctly as an
anomaly; the true negative (TN), which indicates if the label is classified
correctly as not an anomaly; false positive (FP), which indicates that the node
is misclassified as an anomaly; and false negative (FN), which indicates that

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 41

DM method Example evaluation criteria

Classification Accuracy =
Precision =
Recall =
F-Measure =

K
Clustering Sum of squared error = ¦  ¦
i 1 x  Ci
dist 2 (mi , x )

Silhouette coefficient =
ao bo

Association rule mining P(X U Y )

Confidence XoY = P(Y|X) =
P(X)
P(X U Y )
Support = XoY =
P(X U Y )
Lift =
P(X) P(Y)

Figure 3.10 Example measures of evaluation.

the anomaly is misclassified as not an anomaly. There are several performance

measures that can be calculated based on the labeled anomalies as shown in
Figure 3.10.
There are additional measures that can be used, such as (1) true positive
rate = TP/(TP + FN), which shows the percentage of positive instances
correctly classified; (2) true negative rate = TN/(TN + FP), which shows the
percentage of the negative instances correctly classified; (3) false positive
rate = FP/(TN + FP); and (4) false negative rate = FN/(FN + TP). Overall
accuracy is a commonly accepted performance measure for classification
performance. However, accuracy alone can be misleading, for example in a
setting where anomalies are small percentage in the data. In such instances, a
combination of precision and recall in the form of F-measure can be used. The
additional measures allow us to validate the results from multiple perspectives.
In case of association rule mining, measures such as lift, confidence, and
support are used to measure the quality of results of association rules of the
form X!Y.
Lift is a correlation measure that assesses the degree to which the
occurrences of one lifts the occurrences of the other. The lift between the
occurrences of X and Y items in a transaction can be measured as lift(X!Y) =
P(X∪Y)/P(X)P(Y), where numerator defines the occurrences of the X and

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 42 Introduction to Data Mining

Y instances together and the denominator indicates the occurrences of the

instances independently. The preceding equation can be extended to more than
two domains.
If lift > 1, then X and Y are positively correlated, and if it is less than 1, then
they are negatively correlated. If the lift = 1, then X and Y are independent and
there is no correlation between them.
The support of the rule is the probability of P(X∪Y)/total number of tuples.
Confidence of the rule is the conditional probability of Y given
X (i.e., P(Y|X)) = P(X∪Y)/P(X). Each rule is associated with a confidence
that measures the degree of association.
Clustering quality can be measured through the sum of squared error (SSE),
which measures the deviation of every point in the cluster from its mean. The
SSE of all clusters together gives a measure of the quality of the clustering.
Another commonly used measure is the silhouette coefficient, which meas-
ures the shadow of a cluster over another cluster by measuring the distance
between each point of a cluster to its neighboring clusters.
These measures are shown in Figure 3.10.

3.3.3 Clustering Algorithms

Clustering is a technique used for finding similarly behaving groups in data or
to identify preexisting structure in the data. Ideally, a clustering method should
lead to a crisp demarcation of the data such that objects in the same group are
highly similar to each other and objects in different groups are very dissimilar
to each other, that is, intercluster distance should be maximized and
intracluster distance should be minimized.
In addition to distinguishing between various types of objects, clustering can
be used as an exploratory process and as a precursor to another mining task
such as anomaly detection. When a cluster is discovered, the objects outside of
this structure could be of interest (e.g., anomaly detection). In other cases, the
cluster itself could be of interest.
Clustering has been a well-studied problem in the statistical literature, and
there exist several well-developed algorithms already tried and tested
(Kotsiantis and Pintelas 2004, Jain 2010). Clustering approaches can be
roughly divided into several subclasses of algorithms, as shown in
Figure 3.11. Partitioning approaches divide the data into partitions with
optimal partitions being defined based on some heuristics as in the case of
K-Means. Hierarchical algorithms divide the data into a hierarchy starting
from all data points into one cluster and dividing up the data points such that
each data point is in its own cluster. Density-based approaches divide the data

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 43

K-means (MacQuffn, 1967)

PAM: Partioning around medoids/K-medoids (Kaufman & Rousseeuw 1987)

Partitioning
CLARA: Clustering large applications (Kaufman & Rousseeuw, 1990)
CLARANS: CLARA based on randomized search (Ng & Han 1994)

AGNES: Agglomerative nesting(Kaufman & Rousseeuw, 1990)

DIANA: Divisive analysis (Kaufman & Rousseeuw, 1990)

Hierarchical BIRCH: Balanced iterative reducing and clustering using hierarchies (Zhang et al. 1996)

CURE: Clustering using representatives ((Guha et al. 1998)

CHAMELEON (Karyphis et al. 1999)

DBSCAN: Density-based spatial clustering of applications with noise (Ester et al. 1996)

OPTICS: Ordering points to identify clustering structure (Ankerst et al. 1999)

Clustering

DENCLUE: Density-based clustering (Hinnenberg & Keim 1998)

GDBSCAN: Generalized DBSCAN (Sander et al. 1998)
Density based
PDBSCAN: Parallel DBSCAN (Xu et al. 1999)

DBCLASD: Distribution-based clustering of large spatial databases (Xu et al. 1998)

Jarvis Patrick algorithm (Jarvis & Patrick 1973)

FDC: Fast density-based clustering (Zhou et al. 1999)

CLASSIT (Gennari et al. 1989)

AutoClass (Stutz & Cheeseman 1996)

Model based
EM: Expectation maximization (Dempster et al. 1977)

RBMN: Recursive Bayesian multinets (Peña et al. 2002)

STING: Statistical information grid (Wang et al. 1997)

Grid based CLIQUE: clustering In Quest (Agrawal et al. 1998)

ROCK: Robust clustering using links (Guha et al. 2000)

Figure 3.11 Clustering approaches.

into groups based on density of the data points in a user-defined neighborhood

range, such as density-based spatial clustering of applications with noise
(DBSCAN). We focus on a few selected and well-used algorithms to illustrate
the concepts of clustering across some of these types of approaches.

3.3.3.1 Partitioning Algorithms: K-Means

The primary approach of partitioning algorithms, such as K-means, is to form
subgroups of the entire set of data, by making K partitions, based on a
similarity criterion. The data in the subgroups are grouped around a central
object such as a mean or medoid. However, they differ in the way that they
define the similarity criterion, central object, and convergence of clusters.
K-means (MacQueen 1967) is based on finding partitions in the data by
evaluating the distance of objects in the cluster to the centroid, which is the
mean of the data in the cluster. The algorithm, depicted in the example in
Figure 3.12, starts with a set of points and a K value for number of clusters to
form and seed centroids that can be selected from the data or randomly
generated. The algorithm begins by partitioning points around the seed cen-
troids such that the point is allocated to the centroid to which its distance is

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 44 Introduction to Data Mining

Figure 3.12 K-means clustering.

smallest. At the end of the first round, we have K clusters partitioned around
the seed centroids. Now the means or centroids of the newly formed clusters are
computed and the process is repeated to align the points to the newly computed
centroids. This process is iterated until there is no more reassignment of the
points moving from one cluster to the other. Thus, K-means works on a
heuristic-based partitioning rather than an optimal partitioning. The quality of
clusters can be evaluated using the sum of squared errors (Figure 3.10), which
computes the distance of every point in the cluster to its centroid. Intuitively, the
bigger the error, the more spread out the points are around the mean. However, if
we plot the SSE for K ¼ 1 to K ¼ n (n number of points), then SSE ranges from
a very large value (all points in one cluster) to 0 (every point in its own cluster).
The elbow of this plot provides an ideal value of K.
K-means has been a well-used and well-accepted clustering algorithm due to
its intuitive approach and interpretable output. However, K-means does not
work very well in the presence of outliers and does not form nonspherical or
free-form clusters.

3.3.3.2 Partitioning Algorithms: CLARANS

Clustering large applications with randomized search (CLARANS), proposed
by Ng and Han (1994), is a type of partitioning approach. It extends the ideas
proposed in partitioning around medoids (PAM-K-medoid) and clustering
large applications (CLARA) proposed by Kaufman and Rousseeuw (1987,
1990). This algorithm draws samples with randomness at each step of the

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 45

Figure 3.13 CLARANS process.

process. The quality of clustering is dependent on the sampling technique being

used. Similar to other partitioning techniques, K is fixed. Each cluster is
represented as a graph node containing the solution of clustering in terms of
medoids or the central representative object around which other objects are
clustered. The neighboring node differs by one object from its neighbor, and
they are connected by an edge. For example, a node with objects 2, 3, and 4 and
a node with objects 3, 4, and 5 differ in only one object so they are neighbors
and are joined by an edge. It selects a sample of neighbors unlike CLARA,
which selects a sample of the original data. For each node selected, the squared
error is calculated. CLARANS also uses the technique to determine the natural
clustering using the silhouette coefficient (Kauffman and Rousseeuw 1990).
The algorithm works as shown in Figure 3.13.
The edge graph looks as shown in Figure 3.13a. It starts by setting parameter
Current to an arbitrary node and set S to a random neighbor of Current. It finds the
cost difference between Current and S. If S has a lower cost, then it sets Current to
S and resets j ¼ 1. Now if S has cost more than Current, then a counter is
incremented. If J is less than the parameter maxneighbors, then the algorithm
finds a new random neighbor of Current; otherwise, it compares the cost of current
with minCost threshold; if it is less than mincost, then it sets a flag for bestnode to
current. If threshold for number of iterations is met, then bestnode is output;
otherwise, a new random current node is selected to repeat the process.

Silhouette Coefficient CLARANS uses the silhouette coefficient (proposed in

Kaufman and Rousseuw 1987) to identify the quality of clustering results in
terms of structure and its silhouette or shadow on other clusters.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 46 Introduction to Data Mining

Given object o in a cluster A, then a(o) is the average distance between the
object o and the objects in cluster A. b(o) is the average distance between the
object o and the objects in the second-closest cluster B. The silhouette of o is
then defined as
bð oÞ  að oÞ
SðoÞ ¼
max faðoÞ, bðoÞg
where
SðoÞ = –1: not well assigned, on average closer to members of B
SðoÞ = 0: in between A and B
SðoÞ = 1: good assignment of “o” to its cluster A
The silhouette coefficient of a clustering is thus the average silhouette of all the
objects in the cluster.
CLARANS provides an efficient way to deal with the issues related to PAM
(K-Medoids) that involves a large number of swapping operations, leading to
very high complexity. It also improves upon CLARA. CLARANS is similar to
CLARA in terms of sampling, but here the sampling is done at every step.
However, objects are stored in main memory and the focus on representative
objects is missing. This issue has been addressed using spatial data structures
such as R* trees and focusing techniques. Although CLARANS enables outlier
detection, it is not primarily geared toward the process of detecting anomalies.

3.3.3.3 Density-Based Algorithms: DBSCAN

Density-based algorithms can be useful for traditional cybersecurity data, such
as network traffic data, to identify dense clusters in Euclidean space. In
addition, they are also very useful for geospatial datasets. For example, in
studying threats over a geospatial region, this could be a useful method.
Similarly, for sensor networks that rely a lot on the spatial layouts, it may be
useful to identify similarly behaving dense clusters and potential anomalies
that stand apart from sensors in proximity.
Density-based algorithm for discovering clusters in large spatial databases
with noise, proposed by Ester et al. (1996), is the foundation of several other
density-based clustering algorithms. The basic idea is that we see clusters as
dense or sparse regions based on the number of points in the neighborhood
region of a point under consideration. They formalize these ideas with a certain
radius and a minimum number of points in the region (parameters: ε, MinPts).
The minimum number of points in a radius of ε facilitates the measurement of
density in that neighborhood. ε Neighborhood (in Euclidean point space) is the
neighborhood of a point that contains the (minPts) points. Variations of

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 47

Example: MinPts=3, ε=1 (b)

Core Points: p2, p4, p6, and p7
p10 p1 is directly density reachable from p2
p6 p3 is directly density reachable from p4
p1 p4
p5 is directly density reachable from p6
p7
p8 p6 is directly density reachable from p7
p3
p2 p5 p9
p8 is indirectly density reachable from p7 as p8
DDR from p6 and p6 DDR from p7 so by
transivity p8 is indirectly density reachable
(a)
from p7
p1, p3, and p5 are all density connected points
p10 is an outlier

Figure 3.14 DBSCAN process.

DBSCAN identify ε through computing K distance for every point, i.e., if

k = 2, the distance between the point and its second-nearest neighbor, and
sort this in descending order. The K-Distance is plotted and a valley or dip in
the K-distance plot along with a user input of the expected noise determines
the ε neighborhood. The algorithm defines density in terms of core points.
A core point is an object that has the minimum points in its neighborhood
defined by ε.
The algorithm works as shown in Figure 3.14. The algorithm begins by
identifying core points and forming clusters through directly density reachable
points from this core point. It then merges the cluster with such points. The
process terminates when no new point can be added.
Thus, a density-based cluster is a set of density-connected objects. Every
object not contained in any cluster is considered to be noise.
This approach identifies connected cluster structures, which can identify
arbitrarily shaped clusters unlike the other partitioning algorithms. It takes into
consideration a neighborhood of an object based on the distance and count of
the objects in the neighborhood. The initial ε distance is calculated using the
K nearest neighbor (K-NN) query with some level of user input. Once the
clusters start emerging, they can be merged using the concepts of density
connectivity and density reachability. However, this may lead to forming of
big global clusters (single link effect) without identifying local cluster struc-
tures. This does not truly reflect the density of the cluster in terms of the local
density structure. This algorithm is also sensitive to input parameters such as
the minimum objects in the neighborhood and the value of K for the K-NN
query. Moreover, this is approach is not well suited to high-dimensional
datasets since the density connectedness is not directly applicable and intuitive

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 48 Introduction to Data Mining

for all the dimensions. It could be possible that objects in some dimensions are
similar, but very dissimilar in some other dimensions. Several variations of
DBSCAN, as shown in Figure 3.11, have been proposed addressing these
limitations.

3.3.3.4 Density-Based Algorithms: OPTICS

Ordering points to identify the clustering structure (OPTICS), proposed by
Ankerst et al. (1999), is a density-based algorithm. Many real-time datasets
cannot be characterized by global density parameters. Different local densities
may be needed to identify clusters in different regions of the data space.
Mostly the clusters that are formed are large without any segregation of smaller
clusters that may be contained in these.
Density-based clustering may miss detecting these smaller structures within
a dense cluster as there can be several possible parameter values that can be
used. For a constant MinPts value, density-based clusters with respect to
higher density (or low radius) are completely contained in density-connected
sets with respect to a lower density (or high radius).
OPTICS extends the DBSCAN algorithm to generate clusters with multiple
densities simultaneously. Moreover, a specific order is followed in which
objects are processed. More dense clusters (low radius) are processed first. It
does not assign cluster memberships but just stores the order in which objects
are processed. This information is later used to assign memberships and is
provided as two values: core distance and reachability distance. Core distance
is the smallest distance ε0 between a point p and an object in its ε-neighbor-
hood, such that p would be a core object with respect to ε0 . In Figure 3.15,
distance core (o) makes o a core object. This distance in the example is less
than the ε neighborhood. Reachability distance of an object p with respect to
another object o is the smallest distance such that p is directly density reach-
able from o if o is a core object. The reachability distance cannot be smaller

Figure 3.15 Core and reachability distance.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 49

than the core distance of ‘o’. The reachability distance of p depends on the core
object with respect to which it is calculated.
A low reachability indicates an object within a cluster; high reachability
indicates a noise object or a jump from one cluster to another. Ankerst et al.
(1999) propose a technique to view the cluster structure graphically using the
reachability plot. This is a graph of the reachability distances for each object in
the cluster ordering. This cluster ordering is facilitated by a data structure,
namely the access control list. An object p1 is selected and the core distance is
selected around p1. This leads to the identification of objects in this neighbor-
hood, and their reachability distances are measured and stored in the access
control list. This process is performed iteratively in the neighboring objects as
well. Subsequently, the distance is plotted on the reachability plot.
OPTICS addresses the issue of not only traditional clustering but also
intrinsic clustering structure, which other clustering techniques ignore
(single-link effect). Many clustering techniques have global parameters produ-
cing uniform structures, and many times the smaller clusters are merged in the
nearby clusters connected by few points into one big cluster. Thus, OPTICS
addresses this issue of intrinsic clusters that helps in identifying dispersion of
data and correlation of the data. However, this approach is still somewhat
sensitive to input parameters such as minpts in a neighborhood K value for the
KNN query. There could be a scenario when a small cluster itself could be
outlying with respect to a global structure. This book does not address such
a scenario.

3.3.4 Classification
Given a collection of records, the goal of classification is to derive a model that
can assign a record to a class as accurately as possible. A class is often referred
to as a label; for example, a data point is anomaly or not. For example,
Anomaly and Not an anomaly can be two classes. Objects can also be
classified as belonging to multiple classes. Classification is a supervised
method that is based on a well-formed training set, i.e., prelabeled data with
samples of both classes, to create a classifier model that identifies previously
unseen observations in the test set for labeling each instance with a label based
on the model’s prediction. The overall process of classification is shown in
Figure 3.16. In general, the labeled data, which is a set of database tuples with
their corresponding class labels, can be divided into training and test data. In
the training phase, a classification algorithm builds a classifier (set of rules) for
each class using features or attributes in the data. In the testing phase, a set of
data tuples that are not overlapping with the training tuples is selected. Each

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 50 Introduction to Data Mining

Labeled data

Training
data

Classification
Accuracy
Test metrics
Model data
generation

Expert Real-time prediction

systems

Figure 3.16 Classification process.

test tuple is compared with the classification rules to determine its class. The
labels of the test tuples are reported along with percentage of correctly
classified labels to evaluate the accuracy of the model in previously unseen
(labels in the) data. As the model accuracy is evaluated and rules are perfected
for labeling previously unseen instances, these rules can be used for future
predictions. One way to do that is to maintain a knowledge base or expert
system with these discovered rules, and as incoming data are observed to
match these rules, the labels for them are predicted.
Several classification algorithms have been proposed that approach the
classification modeling using different mechanisms (see Figure 3.17 and
Kotsiantis et al. 2006). For example, the decision tree algorithms provide a
set of rules in the form of a decision tree to provide labels based on conditions
in the tree branches. Bayesian models provide a probability value of an
instance belonging to a class. Function-based methods provide functions for
demarcations in the data such that the data are clearly divided between classes.
In addition to some of the basic methods there are also classifier combinations,
namely ensemble methods, which combine the classifiers across multiple
samples of the training data. These methods are designed to increase the
accuracy of the classification task by training several different classifiers and
combining their results to output a class label. A good analogy is when humans
seek the input of several experts before making an important decision. There
are several studies suggesting that diversity among classifier models is a
required condition for having a good ensemble classifier (Li 2008, Garcia-
Teodoro et al. 2009). These studies also suggest that the base classifier in the

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 51

ID3 (Quinlan 1979)

Decision tree based
C4.5 (Quinlan 1993)

Naïve Bayes (Maron and Kuhns 1960)

Bayesian
Bayesian Network (Cooper & Herskovits 1992)

RIPPER K ( Cohen 1995)

Rule based
Classification

PART: partial decision trees (Frank & Witten 1998)

Support vector machine (Cortes & Vapnik 1995)

Funcon based
Logistic regression (Press & Wilson 1978)

K-nearest neighbors (Cover & Hart 1967)

Lazy learning/ instance-based algorithms
K* (Cleary & Trigg 1995)

Bagging (Breiman, 1996)

Boosting: ADA boost (Freund & Schapire, 1996)

Classifier combinations
(ensemble methods) Random forest (Beriman 2001)

MILES: multiclass imbalanced learning in ensembles through selective sampling (Azari et al. 2017)

Figure 3.17 Classification techniques.

ensemble should be a weak learner in order to get the best results out of an
ensemble. A classification algorithm is called a weak learner if a small change
in the training data produces big difference in the induced classifier mapping.
We next discuss a few relevant classification methods.

3.3.4.1 Decision Tree Based: C4.5

Proposed by Quinlan (1993), C4.5 is a well-accepted classification method that
provides an intuitive model in the form of a decision tree based on the
training set.
Let us consider a small internet trace data, shown in Figure 3.18a, including
attributes such as day of the week, average size of packets, and average size of
the flows, and labeled with attack information as to whether such traces
contained attacks or not. The task would be to classify whether there will be
an attack or not given the type of trace data.
The C4.5 decision tree algorithm starts by evaluating attributes to identify
the attribute that gives the most information for making a decision for labeling
with the class. The decision tree provides a series of rules to identify which
attribute should be evaluated to come up with the label for a record where the

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 52 Introduction to Data Mining

Figure 3.18 Example C4.5 execution.

label is unknown. These rules form the branches in the decision tree. The
purity of the attributes to make a decision split in the tree is computed using
measures such as entropy. The entropy of a particular node, InfoA (correspond-
ing to an attribute), in a tree is the sum, over all classes represented in the node,
of the proportion of records belonging to a particular class. The entropy
computation of the data, InfoD and attribute, InfoA (e.g., Info (packets)) is
shown in Figure 3.19. When entropy reduction is chosen as a splitting criter-
ion, the algorithm searches for the split that reduces entropy or equivalently the
split that increases information, learned from that attribute, by the greatest
amount. If a leaf in a decision tree is entirely pure, then classes in the leaf can
be clearly described, that is, they all fall in the same class. If a leaf is highly
impure, then describing it is much more complex. Entropy helps us quantify
the concept of purity of the node. The best split is one that does the best job of
separating the records into groups where a single class predominates the group
of records in that branch.
The attribute that reduces the entropy or maximizes the gain (GainA =
InfoD – InfoA) is selected as a splitting attribute. Once the first splitting
attribute is decided, then recursively the process is repeated in all the attributes
of the subset of the data remaining, resulting in a decision tree such as is shown

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 53

Figure 3.19 Entropy computation for a data set and attributes.

in Figure 3.18d. Several other measures for attribute selection measures can be
used (such as evaluated by Dash and Liu 1997).

3.3.4.2 Classifier Combinations: MILES

In massive datasets, patterns can often be hidden and hard to discern with
individual classifiers. Moreover, if different parts of the data can contribute to a
different aspect of the pattern, a single classifier may not be useful. Let us
consider an example of phishing detection (Azari et al. 2020). If we look at
email messages to predict whether an email could be phishing or not, we can
look at the header information or the body of the email to get different
elements of potential phishing. Moreover, the data may have imbalance in
terms of phishing emails and normal emails. In this case, using an ensemble-
based method, which can handle class imbalance, can be very useful to
consider different parts of the email to flag a phishing message.
Multiclass imbalanced learning in ensembles through selective sampling
(MILES) (Azari and Janeja 2017) is an approach that comprises the following
three components: (i) creating training sets using clustering-based selective
sampling, (ii) rebalancing the training sets, and (iii) combining the classifiers
and forming the ensemble. In ensemble learning, it is important to form

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 54 Introduction to Data Mining

Figure 3.20 Selection of training sets in MILES.

learning sets from different regions in the feature space. This is particularly
important if the data are imbalanced and have distribution of multiple classes
with some majority and minority classes in the feature space. Learning sets that
are dissimilar can produce more diverse results. MILES adopts clustering as a
stratified sampling approach to form a set of learning sets that are different
from each other, and at the same time each one of them is representative of the
original learning set D. It uses K cluster sets to selectively choose the training
samples. It utilizes the concept of split ratio (SR) to split the examples in
clusters into three strata (near, far, and mid) around the cluster centroids as
shown in Figure 3.20.
First, a base classifier H is selected. Here H can be decision tree classifier
such as C4.5, described earlier. Next, the base classifier is trained with training
sets N, F, and M in order to form a set of learned classifiers
π ¼ fH N , H M , H F g. Then the maximum vote rule is used to combine the
votes of the classifiers and build the ensemble H.  H
 is the original ensemble
MILES, which is created by combining the votes that come from
 Under is the variation of MILES combined with random
fH N , H M , H F g. The H
 Over is the
undersampling, where fN Under , M Under , F Under g are utilized, and H
variation of MILES combined with random oversampling, where
fN over , M over , F over g are utilized. Undersampling is useful when there are too
many instances of a class, and oversampling is useful when there are too few
samples of the class.
The training examples could potentially be selected into more than three
training sets, however: (1) If the number of splits is increased to very highly
granular demarcations, then there will be splitting of already identified similar
groups so the diversity of the training set will go down. (2) An ensemble

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 55

requires at least three classifiers to form a maximum vote combination rule, so

a minimum of three sets can be sufficient. (3) A smaller ensemble is faster and
more efficient because there are fewer classifiers to train. Results with MILES
have shown a size of three produces the highest diversity in terms of a diversity
statistic (Q statistic) and also the best classification performance for the
proposed ensemble.

3.3.4.3 Other Approaches

Bayesian methods, such as naive Bayes (Maron and Kuhns 1960), provide a
probabilistic label for a tuple of the test data based on the training set. The
naïve Bayes classifier considers each attribute and class as a random variable.
Now given a tuple with attribute values, the goal is to predict probability of
class label C. Specifically, it is based on finding the class probability given the
attribute values that can be framed as a conditional probability P(C|Attributes).
The naïve Bayes algorithm assumes independence of attributes such that the
P(C|Attributes) becomes P(C|A1) * P(C|A2) * . . . * P(C|An). The idea is to
find class C, which maximizes this probability. While this is a relatively
intuitive method, it relies on the independence assumption. Bayesian networks,
on the other hand, represents dependency among attributes.
Another popular function-based method is support vector machines (SVM)
(Cortes and Vapnik 1995). SVM relies on the identification of support vectors
that linearly partition the data based on the class labels. It looks for a maximal
marginal hyperplane, which maximizes this separation. If the data are not
linearly separable, such that the decision boundary is not clearly defined, it
uses a nonlinear mapping of attributes to transform the original training data
attributes to a higher dimensional space for a well-defined classification.

3.3.5 Pattern Mining: Association Rule Mining

Pattern essentially refers to occurrence of multiple objects in a certain combin-
ation that could lead to discovery of an implicit interaction. A frequent pattern
refers to frequently co-occurring objects. A rare pattern essentially refers to a
combination of objects not normally occurring and that could lead to identifi-
cation of a possible unusual interaction between the objects.
Events become relevant when they occur together in a sequence or as
groups. Some approaches deal with this as a causal problem and therefore rely
on conditional probability of events to determine the occurrence. If events are
seen as itemsets in transactions, focus has been put on finding conditional
relationships such that presence of A causes B with a certain probability.
However, there might be events where there might be no implied causal

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 56 Introduction to Data Mining

Multilevel Apriori (Han & Fu 1995)

Multidimensional Apriori (Srikant and Agrawal 1994)

Apriori (Agrawal & Srikant 1994)
Apriori TID, hybrid (Agrawal and Srikant 1994)

Frequent pattern mining Constraint based (Ng et al. 1998, Pei & Han 2000)

FP tree based (Han et al. 2000)

Tree based
Rapid association rule mining (Das et al. 2001)

Actionable patterns in events data (Ma & Hellerstein, 2002)

Mutually dependent patterns (Ma & Hellerstein 2001a)

Pattern mining

Partially periodic event pattern (Ma & Hellerstein 2001b)

Weighted ARM (Tao et al. 2003)

Rare pattern mining Relative support (RSAA) (Yao et al. 2003)

Variable support based Apriori inverse (Koh & Rountree 2005)

Multi support (MSAPriori) (Liu et al. 1999)

Adaptive Apriori (Wang et al. 2003)

Min hashing (Cohen et al. 2001)

Without support
Confidence-based pruning (Wang et al. 2001)

Rare event sequences (Hu et al. 2014)

Sequence mining Structural paerns (Lee et al. 2015)

Spaal colocaons (Huang et al. 2006)

Figure 3.21 Pattern mining approaches.

relationships between them, but the relationship is based purely on co-occur-

rence. For example, events might have no implied or evident causal relation-
ship but a co-occurrence relationship based on frequency of occurrence
(Agrawal et al. 1994) based on a certain degree of support or confidence
defining the probability of one event given the other. A categorization of the
various pattern mining approaches is shown in Figure 3.21. Several surveys of
patterns such as associations (Zhao and Bhowmick 2003) and rare patterns
(Koh and Ravana 2016) provide an overview of these approaches.
We discuss two approaches to show a contrast between the discovery of
patterns based on frequency of occurrence versus the discovery of patterns
where the frequency of events is very less and thus will require other infor-
mation for discovery of the rare pattern.

3.3.5.1 Frequent Pattern Mining: Apriori

The Apriori algorithm (Agarwal and Srikant 1994) is one of the most well-
established data mining algorithms and the basis for the work of several
frequent pattern mining approaches. It provides the approach to discover
association rules. It is used to identify relationships among a set of items in a

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 57

Figure 3.22 Example run for the Apriori algorithm.

database. An example is shown in Figure 3.22, with a set of transactions

containing IP addresses and levels of alerts raised in an IDS (high, med, and
low). Now given a set of transactions in the database, if IP1 and high are two
items, the association rule IP1 ! high means that whenever an item IP1 occurs
in a transaction, then high also occurs with a quantified probability. The
probability or confidence threshold can be defined as the percentage of trans-
actions containing high and IP1 with respect to the percentage of transactions
containing just IP1. This can be seen in terms of conditional probability where
S
P(high|IP1) = P(IP1 high) / P(IP1).
The strength of the rule can also be quantified in terms of support, where
support is the number of transactions that contain a certain item; therefore,
support of X is the percentage of transactions containing X in the entire set
S
of transactions. Confidence is determined as Support (IP1 high) / Support
(IP1), which gives the confidence and support for the rule for IP1 ! high.
The rule IP1 ! high is not the same as high ! IP1 as the confidence
will change.
The Apriori algorithm is based on finding frequent itemsets and then
discovering and quantifying the association rules. One major contribution of
the Apriori algorithm is that it provides an efficient mechanism of discovering

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 58 Introduction to Data Mining

frequent itemsets. A subset of a frequent itemset must also be a frequent

itemset, i.e., if {IP1, high} is a frequent itemset, both {IP1} and {high} should
be a frequent itemset. It iteratively finds frequent itemsets with cardinality from
1 to k (k-itemset).
Apriori generates the candidate itemsets by joining the itemsets with large
support from the previous pass and deleting the subsets that have small support
from the previous pass. By only considering the itemsets with large support,
the number of candidate itemsets is significantly reduced. In the first pass,
itemsets with only one item are counted. The itemsets with higher support are
used to generate the candidate sets of the second pass. Once the candidate
itemsets are found, their supports are calculated to discover the items sets of
size two with large support, and so on. This process terminates when no new
itemsets are found with large support. Here min support is predetermined as a
user-defined threshold. Similarly, a confidence threshold is predetermined by
the user. For example, in Figure 3.22, the min support is 0.25, and min
confidence is 0.6. The rule IP1 ! high has confidence 1 and support 0.38
and is thus a strong rule.
The Apriori algorithm proposes a highly intuitive technique for identifying
association rules. Several variations, as shown in Figure 3.21, with efficient
versions are proposed, such as FP tree and Rapid ARM. For example, multi-
level associations (Han and Fu 1995) deal with levels of items (such as the
alert level high may have subtypes), and multidimensional associations
(Srikant and Agarwal 1996) discover rules in multidimensional datasets with
different types of discretization for the attributes. Constraint-based rules (Ng
et al. 1998) discuss imposing constraints through user input on the rules and
pruning-based rules on constraints. Spatial colocations (Huang et al. 2006)
discovers rules in spatial datasets that identify colocations of different itemsets
in space
Apriori does not address discovery of rare patterns. If we wish to identify the
rare patterns, they would be ones with very low support, leading to an
extremely large number of meaningless rules. There is a whole class of
algorithms proposed for rare pattern mining (as shown in Figure 3.21) that
devise variations of the support constraints for rare pattern mining.

3.3.5.2 Rare Pattern Mining: Mining Mutually Dependent Patterns

Mining mutually dependent patterns (proposed by Ma and Hellerstein 2001a)
introduces data mining for event bursts, periodicities, and mutually dependent
events based on event management used in computer and communication
systems. The main focus is on the mutually dependent patterns. Mutual
dependencies are discovered as M-patterns or mutually dependent patterns

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 3.3 Data Mining 59

Figure 3.23 M-pattern example.

for mining events that occur together and are mutually dependent. Thus,
occurrence of one event is also an indication of the occurrence of the other
event by a certain probability. The detection of M-patterns can be illustrated
using the example shown in Figure 3.23. Let us say we have event dataset
(transactional data) from IDS logs indicating IP address and alarm text (high,
med, and low).
The algorithm IsMPattern is initiated with the value for minimum depend-
ence threshold, such that 0  minp  1. An expression E (such as IP1, high) is
evaluated in terms of individual supports of both directional rules
ðIP1 ! high and high ! IP1Þ. M-pattern relies on a strong bidirectional
dependence; thus it enables the algorithm to discover items even if the supports
are low. However, when generalized for larger minp, the algorithm also
discovers frequent patterns similar to association rule mining. The itemsets
considered for mining for such event sets or patterns do not eliminate low-
support items (nonzero). However, they do eliminate itemsets that may not be
m-patterns, as shown in Figure 3.23. Now, given the itemsets, let us consider
E = IP1 & high, then PðIP1 ! highÞ ¼ 1 and ðhigh ! IP1Þ ¼ 0:6. If the minp
threshold is 0.5, this set is an m-pattern. Similar to the example shown in
Figure 3.22 (min support threshold = 0.25, and min confidence = 0.6), this is
also a strong confidence rule. However, when we look at E = (IP6, med), then
this is an m-pattern but not a strong confidence rule.
The process of identification of rare patterns is based on a probability level.
Thus, even if the supports are low it can identify the rarity of the pattern. It also
discovers rare patterns such that if an expression is an m-pattern, then the
subsets of the expressions are also m-patterns.

https://doi.org/10.1017/9781108231954.004 Published online by Cambridge University Press

 4
Big Data Analytics and Its Need
for Cybersecurity
Advanced DM and Complex Data Types
from Cybersecurity Perspective

4.1 What Is Big Data?

The global internet population has grown from 2.1 billion in 2012 to 4.6 billion
in 2021 (Statista 2021). These users are generating myriad datasets along with
the network traffic to move these data on the internet. We produce 2.5
quintillion of data every day (Domo 2017), and $2 trillion is projected to be
spent in e-commerce is 2020 (Statista 2020). The nature of big data comes
from the complexity of the data and the mechanisms required to bus, analyze,
and find insights from the data. Several initiatives have been established to
define what we mean by big data (such as Bayer and Laney 2012, Ward and
Barker 2013, and NIST 2015). Based on the different types of definitions, there
are three or more of the qualities of the big data also referred to as V’s, as
shown in Figure 4.1. The most known qualities of big data are the volume,
velocity, and variety of the data.
One thing that is clear that big data does not just refer to the volume or size
of the data. Data can be complex and have the following qualities:
• Velocity: data generated at a rapid pace
• Variety: data consisting of multiple types of heterogeneous data
• Veracity: data that provide trustworthy insights into the domain function
• Value: data that may be able to generate revenue or provide other benefits
• Venue: data that are dynamic with respect to the location
• Volume: large amounts of data being generated
• Variability: some aspect of data changing over time
Big data are the dataset that has a combination of these qualities. In some cases,
data that may not be particularly large but has many of these other qualities, such

60

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.2 Big Data in Cybersecurity 61

Figure 4.1 The V’s of big data.

as variability, variety, and value, may also be considered big data. The term “big”
may refer to the big aspect of the overall complexity of the data. If a dataset is
highly unusual in terms of its complexity, it may qualify as big data under one of
the big data qualities.
In the cybersecurity domain, there are several applications that qualify as big
data. If we consider network data traffic, it is massive; for example, in a midsize
organization, it can range into petabytes per second. If we consider Internet of
Things (IoT), the complexity of data is based on the velocity and variety and in
some cases also the volume. A wireless sensor network or a network of IoT
monitored through distributed IDS (Otoum et al. 2019) can generate a constant
stream of data originating from multiple sensors over time, producing high
variability in the data. This provides a complex and rich dataset.
The value from these big datasets can be derived by sifting through these
complex datasets to derive insightful patterns.

4.2 Big Data in Cybersecurity

According to the World Fact Book maintained by the CIA (2021), the number
of internet users is now in the billions. Figure 4.2 shows the top 25 countries in
terms of internet users.
It is clear that a vast number of users on the internet are generating traffic
from computers, mobile devices, and other things or devices. If we try to
quantify this network traffic data, it is truly complex in terms of the variety of
the data generated, volume of the data, and velocity at which it is generated. In
addition to the individual users, there are also organizations generating the data
and network traffic.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 62 Big Data Analytics and Its Need for Cybersecurity

Internet users
800,000,000

700,000,000

600,000,000

500,000,000

400,000,000

300,000,000

200,000,000

100,000,000

0
es

ria
sia
na

n
a

y
il

an
di

pa
si

ic
az

ge
at
hi

ne
us

ex
In

m
Br

Ja
St
C

Ni
do

M
R

er
d

In

G
te
ni
U

Figure 4.2 Number of internet users by country est. 2018.

Interestingly, the number of attacks (for example, the web attacks discussed
in Akamai 2016) sourced worldwide paints a slightly different picture, as
shown in Figure 4.3.
While majority of the attackers may come from countries with more internet
users, in addition they may also come from the countries with fewer internet
users. For example, the number of internet users in Lithuania in 2018 was a
fraction of that of the United States and is 116th in the list of countries ranked
by the number of internet users. Despite this, Lithuania was placed eighth in
terms of the countries sourcing web application attacks. Similarly, more recent
reports (Akamai 2019) have listed other countries in the top sources of web
attacks even as they do not appear to be a country with highest number of
internet users.
The amount of internet traffic generated may not be an indicator of attack
traffic. There are other factors at play in identifying cyberattacks on the larger
world stage taking into account not only the number of internet users but
sociopolitical factors, crime rates, and other cyberattacks. For example, a
simple search of Lithuania and cyberattacks reveals that Lithuania is also at
the receiving end of many major cyberattacks. Lithuania led a cyber shield

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.3 Landscape of Big Data Technologies 63

12,00,00,000
10,00,00,000
8,00,00,000
6,00,00,000
4,00,00,000 Aacks
2,00,00,000
0
US

UK
Figure 4.3 Sources of web attacks from countries (2016).

exercise to practice procedures for protecting the cyber infrastructure (L24,

2016). Thus, it is not clear whether the sourced attacks, where Lithuania
ranked eighth across countries in the fourth quarter of 2016, are more defen-
sive or offensive.
It is clear that a deeper dive can reveal many more insights into the real state
of affairs on the internet. If we consider an organization and study trends of
usage and simple rules to analyze traffic, we may miss insights of potential
advanced persistent threats. Such threats can be evaluated by not just looking
at data in isolation but in combination with user behaviors, sociopolitical
factors, context of time, and potentially many other contextual features. This
is truly a big data problem as it brings in the volume, velocity, variety,
variability, and venue and can lead to great value with high veracity of the
results. Here testing the veracity is equally important to generating value.
The challenge is to sift through the internet traffic and identify attacks. This
can be at the device level, user level, or the network level at large.

4.3 Landscape of Big Data Technologies

Big data technologies can be selected based on criteria that fit a business need
(as highlighted in Cloud Security Alliance 2014). The set of criteria is shown
in Figure 4.4.
These criteria include the latency in response time, whether the data are
structured, if a SQL-like environment is required, types of analytics required,
specific types of visualizations that may be needed, and the security and

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 64 Big Data Analytics and Its Need for Cybersecurity

Real time
Latency Near real time
Batch processing
Structured
Structure Semistructured
Unstructured

Batch
Compute infrastructure
streaming

Relational
Factors affecng big data Storage infrastructure No SQL
technology selecon New SQL

Analytics Supervised unsupervised semisupervised

Spatial
Visualization
Interactive abstract/summary

Data privacy
Security and Security infrastructure
privacy Data management
Integrity and reactive security

Figure 4.4 Factors affecting big data technology selection.

privacy needs of the business. Several big data technologies can be selected
based on these criteria. A pictorial representation of some of these technologies
is shown in Figure 4.5 (Grover et al. 2015 and Roman 2017). For example, if
the need of the business is low latency and SQL-like processing, then the
business can select tools that have a quick turnaround time for queries in
massive datasets. Of the several big data analytics frameworks present in the
market, the business can select the tools that provide massively parallel
processing (MPP), such as engines on top of Hadoop that have high SQL-
like functionality and portability, such as Apache Hive, Facebook Presto,
Apache Drill, and Apache Spark. Out of these, Presto and Spark have been
shown (Grover et al. 2015) to produce better outcomes for SQL-
like processing.

4.4 Mahout and Spark Comparative Analysis

Of the big data technologies, Mahout, which powered several machine learn-
ing applications,1 and Spark have been identified as a major big data technolo-
gies (Meng et al. 2016). For massive cybersecurity datasets such as traffic data

1
https://mahout.apache.org/general/powered-by-mahout.html.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.4 Mahout and Spark Comparative Analysis 65

Figure 4.5 The big data technology ecosystem.

or sensor network data, even small differences in speedup can mean the
difference in millions of dollars. Thus, in selecting a big data technology, it
is also essential to study the scale-ups with the type of technology used.
Let us consider these two big data technologies, which are used for machine
learning and data mining. Apache Mahout2 works on top of the Hadoop map
reduce framework to provide a data analytics library. Spark provides its own
map reduce framework and provides an integrated Spark machine learning
library (MLLib) for machine learning and analytics tasks.
Hadoop is an open-source mechanism for parallel computation in massive
datasets. It provides the Hadoop File System (HDFS) and the map reduce
framework, which provides the mechanism for parallel processing, as will be
discussed in an example case study for change detection. The key steps require
breaking down the data and processing the data in parallel through mappers
that work on key and value pairs. The pairs are then reduced and the

2
http://mahout.apache.org/.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 66 Big Data Analytics and Its Need for Cybersecurity

Table 4.1 Sample comparisons of speedup between Spark and Mahout

Spark over
Mahout
speedup
Data size Nodes Method Citation achieved
62 MB– 1–2 K-means Gopalani 2 times
1,240 MB and Arora
2015
200–1,000 16 Logistic regression, ridge Meng et al. 3 times
mil data regression, decision tree, ALS 2016
points K-means
50K–50 mil 10 C4.5 K-means Wang et al. 2–5 times
points 2014
Upto 7.5 GB 1 Word count 3–5 times
0.5–4 GB 4 K-means Misal et al. 2 times
2016

aggregated information is produced as a final result. Spark provides its own

implementation of the map reduce processing but also additional inbuilt SQL
and machine learning library (MLLib). For Hadoop to perform the machine
learning, Mahout can work on top of Hadoop. The key speedup in Spark is due
to the resilient distributed datasets (RDD), which are stored in memory; they
are also resilient to loss, as they can rebuild lost data based on where they
originate from. Spark also allows for machine learning algorithms to retrieve
the data from memory iteratively.
Comparative performance of data mining and machine learning algorithms on
Spark and Mahout with Spark have shown several orders in speedup. For example
(Gopalani and Arora 2015), Spark has been shown to have roughly two times
greater speedup over Mahout. Table 4.1 outlines some comparative analysis found
in the literature. In general, Spark achieved a speed up of about two times greater
than Mahout across various datasets and techniques. This is just one type of
comparison when low latency and speedup are desired; there may be other
comparisons that need to be evaluated when other analysis attributes are relevant,
as shown in Figure 4.4. In mission-critical systems where dependable real-time
responses are needed, these speedups can translate to real-world impacts in
analyzing anomalies in massive network traffic datasets (Kato and Klyuev 2017).

4.5 Complex Nature of Data

The premise of achieving a high level of analysis and mining in big data is a
good understanding of the nature and complexity of the data. The mining in

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.5 Complex Nature of Data 67

large heterogeneous data is increasingly becoming challenging due to the

complexity of the data. Many data mining algorithms have been proposed
for traditional (numeric, categorical) datasets. However, the nature of the data
itself renders a challenge to the data mining. Mainly such problems can be
exemplified in the mining of high-dimensional datasets. The quality of the
mining results is affected adversely as the number of dimensions increases.
Here dimension refers to the attributes of an object. This is especially true in
outlier detection, i.e., the identification of anomalous objects. In higher dimen-
sions, data become sparse, and points tend to become equidistant. This
adversely affects the methods based on density of points, unless data follow
certain simple distributions. Thus, the outliers could be hidden in high dimen-
sions. Another fact is that for outlier detection a subset of the entire attribute set
could be used to detect or indicate the outliers. In addition, the nature of the
data – whether it is spatial, temporal, graph, or unstructured data – can also
make the mining process fairly complex. In each of these types of complex
datasets, the data preprocessing, which is already the most intensive part of any
data mining project, can become even more involved and challenging.

4.5.1 Nature of Data: Spatial Data

Spatial data mining deals with identification of nontrivial and useful know-
ledge discovery in a spatial database, where spatial (point, lines, polygons,
location, pixel data) and nonspatial data (e.g., population count) are stored.
Unlike traditional data mining, it is important to address the nature of spatial
data, which renders new challenges in the inherent spatial autocorrelation and
heterogeneity in the data. Spatial data are seen in the cybersecurity domain in
sensor networks placed in a region. These are particularly relevant to cyber-
physical systems (Tan et al. 2009), which have interactions with several
physical sensors that are impacted by the regional variables and also nonspatial
elements, such as the computer networks involved in the communication and
analysis of the data. CPS provides a very comprehensive example of how the
cyber elements interact with the physical elements such that to perform any
type of knowledge discovery, the physical elements along with the environ-
mental variables impacting the physical locations need to be
modeled correctly.
Thus, understanding the nature of data is even more important for spatial
data mining because of the inherent complexities of the spatial data them-
selves. Spatial data consist pixel data, data pertaining to space-
multidimensional points, lines, rectangles, polygons, cubes, and other geomet-
ric objects. The data are in very large volumes; it would essentially take an

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 68 Big Data Analytics and Its Need for Cybersecurity

infinitely large database to capture the real world precisely. Spatial data pose
more of a challenge due to the type of data, volume of data, and the fact that
there is a correlation between the objects and their neighbor or neighborhood.
There can be relationships between spatial objects such as topological rela-
tionships, direction relationships, metric relationships, or complex relation-
ships, which are based on these basic relationships. Moreover, the data could
depict temporal changes, which is inherent in the data themselves. The same
spatial data can also be represented in different ways, i.e., raster and vector
format. For example, georeferenced data include spatial location in terms of
latitude, longitude and other information about spatial and nonspatial attri-
butes. This essentially means that there is some preprocessing on the raw data
such as satellite imagery or tiger files to extract the georeferenced attributes
and some other nonspatial attributes such as vegetation type, land-use
type, etc.
Another view where spatial data can benefit cybersecurity applications is
in characterization of the region with socioeconomic data and cybercrime
data. The approach proposed by Keivan and Koochakzadeh (2012) first
creates a characterization to determine which cities, areas, or countries may
be more likely to send a malicious request. This is based on the number of
complaints received at each location for different types of cybercrimes.
Locations with a number of complaints above the median are classified as
“critical,” while locations with a number of complaints below the median
are classified as “safe.” In addition, demographic data such as housing
(occupied, vacant, owned/rented, etc.), population (gender, age, etc.) are
also used in the characterization. Based on the complaints by location and
the demographic information, they predict the potential of a location as a
cybercrime origin.
Ferebee et al. (2012) utilize spatial data to produce a cybersecurity storm
map that combines several data streams and provides a geospatial interface to
visualize the location of cyber activity. This concept has been studied at length
in other types of disaster recovery domains but not so much in cybersecurity.
Banerjee et al. (2012) highlights the importance of looking at the
cyberattacks and their impact and interaction with the physical environment.
Since CPS is not limited to sophisticated users, usability is also a big consider-
ation, and this can impact the value of the system and eventually the veracity of
the findings if the usability and interpretability of the systems is low. This work
builds on the cyberphysical security (CypSec) concepts that discuss the map-
ping of the cyber and physical components to provide an environmentally
coupled solution (Venkatasubramanian et al. 2011). Additional examples of
using spatial data are described in Chapter 11.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.5 Complex Nature of Data 69

4.5.2 Nature of Data: Graph Data

Graphs represent a set of nodes and links between the nodes. Here nodes can
be IP addresses (source and destination) and the links can be the packets or
communication sent between the two IP addresses. Another type of graph
could be a router network, where the nodes are routers, and links are the
possible paths from the router table. Graphs have been a well-studied area.
With the proliferation of computer network data, graphs are becoming massive
in size and are constantly evolving. This is an example where big data
techniques can facilitate the discovery of novel insights. For example, how
does a graph for computer network traffic change over time in terms of node
level? How can graph-level properties (Namayanja and Janeja 2014) help
determine potential events? Thus, if a node is highly connected, it is an
important node. Similarly, we can study the behavior of the network data
whether it increases the diameter of a graph or density of the graph (Leskovec
et al. 2005). These properties have direct implications on managing the
networks, identifying critical events such as cyberattacks, and evaluating the
impact of events on the graphs.

4.5.3 Nature of Data: Other Types of Complex Data

Cybersecurity data such as network traffic data also has a temporal element.
Later in this chapter, we discuss how traffic data evolving over time can
provide novel insights. Each data point has a time stamp, and if considered
sequentially, the patterns can have additional meaning in terms of the
evolving patterns, such as periodicity or recurrence of patterns over time.
This is particularly useful in studying advanced persistent threats (APT).
Spatial and temporal data analysis for cybersecurity is further discussed in
Chapter 8.
Another complex dataset is unstructured data such as text (for example,
vulnerability descriptions in the NVD3). There is a lot of rich knowledge about
vulnerabilities in this text description. To analyze these data, a lot of prepro-
cessing is required to be able to discover patterns in the text within and across
vulnerabilities. One example is to break down the text into keywords and
translate each vulnerability into a binary-valued feature vector based on the
presence or absence of the keywords. This can help identify similar or dissimi-
lar vulnerabilities (as discussed in similarity matching in Chapter 3).

3
https://nvd.nist.gov/.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 70 Big Data Analytics and Its Need for Cybersecurity

Other types of complex datasets include audio, video, and image datasets,
which are increasingly becoming a key data source of anomaly detection due
to propagation of deep fakes (Agarwal et al. 2019).

4.6 Where Does Analytics Fit in for Cybersecurity?

Traditional cybersecurity methods looking at network traffic data focus on
Intrusion detection, which aims at identifying threats of unauthorized access
using signatures of unauthorized access or attacks. Data analytics can go
beyond this signature-based discovery to identify complex types of attacks
that may be hidden in massive datasets or may be spread out over time and
multiple networks (similar to threat hunting, as mentioned in Chapter 1). Here
we highlight two examples that show the need of using big data analytics in
cybersecurity:
(1) Sampling based change discovery in massive networks (Namayanja and
Janeja 2015)
(2) Big distributed intrusion detection system (Janeja et al. 2014)
We focus on these examples to illustrate the need for big data technologies and
the complexities of the datasets involved.

4.6.1 Change Detection in Massive Traffic Datasets

Computer networks can be seen as graphs of communications between nodes,
where each node is represented by an IP address. If we consider a snapshot of
network traffic data, we can generate a communication graph from it. Studying
these snapshots through time, we can evaluate whether there have been node-
level or network-level changes in the graphs indicating potential cyberattacks,
as shown in Figure 4.6 (adapted from Namayanja and Janeja 2015). Such
detection of unusual time periods can be used to identify major network traffic
changes due to cyber threats. Moreover, this can also help identify nodes that
may have been targeted since their node connectivity changed over time.
Lastly, this type of analysis can also feed into a predictive model to predict
temporal patterns where periodic changes in the network may impact
network performance.
Given that network communication can be massive, it is important to
identify a subset of this traffic to evaluate. This can be done by targeted
sampling. One way is to select the top central nodes where nodes with a very
high degree of connectivity represent the major communication in the traffic data.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.6 Where Does Analytics Fit in for Cybersecurity? 71

Figure 4.6 Multilevel change detection.

Source: Namayanja and Janeja (2015).

If we see these central nodes are consistent through time, represented by bins or
intervals of time, then we can say that these nodes are consistent nodes. On the
other hand, if these nodes are not present in all time periods, then they are
inconsistent nodes.
If there are time periods where several of the central nodes become incon-
sistent, then this time period is a potential time point where an event affected
these nodes and may require further investigation. The subgraphs for the time
intervals can also be studied to evaluate graph-level properties. Thus, if the
densification or diameter does not follow the network properties during time
periods, it can again be an indicator of a network-level event at the time points.
Cybersecurity through network and graph data is further discussed in
Chapter 9.

Example
Let us consider the graphs across three time periods where the degrees of the
nodes are shown in the Figure 4.7. Each graph represents a set of communi-
cation lines between the IP nodes. Note that here the degree is based on the

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 72 Big Data Analytics and Its Need for Cybersecurity

Figure 4.7 Example walkthrough for evaluating central nodes over time.

nonduplicate edges. This can be modified to a weighted degree to count for

multiple times the nodes communicate with each other, for accommodating
duplicate edges. Alternatively, edge weight can be added for the number of
times the nodes are communicating. Here, for simplicity we only consider the
nonduplicate communications. We can see that certain nodes (a, d, e) are
consistently central if the degree threshold for centrality is greater than or
equal to two. On the other hand, we see that node f is consistently low degree.
Node b starts as a high degree but is dropped in time period T2 and again
comes back as a low node degree in T3. The consistent nodes a, d, and e can be
seen as the regularly used nodes if they are consistently central. If such
consistent nodes get dropped after being consistent in several time periods,
then this can prompt further analysis. If several nodes lose their degree in a
time period, then we can also label that time period to prompt further
evaluation. This process helps identify central nodes over time to evaluate
any unusual changes in the communications.
Let us consider the graph analysis in parallel, through big data technologies.
We depict the map reduce process for a single time period. Here the split of the
data can be done within a single time period. Alternatively, the split could be

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.6 Where Does Analytics Fit in for Cybersecurity? 73

Figure 4.8 Map reduce for degree computation.

done by time period. Figure 4.8 shows the process where within a time period
we split the data, and the degree count is done within those chunks where the
key is the node name and the value is the degree, thus forming the key–value
pair. The data are computed for each node in the shuffling and finally the
values are tallied for each node in the reduce phase, producing the final result.
This mechanism of map reduce is common across several big data tools. The
internal workings of how the data are split and the map reduce is performed
may vary.

4.6.2 Multipronged Attacks

Multipronged attacks are attacks that are spread out over time and several
points in the network. Discovering such attacks becomes a challenging task
particularly because the datasets become massive and heterogeneous.
A distributed intrusion detection system (dIDS) provides the infrastructure
for the detection of a coordinated attack against an organization and its
partners’ distributed network resources. However, given the complexity of
multiple attack sources and the massive amounts of data generated for such a
multipronged attack, a multilevel mining framework in the form of a big
distributed IDS (B-dIDS) can be utilized (as proposed in Janeja et al. 2014),
as shown in Figure 4.9. This architecture utilizes IDS logs to sift through

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 74 Big Data Analytics and Its Need for Cybersecurity

Figure 4.9 Distributed processing of IDS alerts.

alarms that may look benign individually but may indicate a critical alert in
combination with other alerts.
In this distributed environment, within each subnet a dIDS agent performs
local intrusion detection and generates IDS log data, as shown in the figure as
IDS1. . . IDSn. Log data from each dIDS agent are sent to a control center,
employing big data tools, where it is stored for aggregated analysis. Each
signature-based agent generates a priority level associated with the alarm when
an attack against a known threat is detected, and generates high, medium, and
low priority alarms for “anomalous” behavior. High-priority alarms can be
clearly labeled; however, the low and medium priority alarm data are very
large, making it difficult for an administrator to perform manual analysis. In
such a scenario, several alarms that are part of a coordinated attack will be
missed. However, if we can show that the high-level alarms have similarities
with low-level alarms, we can propagate the labels to the low-level alarms.
Once we label them as similar to high-level alarms, we can try and study them
carefully for possible breaches that are part of a coordinated attack.
There may be several alarms generated at multiple sources from multiple
sensors (IDS monitors) across networks. These alarms may range from abnor-
mal traffic, unusual port access, and unusual sources. The key idea here is to
connect the anomalies using co-occurrence of alarms at specific time periods,
specific location (node/computer), or similar abnormal behavior. This essen-
tially identifies which have overlaps or similarities in terms of some of the
features as shown in Figure 4.10.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.7 Privacy and Security Issues in Big Data 75

Alarms
Business from
firewall IDS sensors
alarms

Alarms
originating at
business partners

Figure 4.10 Overlapping alarms.

How do we find similarities? This can be done through clustering (as

discussed in Chapter 3) the alarms together, and if alarms fall in the same
cluster, then the labels from a higher-level alarm can be propagated to a lower-
level alarm. Once we have these propagated labels, after domain user valid-
ation, we can use this data for future predictions as well.
Let us consider each agent to provide a training set that is generated after
preprocessing the data through a clustering algorithm. Then classification in
these data can be seen as a class-imbalanced learning problem. This approach
uses an ensemble classification technique to automatically classify the large
volume of aggregated alarm data and to alert a system administrator to a
potential coordinated attack.

4.7 Privacy and Security Issues in Big Data

“Big data” has become a buzzword and a new direction for business avenues
and scientific research. However, there is a trade-off to enjoying the benefits of
big data.
As Figure 4.11 shows, there is a set of pros and cons to adopting big data.
Big data allows the combination of several datasets in a scalable and efficient
manner. It also is able to provide high value since the knowledge discovery is
across multiple datasets, bringing in novel insights. However, because the data
are being merged from multiple distinct datasets possibly from different
domains, domain insight becomes a key to getting the value out of the big
data. This can be highly subjective based on the domain experts, and as a result
the interpretation of the data will have direct impact on how the data are
analyzed and the results discovered. This can reduce or increase the value
depending on the individual domain experts’ insights and collaborative under-
standing in creating a combined solution.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 76 Big Data Analytics and Its Need for Cybersecurity

Figure 4.11 Pros and cons of big data.

Now given that a lot of data are combined in a big data solution, this can
lead to novel insights, but on the other hand can also lead to loss of privacy. In
today’s day and age, when massive amounts of data are being collected,
privacy guarantees are not possible; however, utilizing multiple datasets actu-
ally decreases the privacy. So in essence we can say that there is an inverse
relationship between the utilization of big data technologies and users’ auton-
omy over personal data.
Big data technologies are highly scalable and efficient as the data storage
and processing is distributed across several nodes in a cluster computing
environment. As a result such an infrastructure potentially opens up multiple
access points to the data be it the network or cloud environment. This can
directly impact the security of the data.
While several big data technologies are available, there is a lack of stand-
ardization in technologies, with new technologies evolving every day. As a
result, portability, standardization, and adoption can be impacted.
While it is clear that big data technologies can benefit business applications,
if not used appropriately they can also negatively impact the end users’
experience. So is a frictionless environment, where there are no restrictions
on data integration and knowledge discovery, good? Probably not. Some
friction is good for getting the right information to the right people at the right
time. Big data helps breach information silos; however, some balance is
required to prevent privacy and security breaches. Privacy has a new challenge

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 4.7 Privacy and Security Issues in Big Data 77

in big data: we might share something without realizing how it will be linked
with other data, leading to privacy and security breaches somewhere in the
future. The privacy concerns are also geospatial to some extent and should be
evaluated as such. For example, in Norway the tax return of every individual is
public (The Guardian 2016); however, the same situation would not be
acceptable in the United States, where individual tax returns are private except
when shared with authorization by the individual.
There is also a recent push for open data. Big data in the public domain can
only be of value if more and more datasets are openly available. Thus, open
data initiatives may lead to big data. However, again security and privacy must
be balanced in this environment. There is indeed a high risk if open data is
merged with some restrictive closed data, leading to security breaches. For
example, if the social background of an individual is posted on networking
sites (for example, a picture that is geo tagged or the geo history in iPhone),
this when combined with other knowledge from LinkedIn and other social
media sites can lead to social engineering attacks.

https://doi.org/10.1017/9781108231954.005 Published online by Cambridge University Press

 5
Types of Cyberattacks

This chapter summarizes a taxonomy of the different types of cyberattacks and

the mechanism by which some of these attacks take place. Some of the attacks
will be described in terms of the attack vector, exploit, datasets generated
during the attack, and potential countermeasures for the attack. There are
several types of systems and networks impacted by cyberattacks, as discussed
in Chapter 1. The taxonomy described here focuses on some of the well-known
types of attacks and how they relate to each other.

5.1 Types of Attacks

Several threat taxonomies are proposed, including the European Union
Agency for Cybersecurity (ENISA) threat taxonomy (ENISA 16), the cyber-
physical incident catalog (Miller 2014), and the Attack Vector, Operational
Impact, Defense, Information Impact, and Target (AVOIDIT) cyberattack
taxonomy (Simmons et al. 2009). Figure 5.1 consolidates certain categories
of cyberattacks based on the general characteristics of the attacks (Quader and
Janeja 2021).
Each of the major categories of attacks presents several variations, as show
in the figure. Social engineering relies on understanding the social interactions
of the individual and trying to gain access to user credentials. For example,
through phishing an individual may be motivated to provide their username
and password, which will allow a hacker to gain access to a system. Similarly,
phone phishing deals with phishing but through a phone system where the user
is coaxed to provide their credentials. Malware, on the other hand, deals with
the class of attacks that install a malware, such as virus, spyware, Trojans, etc.,
on a user’s system with the intent of gaining access and information from a
user’s system. Password attacks focus on getting user credentials either

78

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 5.1 Types of Attacks 79

Figure 5.1 Types of attacks.

through a brute force attack or by checking the passwords against a dictionary

of words. Password attack is based on a hacker trying out different passwords
based on a set of passwords such as from a dictionary. To counter such attacks,
complex passwords and passwords that are not repeated from one site to the
next are recommended. Denial of service (DoS) attacks are caused when
malicious users tend to block legitimate traffic by sending too many requests
to a server. Distributed versions of these attacks cause a coordinated attack to
take place, such as in the case of a botnet where several machines are carrying
out these requests being sent. Advanced persistent threats persist through
coordinated long-term attacks and mostly originate from organizations with
long-term interest in the assets on the hacked system. These could be initiated
from nation states. Database or software-based attacks are caused due to bugs
in the software such as in the case of a SQL injection attack where the front-
end user form can be used to initiate a query request to the back-end database
to gain access to the username and password credentials. Mobile ad hoc attacks
can be caused where the information flow is disrupted, such as in the case of
the blackhole attack, where the network traffic is redirected and may be lost. In
a cyberphysical system, there is another class of attacks such as a replay attack,
where valid data are sent maliciously but repeatedly with the intent to cause
delay or block the traffic. Some of these attacks are further explained in
Figure 5.2, in terms of what the attack is, the mechanism of the attack,
countermeasures used, and any datasets that can be harvested to study
the attacks.
While we can focus on understanding each of these attacks individually,
they are inherently linked by the aim of the attacker. For example, an APT is

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 80 Types of Cyberattacks

Figure 5.2 Example attack types.

generally carried out by exploiting a vulnerability or by gaining access to a

system through a social engineering attack. More often than not, spear
phishing is the root cause of many types of attacks. Thus, security is more
than an engineering challenge, as people are an essential part of the critical
infrastructure. Therefore, understanding and addressing human behavior is
essential to building a security culture. Methodologies such as Cyber Kill
Chain (Lockheed Martin 2015), FireEye Mandiant purple team assessment
(Fireeye Mandiant 2019), and Mitre ATT&CK (MITRE 2020) provide add-
itional insights into the mindset of the attacker and the threat method.
Additional human behavioral perspectives in understanding cyberattacks are
discussed in Chapter 10.
We next consider two of the attack types and describe the cyberattack and
data analytics methods to discover characteristics of these attacks. We pick
these two types for their complexity and multiple types of datasets that might
be involved in analyzing them.

5.2 Example: Social Engineering

Social engineering is a mechanism to gain access to computer systems, and the
information that resides there, in an unauthorized manner through human- or

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 5.2 Example: Social Engineering 81

computer-mediated methods targeting the social aspects of a user’s data.

Compromising information about a user’s credentials can be obtained through
social connections or impersonating social connections. An example of imper-
sonating social connections is spear phishing, where an email that appears to
be from a legitimate social connection or authority figure requests credentials.
The goal of social engineering is to obtain information that will allow the
hacker to gain unauthorized access, via human-mediated or computer-
mediated methods, to a system in order to compromise the system’s integrity,
potentially leading to other types of attacks. Human-mediated methods may
include impersonation, impersonating a help desk, third-party authorization,
phishing through impersonating a trusted authority figure, and snail mail, to
name a few (Quader and Janeja 2021).
The question addressed here is how do we detect common social engineer-
ing attacks in the network using computational models? A computational data
model is discussed next to detect the anomaly in the network data as a result of
password theft.

5.2.1 Computational Data Model for Social Engineering

Computational data models are derived on intuitions of how these attacks
affect the patterns in the data collected on the network. To build a computa-
tional data model for social engineering attacks, specifically password theft,
log data from intrusion detection systems, such as Snort log data, can be
analyzed. The objective is to derive a data model by which we can detect an
anomaly in the network that emulates the behavior of a password theft. This
model can be applied to Snort IDS log data or firewall or server log data to
identify a potential password theft attack.
Intuition: Let us consider what happens when a password is stolen with
malicious intent. It is highly likely that if a password is stolen, the intent could
be to generate more attacks inside the network through this compromised
system. Alternatively, if the intent is data theft, then several datasets might
be moved from this compromised system. Thus, in either case, the traffic
would substantially increase after the password was stolen. To distinguish
such a change in data pattern of the user, we can also differentiate it from
other traditional attacks by eliminating any probes conducted on this system,
since traditional hacks into a system are done after a probe and scan in the
network. Therefore, the intuition is that if there is no probe or scan in the
network and the system under study and the network pattern has substantially
changed over time, this then leads to an observation that a password theft may
potentially have occurred. While this is a broadly defined intuition, this is not a

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 82 Types of Cyberattacks

predictive labeling but an exploratory analysis, as a network administrator

would have to evaluate further whether the identified system indeed had a
password theft or there is another event signifying change in the traffic
patterns. These types of intuitions allow the administrator to focus on a
narrower swath of data rather than the entire network traffic at large.
Using this intuition, let us consider how it can be translated into a computa-
tional data model. The conditions of the model (Quader and Janeja 2014),
depicted in Figure 5.3, are about observing the activities on a user’s system and
identifying if there is any probe on that machine to gain access. A probe is an
attempt to gain access to a computer and its files through a known or probable
vulnerability in the computer system. Password theft in this study of social
engineering attack is a nonprobing scenario where the intruder may gain access
to the password via nonprobing means such as getting hold of a password
through human-mediated methods such as impersonation as IT
security personnel.
If the log files are evaluated for frequency of IP addresses, then the count of
the destination IP would be a small number in a situation when there is no
probe. Another condition that can be added is that a single source IP may
access several IP addresses in the network after the password theft.
In Figure 5.3, given a set of IP addresses N = {n1,. . . , nn}, IP n1 is the source
accessing n2 as the destination, with the number of accesses less than a certain
threshold ρ. Threshold computation is elaborated further.
The second condition, as depicted in Figure 5.3, is about significant activity
increase on the same machine (IP) as a result of the intruder having access to
that machine. Once the target machine becomes a victim, the same machine is
used as a source (IP) to potentially attack other machines or send out data.
Hence, the activities on the source IP significantly increase over time as a
result of password theft.

Figure 5.3 Computational data model conditions.

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 5.2 Example: Social Engineering 83

In this model, the behavior of the destination IP at time Tx‒1 when this IP has
normal behavior is compared to the behavior of the same IP as a source at time
Tx when the IP becomes a victim of password compromise. ρ is the threshold,
below which a particular destination is identified as having a nonprobing
scenario. δ is the threshold, below which the source IP depicts normal activities
while the value above the threshold would identify the source to be sending out
a large amount of communication to other destinations (for example, data or
attacks).
Thus, given a set of IP addresses N = {n1,. . ., nn}, a compromised IP n1 is
the source and accessing destination ni more than a certain threshold δ.
Computing Thresholds: The data from log files can be divided into several
bins based on the time stamp and for each bin the count for each destination IP
can be computed and sorted in ascending order. The mean and standard
deviation of the counts can be computed. This mean can be considered as
the threshold value for ρ and δ; alternatively, mean + standard deviation can
also be used as a slightly restrictive threshold. For example, Figure 5.4 shows
example data for both time bins. Thresholds of µ + σ1 and µ + σ2 are computed
to indicate an anomalous communication at time Tx for IP n5. In general, if
subsequent bins show low values as compared to the threshold followed by high
values of the counts as compared to thresholds, then this may indicate a period
worth investigating. While this is not a predictive method, it may provide an
administrator with a way to zoom into unusual communications.
Notice the total count for source IP n1 before the password theft on Tx‒1 is 2,
which is less than our threshold of 2.28 (µ + 2σ). The total count of the source
n1 goes up significantly after the password is potentially compromised on Tx,
which is 5. This count is greater than the threshold of 4.9 (µ + 2σ).
It is important to be able to identify the time periods as well where such
investigations can be initiated. As a preprocessing task, thresholds can be
identified from historic datasets. Overall interesting conditions are when

Figure 5.4 Setting the value for ρ and δ.

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 84 Types of Cyberattacks

Count (destination (n1), Tx‒1) < ρ and Count (source (n1), Tx) > δ. Thus,
before the password theft occurs, the count (activity) for that IP (as a destin-
ation) is going to be lot less at time Tx‒1 and the count (activity) for that same
IP (as source) after the password theft is going to be a significantly high at
Tx time.
Discovering Frequent Rules: Association rule mining (discussed in
Chapter 3) can be used to validate the rules where the activities of the
compromised IP increase significantly with a high LIFT value. Association
rule mining provides the rules where the compromised source IP targets many
destination IPs once the intruder gains access to the source machine via
password theft. This source IP from ARM may also show a correspondence
with results from the computational data model and show this IP reaching out
to many targets (n2, n3, n4), and as a result it should show up in the association
rules. The lift value gives us the performance measure and the confidence for
the ARM rules gathered from the algorithm.

5.3 Example: Advanced Persistent Threat

One of the fundamental goals of an advanced persistent threat (APT) is to
remain undetected on the network for a long period of time in order to carry
out the threat’s mission. The desire to remain hidden, and the sophisticated
measures taken to avoid detection, makes it extremely difficult to identify,
analyze, and categorize APTs. APTs are difficult to detect, prevent intrusion
from, and remove, as explained next.
Difficult to detect: Use of custom tools that do not have a signature for
detection are very challenging to discover. Zero-day exploits avoid detection
by signature-based systems. Legitimate third-party applications leveraged for
command and control hide in plain sight. Evidence that data have been
transferred from the system is deleted or overwritten.
Difficult to prevent intrusion: Attackers target the users of the systems, via
socially engineered spear-phishing emails. Up-to-date systems are still
vulnerable to zero-day exploits. Attackers have been known to target add-
itional, outside organizations in order to access their true intended victim
(Trend Micro 2012).
Difficult to remove: Attackers install backdoors throughout victim net-
works to maintain footholds, and hence they are very difficult to remove.
Here a review of the APTs such as Mandiant (Binde et al. 2011), McAfee
(McAfee 2010, Websense 2011), Symantec (O’Gorman and McDonald 2012),
and Trend Micro (SecDev Group 2009, Leyden 2012) and from security

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 5.3 Example: Advanced Persistent Threat (APT) 85

Figure 5.5 Characteristics of advanced persistent threats.

researchers is discussed. Several properties displayed by different APT groups

are discussed (as outlined in Quader et al. 2015). Also included in these
properties is the pervasiveness of the attacks, victims, the robustness of the
infrastructure, as well as other select attributes. By reviewing these attributes,
APTs are categorized into groups by their technical sophistication, longevity,
persistence, and overall threat.
APTs have specific characteristics, as shown in Figure 5.5, such as attack
vectors through which they are initiated, exploits through which they are
carried out, tools used to establish the threat, targets that make a target inviting
for an APT attack, command and control, and persistence of the attack, which
is longer lasting than other traditional attacks.
There are several reported APTs; however, only a very small set of APTs are
well known, fully understood, and reported to the general public. Of the
known threats, very few holistic pictures are available that detail the tactics,
techniques, and procedures. Here the more prominent APTs discovered over
the last several years are discussed and broken into individual parts of the APT
to study the attack vectors, exploits, tools, targets, command and control, and
persistence of each APT. Each attribute is described in the following subsec-
tions (Quader et al. 2015, Stauffer and Janeja 2017) and shown in Figure 5.5.

5.3.1 Attributes of APTs

5.3.1.1 Attack Vector
APTs typically follow three primary attack vectors for gaining access to a
target’s system. No vector is necessarily more sophisticated than any other;
however, the level of sophistication is inherent in the quality and thoroughness
of the social engineering and spear-phishing techniques.
Several APTs have broadened their potential victim base and use an attack
vector known as a watering hole. Attackers first identify a vulnerable website
and insert malicious code into the site that will redirect visitors to another

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 86 Types of Cyberattacks

malicious site hosting the APT‘s malware (O’Gorman and McDonald 2012).
SQL injection/other is probably the least utilized vector by APTs. Attackers
may use SQL injection to exploit a company’s web servers and then proceed
with the attack.

5.3.1.2 Exploits
Before an attacker can surreptitiously install their chosen backdoor on a victim
machine, vulnerability needs to be available to exploit. Vulnerabilities can be
found in many software applications.
Zero day: A zero-day vulnerability is not known by the software developer
or security community; it is typically only known by the attacker or a small
group of individuals. A zero day is valuable because any victim machine
running that software will be vulnerable to the exploit. Thus, Zero-day vulner-
abilities are difficult to discover (O’Gorman and McDonald 2012).
Known, but unpatched vulnerabilities: A small window of time exists
between the discovery of vulnerability and the time it takes to develop and
release a patch. An attacker may take advantage of this time to develop an
exploit and use it maliciously.
Known, patched vulnerabilities: Since software developers have to release
updates to patch any vulnerability, an opportunity exists for attackers to take
advantage of users who do not regularly update their software.

5.3.1.3 Tools
Many attackers use similar, publicly available tools to navigate a victim
network once they have established backdoors on victim systems. The differ-
ence between APTs lies in the tools they use to maintain access to victims.
Custom tools: Some APTs use custom tools that are not publicly available.
Since a considerable amount of time, expertise, and money is needed to
develop a family of malware, custom tools will highlight the potential expert-
ise and backing of a group.
Publicly available tools: Many APTs use tools that are publicly available
and easily downloadable from many websites dedicated to hacking.

5.3.1.4 Targets
APT targets can be government/military, nongovernmental organizations,
commercial enterprises, or the defense industrial base.

5.3.1.5 Command and Control

This highlights the ways in which an attacker communicates with implanted
machines:

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 5.3 Example: Advanced Persistent Threat (APT) 87

Trojan-initiated callbacks: This type of communication is marked by the

malicious program initiating the connection to an external server. Most attack-
ers use this technique in order to defeat firewalls, which typically allow
outbound connections from computers within the network (Mandiant 2013).
Encryption: Encryption is used to hide C2 communications and make such
communications more difficult to detect.
Third-party applications: Third-party applications, such as MSN
Messenger, are leveraged to communicate with victims. Using trusted applica-
tions makes it more difficult for network defenders to identify C2 communi-
cations (Mandiant 2013).
Standard/default malware connections: Many publicly available back-
doors come with default settings. For example, Poison Ivy uses port 3460 by
default. Communications over default settings are more easily detected
because they are known by network defenders (Mandiant 2013).

5.3.1.6 Persistence
Persistence is characterized by the length of time a particular APT has been
conducting operations, such as more than five years, between two and five
years, or less than two years
Figure 5.6 presents an overview of the characteristics of some well-known
APTs (SecDev Group 2009, Alperovitch 2011, McAfee Labs 2010, 2011,
Chien and O’Gorman 2011, Thakur 2011, Villeneuve and Sancho 2011,
Blasco 2013, Mandiant 2013). Taking into account the tools, exploits, and

Name Attack vector Exploit Tools Targets Command and control Persistence
APT1 Spear phishing Custom Commercial enterprises, defense Trojan-initiated callbacks, More than five
industrial base third party applications years

Shady Rat Spear phishing Known patched Government/military, non Less than two
vulnerabilities governmental organization, years
commercial enterprises, defense
industrial base
Aurora/ Spear phishing Zero day Custom Nongovernmental organization, Trojan-initiated callbacks, Between two
Elderwood Watering hole commercial enterprises, defense encryption and five years
industrial base
RSA Hack Spear phishing Zero day Publicly Commercial enterprises Trojan-initiated callbacks, Less than two
available standard/default malware years
connections
Lurid Spear phishing Known but Publicly Government/military, commercial Trojan -initiated callbacks
unpatched available enterprises, defense industrial base
vulnerabilities
Night Dragon Spear phishing Known patched Publicly Commercial enterprises Trojan-initiated callbacks Between two
SQL injection/other vulnerabilities available and five years

Ghost Net Spear phishing Known patched Publicly Government/military, non- Less than two
vulnerabilities available governmental organization years
Sykipot Spear phishing Known, but Custom Government/military, defense Trojan-initiated callbacks, More than five
Watering hole unpatched industrial base encryption years
vulnerabilities
Nitro Spear phishing Publicly Non-governmental organization, Trojan-initiated callbacks, Less than two
Watering hole available commercial enterprises encryption years

Figure 5.6 APT characteristics.

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 88 Types of Cyberattacks

command and control communications, it appears that Elderwood, APT1, and

Sykipot present the most substantial threats among the currently reported
APTs in terms of the characteristics discussed.

5.3.2 Data Analytics Methods for Persistent Threats

Let us consider how we can use data analytics to discover persistent threat
patterns such as those exhibited in an APT, particularly using the network
traffic data.
For this, it is useful to evaluate the intrusion detection log data along with
other raw log data files and study them over time using data discretization.
Once the temporal bins are created to divide the data into chunks of time to
study persistent behavior in these time chunks, frequent patterns using the
association rule mining can be identified, and the overlapping and nonoverlap-
ping rules across these bins can be determined. The high-priority unusual
persistent threats can be isolated, i.e., the rules that occur frequently in the
temporal bins but are nonoverlapping across time periods. This will help
identify rules that do occur multiple times in a timer period but do not occur
across other time periods.
Intuition: An advanced persistent threat is generally unusual and uses
stealth mechanisms to not be discoverable. APTs also exhibit signs of system
access at unusual times, a high level of backdoor access, and unusual infor-
mation exchange. These events can be unusual yet frequent – but not neces-
sarily across multiple times. This unusualness comes from the times when
these events happen or in some cases the specific systems being targeted.
Association rules that are repeatedly occurring across the multiple time periods
represent the daily chatter in the log files. However, rules that are frequent but
are nonoverlapping across time periods can be deemed unusual and considered
as potential persistent threats. Similar to the social engineering example, here
the idea is to narrow down the data and time window, which can be further
investigated by an administrator. For example, frequent logins at unusual times
might be useful for an administrator to follow up on rather than looking at the
entire massive dataset. Keeping this intuition in mind, let us consider the
mining strategy.
Data discretization: This segregates data into time segments or bins repre-
senting the multiple time periods, which makes it easier to evaluate the
persistence of the threats across time periods represented by bins. Given a set
of temporally ordered alerts A = {a1, . . . , an} (from the IDS log files such as
from Snort logs) where each ai has a set of features fi = {fi1, fi2, . . . , fim}, Bin
is a set of temporally ordered segments from A such that B = {b1 = (a1, . . . , ax),

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 5.3 Example: Advanced Persistent Threat (APT) 89

b2 = (ax+1, . . . , ay), . . . , br = (ai+1, . . . , an)}, where |b1| = |b2| = . . . = |bz| = z,

and where z is the size of the bin. Let us consider equal-sized bins, in this case
where z is fixed frequency.
Suspicious network traffic is captured through IDS in an alert log capturing
the source and destination IP, with time stamps and the priority. Here priority
indicates the severity of an alert (high, medium, or low-level alert). The high-
priority threats can be obvious suspicious activities, and the low-priority
threats can be potential alerts that may or may not be real. Such high-priority
alerts when looked at in combination over a long period of time may turn out to
be persistent. In general, an IDS does not capture threats over a period of time,
but views each individual alert and assigns priority. Therefore, it is important
to evaluate the alert priorities to see, when combined over time, what the
overall threat is.
Identify frequent patterns: Analyzing threats individually may not provide
the overall persistence of a threat. It is important to study the data from the
overall perspective and identify frequent patterns on a given time frame to
discover persistent threats. Thus, the data can be mined using association rule
mining techniques and isolate the persistent high-priority threats (PHPT) from the
individual high-priority threats (Quader et al. 2015). The PHPT are consistent and
may indicate APT behaviors. Now given a set of discrete bins B in the log
data, frequent pattern set P = {p1, . . . , ps} is a set of association rules where each
pi = {fij ! fik, fil ! fim, . . . , fip ! fiq} such that each rule set pi corresponds to
each bin bi and where fi = {fi1, fi2, . . . , fim} are features of alert ai in bi
Persistent threats: Persistent threats (PT) are unusual threats that stay
consistent over a long period of time. The binned datasets and their corres-
ponding frequent patterns can be intersected with each other to isolate the
nonintersecting high-priority threats to detect the potential persistent threats.
Persistent threat patterns have the following key characteristics: consistent
over time (occurs repeatedly); single source (same key players); individually
each is a threat, unusual (the pattern may be repeated at an unusual time of the
day or a single source accessing different sources repeatedly at the same time
of the day); and nonobvious (nonoverlapping).
In Figure 5.7, the PT patterns are located in the nonoverlapping area.
Each circle in the figure represents the association rules for a specific bin.
Now if the association rules overlap, they are consistent across bins and may
be more like probing, which is not consistent with the characteristics of an
APT. However, if the association rules do not overlap across bins and are
unique to certain time periods, then they can be considered as potential
persistent patterns of interest. Note that these are still frequent but not the
same across each time period.

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 90 Types of Cyberattacks

Figure 5.7 Example overlapping and nonoverlapping rules from ARM for Bin1,
2, and 3.

Association rule mining technique with classification-based association can

be used to obtain the implied priority on the right-hand side of the rule. This
helps identify the potential persistent threats with priority alert labels along
with the repetitive nature of the threats. The binning also allows plotting the
priorities on a timeline to associate the PT with time periods when these
events occur.
In Figure 5.7, we see three time periods T1, T2, and T3 with their
corresponding rule sets p1, p2, and p3. For example, p1 has associations in
the form of IP1 ! IP2. This could imply a source destination relationship
where IP1 is sending data to IP2. A union of two consecutive sets is taken and
the intersection is subtracted from this set. As a result, the nonoverlapping or
nonintersecting part of the two sets is discovered. Finally, after subsequent
operations, the most unusual frequent set is identified that is nonintersecting
across multiple bins.
In both these examples, the end result is not predictive but exploratory to
help narrow down the data and time periods that should be further investigated.
Each of this analysis is also based on the intuition of the attack pattern. Even in
supervised methods, the intuition is used in formulating the training set with
a given distribution. This is true to any data analytics task for identifying a
cyberattack. There is no true model that can predict or identify an attack with
a 100% certainty. Each detected pattern helps an administrator investigate the
potential occurrence of the attack or zeroing into the needle in the haystack
of data.

https://doi.org/10.1017/9781108231954.006 Published online by Cambridge University Press

 6
Anomaly Detection for Cybersecurity

In this chapter, we focus on what anomalies are, their types, and challenges in
detecting anomalies. In Chapter 7, we focus on specific methods for
detecting anomalies.

6.1 What Are Anomalies?

An anomaly in a dataset is an observation or a set of observations that are
different and inconsistent with respect to the normal observations. An anomaly
is also referred to as outlier, peculiarity, exception, and discord, among other
terms. “Anomaly” and “outlier” are mostly used interchangeably. An anomaly
is always identified with respect to a frame of reference. The frame of reference
is the normal or baseline set of observations in the data. An anomaly stands
apart from this frame of reference. Therefore, anomaly detection heavily relies
on discovering the normal, the frame of reference, or the baseline against
which an observation is compared to be termed as an anomaly. The further
away the anomaly is from the frame of reference, the more severe is
the anomaly.
Anomaly detection is a well-studied area of research, with several surveys
describing the state of the art (Patcha and Park 2007, Chandola et al. 2009),
and more specifically anomaly detection in areas of cybersecurity (Estevez-
Tapiador et al. 2004, Feily 2009, Ten et al. 2011, Hong et al. 2014, Ahmed
et al. 2016).
In the area of data analytics, anomaly detection has utilized techniques from
supervised, semisupervised, and unsupervised learning to identify anomalies.
Anomalies have been characterized as single point, group of points, and
contextual anomalies (Chandola et al. 2009), as shown in Figure 6.1.
However, these different types of anomalies can be interrelated.

91

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 92 Anomaly Detection for Cybersecurity

Figure 6.1 Types of anomalies.

Figure 6.2 Illustrating the types of anomalies in clusters of network traffic data.

To explain the types of anomalies, let us consider Figure 6.2. The figure
shows clusters of network traffic data that comprise features such as time to
live, datagram length, and total size. Once clustering is performed, we see that
certain points do not fall into any of the clusters. We can evaluate this
clustering as containing potential anomalies.
Figure 6.2 depicts different types of anomalies. A point anomaly is a single
data point that is anomalous with respect to a group of neighboring points.
There can be different points of references such that an anomaly can be
identified with respect to all clusters or with respect to points in a cluster or
in some cases the entire dataset. For example, in the figure we can see that
there is a single point that is a global anomaly, which is an anomalous with
respect to all the three clusters 1, 2, and 3. There some local point anomalies,
which are anomalous with respect to specific clusters such as cluster 2 and 3.

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 6.2 Context 93

In some cases, a small group of points can also be an anomaly, called as a

collective anomaly, such as cluster 4 in Figure 6.2. This cluster is an anomaly
with respect to all the clusters due to the relative size of the cluster and its
distance from other clusters. Individual points inside the cluster may not be
anomalous on their own; however, when considered together as a group they
are anomalous. In the example in Figure 6.2, the point of reference is the other
clusters, such that the size of the clusters 1, 2, and 3 is much bigger and the
distance of the small cluster 4 to the other clusters is much bigger than the
distance between the clusters 1, 2, and 3. Thus, it is clear that a frame of
reference is extremely important in discovering anomalies such that points or
groups of points are not frivolously identified as anomalous.
Another type of anomaly is a contextual anomaly, where the anomaly is
identified with the perspective of a certain context. For example, if the network
traffic in an e-commerce site at a certain time appears to be very high as
compared to the other network traffic points for this site, they may appear to be
anomalous or the time period may appear anomalous. However, if there is a
specific event that is leading to the high network traffic, such as high sales
during holiday time, then this high traffic would not be anomalous. The
context here is that the high amount of traffic is generated due to a marketing
event or a special sale and not because of a flooding attack or a denial of
service attack. In the figure, we provide additional information that the cluster
4 is in fact a group of points where a large file was being uploaded by a specific
network admin, which resulted in the massive size of the transfer on the
network generating the data point that is different from that of the rest of the
traffic. Given the context of the origin of the data, these data should not have
been included in the analysis. Thus defining a context can be very important to
create a frame of reference. This can be done before the anomaly detection,
which is much more robust, to clean up the data of possible frivolous outliers
or as a postprocessing step where the outliers are filtered through any context.

6.2 Context
Context can be in the form of attributes within the data that can be used to
create demarcations in the data so that data are separated into groups for further
analysis. Context can help define homogeneous groups for anomaly detection.
Context discovery is a preprocessing step where the groups are identified using
contextual attributes (such as shown in Figure 6.3), and then anomaly
detection is performed in the contextually relevant groups of data points using
the behavioral attributes that measure the phenomena (such as network traffic

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 94 Anomaly Detection for Cybersecurity

Figure 6.3 Context definition.

density). Context-aware processing allows for accommodating context defin-

ition into the algorithmic process.
Context can be defined in the terms of activity, individuality, spatial or
location context, temporal context, and relations context (Zimmermann et al.
2007). Context definition for an example network data point is shown in
Figure 6.3. For example, a data point, part of the network traffic data, is a
source IP, is sending the data packet at 5 a.m., and is part of the IP address
allocated to the support team. Thus, if all the support team IPs are sending
similarly sized data and at similar time frame, then this point will be considered
normal. However, if this particular data point is sending data at an unusual
time and is originating from a geolocation outside of the expected locations,
then this could be a potential anomaly. Here the frame of reference is defined
in terms of time, IP location, and the team. Thus, data points with a similar
context should be grouped together, and anomaly detection can be applied to
each of the groups for a well-defined anomaly detection. The point, group, and
contextual anomalies are related to each other since a context can apply to both
a point and groups of outliers. Moreover, in some cases individual point
anomalies can collectively be part of a group anomaly. However, this may
not always be the case since data points that are part of the group may or may
not be individually anomalous.
Anomaly detection has several associated challenges, and these will be
clarified next with the help of a motivating example.

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 6.3 Motivating Example: BGP Hijacking 95

6.3 Motivating Example: BGP Hijacking

Network packets follow certain specific traffic routes on the internet using the
announced and available routes maintained in the routing tables. Border
Gateway Protocol (BGP) is primarily responsible for exchange of information
between autonomous systems (AS) for successful transmission over the pub-
lished routes. Several types of attacks lead to hijacking of hosts or servers and
redirecting traffic to anomalous sites or dumping traffic at random sites. BGP
hijacking is one such type of attack. Examples include redirection of Google
and YouTube requests in Turkey (Paganini 2014) and erroneous rerouting of
YouTube traffic (Sandoval 2008), among others.
BGP anomalies (Al-Musawi et al. 2016) prevent successful exchanges
between the AS, leading to loss of data or unauthorized rerouting of the data.
Such anomalies may include route flapping, path announcements that delay in-
routed packets, malicious redirecting of traffic for unauthorized surveillance
purposes, and dropping of packets to unknown destinations. BGP anomalies
where path announcements are made may lead to thousands of anomalous
updates. BGP first initiates a complete routing table exchange and then subse-
quently exchanges updates. Thus, if there is an anomaly, then there will be a
lot more updates than normal. This can potentially indicate the presence of an
anomaly. However, a more specific discovery is needed to identify where (at
what location or specific AS) the anomaly originated, as this can provide
pertinent information for interpretation of the anomaly, whether it was a simple
misconfiguration or a more malicious intent to disrupt or redirect the traffic.
BGP anomalies can also be classified similar to the general anomalies a
point or collective anomaly. A single update can be considered as a point
anomaly if it has an invalid AS number, invalid IP prefix, or a prefix
announced by an illegitimate AS (Wübbeling et al. 2014). A set of updates
can be anomalous (collectively) if there are several BGP updates in a short
amount of time containing longest and shortest paths or substantial changes in
the BGP traffic (Al-Musawi et al. 2015). Such anomaly detection can use data
from BGP update messages or routes that packets use. Several repositories are
available for such datasets to be analyzed and for discovering anomalies, as
shown in Table 6.1. Each of these datasets shows the complexity and massive
nature of the data.
Studies have shown that a large number of attributes can be mined to
discover BGP anomalies (Li et al. 2005, Al-Rousan et al. 2012). Using these
as an example the following section identifies the challenges in understanding
and discovering anomalies.

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 96 Anomaly Detection for Cybersecurity

6.4 Challenges in Understanding Anomalies

As data size increases both in terms of the number of data points and the
features describing them, the patterns in the data get further embedded
deeper into the data space. This is primarily because data are sparse in high
dimensions. This means that the patterns may exist in some subspaces in the
data and not in others. This is referred to as the curse of dimensionality
(Bellman 1961). The curse of dimensionality affects all the data analytics
techniques and is even more relevant for anomaly detection, as anomalies
are already rare.
Several studies have explored the datasets listed in Table 6.1. Li et al. (2005)
used 35 features every one minute to generate decision trees based on past
BGP events. Such data are high dimensional and nearly real-time data. Let us
consider the dimensionality issue in such data. High-dimensional data can be
analyzed in various ways to identify anomalous routes. The full dataset can be
considered for the analysis or alternatively a smaller relevant (contextual)
subset of the data can be considered. It is possible that a route or a single
update can be an outlier in one subset of attributes but not an outlier in another
subset. For example, a route is not an outlier in the updates attribute but is an
outlier in the location attribute.
Thus, it is essential to carefully determine whether all the dimensions are to
be considered or a selected critical attribute set is considered. This selection of
the critical attribute set is also a data mining task in itself, such as through
feature selection.
In the example shown in Figure 6.4, the data are comprised of four attri-
butes: number of BGP announcements, number of BGP updates, number of

Table 6.1 BGP data repositories

Data description Link

RouteViews www.routeviews.org/
Réseaux IP Européens www.ripe.net/manage-ips-and-asns/db
BGPMon https://bgpmon.net/
Internet route registries http://irr.net/index.html
CIDR report (bogon prefix www.cidr-report.org/as2.0/
lists)
MaxMind, GeoIP location www.maxmind.com/en/open-source-data-and-api-for-
data ip-geolocation
IP2location www.ip2location.com/
CAIDA traceroute www.caida.org/proiects/ark/
platform

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 6.4 Challenges in Understanding Anomalies 97

Figure 6.4 Example curse of dimensionality.

new paths announced, and number of withdrawals after announcing the same
path. The two-dimensional plot of the data considering all the attributes
together (Figure 6.4a) depicts clear anomalies that stand out from the other
data points. This anomaly corresponds to observation 4, which shows high
values for all attributes with respect to all the other observations. Figure 6.4b–e
shows the plot for each attribute. It is clear that observation 4 dominates plots
shown in Figure 6.4b–d. However, in Figure 6.4c we see that observation 3,
which has a high number of BGP updates, is also deviant with respect to the
rest of the data points. Finally, in Figure 6.4d, which plots the attribute AW,
which is the number of withdrawals after announcing the paths, it is evident
that observation 5 is highly deviant compared to the rest of the data points.
This anomaly did not show up in Figure 6.4a and was masked by the other
highly deviant values. This clearly shows the challenge of high dimensionality
when the patterns are hidden in the subspaces.
Thus, anomaly detection is not only focused on the mechanism of labeling
an outlier or anomaly but also carefully identifying the frame of reference in
terms of the objects to compare against and also the subset of attributes or
subspaces to consider.
Once the frame of reference or context in terms of the data and set of
dimensions to be used is determined, then the data mining and analysis could
be used to identify the deviation of the route from the normal.
This type of analysis based on a clearly defined context could be useful in
selecting the suspicious routes for further investigation. Such analysis can help

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 98 Anomaly Detection for Cybersecurity

answer the following: What routes or updates should a point be compared

against? What are the critical dimensions that facilitate the discovery of
anomalous routes? How can we inform why a certain route is an outlier by
identifying the dimensions contributing to the outlierness? If all dimensions
are being used for the discovery, then it is also important to explore subsets of
the attributes to account for contribution from dimensions and also identify the
(outlierness) contributing dimensions.

6.5 Interpretation
Anomalies are in the eyes of the beholder. Anomalies can be a nuisance for an
operator or a user who does not want too many alerts to slow them down. For
example, in some domains such as the emergency operating centers where
complex and multiple systems are used, too many alerts can lead to alert
fatigue, desensitizing the end users to the numerous pings from their systems.
Indeed, in healthcare environments, this type of alert fatigue has been studied,1
and it could lead to potential threats such as ransomware, etc. This can affect
the overall performance of a system, but more importantly this can desensitize
users to click when they should not be clicking. Alternately, for a security
officer each anomaly may be a critical alert to portend the arrival of a major
attack or a major misconfiguration leaving a gaping hole in a system.
Anomaly interpretation may also change from one domain to another. What
appears as a blip in an instrument reading in one domain (such as sensors
monitoring water monitoring quality in a stream) may appear to be a major
alert in another domain (such as sensors monitoring water in the nuclear power
plant). Another example is password attempts. How many failed password
attempts are allowed into a system will depend on the function of the system
being accessed. A bank may only allow three failed attempts before it locks out
the user, whereas an emailing system may allow more attempts. In each of
these systems, predicting whether there is a guess password attack will be
based on different criteria based on the domain context. Figure 6.5 depicts
some of the priorities and discussions at various levels that can impact the
discovery and interpretation of an anomaly. The organization’s priorities may
be connected to the business bottom line, such as reputation, when an attack
happens that can translate to millions of dollars lost (Drinkwater 2016).
Similarly, an end user is highly driven by the day-to-day functions and how

1
https://psnet.ahrq.gov/primers/primer/28/alert-fatigue Last accessed June 2017.

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 6.6 Treating Anomalies 99

Figure 6.5 Factors governing the interpretation and discovery of anomalies.

fast the functionality can return to normal. The system and security priorities
are more directly related to the anomaly detection and scoring. However, these
must work in conjunction to the other players’ priorities.
Thus, interpretation of anomalies will vary from one type of user to another,
one type of system to another, and one domain to another based on the context
defined and agreed upon by each.
Many times, two users in the same organization may also not agree on the
definition of an anomaly. It is critical to have a well-crafted definition of an
anomaly to (a) facilitate the discovery of an anomaly, (b) reduce the false
positives, and (c) associate importance to the level of the anomaly for faster
recovery and remediation.

6.6 Treating Anomalies

Traditional statistical outlier detection (Barnett and Lewis 1994) has addressed
how outliers can be treated once they are discovered, whether they should be
accommodated into the analysis or rejected before the analysis.
Accommodation tends to detect the outlier and accommodate the value in
the dataset so that the whole set of observations is still intact. It tends to
manipulate the observation so that even if an anomaly is in the dataset, the data
do not become skewed because of this rogue observation. The idea is that
outliers are part of the data: although their removal fits the data into some

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 100 Anomaly Detection for Cybersecurity

Figure 6.6 Treatment of outliers (adapted from Barnett and Lewis 1994).

distribution or normalizes the data, the outliers’ presence could be equally

important for the analysis. Hence, the idea is to not reject outliers, but to
accommodate them so that the analysis can still be safely performed without
any loss of data or observations. Thus, it aims at the robustness of the analysis
in the presence of outliers.
Rejection is basically testing an outlier purely from the point of view of
rejecting it from the dataset or of identifying it as a feature of special interest
and removing it before further analysis is done.
Figure 6.6, adapted from Barnett and Lewis 1994, summarizes the treatment
of outliers based on the source of variation and the nature of outliers. If the
anomaly is caused due to a measurement error, it is either rejected or corrected,
or the detection is repeated. However, if the variation is inherent or due to
some real measurement of an event, then the anomaly detection is used, but
again in this case the outlier is incorporated in the revised model, identified for
a separate study of origin and form. Also, if the anomaly is due to some
random reason, then it could be accommodated to see the overall analysis
rather than eliminating the observation altogether.
In terms of cybersecurity, in the majority of the cases, the anomaly must be
eliminated. For instance, if there is a Trojan discovered in a computer system
that disrupts the regular functionality, it has to be eliminated. Similarly, if the
BGP updates appear to be anomalous, they need to be corrected for proper
communication and movement of the traffic. However, in some cases the
anomaly needs to be studied, in which case the anomaly may be detected
and traced. This was a concept carried out in a setting called “Jail,” where an
attacker was traced through the system access and observations collected to
learn about the role and activities of the attacker (Cheswick 1992). Similarly,
honeypots are set up to collect data and activities of attackers (Spitzner 2003).
We next discuss methods for anomaly detection in Chapter 7.

https://doi.org/10.1017/9781108231954.007 Published online by Cambridge University Press

 7
Anomaly Detection Methods

As discussed in Chapter 6, anomalies or outliers are objects or groups of

objects that are deviant with respect to other normal objects. This anomalous
behavior can be caused by extreme values in some dimensions, which could be
an inherent rare property of the anomalous objects. In high-dimensional data,
this rarity increases with the increase in the number of dimensions. This is
mainly because data become sparse in higher dimensions; this will further
aggravate the discovery of outliers. Other situations may be caused due to
measurement errors or phenomena that are causing the deviations. This chapter
focuses on techniques for anomaly detection addressing these issues and their
applicability to cybersecurity problems.
Figure 7.1 shows a broad categorization of anomaly detection techniques to
give an idea of the development in this field. Several recent surveys provide a
lot more categories and discuss the works in detail (Patcha and Park 2007,
Chandola et al. 2009, Garcia-Teodoro et al. 2009, Bhuyan et al. 2014).
Outlier detection techniques have been adopted from the statistical literature
and treat the outliers as parametric outlier detection, in which the data are
assumed to fit a certain distribution. The tail region of the distribution, such as
in a normal distribution, would constitute the outliers. However, in real-world
datasets, the data do not fit into a standard statistical distribution. Certain other
techniques detect outliers using metrics such as distance, distance from nearest
neighbors, density, deviation, and other such measures. Other methods are a
byproduct of clustering methods. The data are first clustered using some
clustering technique, and the data points, which lie outside of a cluster, are
considered to be outliers. There could also be a scenario that part or all of a
cluster of data points could be outliers based on the problem being studied.
Classification-based techniques identify and predict anomalies based on his-
toric anomaly signatures.

101

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 102 Anomaly Detection Methods

Classification
based
• Neural networks,
classifiers,
ensembles

High dimensional
outlier detection Nearest
• Subspace clustering neighbor
based
• Discovery in large • Density based
and high-dimensional • Distance based
datasets
Anomaly
detecon

Statistical Clustering
• Discordancy test based
for distributions
• Interquartile • Outliers as by
range based product of
• Standard clustering
deviation based methods

Figure 7.1 Broad categorization of anomaly detection methods.

The discussion in this chapter is limited to key anomaly detection techniques

since they are the foundational approaches that are adopted as variations of
these methods, such as in the distance-based techniques and discovery of
anomalies in large, multidimensional datasets.

7.1 Statistical Outlier Detection Tests

The distribution-based parametric methods are discussed in the statistical
outlier detection literature, which assumes that data fit into a prespecified
distribution, such as normal distribution, and then identifies the outliers with
respect to the model using discordancy tests (Barnett and Lewis 1994). The
discordancy tests are developed based on the properties of the distribution.
Similarly, the Inter Quartile Range based outlier detection method is developed
using statistical dispersion.

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 7.1 Statistical Outlier Detection Tests 103

7.1.1 Discordancy Tests

Discordancy tests evaluate the fit of the data to an assumed distribution. It has
two main hypotheses: a working hypothesis and an alternative hypothesis. It
assumes a standard distribution of the dataset, and if the discordancy test finds
the outliers, then it tries to fit the data to the alternative distribution; however,
the underlying assumption is that the data fit a distribution, which might not be
the case always, especially for multidimensional real-world data.
As an example, if there is a normal distribution where x1 and xn are the
extremes, a discordancy test will check if xn, apart from being an extreme, is
also statistically unreasonable; if it is, then it is called a discordant upper
outlier. The next step would be to show that the outlier comes from the
alternative distribution and not the working distribution. However, this is a
difficult task in itself. The key thing to note here is that the discordancy tests
are statistical measures used to see if the data element being tested fits a
distribution. A key metric is to see the fit of the data point using measures
such as deviation/spread statistics such that d ¼ μx1 s , where µ is a measure of
central tendency; in this case, the mean and s can be a known measure of
spread, such as range. For example, if d  1, then x1 is not an outlier; on the
other hand, if d > 1, it is an outlier.
There are two major procedures for outlier detection in statistical studies, as
shown in Figure 7.2: block procedures and consecutive (or sequential)
procedures.
In consecutive (or sequential) procedures, sample size is not fixed, but
determined in each cycle in relation to the values of earlier observations.

Anomaly

Collective
Point Contextual
or group

Statistical
outliers

Block
Sequential
procedure
Figure 7.2 Types of anomalies and statistical methods of detection.

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 104 Anomaly Detection Methods

Each observation is screened for discordancy in relation to the already

screened set of observations. This procedure can be done in two ways: outside
inward, which means that the extreme outlier is tested first, then second most
extreme, and so on, but more widely preferred way of doing it is the other way
around (starting with the least likely element to be an outlier). In block
procedures, two or more observations are tested together for outlierness.
These procedures have similarity to the point anomaly detection and col-
lective anomaly detection discussed in the data mining literature (Chandola
et al. 2009).
Consecutive procedures can suffer from challenges such as masking and
swamping (Barnett and Lewis 1994). Masking is an erroneous judgment
arising out of a single outlier test. Suppose if through a discordancy test
observation xn is declared an outlier with respect to all the n ‒ 1 values, then
xn‒1 is further tested. But if xn is not declared an outlier with respect to all n ‒ 1
values then xn‒1 is not tested. This could be a problem in a scenario where xn is
not an outlier with respect to the adjacent value of xn‒1, but both xn and xn‒1 are
outliers when tested together with respect to all n ‒ 2 values. Let us consider a
small sample values: 8, 10, 11, 13, 1,000, and 1,005. Now, for example, if
tested individually 1,009 will not be declared as an outlier since it is close to
1,000, but if 1,000 and 1,005 are tested together against other values, then they
are outliers with respect to all other n ‒ 2 values.
Swamping, on the other hand, results from using a block or consecutive
procedure to test for discordancy, i.e., testing two values together, as in a
sample such as 5, 8, 10, 12, 1,000.
If two values are tested together, then they are considered outliers, but value
12 is not an outlier, so in essence 12 is swamped by 1,000.
This behavior is also seen in majority of outlier detection techniques when
the incorrect context for comparison can lead to missing outlier values. This is
very relevant to cybersecurity applications where data is temporal and needs to
be evaluated as a sequence of time series measurements such as network
traffic data.

7.1.2 IQR-Based Outlier Detection

Interquartile range (IQR)–based outlier detection, shown in Figure 7.3, aims to
identify the range of the data and find extreme values that are far away from the
range. First, data are sorted in increasing order. The median of the data is
identified, which becomes the second quartile. All the numbers to the left are
smaller than the median, and all the numbers to the right are greater. The
median of these left- and right-hand sides are identified as first quartile and

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 7.2 Density-Based Outlier Detection 105

25th percentile or 1st quartile 75th percentile or 3rd quartile

IQR
Median or 50th percentile

Sorted data
Outliers are 1.5 x IQR above Q3 or below Q1

Figure 7.3 IQR-based outlier detection approach.

third quartile respectively. The data are now divided into four chunks of data
ordered by the quartiles. The difference between the third quartile and the first
quartile is the IQR. IQR-based outlier detection assumes that the IQR should
have the majority of the data distribution measuring the spread of the data.
Now if this IQR range is further stretched beyond the first and third quartile
such that the lower threshold is the first quartile minus IQR times 1.5, and the
upper threshold is the third quartile plus the IQR times 1.5. This stretches the
range or spread of the data. Points that fall beyond these thresholds at the lower
extreme or the upper extreme are the outliers.
IQR is primarily a univariate outlier detection method. An example of IQR-
based outlier detection is shown in Figure 7.4, in the network traffic data
sample for the attribute “packet size.”
In general, statistical outlier detection is a highly developed field and has
been well established for a long time. However, for outlier detection using
statistical methods, the distribution of the dataset needs to be known in
advance and the knowledge of the distribution is the underlying assumption.
The properties of the distribution need to be known in advance such as the
mean, standard deviation, etc. Outlier detection also has to take into consider-
ation the masking and swamping effects for sequential and block procedures.
In general, the modeling process is more complicated, especially in case of
multidimensional datasets. The statistical outlier detection mainly works well
for univariate variables. Moreover, these techniques do not work well with
incomplete or inconclusive data, and computation of the results is expensive.

7.2 Density-Based Outlier Detection: OPTICS-OF

Identifying Local Outliers
OPTICS-OF (Breunig et al. 1999) is an extension of the OPTICS density-
based clustering algorithm as discussed in Chapter 3. It discovers outliers
relative to their neighborhood. Distance-based outliers are good in a global
sense, but when different densities of objects exist within the data, distance-
based discovery is not adequate. Hence, each outlier is assigned with a degree

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 106 Anomaly Detection Methods

Figure 7.4 Example of IQR-based univariate outlier detection.

by which the object is an outlier. Outliers, relative to their local surrounding

space, motivate a formal definition of local outliers. An object is mostly
defined as an outlier or not an outlier; however, there could be a degree of
outlierness attached to each object based on the local density where position
of each object affects the other. The basic concepts of ε-neighborhood and
K-distance are adapted from the algorithm DBSCAN, and the concepts of core
distance and reachability distance are from the algorithm OPTICS.
OPTICS-OF defines a concept called local reachability, which is measured
in terms of the local reachability density of point p given as Lrd(p) = 1/average
reachability distance of minpts nearest neighbors of P.

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 7.2 Density-Based Outlier Detection 107

Figure 7.5 Example of OPTICS-OF parameter computation.

Let us consider the example shown in Figure 7.5. Here the minpts are set to
3, although there are more than 3 points in the neighborhood. Only reachability
distance to the 3 points needs to be measured. Thus, Lrd(p) = 1/((1 + 1 + 2)/3),
which is 0.75.
The outlier factor of p is measured in terms of the uniformity in the density
of p’s neighborhood. Thus, it takes into consideration the local reachability
density of all the minpnts’ nearest neighbors of p. In the preceding example,
these are M1, M2, and M3. The outlier factor of p is as follows:
P
lrd ðoÞ
lrd ðpÞ
OF ðpÞ ¼ N
So in the example from Figure 7.5, the outlier factor is as follows: (0.75 + 1 +
0.5/0.75)/3 = 1.
The OF = 1 represents a cluster of uniform density. The lower p’s Lrd is, and
the higher the Lrd of p’s minpts’ nearest neighbors is, the higher will be p’s
outlier factor.
This approach is one of the first and few approaches that quantifies the
outlierness of an object. In the case of multidimensional datasets, however, the
degree of the outlierness also depends on dimensions that are contributing to
the outlierness and possibly the cause of the outlierness. This approach is able
to identify the local or global outliers with respect to a set of points. For some
applications, both global and local outliers must be pointed out. Moreover, if
only distance from the neighbors or minpts in the neighborhood are con-
sidered, then we overlook the situation when several points are collectively
an anomaly compared to others. If several points deviate from the standard,

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 108 Anomaly Detection Methods

simply relying on the neighborhood or distance measure is not effective. For

this, a clustering-based approach is needed where unusual clusters can be
identified as anomalous. In graphs and spatial data, other types of approaches
look at the collective structure of the anomalous subset with respect to the
other data points.

7.3 Distance-Based Outlier Detection

In data mining, distribution of attributes is mostly unknown. Distance-based
outlier detection algorithms initially proposed by Knorr et al. (2000) continue
the idea of a unified notion of mining outliers, which extends statistical-based
outlier detection.
Such an approach generalizes the notion of outliers provided by many of the
discordancy tests for known distributions, such as Normal, Poisson, Binomial,
etc. Although it proposes a partitioning-based technique for distance-based
outlier detection, it also elaborates on how this approach can unify the various
discordancy tests available in the statistical/distribution-based outlier detection
techniques. The time complexity is linear with respect to number of objects,
and the algorithms are efficient for low-dimensional datasets. The key idea
that outliers are far away from their neighbors in terms of a distance threshold
was proposed and has been adopted by many recent works (for example,
Zhang et al. 2009, Sugiyama and Borgwardt 2013, Cao et al. 2014).
Ramaswamy et al. (2000) treats the distance-based outlier detection as one
relative to K-nearest neighbors.
The distance-based outliers are identified with two parameters, p and D,
namely DB (p, D). An object O in a dataset T is a DB (p, D) outlier if at least a
fraction p of the objects in T are at a greater distance D from O. For example, in
Figure 7.6, the fraction of points p is 3 and distance threshold is 2. Thus, the
point O is at least greater than distance 2 from at least 3 points and can be
labeled an outlier.
Essentially, these outliers are objects that do not have enough neighbors
within a certain distance. This eliminates the computation for fitting the

p=3
D=2
o
3.5 2.5

Figure 7.6 Example of a distance-based outlier.

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 7.3 Distance-Based Outlier Detection 109

observed distribution into a standard distribution and in selecting the

discordancy tests. For many discordancy tests for standard distributions, if an
object is shown to be an outlier, then it is also a distance-based outlier.
Parameters p, D are set by a user in case of nonstandard distributions; this is
a trial-and-error approach, where the user interactively changes the p, D
variables to evaluate the outcome.
This approach identifies outliers based on the distance of a candidate outlier
from a set of neighboring points; it is not dependent on the clustering structure
of the dataset. The outliers detected, however, are dependent on user input.
Different algorithms are proposed that are suitable for different types of
datasets. But these algorithms are sensitive to user input such as the
distance threshold, or the value of p to identify the number of points for
comparison to the candidate point. This approach is not suitable for high-
dimensional data since distance measures (such as the Euclidean distance)
do not scale well over high dimensions and average out the behavior of the
dimensions. The distance metric is a point metric; therefore, it detects
outliers that are at a distance greater than a parameter, and it does not take
into consideration outliers that could be as a result of a very small distance
measure in a subset of the dimensions. This technique is also mainly geared
toward detecting global outliers.
Ramaswamy et al. (2000) eliminates the dependency in terms of the user
input required for distance threshold. It also proposes outlier detection for large
datasets using a pruning technique. This issue is addressed using a structure
such as R Tree and CF Tree. This is a modified way of detecting outliers based
on a distance threshold; it uses K-NN query and detects global outliers. This
technique provides a pruning technique to deal with large datasets. However, it
does not handle multidimensional datasets and is unable to account for attri-
butes contributing to outlierness.
In general, distance-based outlier detection methods suffer from the curse of
dimensionality (as discussed in Chapter 6). In higher dimensions, data become
sparse, and points tend to become equidistant. This adversely affects the
methods based on spatial density, unless data follow certain simple distribu-
tions. For outlier detection, a subset of the entire attribute set should be enough
to detect or indicate the outliers. Aggarwal and Yu (2001) proposed the idea to
examine specific projections of high-dimensional data instead of examining all
the dimensions. If we consider cross sections of the data in various dimensions,
we can find outliers better than in other cross sections. Aggarwal and Yu
(2001) define the abnormal lower-dimension projections. There can be many
abnormal projections, and this work facilitates the discovery of the best
abnormal projections.

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 110 Anomaly Detection Methods

Figure 7.7 Clustering in SANS 1,000-day data.

7.4 Outlier Detection through Clustering

Some clustering algorithms, discussed in Chapter 3, such as DBSCAN, iden-
tify outliers as a byproduct of clustering. However, majority of clustering
algorithms have to be postprocessed to discover outliers. Thus, outliers can
be seen as a byproduct of the clustering process.
In traditional clustering, such as K-Means, by the end of the cluster analysis,
we assign each data point to one of the clusters, so basically there are no
unassigned data points. An example of clustering the SANS data,1 from IDS
sensors around the world for 1,000 days, including attributes of records,
targets, and sources of attacks (from 2017) is shown in Figure 7.7. This
clustering was generated using Weka2 Simple Kmeans clustering. Since all
the data are assigned to clusters, does that mean that there are no outliers in the
data? If there are still outliers, how do we filter out the outliers? There are a few
possibilities in a clustering output: (a) A point may be placed in a cluster by
itself. (b) A data point may be part of a cluster but far from the centroid. (c)
A small cluster with a lot fewer points than other clusters can be far from the
other clusters. (d) In DBScan-type clustering, points will not be connected or
in a cluster but a distinct outlier.
If a point is placed in its own cluster or is identified by the algorithm as an
outlier, such as in cases (a) and (d), then clearly a point anomaly has been

1 2
http://isc.sans.edu/submissions_ascii.html, 2017. www.cs.waikato.ac.nz/ml/weka/.

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 7.4 Outlier Detection through Clustering 111

detected. In the other two cases, postprocessing of the analysis will need to be
performed to identify anomalies:
In the case of small clusters, if the data are clustered into groups of different
density and if there is a cluster with fewer points than the other clusters,
this cluster can be a candidate for a collective anomaly. In this case, the
density of the candidate cluster is highly deviant than the other clusters.
There can be other measures, such as the distance between the candidate
cluster and noncandidate cluster, and if the candidate cluster is far from
all other noncandidate points, they are potential outliers. In Figure 7.7,
we can see a density distribution for each cluster, and cluster 5 appears to
have very low density compared to the other clusters. When we consider
the three points in this cluster, they represent a very high number of
targets, sources, and records, indicating a high level of attack activity.
Each of the observations has a time stamp, i.e., the date. If we further
investigate the date by a simple Google search for the time period and
keyword cyberattack, we see reports of relevant attacks occurring around
that time period indicating the unusual activity during the time period.
This type of evaluation helps with exploratory analysis of the dataset.
In the case of larger clusters with scattered points, there may be bigger
clusters where the points in the cluster are not tightly bound around the
centroid. In this case, the distance of every point can be measured to the
centroid, and the points farthest from the centroid can be identified. This
is similar to a distance-based or a K-nearest neighbors–based outlier
detection inside such clusters. These clusters can be identified by looking
at the sum of squared error (SSE) of the clusters, and the clusters with a
high SSE can be candidates for postprocessing for identifying distance-
based outliers. For example, in Figure 7.7, clusters 5 and 6 are potential
clusters for postprocessing where the circled points could be evaluated as
potential outliers.

https://doi.org/10.1017/9781108231954.008 Published online by Cambridge University Press

 8
Cybersecurity through Time Series
and Spatial Data

As discussed in Chapter 4, the nature of data and data complexity must be

addressed carefully in any data analytics methods. This chapter presents a
spatial and temporal view of cybersecurity issues. Spatial data generally refer
to geospatial data in the context of cybersecurity. The spatial context comes
from the origination of the attacks or sources of attacks. In addition, attributes
related to the geospatial location may be important to consider, such as socio-
economic and geopolitical factors at play in the region. The network layout can
also have a spatial context. One example of spatial context in cybersecurity is
the geo-origination of anomalies such as massive spam originating from
certain spam hubs around the globe. Identifying geolocations for the origins
of attacks can provide insightful knowledge to address the defense mechan-
isms (Hu et al. 2015). Temporal context refers to the time reference of an
attack or network traffic in general over a period of time. This could refer to
how network traffic evolves over time or how certain trends over time could
indicate threats.
This chapter revisits the spatial and temporal nature of the data and the
challenges emanating from the data. Spatial and temporal approaches relevant
to anomaly detection in cybersecurity are discussed. Finally, examples of
spatiotemporal discovery in cybersecurity are discussed.

8.1 Spatial and Temporal Data

8.1.1 Spatial Data
As discussed in Chapter 4, understanding the nature of data is critical for
spatial data mining because of the inherent complexities of the spatial data,
which consist of data pertaining to space, multidimensional points, lines,

112

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.1 Spatial and Temporal Data 113

rectangles, polygons, cubes, and other geometric objects. Spatial data have
large volumes, so much so that they may require an infinitely large database to
capture the real world precisely. Spatial data not only pose challenges due to
the type of data and volume but also due to the correlation between the objects
and their spatial neighbors with a neighborhood. There can be relationships
between spatial objects such as topological relationships (e.g., disjoint, meet,
overlap, etc.), direction relationships (e.g., north, south, above, below, etc.),
metric relationships (e.g., distance greater than), or complex relationships (e.g.,
overlap with a certain distance threshold in the north), which are based on
these basic relationships. Moreover, the data can depict temporal changes,
which are inherent in the data.
For the purposes of cybersecurity, we focus on georeferenced data, which
include spatial location in terms of latitude, longitude, and other information
about spatial and nonspatial attributes. This essentially means that there is some
preprocessing on the raw data to extract the georeferenced attributes and some
other nonspatial attributes such as geopolitical attributes, other regional attributes
affecting the network trends such as network bandwidths, etc. Chapter 11 dis-
cusses one such specific geospatial example. Figure 8.1 shows a list of various
cyberattack maps that capture this information in a georeferenced manner.

8.1.2 Spatial Autocorrelation

Autocorrelation (Griffith 1987, Haining 2003) refers to the correlation of a
variable to itself. Spatial autocorrelation refers to the correlation of the variable
with itself in space. It can either be positive (spatial clusters for high–high or
low–low values) or negative (high–low or low–high values). Everything is
related to everything else but nearby objects are more related than distant
objects (Tobler 1970). This is considered to be the core of spatial
autocorrelation statistics. Any statistical measure (such as Moran’s I, Geary’s
ratio, or Joint Count; Griffith 1987, Haining 2003) of spatial autocorrelation
relies upon this (Tobler’s) first law of geography.

8.1.3 Spatial Heterogeneity

Spatial heterogeneity is the spatially varying autocorrelation where autocorre-
lation can be positive or negative. It refers to the variation in the region that
may generate characteristic spatial patterns, which could be due to the under-
lying geographical processes and geographical features in the region. This
behavior can be understood in different scales of spatial variation (Haining
2003), including large (macrovariation as autocorrelation) and medium/small

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 114 Cybersecurity through Time Series and Spatial Data

Figure 8.1 Cyberattack maps, based on geolocations of attack origins

and destinations.

(mesovariation as heterogeneity), as well as error (independent noise or

anomalies).
Based on the concept of spatial autocorrelation and heterogeneity, it may be
possible that the geospatial trends within United States may be similar due to
spatial autocorrelation; however, there may also be inherent variability due to
heterogeneity. Let us consider the number of attacks coming into the United
States as shown in the various maps in Figure 8.1. However, every location in
the United States will not be targeted with the same frequency. While the
number of incoming attacks are generally high for the United States, depicting
spatial autocorrelation, there is also a spatial heterogeneity due to the relative
variation of the attacks to different locations inside the United States.
Figure 8.2 shows that United States is the top targeted country; however, the

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.1 Spatial and Temporal Data 115

Figure 8.2 Illustration of spatial autocorrelation and heterogeneity.

number of attacks on the East Coast are much larger than the Midwest due to
the richness of targets in the East Coast. Thus, even though there is a bigger
class of locations, for a more accurate analysis the locations will need to be
subdivided to better evaluate and better allocate resources.

8.1.4 Temporal Data

Let us consider a location or an entity identified by an IP address. This location
or entity may have measurements for a variable separated by a regular interval
of time. For example, the traffic information originating from the IP address is
generated continuously over time (separated by seconds or milliseconds).
These traffic data are linked to the geolocation generating the traffic. These
data by nature are temporal, as they are time ordered and each measurement is
separated by a time tick (seconds, milliseconds, or nanoseconds). The attribute
or variable being measured may be the datagram length (dgmLen) or size of
the packet. There is also a possibility that the object or entity is the network
itself and the variable being measured over time is the number of incoming or
outgoing connections to see the health of the network in terms of the network
traffic. Similar to spatial data, temporal data also have issues of autocorrelation
and heterogeneity.

8.1.5 Temporal Autocorrelation

Two instances of an object in time occurring one after the other (reflected as
two attribute values for this object) have a temporal relationship of adjacency.

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 116 Cybersecurity through Time Series and Spatial Data

Since these instances occurred one after the other, they are possibly similar or
are related in terms of their values (quantified as temporal autocorrelation). The
time difference between the values of the attribute is referred to as lag. For
example, for the variable “packet size” for an IP address, the lag is 1 millisec-
ond. An autocorrelation coefficient measures the correlations (such as using
Pearson’s correlation) between temporal values a certain temporal distance
apart. Here autocorrelation measures the correlation (positive or negative) of
the data variable to itself with a certain time lag. A few approaches (Zhang
et al. 2003) have accounted for this aspect in temporal data mining.

8.1.6 Temporal Heterogeneity

The overall autocorrelation of a temporal data sequence may be positive and
high. However, there may be underlying processes that may bring about some
patterns, which are averaged out over larger time periods. Although temporal
mining accounts for issues such as seasonality, it does not account for the
bigger challenge of heterogeneity. Heterogeneity essentially is the property of
the temporal data due to which it behaves differently even in close time
proximity (Raveh and Tapiero 1980). So a temporal sequence that depicts an
ascending trend may give a high value for autocorrelation when considered for
a larger time window, and on close observation we may find pockets of
variations in these data due to their inherent variability. If autocorrelation
exists in temporal data, it can be assumed that data are coming from the same
underlying process for the points in a time window. However, heterogeneity
indicates that even though the data may be autocorrelated, it may be
originating from multiple processes (Basawa et al. 1984) depicting
microscale variations.

8.2 Some Key Methods for Anomaly Detection in Spatial

and Temporal Data
As discussed in Chapters 6 and 7, anomaly detection deals with discovering
nontrivial and intriguing knowledge, in the form of unusual patterns, objects,
and relationships. Such a discovery works on the principle of identifying
anomalies with respect to similarly behaving partitions in the data.
Accounting for heterogeneity within the data in creating these partitions is
critical to accurately identify anomalies. To understand the need for such
refined anomaly detection, the current approaches to anomaly detection in

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.2 Anomaly Detection in Spatial and Temporal Data 117

spatial and temporal data are discussed with an emphasis on identifying the
correct partitioning, referred to as neighborhood discovery, in spatial and
temporal data. The neighborhoods can also be used as a characterization such
that objects that deviate substantially from this characterization can be identi-
fied as suspicious. This process is similar to anomaly detection using clustering
as discussed in Chapter 7; the key difference is that the preprocessing to form
the similar objects into a neighborhood is different for spatial and temporal
data due to the properties of the data outlined previously. So essentially the
anomaly detection process is similar for all data types, and the key difference is
the preprocessing based on different types of data. The next sections discuss
this process of preprocessing to form the neighborhoods.

8.2.1 Spatial Anomaly Detection

We discuss the spatial anomaly detection in two parts: (1) spatial outlier
detection and (2) anomalous window detection. These two aspects of anomaly
detection are representative and important since in one case an individual
object or a set of objects can be outliers with respect to neighboring objects,
whereas in the other a set of contiguous spatial objects are outliers with respect
to the entire data. This follows the types of anomalies discussed in Chapter 6,
namely point and collective anomalies. In addition, here the context of location
plays a key role in preprocessing the data before the anomaly detection can
take place.

8.2.2 Spatial Outlier Detection

A spatial outlier is an object that is deviant in its behavior from its neighboring
objects. This behavior is quantified in terms of significant difference in attri-
bute values of the object with similarly behaving objects in a certain proximity
(neighbors based on spatial relationships). Thus, the neighborhood definition
(Ester et al. 1997, 1998) is a critical aspect for any spatial data mining task. In
general, a graph-based neighborhood definition facilitates the accounting of
autocorrelation (Shekhar et al. 2001). However, traditional approaches ignore
heterogeneity, i.e., when nearby spatial objects may behave differently. Other
approaches address both autocorrelation and heterogeneity (Janeja et al. 2010).
The issue of outlier detection has been addressed in several approaches and has
been extensively studied (Aggarwal 2017). Several techniques consider neigh-
borhoods that are based primarily on spatial relationships, thus they only
account for autocorrelation but ignore spatial heterogeneity. Second, the
neighborhood formation is not order invariant. Thus, a neighborhood based

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 118 Cybersecurity through Time Series and Spatial Data

on cardinality (number of objects in the neighborhood) will lead to different

neighborhood formation and outlier detection outcomes with different starting
points. Third, the outlier detection considers the deviation in terms of a single
attribute only or Euclidean distance between multiple attributes and not inter-
relationships between attributes. Some approaches specifically discuss these
issues and consider the treatment of both autocorrelation and heterogeneity
(Janeja et al. 2009, 2010).
Anomalous window detection deals with the discovery of contiguous set of
objects forming an unusual window. In most clustering techniques (such as
DBSCAN (Sander et al. 1998), OPTICS (Ankerst et al. 1999), the focus is on
the identification of the major groupings in the data, as opposed to unusually
ordered groupings. Similar to clustering, outlier detection identifies individual
outliers and does not quantify the degree of unusualness of an outlier.
Scan statistic approaches (Naus 1965, Glaz et al. 2001) detect a group of
objects behaving anomalously with respect to all the objects in the dataset and
detect quantifiable anomalous behavior in multiple attributes in relation to one
another. A variation of the simple scan statistic is the spatial and spatiotemporal
scan statistic (Kulldorff 1997, Kulldorff et al. 1998), which detects unusual
space or space-time windows using circular or cylindrical windows respectively.
However, this approach does not account for the combination of both
autocorrelation and heterogeneity. Most of these approaches use circle
(Openshaw 1987, Besag and Newell 1991, Kulldorff 1997), square, rectangular
windows (Neill et al. 2005) and several complex shapes (pyramid with square
cross sections) (Iyengar 2004). FlexScan (Tango and Takahashi 2005) identifies
flexible shaped scan windows and only considers adjacency as the factor for
moving to the next spatial node in the K 1 nearest neighbors, and therefore
overlooks heterogeneity. Duczmal and Renato (2003) use simulated annealing to
identify arbitrarily shaped spatial clusters, by absorbing not only the surrounding
regions with nonelevated risks but also faraway regions with nonelevated risk,
extended with a compactness measure to penalize very large-sized windows.
The linear semantic–based scan statistic (LS3) (Janeja and Atluri 2005) extends
the basic scan statistic approach. Additionally, Janeja and Atluri (2005, 2008),
Shi and Janeja (2009) and Tang et al. (2017) propose approaches accounting for
heterogeneity with improved results as compared to several other approaches.
Thus, most of the scan statistic–based approaches suffer from two major
issues. First, the window shape is not free-form. Second, the anomaly detection
is primarily based on autocorrelation only and does not take into account
heterogeneity. It is interesting to note that recently spatial heterogeneity has
been identified as a major issue in various spatial applications, but its quantifi-
cation has suffered from the absence of a clearly outlined definition.

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.2 Anomaly Detection in Spatial and Temporal Data 119

The traditional anomaly detection methods, for example anomalous window

detection (window shapes such as a circular window), assume autocorrelation
by virtue of the proximity of the spatial objects. Subsequently, the deviation
of these objects is compared to the rest of the data. However, spatial analysts
have recognized that every location has an intrinsic degree of uniqueness
(Miller 2004).

8.2.3 Spatial Neighborhood

The goal of this step is to identify spatial neighborhoods accommodating both
spatial autocorrelation and heterogeneity in the region. The neighborhood
should not be identified simply by the change in geospatial features or solely
on the basis of spatial proximity determined using spatial relationships such as
adjacency of objects. Spatial neighborhoods can be considered a characteriza-
tion of the region (as shown in Janeja et al. 2010). This approach consists of
the following distinct steps:
(1) Generate the immediate neighborhood of an object, called a
microneighborhood, as shown in Figure 8.3a, using Voronoi tessellations.
This captures the knowledge about the immediate neighborhood of this
object such as the presence or absence of spatial features (e.g., proximity
to geopolitical events or prior known cyberattacks from this location) in its
proximity and attributes of the object. This will allow us to associate
features and attributes to these objects, which can be used to determine
similarity or dissimilarity across spatial objects.
(2) Identify spatial relationships (such as adjacency) using the spatial attri-
butes of the microneighborhoods to accommodate autocorrelation.

Figure 8.3 (a) Micro neighborhood; (b) macro neighborhood.

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 120 Cybersecurity through Time Series and Spatial Data

m2

IP1
m1

IP2
mi
Macro neighborhood

Figure 8.4 Neighborhood as an underlying characterization.

(3) Capture spatial heterogeneity by identifying semantic relationships

between the nonspatial attributes and features in the microneighborhoods.
A semantic relationship is defined using similarity coefficients such as
Jaccard and Silhouette coefficients. If there is a spatial and semantic
relationship between spatial objects, the microneighborhoods are merged
to form macroneighborhoods as shown in Figure 8.3b. This facilitates the
determination of the local heterogeneous patterns across neighborhood in
terms of the attributes associated with each spatial object. Thus, the
macroneighborhood captures both autocorrelation and heterogeneity in
the neighborhood definitions by not only considering proximity but also
the change in the localized features and attributes in the region.
This neighborhood definition can be used as a precursor to the anomaly
detection in reference to this neighborhood. Similarly, if an anomaly or an
attack event is discovered in one part of this neighborhood, then chances are
that the attack can propagate to other parts of the neighborhood given the
similarity and homogeneousness in this neighborhood.
Neighborhoods can be seen as an underlying characterization as well. For
example, if two IP addresses are communicating with each other (e.g., IP1
sending a packet to IP2), then using geolocation (through databases such as
Maxmind1), these IP addresses can be located and overlaid on top of the
neighborhood definitions, as shown in Figure 8.4. For example, if
macroneighborhood m1 has a series of features indicating that it has known
prior attacks originating and is a source of geopolitical unrest, then this IP1
gets a certain reputation score and should be monitored carefully. This will
help network administrators go beyond the white and black listing of IP

1
www.maxmind.com/en/home.

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.2 Anomaly Detection in Spatial and Temporal Data 121

addresses and use geolocated information sources as well. One potential

challenge to using geolocation information of IP addresses is the vulnerability
of an IP address to be spoofed such that an attacker can illicitly impersonate
another host by using a forged IP address. We envision geospatially enhanced
models to work as a supplementary model that bolsters a network analyst’s
existing knowledge. Geospatial models may not effectively work for network
traffic with high probability of being spoofed, for example in UDP traffic.
However, using geospatially enhanced data helps provide geocontextual infor-
mation as well, potentially enhancing detection accuracy.

8.2.4 Temporal Anomaly Detection

There is a rich literature in temporal data mining and statistical time series
analysis. The aspects closest to anomaly detection in temporal data are
reviewed here. Specifically, this section looks at approaches that facilitate
the discovery of relevant segments in temporal data (equivalent to the neigh-
borhood definition in spatial literature) and discovery of anomalies such as
outliers and discords in such datasets are discussed.
Change point detection addresses the discovery of time points at which the
behavior of time series data change. Several techniques (Aminikhanghahi and
Cook 1994) have been proposed to detect change points in time series data.
However, most of these approaches assume that a phenomenon can be
approximated by a known model whereas in reality several dynamic models
may be in play. Yamanishi and Takeuchi (2002) proposes an approach for
outlier and change point detection in nonstationary time series data. Guralnik
and Srivastava (1999) outline an approach for event detection from time series
data by detecting the change of model or parameters of the model that describe
the underlying data.
Discretization has been extensively studied, and many efficient solutions
exist for simple discretization techniques, including equal width or equiprob-
able discretization and piecewise aggregate approximation (PAA) (Lin et al.
2003, Duchene et al. 2004) Although some approaches (Tim et al. 1999,
Duchene et al. 2004) have looked at the problem of variable width discretiza-
tion, they do not address variable number of widths across multiple time series
generated in a spatial proximity.
Time series discords are subsequences of a longer time series that are
maximally different to all the rest of the temporal subsequences. Discord
discovery was introduced by Keogh et al. (2005). Ameen and Basha (2006)
propose improvements on the discord discovery algorithms. Wei et al. (2006)
describe an approach for finding discords in shapes by converting shapes into

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 122 Cybersecurity through Time Series and Spatial Data

time series data. Yankov et al. (2007) discover unusual time series or discords
with a disk-aware algorithm.
First, most of these approaches consider autocorrelated segments as true
segments for pattern discovery. This may ignore heterogeneity in the discovery
process, possibly in the form of unequal width segments. Second, most of
these approaches do not quantify the unusualness of a patterns discovered in
terms of how distinct the pattern is from the data. Similar to spatial behavior,
temporal autocorrelation, and heterogeneity are a challenge.
Finally, in the area of spatiotemporal mining, several approaches have
explored the spatiotemporal nature of the data (such as Abraham and
Roddick 1999, Roddick and Spiliopoulou 1999, and Roddick et al. 2001).
As such, there does not exist a generalizable framework that facilitates the
accommodation of the properties of space and time in an effective and efficient
manner. Additionally, there is no consensus on spatial-dominant or temporal-
dominant analysis of spatiotemporal data. Thus, there is a need for a shift that
facilitates the analysis based on the properties of space and time (spatial,
temporal autocorrelation, and heterogeneity) in a unified framework.

8.2.5 Temporal Neighborhoods

Let us consider a computer network and its connectivity as a time series dataset
where we have time stamps associated with a set of attributes such as number
of nodes, number of edges, and total traffic. Similar to spatial neighborhood
discovery, temporal neighborhoods can be seen as a precursor to anomaly
detection. Let us now consider these network properties over time to demarcate
temporal neighborhoods.

8.2.5.1 Stationary Distribution-Based Merging

If we consider a time series, with many small time intervals, the end goal is to
generate merged intervals (or neighborhoods). The idea here is to generate
similarly behaving time segments so that these neighborhoods can be used for
anomaly detection within these time segments. One approach models the time
series data as a Markov model (Dey et al. 2009, 2014). This approach starts
with many small equal frequency bins in the data as the states of a Markov
model (Figure 8.5a). The similarity between the bins is computed using a
distance measure. A transition matrix based on these similarities is generated
for this Markov model and subsequently normalized to obtain a row-stochastic
matrix. In order to form the larger time segments or similar temporal neigh-
borhoods, an intuition is used that the adjacent bins in the transition matrix

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.3 Cybersecurity through Spatiotemporal Analysis 123

Figure 8.5 Markov stationary distribution-based discretization.

having a high degree of probability of transition should be merged in order to

obtain the unequal depth bins.
Thus, a Markov stationary distribution-based merging is used to form the
stationary distribution vector (Figure 8.5b, e). The split points in this vector are
identified by finding spikes in the stationary distribution vector plot
(Figure 8.5e) (Brigham 2002).
This temporal neighborhood indicated by the split time series in Figure 8.5d
can be used as a precursor to anomaly detection to identify anomalies in the
temporal neighborhoods, which are more homogeneous for an accurate identi-
fication of unusual events. Moreover, the spikes in the distribution also may
correspond to events of interest. Thus, if we apply this method to the network
traffic data or even data about the network connectivity, we can identify the
more homogeneous neighborhoods as the regular traffic patterns and the
smaller spikes as unusual activity in the network. This approach is fairly robust
to small-scale noise or data-to-day fluctuation in the data.

8.3 Cybersecurity through Spatiotemporal Analysis

8.3.1 Spatial Anomalies in Cybersecurity
While cyber threats that target networks are directed toward specific nodes
during specific time periods, these nodes are not randomly selected but may be
located in specific geographical regions. According to Davies (2012),

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 124 Cybersecurity through Time Series and Spatial Data

Kaspersky reported that Iran is the most common place where the Flame
malware was discovered. Cases were also reported in Israel, Palestine,
Sudan, Syria, Lebanon, Saudi Arabia, and Egypt. Out of all computer systems
that were affected worldwide by Stuxnet, 60% of them belonged to Iran
(Halliday 2010, Burns 2012). Also, a recent analysis indicates that most online
schemes occur in Eastern Europe, East Asia, and Africa. These schemes and
phishing attacks are shifting to other countries like Canada, United States,
France, Israel, Iran, and Afghanistan (Bronskill 2012). Thus, the spatial elem-
ent plays a key role in strengthening the cyber infrastructure, not only com-
puter networks but also in power grids and industrial control systems.
Therefore, while identifying locations of cyberattack victims is important in
a power grid or computer network, it is also crucial to determine the location of
those responsible for these attacks, as it may have serious national security
implications. Cyberattacks may be targeted toward specific countries due to
sociopolitical factors. It is therefore important to identify the association
between a cyberattack, its target, and the origin location.
A spatial aspect to cyber data analysis can be brought through (a) identifi-
cation of spatial neighborhoods in the network structure to discover the impact
of an event in space, (b) studying influence of events in neighborhoods, and (c)
studying correlation and associations of events over time.

8.3.2 Neighborhood-Based Anomalies

Several works have studied spatial neighborhood definition in sensor networks
(Janeja and Alturi 2008, McGuire et al. 2008, 2012, Janeja et al. 2010, Sun
et al. 2012). The neighborhood discovery can be used to describe influence
flow among spatial objects. In the network, vertices represent spatial objects,
and edges represent the possible influence flow. The weight on each edge
indicates the hardship the influence faces to reach from one spatial object to
another. Such a weight takes into account the spatial distance and other factors
such as barriers, which can be network properties such as number of hops as
the distance measure, geopolitical factors such as IP domain, country, etc. So,
for instance, Mexico and United States, even though they are in proximity, will
have different rates of spread of an event due to differences in the cyber
infrastructure. Thus, even though they are in proximity, their links’ weight
will be less if we are studying event propagation. This was made evident even
within the United States when a major breach of social security numbers
(Kosner 2012) occurred in South Carolina that did not affect the nearby or
any other states. It is obvious that there are other organizational contexts at
play that create geospatially heterogeneous behaviors.

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.3 Cybersecurity through Spatiotemporal Analysis 125

Let us consider a network of nodes in a spatial neighborhood, where the

spatial nodes are connected by edges, based on spatial relationships, with
weights computed from the distance between the spatial and nonspatial attri-
butes through the previously discussed similarity coefficients. So given an
event being studied, such as a specific attack if we want to study the propaga-
tion of the attack, influence distance can be used to quantify it. The influence
distance from one spatial object IPp to object IPq is the sum of the weights of
the constituent edges of the shortest paths from IPp to IPq. If IPp and IPq are
unconnected, then the influence distance is ∞, such as in the example of South
Carolina. If IPp to IPq are the same, then the distance is 0.
The difference of the influence distance between events can also be com-
pared. For example, an event such as a local breach only has an influence on
spatial objects located on a direct connection, whereas a global event can
impact a much larger part of the network.
For example, a phenomenon such as a distributed denial of service (DDoS)
attack will have a much larger influence than a single computer infected by a
virus. One can argue that even a single infected computer can also lead to a
larger impact; however, if we model the network appropriately, then that
should be captured in the weights between the edges and subsequently on
the influence distance of this infected system.
These concepts can also be used to model attacks in nuclear power plants or
highly networked organizations. Similarly, multiple organizations using the same
software with a vulnerability can also be modeled using the influence distance.

8.3.3 Spatio temporally Clustered Anomalies

A network can also be seen as in a spatial layout to discover various types of
clustered network anomalies. This can be done based on spatial window-based
techniques discussed earlier (Kulldorf 1997). Figure 8.6 outlines an example
using a scan statistic–based approach. A network can be laid out on a spatial
grid based on relationships of network bandwidth usage and usage of the
network, as shown in Figure 8.6. Subsequently, using any tcpdmp software
the packet data can be extracted into population and case files, where the
population file has the overall network usage information extracted from
historic data and the case file has the current usage information. Here network
usage can be measured in terms of a variable of interest, such as number of
packets every minute or average packet size or total packet size. Once the files
are created from the raw data, spatial and temporal anomalies can be identified.
These anomalies are in the form of (a) nodes on the network that behave
anomalously with respect to other nodes (or geospatial points) on the network,

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 126 Cybersecurity through Time Series and Spatial Data

1.2

ISP switch

1
Server 2

0.8 CISCO 1 Server 3 Media server

USAGE LOW TO HIGH

0.6 Web server

Sharepoint
0.4
server
Database
server
0.2

0
0 0.2 0.4 0.6 0.8 1 1.2
RELATIVE DISTANCE AS BANDWIDTH USAGE

Figure 8.6 Spatio temporally clustered anomalies.

and (b) cluster of network traffic points that are anomalous as a temporal
cluster with respect to other neighboring servers.

8.3.4 Insights through Geo-Origination of IP Addresses

The geolocation of IP addresses sending emails can provide insights into
origination of spam or phishing activities. Certain parts of the world are more
known to be related to spam or phishing activities (for example, top 10
countries2 that are the worst spam enablers or top phishiest countries3).
Geolocation4 of the (nonspoofed) IP address identifies the physical location
down to latitude and longitude. Such information can be utilized in tools
providing actionable information of user emails. Keim et al. (2005) describe
a visualization tool to mine and analyze spatiotemporal aspects of emails.
A histomap provides a visualization of spam origination from across the world
and within the United States. Email routes can be visualized to discover

2
www.spamhaus.org/statistics/countries/. The world’s worst spam enabling countries, last
accessed March 2021.
3
https://news.netcraft.com/archives/2012/12/13/world-map-of-phishing-attacks.html. Phishiest
countries, last accessed March 2021.
4
http://geoiplookup.net/, last accessed August 2017.

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 8.4 Temporal Behaviors in Evolving Networks 127

unusual locations and routes. This can help users identify potential threat
locations, particularly when users are in a sensitive job or secure location.

8.4 Temporal Behaviors in Evolving Networks

Computer network traffic over a period of time can be studied by utilizing the
temporal neighborhoods and help draw out several forms of communication
patterns, as discussed in Chapter 4. For example, one can identify if some nodes
are popular on the network, either as a common source point to all other nodes or
consistently a common destination point from all other nodes on the network.
One can also determine if the popularity of a node either increases or decreases
over time. Studies have discussed various properties in networks. These proper-
ties include but are not limited to the centrality of nodes in a network, the
densification of a network, and the diameter of a network (Leskovec et al. 2005,
Leskovec and Faloutsos 2007, Leskovec 2008, Kang et al. 2009).
Studies have portrayed that networks tend to follow the properties of
densification and shrinking diameter of a network (Lescovec et al. 2005), they
have been applied to study of communication networks, specifically network
traffic flows (Namayanja and Janeja 2015, 2019). More so, these studies have
accounted for the importance of nodes on the network and how this may affect
the densification of the network and the shrinking diameter over time. These
approaches determine how change in the behavior of a central node on the
network may impact the overall behavior of the network over time.
Therefore, the mining of large networks can be performed with the aim of
understanding the following: Does the structure of the network change over
time? Can we define what changes occur as the network evolves? And most
importantly, can we identify when these changes in the network take place? By
detecting changes in a temporally evolving network, one can drill down and
identify the underlying cause of this change in the network, which could have
come forth as a result of a threat to the network. For example, the unexpected
unavailability of a server system from the network could be due to an unknown
denial of service attack. Using a temporal approach can discover inconsist-
encies in the network and identify times when these inconsistencies occur.
This is in turn useful in the efficient allocation of resources for monitoring
network activity. Chapter 9 discusses graph-based techniques and using tem-
poral evolution of the network as a way to measure the threat propagation
over time.

https://doi.org/10.1017/9781108231954.009 Published online by Cambridge University Press

 9
Cybersecurity through Network and Graph Data

Graphs (as discussed in Chapter 4) depict the relationships between various

entities. A graph is comprised of nodes and links (or vertices and edges). The
node or vertex represents an entity such as an IP address, and the edge or link
represents the relationship between the two entities, such as a message sent
from one IP address to another. Thus, network traffic data lend themselves to
graph representation, which opens up a vast array of graph-based analytics that
can be applied to studying cybersecurity issues.
Figure 9.1 depicts the cybersecurity applications that can utilize graphs.
These are only some of the major examples from research and application
themes:
• Communication graphs: Network traffic has been extensively studied as a
graphs; here source and destination IP addresses are vertices, and the
communication between them is the edge. Various properties such as cen-
trality, density, and diameter have been used to evaluate the state of the
graph and study communication patterns over time (Lescovec et al. 2005,
Kang et al. 2010, Tartakovsky et al. 2013, Namayanja and Janeja 2014).
• Attack graphs: These graphs (Wang et al. 2007a, b) are used to evaluate
exploits and their pre- and postconditions. Here the vertex can be an exploit,
and the edges connect to the conditions that occur before or after the exploit.
Attack graphs can also be used to evaluate conditions in a network model
that violate security policies (Jha et al. 2002). Thus, such graphs help in
studying various paths that can be taken in a graph that propagate an exploit.
Attack graphs should ideally be exhaustive and yet succinct to provide a
comprehensive view of the network health in terms of the security policy.
Violations discovered can help make a network more robust through
enforcement of more stringent security policies.

128

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 Cybersecurity through Network and Graph Data 129

Traffic Graph properties : centrality,

Source and desnaon IP as density, diameter to measure the
verces; communicaon as edge level of communication in the traffic

Security policies Network model M, security policies

Graphs in computer network–based
P, all executions of M violating P,
cybersecurity applicaons Modeled in aack graphs exhaustive and succinct

Vulnerabilities and threats Evaluating similarity between

graph verces are threats, edge is threats, identify morphed
similarity between them vulnerabilities

Figure 9.1 Graph-based studies for cybersecurity applications.

• Threat similarity–based graphs: Graphs have been used to find relation-

ships between known vulnerabilities. In such graphs, known vulnerabilities
are the nodes, and the similarity between two vulnerabilities is the edge. The
weight on the edge depicts the level of similarity. These types of graphs can
be augmented with semantic and contextual information (AlEroud and
Karabatis 2013). These graphs can be queried when a new unknown threat
is potentially observed, such as a zero-day attack. By computing the simi-
larities between the zero-day attack and the known attacks, a score can be
provided to indicate the level of threat from this zero day (AlEroud and
Karabatis 2014).
• Threat propagation: A cyber threat not only impacts an individual entity
(one single computer or IP address) but can easily propagate to other
connected entities (Ingols et al. 2006, Trinius et al. 2009, Ohm et al.
2020). Studies have been conducted to identify the spread and prevention
of threats such as in high-risk groups of IP addresses. Some of these studies
and approaches are restricted to individual IP behavior, such as black and
white listing IP addresses, which limit the understanding of the dynamics
involved in the spread of the threat. Apart from individual behavior, high-
risk behavior can also be captured at two other levels: the microsocial level
(e.g., personal network), and macro level (through analysis of the network
structural factors influencing the transmission mechanism). Computer net-
works and communication across the networks provide a framework to study
the spread and prevention of high-risk behavior for widespread cyberattacks.
Several relevant graph-based anomaly detection applications are also dis-
cussed in Akoglu et al. (2015). In the following sections, this chapter evaluates
basic graph properties that can be utilized and discusses some specific types of
methods within these themes.

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 130 Cybersecurity through Network and Graph Data

9.1 Graph Properties

Certain graph properties are central to the graph mining used in cybersecurity
applications. For a general introduction to graph theory and those properties
relevant to anomaly detection, several texts are available, such as Rivest and
Vuillemin (1976), Wilson (1986), and Akoglu et al. (2015). Some key proper-
ties are described in this section.
Adjacency matrix: A graph can be represented as an adjacency matrix. Let
us consider a graph G with E edges and V vertices. The adjacency matrix of
G is a V  V matrix where each cell in the matrix has a 0 for no edge between
the pair of vertices or 1 for an edge between the pair of vertices. If there are
multiple edges between two vertices, then the value of the matrix cell is the
number of edges. A vertex degree is the number of edges incident on a node or
vertex. In an undirected graph, the matrix is asymmetric, as shown in
Figure 9.2a. The sum of the row or column in the asymmetric matrix is the
out degree (edges going outward from a node, such as outbound network
communication from an IP address) and in degree (edges going inward to a
node, such as inbound network communication to an IP address) respectively.
In an undirected graph, the matrix is symmetric, as shown in Figure 9.2b. The
sum of the row or column is the same, which is the degree of a vertex (such as
the total of the inbound and outbound communication for an IP address).
Walk: In a graph, the sequence of edges leading from one vertex to another
is called a walk. A walk where no vertex appears more than once is a path. The

Figure 9.2 (a) Directed and (b) undirected graphs.

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 9.1 Graph Properties 131

distance between two vertices is the shortest path between the two vertices,
also referred to as geodesic path. A graph can be disconnected when some
parts of the graph are not connected by a path, or if a single node is discon-
nected it means that there is no edge connecting it to the graph. For example, in
the graph in Figure 9.2b, there are three separate (disconnected) components.
Centrality: Centrality in a graph structure indicates prominence of a node in
a graph. A node with high centrality will potentially be more influential.
A simple type of centrality is the degree centrality (Freeman 1978), which
indicates the total number of edges incident on a node in an undirected graph.
This can be divided into in-degree and out-degree for directed graphs.
Betweenness centrality (Freeman 1978) of a vertex v measures the number
of geodesic paths connecting vertices ij that are passing through v, divided by
the total number of shortest paths connecting ij. This is summed across all ij to
get the overall betweenness centrality of v. Akoglu et al. (2015) provides a
detailed survey of several different types of centrality, such as betweenness,
closeness, and eigenvector centrality, among others.
Cliques: Traditionally, a clique is defined as any complete subgraph such
that each pair of the vertices in the subgraph are connected by an edge, for
example nodes b, n, and g in Figure 9.3. Cliques may be mapped to coteries
(such as high-risk groups in the context of spreading a virus or DoS attack).
More importantly, the nodes among different cliques may overlap, which
indicates a higher relevance of those nodes in high-risk groups.

Figure 9.3 Graph example.

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 132 Cybersecurity through Network and Graph Data

Figure 9.4 Using graph properties to evaluate consistency of nodes across time.

Given the adjacency matrix for the graphs in Figure 9.3, the visualization is
generated using NodeXL.1 Figure 9.3a presents a directed graph view, and
Figure 9.3b shows an undirected view of the graph. In Figure 9.3a, it is evident
that node k has a high out-degree, and nodes h, n, r, and t have a high in-
degree. Overall, vertex n has the highest degree. In terms of centrality in an
undirected graph vertex, n will be considered the most central node.
Another example of using graph properties is to evaluate consistency of
nodes over time. For example, Figure 9.4 depicts three time periods, T1, T2,
and T3, with their corresponding graphs. Each time period has a set of nodes
a–f that communicate with each other as depicted by the edges. The corres-
ponding graphs and degrees are shown in the figure. A degree threshold can be
used such that anything less than or equal to the threshold is not central. For
example, in this dataset, the threshold is two, then the central nodes are shown
that are equal or greater than this threshold. Node c, which is central in time
period T1, does not show up in the other time periods. As discussed in
Chapter 3, association rules and their related metrics of confidence and sup-
port, showing the likelihood and strength of associations or frequent co-
occurrences, can be utilized here. Association rules for each time period are
shown that provide a complementary view of the communication in this
example. For example, f, which is not a central node, is also not frequent, as
support is low. However, in all time periods, the rule f  e is always present.

1
https://nodexl.com/.

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 9.2 Understanding Evolving Network Communication 133

Similarly, the nodes a and d are consistently central, and the rule a  d is
consistently a strong rule in terms of support and confidence. On the other
hand, a  e has high confidence, and both nodes are central; however, the
support is low compared to other rules with central nodes. Thus, multiple types
of analysis can be used to analyze the network traffic data to provide
complementary knowledge.

9.2 Graphs for Cybersecurity: Understanding Evolving

Network Communication
Utilizing the temporal neighborhoods, as discussed in Chapter 8, communi-
cation networks over a period of time can be evaluated as shown in the
example in Figure 9.4. Several forms of communication patterns can be
discovered. Examining certain key graph properties can help us understand
the overall behavior of a network, as discussed next.
Centrality over time: The level of connectivity to or from a given node can
be used to measure the centrality of the node. This level of centrality can be
used to determine the importance of a node and how critical this node is on the
network, by virtue of its high degree of connectivity in the graph. For example,
as discussed earlier, the degree centrality is the number of edges incident to a
given node. Page rank provides the number of highly central nodes that are
adjacent to a given node. The number of nodes that are connected to a given
node and the quality of the nodes that connect to it can both be identified.
Intuitively, one can expect to find certain nodes to be consistently central over
time periods. Alternatively, a node that is central only at one or few time
periods is interesting to study as it may be a source or destination of an attack,
or alternatively, due to the period in time, may have taken an important role in
the network. The idea here is to not predict an anomalous central node but to
narrow down nodes of interest, especially in massive network traffic data,
which may lead to a very complex graph.
Densification over time: Studies have indicated that an increase in the
number of nodes on the network over a period of time also leads to an increase
in the number of edges. This process is referred to as densification of a network
(Lescovec et al. 2005). However, these studies have looked at densification of
all nodes and edges on the network without looking at evolving densification
(Faloutsos et al. 1999, Phillips 1999, Fabrikant et al. 2002, Kang et al. 2010).
Densification can also be evaluated in terms of the highly central nodes.
Determining the change in the behavior of a central critical node – for

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 134 Cybersecurity through Network and Graph Data

example, its unexpected absence from the network – can identify whether this
affects the densification of the network. If densification is tracked over a period
of time, it is possible to identify time points where the densification is
anomalous compared to the overall trend in the network. Such time periods
may indicate unusual changes in the network. Now if such changes are
consistent over multiple, large time periods (depicting periodicity, for example
every year in December), then this can be related to a trend; however, if this
happens in only one sample year, then it may indicate an anomalous
time period.
Diameter over time: Densification draws nodes closer to each other such
that the distance between those nodes that are far apart becomes smaller. As a
result, the diameter of the network reduces, a concept commonly described as
the shrinking diameter (Lescovec et al. 2005). Similar to densification,
changing the diameter of the network over time can provide insights about
the health of the network traffic. This analysis can be performed for the central
nodes, which is particularly helpful for massive datasets. The change in the
behavior of a critical node, for example its unexpected absence from the
network, can impact the diameter of the network. Thus, drastic changes in
the diameter, such as expanding diameter, can indicate malicious effects on the
central nodes.

9.3 Graphs for Cybersecurity: Similarity-Based Redundancy

in Router Connectivity
Router configuration can be a difficult and complex task, especially on larger
networks, which can typically have hundreds of routers (Caldwell et al. 2004).
There is a clear need for automatic router configurations; however, many
organizations continue to rely on manual configurations and reconfigurations,
leading to several network outages. After several reconfigurations, it is pos-
sible that the network topology may be considerably different from the origin-
ally planned topology. As a result, problems such as bottlenecks or redundancy
may crop up.
A bottleneck router can be a router that is dissimilar to other routers on the
network in terms of its connectivity, such that it has a unique function to
perform through its unique set of connections. Other parts of the network will
rely on it to connect to keep certain parts of the network connected, and if this
router is impacted in any way to perform its functionality, then the overall
network connectivity may suffer.

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 9.3 Similarity-Based Redundancy 135

Figure 9.5 Network redundancy.

A redundant router, on the other hand, can be highly similar to other routers
in terms of its connectivity. Taking a redundant router offline would probably
have little effect on the network connectivity. Such a redundant router can act
as a fail-safe if it is similar to some critical network routers. On the other hand,
this redundant router can be used for load balancing or reallocating it to other
useful roles.
Router connectivity can be considered as a graph, and similarity coefficients
can provide an intuitive method of determining whether a router is similar or
unique in relation to the other routers connected on the graph representing the
network. As discussed in Chapter 3, several similarity coefficients have been
studied (Lewis and Janeja 2011). An example use is shown in Figure 9.5 using
the Jaccard coefficient (JC). An example graph with various routers connected
is shown in Figure 9.5a. The corresponding adjacency matrix is also depicted
in Figure 9.5b. Pairwise similarity between each pair of the routers is com-
puted using JC, which is given as positive matches (1–1 match) divided by the
sum of positive matches and positive mismatches (0–1 or 1–0 mismatch) as
shown in Figure 9.5c. Thus, looking at the two rows of connectivity in the
adjacency matrix for routers R4 and R3, we can see that they have perfect set
of matches (1–1) and no mismatches (0–1 or 1–0). Thus, JC for similarity
between R3 and R4 is 1, and they can be seen as redundant routers that are
perfectly similar to replace each other’s role in terms of their connectivity. On
the other hand, when we look at the total similarity of each node and its degree
in Figure 9.5d, e respectively, we can see that nodes R5 and R6 have lowest

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 136 Cybersecurity through Network and Graph Data

degree and lowest similarity, thus they are the most vulnerable nodes in terms
of connectivity. Similarly, nodes R1 and R2 are highly central but have only
some similarity. Thus, they are possible bottleneck nodes due to their unique
connectivity and highly central nature.
Such preemptive exploratory analysis using basic graph properties can help
facilitate the network management for better response to threats and recovery
from network impact. For instance, in the preceding example, if there is better
planning in terms of identifying bottlenecks and redundant nodes, the reme-
diation and recovery from a network-based cyberattack can be facilitated.

https://doi.org/10.1017/9781108231954.010 Published online by Cambridge University Press

 10
Human-Centered Data Analytics
for Cybersecurity

Humans are a key part of the security infrastructure (Wybourne et al. 2009).
Several studies have looked at the human aspect of cyber threats from the user
or attacker perspective. Given that more and more users use home computers,
systems are becoming vulnerable and susceptible to more cyberattacks,
Anderson et al. (2010) hypothesized that psychological ownership of one’s
computer is positively related to behavioral intentions to protect one’s own
computer. More and more research has suggested that social, political, eco-
nomic, and cultural (SPEC) conflicts are a leading cause for cyberattacks
(Gandhi et al. 2011). This chapter looks at the different key players’ view-
points in the cyberattack and how they can be strengthened to prevent and
mitigate the risks of the attack.
In general, there are two key parties at play in a cyberattack: a user who is a
victim and the attacker. In a cyberattack against a business, there are three key
parties in play: the business, a user or employee, and an attacker. In addition,
there may also be other types of employees involved, such as network adminis-
trators or system administrators who are configuring the systems, and pro-
grammers who are programming functionalities.
While this is a compartmentalized view of the real-world fluid scenarios, it
helps us understand the dynamics of an attack and study ways to prevent it.
Each of these parties have a set of constraints and factors that help facilitate or
restrict the perpetration of the attack.
A user or employee in a business comes with a certain educational back-
ground, experiences, and any lessons learned in the past. In addition, a user
also has certain psychological traits that impact how they respond to certain
types of requests, from clicking of links, downloading content, or providing
their credentials.
Businesses have to follow their business model, which could be highly
secure or more open. They have a mission and a set of principles they abide by.

137

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 138 Human-Centered Data Analytics for Cybersecurity

Figure 10.1 Key players in a cyberattack.

Moreover, a business may or may not have a well-defined set of security policies.
Lastly, the types of business applications being run in a business may have their
own set of complexities and challenges.
In addition to having technological know-how, an attacker has certain
motivations and an end goal to achieve through perpetrating the attack. An
attacker may have certain psychological traits that lead them to instigate these
attacks not unlike other types of crimes. Greitzer et al. (2011) has implemented
a psychological reasoning for human behavior in cyberattacks and categorized
the attackers with disgruntlement, difficulty accepting feedback, anger man-
agement issues, disengagement, and disregard for authority.
Additionally, depending on the amount of resources an attacker may have,
the extent of the attack may vary from a single computer to a network or entire
business being targeted. An attacker targets a user of the business systems, be
it a technology worker or an end user, utilizing the business systems. Some of
these aspects of the key players in a cyberattack are depicted in Figure 10.1.
This chapter addresses certain types of cyberattacks that have a major influence
of the human perspective, such as phishing and insider threats. It outlines data
analytics methods to study human behaviors through models extracted from data.

10.1 Human Perspective to Cybersecurity: Phishing

A cyber threat can come in several forms, such as phishing, viruses, and
vulnerabilities, to name a few. Let us consider the case of spear phishing,

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 10.1 Human Perspective to Cybersecurity: Phishing 139

where an individual receives an email claiming to be from an organization

known to the individual and is asking for the individual’s username and
password or interacting with the user through a malicious link in the email.
While users with knowledge of security threats may be inclined to check the
sender’s email address and the validity of the link for authenticity, others may
interact with the message, compromising security.
In general, cybersecurity can be aimed toward hardening our systems or taking
offensive approaches. However, in this case, cybersecurity cannot simply be seen
as an organization-centric issue but an individual user-level issue. In social
engineering–type attacks, such as phishing, indeed it is an individual user who
may lead to opening up an organization to bigger risks (for example, by clicking
on an unsafe link or providing the users credentials). Recent cases of cybersecurity
breaches in the commercial (Rashid 2013) and government sectors (Hoffman
2011, Zetter 2011) indicate that in the majority of cases, a user’s choice can be
attributed as the root cause of the cyberattack. A recent study by Trend Micro
found that 91% of all targeted attacks are initiated with a spear-phishing email,
potentially leading to advanced persistent threats (APT) (Trend Micro 2012).
Therefore, the human perspective is critical to evaluate and address questions
such as: what makes a user click on a potentially malicious link? Why would
a user open a malicious website or provide their credentials? How do users
determine whether an email appears to be fraudulent?
There is also an organizational perspective in a user’s behavior. It’s import-
ant to consider if an organization has checks in place to prevent phishing
emails reaching from email inboxes. Are there ways that the organization
labels emails? For example, many organizations include messages or tags
indicating that the email is external so that it is not incorrectly seen as
originating from a supervisor or a colleague in the organization – which is
generally the mask in spear phishing emails. This is a simple way to tag emails
but may be quite effective.
An evaluation of this perspective can lead to insights into preventing future
cyberattacks. Built on prior psychological and behavioral research on phishing
(Grazioli 2004, Davinson and Sillence 2010, 2014, Wright and Marett 2010), it
can be hypothesized that end users’ experience and perception play critical roles in
the behavioral process of phishing detection. However, the perceptions of the user
can be changed. Specifically, a user’s perception of security can be changed based
on targeted education and awareness of ways to address cybersecurity threats.
Studies to evaluate user perception can include the following:
(a) Evaluating behavioral factors affecting end users’ ability to detect phish-
ing, such as through survey-based methods to study users’ psychological

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 140 Human-Centered Data Analytics for Cybersecurity

factors related to phishing detection (Jakobsson 2007, Wright and Merett

2010, Fischer et al. 2013, Pfeiffer et al. 2013).
(b) Evaluating the effectiveness of a simply designed educational intervention
and awareness program. The intervention can involve presentation of
vignettes, carefully designed to support a diverse user group (those with
or without technical background). Zimmermann and Renaud (2021) evalu-
ate the impact of user nudges on decision making, particularly when
making choices of picking a secure Wi-Fi channel.
(c) Eye-tracking to determine areas of an interface where attention is focused
when performing tasks where threats may be evident (e.g., when interact-
ing online, the user may focus on the security icon on the toolbar). There
has been some research in evaluating toolbars in preventing phishing (Wu
et al. 2006) and using eye-tracking data to evaluate browser security
indicators (Darwish 2012).
(d) Data mining to evaluate patterns observed in the user’s eye gaze data and
how those patterns affect phishing detection or interaction with threat-
prone areas on the user interface. This can include evaluating the focus of
the user and associate different foci for different types of users (based on
psychological profiles, background, and demographics).

10.2 Human Perspective to Cybersecurity: Insider Threat

Insider threats consist of an authorized user of a system who perpetrates an
attack compromising the system’s confidentiality, integrity, and availability,
which may include illegal copying of data, sending data to unauthorized users,
or providing access to unauthorized users. In such scenarios, there is a need for
a deeper understanding of cybersecurity from multiple perspectives. Insider
threat detection and masquerade detection (Salem et al. 2008, Salem and Stolfo
2011, Garg et al. 2013) can be categorized based on the approaches used
namely: system call analysis, graph-based analysis, network mapping/top-
ology analysis, structural anomaly detection, and rule-based analysis.
Brdiczka et al. (2012) propose a hybrid system that combines structural
anomaly detection from social and information networks (graph analysis,
dynamic tracking, and machine learning) with psychological profiling
(dynamic psychological profiling from behavioral patterns). Eberle et al.
(2010) describe a graph-based anomaly detection approach for insider threat
detection. The approach is quite robust with few false positives, but falls short
on dynamic graphs and while handling numeric values in structural analysis.
Across the multiple areas of study discussed here, there is no work that brings

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 10.2 Human Perspective to Cybersecurity 141

together multiple facets from data and behavioral perspectives for discovering
insider threats. Insider threats remain one of the hardest types of threats to detect.
Studying how network traffic changes over time, which locations are the
sources, where is it headed, how are people generating this traffic, and how
people respond physiologically (such as through stress indicators) when
involved in these events – all these aspects become critical in distinguishing
the normal from the abnormal in the domain of cybersecurity. This requires
shifting gears to view cybersecurity as a people problem rather than a purely
technological problem.
As discussed in Chapter 2, several features can be utilized from disparate
domains, as shown in Figure 10.2, such as computer usage including CPU,
memory, and kernel modules, network traffic features including source, des-
tination IP, port, protocol, derived geolocation, and other location-related
features, such as geopolitical event information and physiological sensors
providing knowledge of affective behaviors, including features such as emo-
tion and stress variations. Each of these domains provides insights into the
workings of a networked information system over a period of time. Each
domain individually is not sufficient to indicate an insider threat. When
combined, these disparate data streams can facilitate detection of potentially
anomalous user traffic for deeper inspection. These features can be evaluated
individually and in conjunction to provide knowledge of potential insider
threats.

Figure 10.2 Integrating multiple domain datasets for insider threat detection.

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 142 Human-Centered Data Analytics for Cybersecurity

Figure 10.2 indicates relating multiple factors for insider threat evaluation:
• A user’s systems usage changes over time; similarly, network traffic evolves
over time, and communication patterns change over time. These key
changes, which are deviant from the normal changes, can be associated with
anomalies in the network traffic and system usage.
• Any type of attack has common underpinnings of how it is carried out; this
has not changed from physical security breaches to computer security
breaches. Thus, data representing the user’s behavior from both the usage
of the systems and affective behaviors (such as stress indicators) provide
important knowledge. This knowledge can be leveraged to identify behav-
ioral models of anomalies where patterns of misuse can be identified.
• Studying multiple types of datasets and monitoring users through the appli-
cation of affective computing techniques (Picard 2003) can have ethical and
privacy implications. Effective adversarial monitoring techniques need to be
developed such that they are ethical and respect user privacy.
Utilizing data-based and human behavioral aspects of learning, new know-
ledge from the vast array of processes can lead to new insights of understand-
ing the challenges faced in this important domain of cybersecurity.

10.3 User/Employee Viewpoint

A cyberattack victim may be a technology-savvy user or a user who is not as
well versed with technology use. While one may assume that those users less
familiar with technology would be more prone to cyberattacks, technology-
savvy users are equally prone to attacks.
Goldberg (1990) and Roddick and Spiliopoulou (1999) outline the big five
factors that describe an individual’s personality traits, as shown in Figure 10.3.
Shropshire et al. (2006) and McBride et al. (2012) propose to create a link
between the personality traits and users who are likely to commit security
infractions. These could be unknowingly or through malicious intent.
Shropshire et al. (2006) propose the hypothesis that conscientiousness and
agreeableness are positively related to IT security compliant behavior.
McBride et al. (2012) establishes that individuals react differently to different
scenarios, and therefore the cybersecurity training approach should be adopted
to differentiate between personality types. McBride et al. (2012) also provide a
fine-grained evaluation within the personality types. For example, agreeable
individuals with a low sense of sanction (fear of receiving a reprimand) are
more likely to violate security policies.

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 10.3 User/Employee Viewpoint 143

Figure 10.3 Big five personality factors.

As such, there is a very limited number of studies that establish the relation-
ships between personality traits and IT security incidents where an insider or
an unknowing user was responsible for the perpetration of the attack.
Certain key aspects of human behavior and personality traits can manifest in
technological viewpoints, as follows:
• User interaction: How a user interacts with a system is telling of their
personalities and their preferences. One common example is setting pass-
words and saving passwords, which is considered to be one of the biggest
points of risk for systems (Yildirim and Mackie 2019). Setting the same
passwords for multiple applications is a big security threat and in some
regards a single point of failure. Saving passwords on the system, or on a
piece of paper, while convenient, may be a security threat, especially if the
device is located in a somewhat nonsecure location. Similarly, use of
mechanisms such as two-factor authentication have to be carefully evaluated
for the types of proposed users of the systems (Das et al. 2020).
• Interface design: A very busy interface can leave average nontechnical
users confused. This is true for new users who may be starting to use an
interface in a new job or for experienced users after a change from the
systems they are used to. In some cases, it is also possible that highly
technical systems with several parameters may leave users vulnerable to

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 144 Human-Centered Data Analytics for Cybersecurity

mistakes. Simple-to-use interfaces can be helpful in alleviating this. Studies

(Abdellahi et al. 2020) have also presented user interface security assessment
models that can help evaluate user interactions from a security perspective.
• Where does the user look: Studies have been performed (Katsini et al.
2020) where user views are tracked to identify the level of interest of a user
while interacting with an interface such as a browser. There is indeed a trade-
off between accuracy of eye gaze–based solutions for security. Moreover,
there may be privacy concerns to use such methods.
• Education: A basic cyber education can be critical in averting the more
common types of situations such as phishing attacks. Similarly, educating
users to keep their systems up to date is also very important to avert known
attacks. Nudging users in the right direction and giving indicators in decision
making (Zimmerman and Renaud 2021) are indeed helpful ways to increase
user awareness.
While these aspects appear to be straightforward, they need to be enumerated
and carefully evaluated in an organization’s context to avoid cyberattacks that
can be averted with better design principles from a user interaction perspective.

10.4 Attacker Viewpoint: Anomaly Detection Methods

A data-oriented approach can be developed to understand how an attack is
carried out or what the behavioral aspect of an attack is. In cyberattacks, the
intruder will first choose a target and then initiate attacks. Before the attack
succeeds or fails, the intruder will have to try multiple times, and they will
continue trying different methods until they are successful or eventually give
up. This is the nature of the cyberattacks and almost universal in all attacks.
This may vary for attacks by insiders where they may know the system well
and do not need to scan the system and try multiple methods.
Another behavioral factor in cyberattacks is that the attackers will target
relatively few targets each time, and most times only attack one target at a
time. This provides a clue that when the attack happens, usually they will focus
on one target in a short period of time. The attackers assumed to be human will
usually not spend a prolonged time on one target. After they have tried
extensively on one target in a short period of time, they will switch targets.
These behavioral factors provide hints in developing behavioral models of
anomalies. The intruders will essentially leave a trace of activities in network
monitors such as in the intrusion detection system (IDS) logs. IDS logs (for
example, snort alert logs) can be utilized to generate behavioral models
of anomalies.

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 10.4 Attacker Viewpoint: Anomaly Detection Methods 145

Figure 10.4 Models describing attacker behaviors.

Let us consider all the visitors V of a network N, for each visitor of the
network, V in N, a source IP address, IPsource, and a target IP address, IPtarget,
are displayed in the IDS log. The alert information indicates the actions from
the source IP addresses. The IDS logs do not always categorize the attacks as
such since they observe it in isolation. The alerts in combination may be
potentially related to identify the collective behavior of attackers. Let us
consider three user patterns and build them into three models to exhibit
attacker behavior (Chen and Janeja 2014), as shown in Figure 10.4.
• Model 1: When a cyberattack occurs, the attacker usually will not be
successful the first time. The attacker will attempt different methods in order
to gain access to the target. As discussed in Liu et al. (2008), when an attack
happens, it takes multiple steps: probe, scan, intrusion, and goal. Let us say
each attacker is represented by a unique IP source, and each attempt is
differentiated by the alert messages. If the time when the attack will happen
is not known, then data can be divided into temporal neighborhoods, as
discussed in Chapter 3. This will reduce the number of instances analyzed
each time. It also helps to show the attacks in a smaller subset of the data. If
the count of number of unique attempts between IP1 and IP2 in a given
Interval(i) is greater than a threshold, then this action can be considered
anomalous and this source IP address is a potential attacker. Here
C represents empirical criteria to differentiate attacks and nonattacks. This
is based on the context of a network. If it is at a commercial network, C will
be larger than in a private network. Heuristically, we can consider C as three
times the average attempts of regular users. If an IPsource is identified as an
attacker IP address, its activities before or after the actions will be considered
part of the attack because its other incidents are likely to be in the probe or
scanning stage and in preparation for the following attacks.
• Model 2: When an attack happens, the target address usually is unique or
relatively a few addresses. The attackers will target the unique targets
persistently until success or failure. If a target IP address is accessed by a

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 146 Human-Centered Data Analytics for Cybersecurity

much higher number of unique IP addresses than usual within a short period
of time, this target IP address is potentially being accessed anomalously.
• Model 3: A common attack method is when there are a massive number of
attempts in a short time in order, for example, to obtain the password
information. Hence, if an IPtarget is experiencing much higher than normal
traffic from a single or a few IPsource, this IPtarget is under attack. This
represents the percentage of total traffic between two IP addresses in a given
interval (i). If this exceeds a threshold C, an alert can be raised. C can be an
empirical criteria to differentiate attacks and nonattacks, and this can be
based on experimental assessments and evaluating historic traffic patterns.
All these models are based on intuitions on potential attacker behavior,
represented by data on a network. Other models can be developed based on
psychological behavioral models such as those discussed in Chapter 5 for
social engineering threats.

https://doi.org/10.1017/9781108231954.011 Published online by Cambridge University Press

 11
Future Directions in Data Analytics
for Cybersecurity

This chapter discusses some key future directions and new and emerging
threats in a complex landscape of machine learning and AI models. This
chapter outlines cyberphysical systems, with a subcategory of the Internet of
Things (IoT), multidomain mining, and how outputs from disparate systems
can be fused for an integrative anomaly detection. This chapter also outlines
deep learning and generative adversarial networks and discusses emerging
challenges in model reuse. These areas are selected due to the emerging threats
and technology landscapes. At the end of this chapter, we also present an
important topic of ethical thinking in the data analytics process.

11.1 Data Analytics in Cyberphysical Systems

11.1.1 Cyberphysical Systems
Cyberphysical systems are formed as a result of an integration and interaction
of the cyber and physical systems. This introduces many interesting issues in
managing, monitoring, and analyzing these interactions and interfaces. One
major category of CPS is the industrial control systems (ICS).
Traditionally, ICS is comprised of supervisory control and data acquisition
(SCADA) systems, distributed control systems (DCS), and programmable
logic controllers (PLC) found in the industrial sectors and critical infrastruc-
tures. These are used for industries such as electrical, water, oil, gas,
and energy.
Traditionally, ICS has been considered a much more physical process-
oriented space as compared to traditional information technology systems
(see Figure 11.1). As a result, a cyberattack on an ICS will have much bigger
impact on the health and safety of people and impact the economy. These

147

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 148 Future Directions in Cybersecurity

Figure 11.1 ICS vs. IT systems.

systems will also have a higher emphasis on reliability and accurate perform-
ance. System integrity and availability have much more emphasis, as these
systems cannot be offline for long periods. In addition, there are redundancy
requirements in events of failure. All these are also brought into focus with the
interactions with the physical domains (Stouffer et al. 2009).
One of the key vulnerability identified in the Guide to Industrial Control
Systems (ICS) Security (Stouffer et al. 2009) is a lack of redundancy in critical
components that could lead to a single point of failure. The signals sent
between the various sensors and the control units can be mapped similar to a
network communication flow between IP addresses. Let’s consider the scen-
ario in Figure 11.2, where the various signals or signal pathways are shown.
The ICS systems may be comprised of sensors measuring a signal and
actuators or movers acting on the signal. Anomaly detection is an important
security application to evaluate risks in ICS, such as water monitoring sensor
networks (Lin et al. 2018).
Sensor networking (such as in spatial data, discussed in Chapter 8) and its
applications can be adapted to industrial control systems (ICS) and computer
networks for supporting cybersecurity. Let us consider a part of the sensor
network where the nodes A, B, C, and D (comprising of sensors and actuators,
with the lines indicating the potential connections and communication between
them) are shown in Figure 11.2a, b. We evaluate the nodes based on connec-
tions or number of links in Figure 11.2a. Using concepts of graph-based
metrics, as discussed in Chapter 9, we can evaluate these communication
patterns. Here nodes A and C can be considered important as they are
connecting hubs, where a large number of edges are incident, indicating a
high degree of communication. Similarly, node B can be considered important

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.1 Data Analytics in Cyberphysical Systems 149

Figure 11.2 Temporal evaluation of a sensor network.

because it is a connector between the two hubs. However, in Figure 11.2b we

see a different scenario, where it is difficult to determine the clear importance
based on links. Thus, simply using metrics such as centrality, focusing on the
number of edges incident on a node, may not be the best approach or even
feasible. Evaluating the critical nodes based on their behavior over time is much
more useful here. This behavior can be captured in terms of relationships based
on edges between nodes, which could represent data transfers between nodes,
such as in a computer network or an ICS network. For example, in Figure 11.2c,
the data transfers at time t1, t2, t3 for the network in Figure 11.2b are shown.
These can now be used for link-based or node-based patterns to produce a
ranking of links and nodes by importance, in terms of how many times they
appear in the temporal windows, as shown in Figure 11.2d. These can be mined
in an association rule–based method. Certain time periods may be more import-
ant, and thus may be given more weight, which may be mined using quantitative
association rule–based methods. This brings together concepts of temporal
evolution of a network, as discussed in Chapter 9.

11.1.2 Internet-of-Things (IoT)

IoT is interchangeably used and considered complementary to CPS
(Simmon et al. 2015, Blasch et al. 2017). Other studies have distinguished

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 150 Future Directions in Cybersecurity

IoT as a class of CPS (Nunes et al. 2015) or seen it as intersecting with CPS
(Calvaresi et al. 2017).
IoT has indeed become a big share of the CPS space due to the explosion of
the number of devices and advancements of smart devices and sensors. Even
though estimates vary, they are projected to grow upto 31 billion devices
worldwide by 2025 (Statista 2021). This is a very large number of devices
by any estimate and creates more security challenges given the highly unregu-
lated and nonstandard market for IoT devices. One such space where smart
devices have created this interesting intersection between cyberphysical and
Internet of Things is a smart car. We are making our vehicles smart: fully
connected with the internet to view real-time traffic and weather, talk on the
phone with Bluetooth, listen to the radio, watch video, as well as get real-time
status of automobile’s mechanical functions. However, this smart interface
comes at a price, which is the vulnerability to threats as well as malfunctions to
mechanical parts of the vehicle.
Other areas where IoT has taken a big role is in home security systems with
fire alarms, CO monitors, door and garage sensors, alarms, and temperature
control sensors. Figure 11.3 depicts the complex landscape of IoT.
Let us consider a smart home with a series of devices measuring various
phenomena around them. In such a setting, depicted in Figure 11.3, several
smart devices collect the information from a location with different levels of

Figure 11.3 IoT examples: car and home sensors.

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.2 Multidomain Mining 151

precision, collecting different data streams, perhaps using different standards.

We may want to evaluate behavioral aspects, such as, how are we using our
devices? Are there behavioral trends? We may also want to evaluate aspects of
something anomalous in this complex space, such as, are there deviations from
trends? How do we discover threats and attacks?
The data collected in the IoT space are often of different modalities,
including spatial, temporal, image, and text data. The data can also sliced in
a multidimensional view over time. The attack space is simply unmanageable.
Let us consider an example of a case study evaluating an incident at a
university that had close to 5,000 systems making DNS lookups every 15 min-
utes (Verizon 2017). In our current connected environments, edge devices,
vending machines, environmental systems, alarm systems, and every other
connected device on a university campus can lead to a massive attack space.
This truly becomes a needle in the haystack problem.
Unlike companies and businesses where the responsibility of preventing and
mitigating an attack is on the system administrators, where does this responsi-
bility lie in an IoT scenario? It lies both with the device maker and the user.
This still creates difficult scenarios where IoT use is occurring in a public or
shared space. This landscape is one of the frontiers of cybersecurity and
development of novel data analytics solutions for such a space.

11.2 Multidomain Mining

Data in real world are generated by multiple sources and are often
heterogeneous in terms of the types of attributes in each dataset. To be
preemptive and provide actionable insights, data from multiple sources need
to be analyzed to. This type of mining is referred to as multidomain mining,
where “domain” refers to distinct sources of data, and these distinct sources
may be completely disparately generated. Here a couple of examples are
discussed to highlight the challenges and potential solutions to analyzing
disparate data sources to provide actionable knowledge for events.

11.2.1 Integrating Multiple Heterogeneous Data

In a computer network, there are various mechanisms to allow for analyzing
the network traffic data. However, there may be scenarios where we want to
expand the decision criteria, especially when we may not have access to any
traffic data, such as payload, but only header information in the network traffic.
If we want to detect IP addresses with malicious intent, it becomes harder, as

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 152 Future Directions in Cybersecurity

the data would be limited. In such a scenarios, we can augment the header
information with other types of data. One such view point is that of geospatial
data, which can enhance the knowledge of the IP session or even the IP
reputation score itself.
Current reputation systems pursue classification into a white and black list,
i.e., binary categorization. Separate lists for URLs and IP addresses are
maintained. Some tools that provide rudimentary reputation services include
Cisco SenderBase (www.senderbase.org/), VirusTotal IP reputation (www
.virustotal.com/), and Spam and Open Relay Blocking System (SORBS)
(www.sorbs.net/). Most of these tools and lists are based on smaller feature
sets from one source, with no correlation across other data sources. Such a
shortcoming degrades a system’s effectiveness for detecting sophisticated
attacks and terminating malicious activities, such as zero-day attacks.
The set of attributes that the reputation scoring considers can be enriched,
providing an expressive scoring system that enables an administrator to under-
stand what is at stake, and increasing robustness by correlating the various
pieces of information while factoring in the trustworthiness of their sources. IP
reputation scoring model can be enriched using network session features and
geocontextual features such that the incoming session IP is labeled based on
the most similar IP addresses, both in terms of network features and geocon-
textual features (Sainani et al. 2020). This can provide better threat assessment
by considering not only the network features but also additional context, such
as the geospatial context information collected from external sources.
Indeed, in some countries, networks may encounter or even host large
quantities of attacks as compared to others. This may be due to shortage of
cybersecurity expertise, level of development, the abundance of resources,
corruption levels, or the computing culture in these countries (Mezzour
2015). Identifying these factors and quantifying them can provide insights into
security policies and have a positive impact on the attack incidents. These
scenarios not only impact the countries facing such cybersecurity crises but
also impact other countries and end users due to the level of connectivity in
today’s day and age. In addition, studies have also identified regions across the
world that are prone to hosting certain types of attacks. For example, studies
have indicated that Trojans, worms, and viruses are most prevalent in Sub-
Saharan Africa (Mezzour 2015), some families of malware preferentially target
Europe and US (Caballero et al. 2011). Yet other studies (Wang and Kim
2009) have explained broad categories of worldwide systemic risks and
country-specific risks where country-specific risks include aspects of econ-
omy, technology, industry, and international cooperation in enforcement of
laws and policies.

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.2 Multidomain Mining 153

Figure 11.4 Sources of geospatial data relevant to cybersecurity events with

geopolitical aspects.

Thus, geospatial data not only provide additional context but offer a frame-
work to accommodate additional geopolitical information that often plays a big
role in hactivism or politically inspired attacks. Figure 11.4 provides a set of
rich sources to access geospatial data for countries and in some cases even at a
granular level of cities. Some of these sources, such as Ip2location, provide a
way to identify a user location based on IP address in a nonintrusive manner.
Several other data sources such as World Bank Open Data (https://data
.worldbank.org/), PIOR-GRID (http://grid.prio.org/#/), and the Armed Conflict
Location and Event Data Project (ACLED) (www.acleddata.com/) data provide
sociopolitical and geopolitical conflict data. Such data can be used to create
geospatial characterization of regions (for example, using methods proposed by
Janeja et al. 2010, as discussed in Chapter 8). When an IP address is encoun-
tered, it can be geolocated using the IP location databases such as Ip2location or
Maxmind. As shown in Figure 11.5, based on its geolocation the location score
from the characterization can be attributed to it. In addition the geospatial
attributes for this region can be appended to the network attributes for this IP
(Sainani 2018, 2020). Any additional security intelligence can be appended to
provide an aggregate reputation score to this IP.
The data heterogeneity in terms of types of attributes, namely categorical
versus continuous, can be addressed using methods that are capable of hand-
ling mixed attribute datasets (such as Misal et al. 2016).

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 154 Future Directions in Cybersecurity

(Offline) Precomputed geo reputation

Evaluate
sources Geospatial Data preprocessing
Geospatial
sources features

Geospatial
Geospatial
Feature selection characterization reputation

(Online) online geo reputation attribution

IP address IP geolocation
(MaxMind etc.) Geospatial score

Aggregate reputation
Security intelligence Multi-attribute
score
reputation scoring

Figure 11.5 Reputation scoring with geospatial context.

11.2.2 Integrated Alerts from Multiple Sources

Computer networks are increasingly facing the threat of unauthorized access.
In addition to computer networks, other networks, such as sensor networks and
industrial control systems, also face similar threats. Intrusion detection aims at
identifying such threats using signatures of unauthorized access or attacks.
There are very few systems that address the issue of zero-day attacks, where
the attack signature is not known before-hand.
Let us consider a scenario, as outlined in Chapter 1, where the threat is two
pronged: first, there is an attack on the organization; and second, there is an
attack on a partner that shares key resources. In the first part of the attack,
intruders take advantage of vulnerabilities in public-facing web servers. In
addition, hackers secretively scout the network from compromised worksta-
tions that have already been targeted beforehand as part of a coordinated
prolonged attack. The second part of the attack starts with spear-phishing.
Instead of casting out thousands of emails randomly, spear phishers target
select groups of people with something in common, such as a common
employer, a similar banking or financial institution, the same college, etc.
The emails are deceptive since they appear to be from organizations from
which victims are expecting emails. So potentially a second group of hackers
institutes a spear-phishing attack on the organization’s major business partners,
with which it shares network resources. The hackers are able to obtain a

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.2 Multidomain Mining 155

Soware Soware
Agency A module 1 Agency B module 2
Level1

Alarms Alarms Alarms Alarms

Alarms
originating from
Alarms
Level 2

agency B
originating for
agency A
Iterative
alarm fusion
Alarms
originating
from agency
C

Alert Alert

Figure 11.6 Alarm data fusion.

privileged account and compromise a root domain controller that is shared by

the organization and its partner. When the intruders try to recreate and assign
privileges, they trigger an alarm.
In this scenario where the attack is distributed, it may be difficult to integrate
these alarms to identify if this is a single malicious attack or if the attack was
spread out over time. This requires the discovery of an integrated alert from
alarms generated from disparate sources across a prolonged period of time, as
shown in Figure 11.6. Each alarm generated has a set of attributes such as the
source and destination of the data, port, length of data packet, etc. Each source
generating the alarm is also described through a set of attributes, such as type,
port frequently used, etc.
The series of alarms received in such a complex application domain is
composed of the raw messages received from multiple sources, and an alert
is the processed and aggregated output generated from these alarms. Such an
approach for the alarm data fusion would include alarm source characterization
and alarm clustering using source characterization and temporal lag determin-
ation. The alarms could be generated from a single source as a series of bursts,
or from multiple sources such as multiple agencies or multiple software
modules, as shown in Figure 11.6. Another scenario (Janeja et al. 2014) is
when the alarms are being generated from multiple IDS systems.

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 156 Future Directions in Cybersecurity

We can perform clustering of the IDS log data after preprocessing. The data
are in the form of parsed IDS alarms, where each alarm is a data point with a
priority level: high, medium, or low. These data are collected from multiple
IDS sources. In this case, the premise is that a low-priority alarm may not
really be a low-priority alarm when seen in conjunction with some of the other
alarms from different IDS systems. Since we are looking at all alarms from
multiple IDS systems, we have the opportunity to study the similarity between
alarms and then judge whether an alarm is truly a low priority or could be
potentially high priority. The end goal is that multiple similar low-priority
alarms could potentially indicate cyberattacks.
To distinguish whether these alarms are true alerts, we would need to
examine if the sources have similar characteristics. A source characterization
based on the features associated with sources can be performed. Here “source”
refers to either IDS systems associated with various parts of an organization or
even multiple organizations; similarly, “source” could also refer to an applica-
tion source or even an agency in a multiagency scenario. Sources could also
refer to sensors in an ICS example, as discussed earlier in this chapter. This
approach is adaptable if the alarm can be sufficiently preprocessed in a uniform
manner across different types of alarm data.
Clusters of alarms based on attributes can be generated. For each cluster, the
set of sources associated with each cluster and the alarm cardinality of each
cluster, which is the number of alarms in a cluster, are identified:
• If only one source is generating alarms in a cluster, then the source is flagged
for investigation. If alarm cardinality of a cluster is greater than a preset
threshold and if all of the alarms in the cluster are from one source, then this
can lead to raising an alert. Let us consider a sensor generating alarms in an
ICS. If a single sensor is generating alarms, then this sensor should be
investigated. Moreover, if data from the sensors are clustered and readings
from one sensor dominates the cluster, this could further raise the alert.
• Next, for every source in a particular characterization (based on similarly
behaving sources) such that this source is not equal to the source in the
cluster, if the source s is in not in any other cluster it implies that potentially
the other sources in the characterization as well are not generating alarms.
This cluster can be flagged as a possible false positive for further investi-
gation. This really means that if a source is part of a cluster but the other
sources from this characterization are not in any of the clusters, then this
source may be a false positive.
After removing the clusters representing the false positives, among the
remaining clusters of alarms, two cases may occur: (a) a cluster comprises a

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.3 Advanced Machine Learning Models 157

significant number of alarms, but these alarms do not belong to one single
source in a characterization; and (b) no single cluster has a significant number
of alarms:
• In the case of the first scenario, an aggregated alert can be generated if there
is any cluster that has alarms greater than a preset threshold.
• In the case of the second scenario, we can identify the overlap between the
clusters, using measures such as silhouette coefficient and entropy, to find
mutually related alarms to generate an aggregated alert. The greater the
overlap, the more strongly related are the alarms.
Once an alert is generated, sources can be associated with it. The set of
sources of the alarms is the sources of the alarms in the cluster, or in the other
scenario, the sources of the alarms in the overlap. These sources can be
aggregated by grouping the feature vectors of all the sources, thus generating
a composite feature vector of the source of the alert.

11.3 Advanced Machine Learning Models

11.3.1 Deep Learning
The output of a machine learning algorithm is only as good as the data that are
fed to train the model. Machine learning models such as classification learn to
predict the outcome of a task based on a set of features that are provided to the
training algorithm. Essentially, the model is trained based on the set of
features. For example, based on the packet features, can a packet be labeled
as anomalous or benign? Here feature selection is a major task. The features
provide a representation of the data. If the features or representations of the
data are not correctly selected, this can lead to erroneous results.
Deep learning is a type of machine learning that learns the features through
multiple layers of abstractions. For example, if the task is learning to recognize a
picture of an individual, the deep learning model may start with various levels of
abstractions, starting with the most basic pixels in an image, to an abstraction of
an outline of the nose to the outline of the facial structure. Deep learning
algorithms compute the representations of one layer by tuning the parameters
from the previous layers (Le Cun et al. 2015). As the amount of data increases,
performance of most machine learning algorithms plateaus. However, perform-
ance of deep learning algorithms increases as the amount of input data increases.
An example deep learning model is shown in Figure 11.7, where the input is
translated into several features or representations in layer 1. Some of these

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 158 Future Directions in Cybersecurity

Figure 11.7 Example of a deep learning model.

representations can be dropped in subsequent layers, and throughout the layers

the representations are weighted based on the reduction in a loss function until
the model converges to the output, which is the prediction task. Deep learning
has found a major application in computer vision, where images can be labeled
based on their most basic of features and abstracting to the higher-level
composition of the images. Deep learning has also found applications in
anomaly detection (Naseer et al. 2018).
Deep learning emulates how human brain learns through connections of
neurons. The most fundamental level of learning comes from neural networks
that were in vogue in the early 1960s and have now had a renewed interest due
to deep learning algorithms. The difference is now we have the availability of
massive amounts of data and computing capacity, which has resulted in
stronger models and learning algorithms. In cybersecurity applications, higher
accuracy and lower false positive rates are very important to label truly
anomalous events. Deep learning methods such as deep belief networks,
recurrent neural networks, and convolutional neural networks have the cap-
acity to learn from unlabeled data and produce better outcomes (Xin
et al. 2018).
However, deep learning suffers from several challenges, as summarized in
Figure 11.8. Deep learning models have several hyperparameters that need to

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.3 Advanced Machine Learning Models 159

Figure 11.8 Deep learning challenges.

be predetermined and tuned, including number of layers, number of nodes in

each layer, network weights, and dropouts in each layer. Some of these factors
are also interdependent and can also impact the learning rate of the model.
Moreover, if the input data are not large, the model cannot be trained well.
Thus, the true strength of deep learning is possible in massive datasets and
requires heavy parameter tuning. The challenge also comes in with
explainability of how these parameters are impacting the outputs and the
interpretation of the final outcomes. However, this area is constantly evolving,
and methods such as deep reinforcement learning (Nguyen and Reddi 2019)
are finding use in cybersecurity applications to provide adaptive, responsive,
and scalable methods of threat detection.

11.3.2 Generative Adversarial Networks

Increasingly, cybersecurity challenges have started emerging with image and
data manipulation. As machine learning models are increasingly being used for
cybersecurity applications, so are ways to manipulate these systems from an

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 160 Future Directions in Cybersecurity

adversarial standpoint. Attacks may increase the chances of misclassification

of outcomes from a predictive model; for example, poison attacks may add bad
data into the training samples, and evasion attacks may distort input data so
that they are not correctly classified (Yinka-Banjo and Ugot 2019). Attacks
may also emerge due to model reuse (as discussed in the next section). It is
imperative that such types of generative samples are part of the learning
strategy for accurate detection of truly anomalous samples.
Generative models have been explored with pioneering work in 1996
(Revow et al. 1996), where an expectation maximization algorithm was used
to identify the model that generated an image of a handwritten digit. This
approach not only labeled the data but also explained the model generating it.
Recent work (Goodfellow et al. 2014) formalized the idea of a generative
adversarial network (GAN) as a model comprising two components: a genera-
tive model and a discriminative model. As shown in Figure 11.9, each of these
models has a different function. A discriminator takes input and labels the
input as belonging to a certain class. In this example, the input can be images
and the output is the label for the images as belonging to a certain class. On the
other hand, the generator can take the images (or distributions as a starting
point) and learn from the distribution to produce new outputs, in this case new
images. The new images may not be realistically comparable to the original but
emulate the distribution.

Figure 11.9 Generative vs. discriminative functions.

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.3 Advanced Machine Learning Models 161

Figure 11.10 Generative adversarial network.

In a GAN, the two components interact in a two-player game where each is

trying to win over the other maximizing the similarity to the original distribution.
Here the generative model emulates the probability distribution to match the
original data, and the discriminative model tries to distinguish whether the gener-
ated sample came from the original distribution or from the generative model.
As shown in Figure 11.10, we can see that the two components are pitted
against each other. Here an input (a combination of noise and random images)
is provided to the generator, which generates samples. On the other hand, the
discriminator, which is trained on the real-world images, examines these
generated samples. These samples are labeled as real or fake. In addition, the
discriminator also learns from the loss of labeling the samples incorrectly and
corrects the weights in the discriminator. This process can iteratively improve
the learning from a discriminator.
Based on game theoretic concepts, an “adversarial” concept is generally
used; however, this could also be a “cooperative” setting where the discrimin-
ator can share the information with the generator (Goodfellow 2016). Other
types of adversaries, such as adaptive adversaries, have also been discussed in
a request-answer game theoretic setting (Ben-David et al. 1994, Gormley et al.
2000). While such adaptations have not been shown in computational deep
learning models yet, these types of models can be used to make the machine
learning samples more robust against adversarial attacks.

11.3.3 Model Reuse

As machine learning and in general data analytics models become more easily
available, they are being pervasively used and reused. Many repositories have

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 162 Future Directions in Cybersecurity

Figure 11.11 Model reuse attacks.

emerged where models are checked in by researchers and others are reusing
these. Examples include the Open Neural Network Exchange (ONNX)
format1, Azure AI gallery2, and Caffe Model Zoo3. It is only a matter of time
before models are as easily available and searchable as public datasets.
However, there is a clear danger in reusing machine learning models without
understanding the provenance of the models and the model pipelines. This has
been a well-studied problem in software reuse (Paul and Taylor 2002, Kath
et al. 2009). Vulnerability databases have been established to study known
software defects to prevent propagating them to other users and applications. If
reuse of vulnerable software and code in general is not prevented, it can lead to
massive disruptions and attacks such as in the case of the heart bleed bug
(Carvalho et al. 2014).
This level of understanding has been well established in machine learning
models. Recently some studies have shown that manipulating even simple
building blocks of a machine learning model can lead to model reuse attacks
(Ji et al. 2018) where corrupted models can lead the host systems to behave
anomalously when certain types of input triggers are introduced into the
system. These threats in model reuse and training data manipulation (Ji et al.
2018) can be summarized as shown in Figure 11.11.
To understand the impacts of such an attack, consider an autonomous
vehicle that has to evaluate several images from multiple camera inputs in an
ensemble learning system. In such a complex system, if a trigger results in an
anomalous decision, this can have far-reaching impacts for the immediate

1 2
http://onnx.ai/ https://gallery.azure.ai//
3
http://caffe.berkeleyvision.org/model_zoo.html

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 11.4 Ethical Thinking in the Data Analytics Process 163

vicinity where the vehicle is driving and also in the long term for the company
that has developed the autonomous vehicle and the systems processing the
images. Indeed, studies have shown how researchers have hacked into driver-
less cars (Versprille 2015) and studied potential risks in smart car functional-
ities (Weimerskirch and Dominic 2018). Model reuse attacks can bring in a
new wave of cyberattacks where complex systems that rely on machine
learning behave erratically when certain trigger inputs are introduced.

11.4 Ethical Thinking in the Data Analytics Process

Data analytics, developed using data-driven technologies from artificial intelli-
gence, machine learning, and statistics, are extremely beneficial for cyberse-
curity if used well. These systems can lead to substantial benefits in terms of
the speed, cost, and scalability of decision making compared to human deci-
sion making, and they often lead to better and more reliable decisions, espe-
cially in cases where system administrators are flooded with data from massive
organizational computer networks. If implemented appropriately, these
systems also have the potential to be more robust in identifying accurate alerts.
Since the same procedure is applied across massive amounts of data, achieving
such robustness may simply not be possible for a network administrator.
However, because of the far-reaching scalability of the data analysis, data
and algorithms being implemented also need to be considered from
ethical viewpoints.
Ethics generally works on the principles of do no harm (Janeja 2019).
Although research protocols to protect human beings have been in place for
a while now, the pervasiveness of multiple types of data and their use make it
less clear where the impact on human beings is in the data analytics processes.
Thus, harm is not only direct based on exposing identifiable data for individ-
uals, but also indirect resulting from the reuse of easily available data and
combining multiple datasets, as we have discussed in various examples of data
manipulation throughout the book.
In particular for data analytics, there is a need to develop ethical critical
thinking while analyzing the data. Throughout the entire life cycle of the data
analytics in the knowledge discovery process, there are many opportunities for
ethical decision making that need to be evaluated to consider the human and
societal impacts.
Here the harms are not only in terms of privacy, for example through
identification of personally identifiable information (PII), but also other types
of data and algorithms that may not necessarily fall under the direct purview of

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 164 Future Directions in Cybersecurity

traditional privacy protections. There may be yet other types of data that may
have impacts for the society as a whole. Consider data such as location data
collected on a phone, driving data collected through on board diagnostics
(OBD), or atmospheric data collected through sensors in a community. Now
let us consider the types of patterns one may discover if they have access to
such data. For example, the GPS data on the phone may pinpoint the geospatial
coordinates of secure locations where an individual is carrying the phone;
OBD driving data may provide insights into driving behavior; systems may
falsely label anomalous behavior. Data collection is another very important
element where ethical considerations (in addition to legal considerations)
are important.
As we think through the examples, we can also think of the uses of the data
while integrating them with data from other sources. For example, if the data of
IP addresses are combined with geographical distribution of demographics,
does that identify vulnerable populations, especially if the system being used is
not robust to false positives? Can the driver behavior, discovered through OBD
or the information on the neighborhoods that the driver passes through, lead to
additional security risks? Even in these limited examples we can see begin-
nings of questions of bias and risk to individuals as more and more systems
relying on data analytics are being used. Indeed, such systems are beginning to
be used in augmented decision making that can impact lives and livelihoods.
We all have come to accept recommendations made by data-driven analytics
algorithms in our day-to-day lives, such as in spam labels for emails and fraud
alerts from our credit card companies. Things start getting tricky if these data
analytics systems are being used to make decisions that would generally take a lot
of thought and deliberation – for example, who should receive bail or not, who
could be a criminal or not, or if a facial recognition system should be used or not.
The vast complexity of data, the data’s easy availability, and algorithmic
tools requires vigilance around access, privacy, provenance, curation, inter-
operability, issues of fairness, accountability, and transparency. In studying
data analytics algorithms and systems such as for the security alert mechanisms
discussed in this book, it is clear that there are several checkpoints where a
system follows a pathway based on the choices that the creator of that system
made or the user of the system made. It is important to know what those
choices are and how they can impact lives and livelihoods. It is also important
to explain the outcomes from such systems and ensure they are not dispro-
portionately impacting certain groups or individuals. Just like we would
carefully hire an employee, supervise them, and give them advice in making
critical decisions, we need to have sensible extensions so that we do the same
for data analytics algorithms and systems.

https://doi.org/10.1017/9781108231954.012 Published online by Cambridge University Press

 References

Abad, C., Taylor, J., Sengul, C., et al. (2003). “Log correlation for intrusion detection: a
proof of concept.” 19th Annual Computer Security Applications Conference, 2003.
Proceedings., 2003, pp. 255–264, doi: 10.1109/CSAC.2003.1254330.
Abdellahi, S., Lipford, H. R., Gates, C., and Fellows, J. (2020). “Developing a User
Interface Security Assessment method.” USENIX Symposium on Usable Privacy
and Security (SOUPS) 2020. August 9–11, 2020, Boston.
Abedin, M., Nessa, S., Khan, L., Al-Shaer, E., and Awad, M. (2010). “Analysis of
firewall policy rules using traffic mining techniques.” International Journal of
Internet Protocol Technology, 5(1–2), 3–22.
Abraham, T. and Roddick, J. F. (1999). “Survey of spatio-temporal databases.”
GeoInformatica, 3(1), 61–99.
Aggarwal, C. C. (2017). “Spatial outlier detection.” Outlier Analysis. Springer
International Publishing, 345–368.
Aggarwal, C. C. and Yu, P. S. (2001, May). “Outlier detection for high dimensional
data.” ACM Sigmod Record, 30(2) 37–46.
Agrawal, R., Gehrke, J., Gunopulos, D., and Raghavan, P. (1998). “Automatic sub-
space clustering of high dimensional data for data mining applications.”
Proceedings of the 1998 ACM SIGMOD International Conference on
Management of Data. ACM Press, 94–105.
Agrawal, R. and Srikant, R. (1994). “Fast algorithms for mining association rules.”
Proceedings of the 20th International Conference on Very Large Data Bases,
VLDB, J. B. Bocca, M. Jarke, and C. Zaniolo (Eds.). Morgan Kaufmann, 487–499.
Agarwal, S., Farid, H., Gu, Y., et al. (2019, June). “Protecting world leaders against
deep fakes.” Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR) Workshops, 38–45.
Ahmed, M., Mahmood, A. N., and Hu, J. (2016). “A survey of network
anomaly detection techniques.” Journal of Network and Computer Applications,
60, 19–31.
Akamai. (2016). “Q4 2016 state of the internet/security report.” www.akamai.com/
newsroom/press-release/akamai-releases-fourth-quarter-2016-state-of-the-inter
net-connectivity-report. Last accessed November 2021.

165

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 166 References

Akamai. (2019). “Web attacks and gaming abuse, 2019.” www.akamai.com/us/en/

multimedia/documents/state-of-the-internet/soti-security-web-attacks-and-
gaming-abuse-report-2019.pdf. Last accessed November 2021.
Akoglu, L., Tong, H., and Koutra, D. (2015). “Graph based anomaly detection and
description: a survey.” Data Mining and Knowledge Discovery, 29(3), 626–688.
Al-Musawi, B., Branch, P., and Armitage, G. (2015, December). “Detecting BGP
instability using Recurrence Quantification Analysis (RQA).” Computing and
Communications Conference (IPCCC), 2015 IEEE 34th International
Performance. IEEE, 1–8.
Al-Musawi, B., Branch, P., and Armitage, G. (2016). “BGP anomaly detection tech-
niques: a survey.” IEEE Communications Surveys & Tutorials, 19(1) 377–396.
Al-Rousan, N. M. and Trajković, L. (2012, June). “Machine learning models for
classification of BGP anomalies.” 2012 IEEE 13th International Conference on
High Performance Switching and Routing (HPSR). IEEEV, 103–108.
Al Shalabi, L., Shaaban, Z., and Kasasbeh, B. (2006). “Data mining: a preprocessing
engine.” Journal of Computer Science, 2(9), 735–739.
Aleroud, A. and Karabatis, G. (2013). “A system for cyber attack detection using
contextual semantics.” 7th International Conference on Knowledge Management
in Organizations: Service and Cloud Computing, L. Uden, F. Herrera, J. B. Pérez,
and J. M. Corchado Rodríguez (Eds.). Springer, 431–442.
Aleroud, A. and Karabatis, G. (2014, September). “Detecting zero-day attacks using
contextual relations.” International Conference on Knowledge Management in
Organizations, L. Uden, D. Fuenzaliza Oshee, I. H. Ting, and D. Liberona
(Eds.). Springer, 373–385.
Alperovitch, D. (2011). “Revealed: Operation Shady RAT.” www.csri.info/wp-content/
uploads/2012/08/wp-operation-shady-rat1.pdf. Last accessed November 2021.
Alseadoon, I., Chan, T., Foo, E., and Gonzales, N. J. (2012). “Who is more susceptible
to phishing emails? A Saudi Arabian study.” Proceedings of the 23rd Australasian
Conference on Information Systems. ACIS, 1–11
Ameen, J. and Basha, R. “Mining time series for identifying unusual sub-sequences
with applications.” ICICIC’06: Proceedings of the First International Conference
on Innovative Computing, Information and Control. IEEE Computer Society,
574–577.
Aminikhanghahi, S. and Cook, D. J. (2017). A survey of methods for time series change
point detection. Knowledge and Information Systems, 51(2), 339–367.
Anderson, L. C. and Agarwal, R. (2010, September). “Practicing safe computing: a
multimethod empirical examination of home computer user security behavioral
intentions.” MIS Quarterly, 34(3), 613–643.
Ankerst, M., Breunig, M. M., Kriegel, H. P., and Sander, J. (1999, June). “OPTICS:
ordering points to identify the clustering structure.” Proceedings of the 1999 ACM
SIGMOD International Conference on Management of Data (SIGMOD '99).
ACM, 49–60. doi: 10.1145/304182.304187.
Azari, A., Janeja, V. P., and Levin, S. (2017). “MILES: multiclass imbalanced
learning in ensembles through selective sampling.” ACM Symposium on
Applied Computing, Data Mining. ACM, 811–816. doi: 10.1145/
3019612.3019667.
Azari, A., Namayanja, J. M., Kaur, N., Misal, V., and Shukla, S. (2020, May).
“Imbalanced learning in massive phishing datasets.” 2020 IEEE 6th

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 167

International Conference on Big Data Security on Cloud (BigDataSecurity), IEEE

International Conference on High Performance and Smart Computing (HPSC)
and IEEE International Conference on Intelligent Data and Security (IDS). IEEE,
127–132.
Banerjee, A., Venkatasubramanian, K. K., Mukherjee, T., and Gupta, S. K. S. (2012).
“Ensuring safety, security, and sustainability of mission-critical cyber–physical
systems.” Proceedings of the IEEE, 100(1), 283–299.
Barnes, R. (2013, August 8). “Geocoding router log data.” http://resources
.infosecinstitute.com/geocoding-router-log-data/#gref. Last accessed November
2021.
Barnett, V. and Lewis, R. (1994). Outliers in Statistical Data. John Wiley and Sons.
Basawa, I. V., Billard, L., and Srinivasan, R. (1984). “Large-sample tests of homogen-
eity for time series models.” Biometrika, 71(1), 203–206.
Bellman, R. (1961). Adaptive Control Processes: A Guided Tour. Princeton University
Press.
Ben-David, S., Borodin, A., Karp, R., Tardos, G., and Wigderson, A. (1994). “On
the power of randomization in on-line algorithms.” Algorithmica, 11(1),
2–14.
Ben Salem, M., Hershkop, S., and Stolfo, S. J. (2008). “A survey of insider attack
detection research.” Insider Attack and Cyber Security: Advances in Information
Security, vol. 39, S. J. Stolfo, S. M. Bellovin, A. D. Keromytis, S. Hershkop, S. W.
Smith, and S. Sinclair (Eds.). Springer, 1–19.
Ben Salem, M. and Stolfo, S. J. (2011). “Modeling user search behavior for masquerade
detection.” Insider Attack and Cyber Security: Advances in Information Security,
vol. 39, S. J. Stolfo, S. M. Bellovin, A. D. Keromytis, S. Hershkop, S. W. Smith,
and S. Sinclair (Eds.). Springer. doi: 10.1007/978-0-387-77322-3_5.
Besag, J. and Newell, J. (1991).“The detection of clusters in rare diseases.” Journal of
the Royal Statistical Society Series A, 154, 143–155.
Beyer, M. A. and Laney, D. (2012). The Importance of “Big Data”: A Definition.
Gartner.
Bhuyan, M. H., Bhattacharyya, D. K., and Kalita, J. K. (2014). “Network anomaly
detection: methods, systems and tools.” IEEE Communications Surveys and
Tutorials, 16(1), 303–336.
Binde, B. E., McRee, R., and O’Connor, T. (2011, May). “Assessing outbound traffic to
uncover advanced persistent threat.” doi: 10.13140/RG.2.2.16401.07520.
Blasch, E., Kadar, I., Grewe, L. L., et al. (2017, May). “Panel summary of
cyber-physical systems (CPS) and internet of things (IoT) opportunities with
information fusion.” Signal Processing, Sensor/Information Fusion, and Target
Recognition XXVI, vol. 10200, International Society for Optics and Photonics,
02000O.
Blasco, J. (2013, March 21). “New Sykipot developments.” https://cybersecurity.att
.com/blogs/labs-research/new-sykipot-developments. Last accessed November
2021.
Brdiczka, O., Liu, J., Price, B., et al. (2012). “Proactive insider threat detection through
graph learning and psychological context.” 2012 IEEE Symposium on Security and
Privacy Workshops (SPW). IEEE, 142–149.
Breiman, L. (1996). “Bagging predictors.” Machine Learning, 24(2), 123–140.
Breiman, L. (2001). “Random forests.” Machine Learning, 45(1), 5–32.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 168 References

Breunig, M. M., Kriegel, H. P., Ng, R. T., and Sander, J. (1999, September).
“Optics-of: identifying local outliers.” Principles of Data Mining and
Knowledge Discovery. PKDD 1999, J. M. Żytkow and J. Rauch (Eds.)
Lecture Notes in Computer Science, vol. 1704. Springer. doi: 10.1007/978-3-
540-48247-5_28.
Brigham, E. O. (2002). The Fast Fourier Transform. Prentice Hall.
Bright, A. (2007). “Estonia accuses Russia of ‘cyberattack’.” www.csmonitor.com/
2007/0517/p99s01-duts.html. Last accessed March 2020.
Bronskill, J. (2012, November 9). “Govt fears Canada becoming host country for cyber-
attacker.” https://winnipeg.ctvnews.ca/canada-becoming-host-country-for-cyber-
attackers-government-fears-.1032064. Last accessed November 2021.
Burns, C. (2012, June 1). “Stuxnet virus origin confirmed: USA and Isreali govern-
ments.” www.slashgear.com/stuxnet-virus-origin-confirmed-usa-and-isreali-gov
ernments-01231244/. Last accessed November 2021.
Caballero, J., Grier, C., Kreibich, C., and Paxson, V. (2011, August). “Measuring pay-
per-install: the commoditization of malware distribution.” Proceedings of the 20th
USENIX Conference on Security (SEC'11), 13.
Cai, L. and Hao, C. (2011). “TouchLogger: inferring keystrokes on touch screen from
smartphone motion.” HotSec, 11(2011), 9–9.
Caldwell, D., Gilbert, A., Gottlieb, J., et al. (2004). “The cutting EDGE of IP router
configuration.” ACM SIGCOMM Computer Communication Review, 34(1),
21–26.
Calvaresi, D., Marinoni, M., Sturm, A., Schumacher, M., and Buttazzo, G. (2017,
August). “The challenge of real-time multi-agent systems for enabling IoT and
CPS.” Proceedings of the International Conference on Web Intelligence. ACM,
356–364.
Cao, L., Yang, D., Wang, Q., Yu, Y., Wang, J., and Rundensteiner, E. A. (2014,
March). “Scalable distance-based outlier detection over high-volume data
streams.” 2014 IEEE 30th International Conference on Data Engineering
(ICDE) . IEEE, 76–87.
Carvalho, M., DeMott, J., Ford, R., and Wheeler, D. A. (2014). “Heartbleed 101.” IEEE
Security & Privacy, 12(4), 63–67.
Chandola, V., Banerjee, A., and Kumar, V. (2009). “Anomaly detection: a survey.”
ACM Computing Surveys (CSUR), 41(3), 15.
Chandrashekar, G. and Sahin, F. (2014). “A survey on feature selection methods.”
Computers & Electrical Engineering, 40(1), 16–28.
Chapman, P., Clinton, J., Kerber, R., et al. (2000). CRISP-DM 1.0, Step-by-Step Data
Mining Guide. CRISP-DM Consortium; SPSS: Chicago, IL, USA.
Check Point. (2020). “Threat map.” https://threatmap.checkpoint.com/. Last accessed
March 2020.
Chen, M., Mao, S., and Liu, Y. (2014). “Big data: a survey.” Mobile Networks and
Applications 19(2), 171–209.
Chen, S. and Janeja, V. P. (2014). “Human perspective to anomaly detection for
cybersecurity.” Journal of Intelligent Information Systems, 42(1), 133–153.
Cheok, R. (2014). “Wire shark: a guide to color my packets detecting network recon-
naissance to host exploitation.” GIAC Certification Paper. SANS Institute Reading
Room.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 169

Cheswick, B. (1992, January). “An evening with Berferd in which a cracker is lured,
endured, and studied.” Proceedings of the Winter USENIX Conference, San
Francisco, 20–24.
Chi, M. (2014). “Cyberspace: America’s new battleground.” www.sans.org/reading-
room/whitepapers/warfare/cyberspace-americas-battleground-35612. Last
accessed November 2021.
Chien, E. and O’Gorman, G. (2011). “The nitro attacks, stealing secrets from the
chemical industry.” Symantec Security Response (2011), 1–8.
CIA. (2021). “World Factbook.” Last accessed November 2021.
Cleary, J. G. and Trigg, L. E. (1995). “K*: an instance-based learner using an entropic
distance measure.” Machine Learning Proceedings 1995, A. Prieditis and S.
Russell (Eds.). Morgan Kaufmann, 108–114.
Cloud Security Alliance. (2014). Big data taxonomy. https://downloads
.cloudsecurityalliance.org/initiatives/bdwg/Big_Data_Taxonomy.pdf. Last accessed
April 13, 2017.
Cohen, E., Datar, M., Fujiwara, S., et al. (2001). “Finding interesting associations
without support pruning.” IEEE Transactions on Knowledge and Data
Engineering, 13(1), 64–78.
Cohen, W. W. (1995, July). “Fast effective rule induction.” Proceedings of the Twelfth
International Conference on Machine Learning. Morgan Kaufmann, 115–123.
Cooper, G. F. and Herskovits, E. (1992). “A Bayesian method for the induction of
probabilistic networks from data.” Machine Learning, 9, 309–347.
Cortes, C. and Vapnik, V. (1995). “Support-vector networks.” Machine Learning, 20
(3), 273–297.
Cover, T. and Hart, P. (1967). “Nearest neighbor pattern classification.” IEEE
Transactions on Information Theory, 13(1), 21–27.
Dark Reading. (2011). “PNNL attack: 7 lessons: surviving a zero-day attack.” www
.darkreading.com/attacks-and-breaches/7-lessons-surviving-a-zero-day-attack/d/d-
id/1100226. Last accessed March 2020.
Darwish, A. and Bataineh, E. (2012, December 18–20). “Eye tracking analysis of
browser security indicators.” 2012 International Conference on Computer
Systems and Industrial Informatics (ICCSII), 1, 6. doi: 10.1109/
ICCSII.2012.6454330. Last accessed November 2021.
Das, A., Ng, W.-K., and Woon, Y.-K. 2001. “Rapid association rule mining.”
Proceedings of the Tenth International Conference on Information and
Knowledge Management. ACM Press, 474–481.
Das, S., Kim, A., Jelen, B., Huber, L., and Camp, L. J. (2020). “Non-inclusive online
security: older adults’ experience with two-factor authentication.” Proceedings of
the 54th Hawaii International Conference on System Sciences, 6472.
Dash, M. and Liu, H. (1997). “Feature selection for classification.” Intelligent Data
Analysis, 1(1–4), 131–156.
Davies, C. “Flame cyber-espionage discovered in vast infection net.” (2012, May 28).
www.slashgear.com/flame-cyber-espionage-discovered-in-vast-infection-net-
28230470/. Last accessed November 2021.
Davinson, N. and Sillence, E. (2010). “It won’t happen to me: Promoting secure
behaviour among internet users.” Computers in Human Behavior, 26(6),
1739–1747.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 170 References

Davinson, N. and Sillence, E. (2014). “Using the health belief model to explore users’
perceptions of ‘being safe and secure’ in the world of technology mediated
financial transactions.” International Journal of Human–Computer Studies, 72
(2), 154–168.
Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). “Maximum likelihood from
incomplete data via the EM algorithm.” Journal of the Royal Statistical Society.
Series B (Methodological), 39(1), 1–22.
Deokar, B. and Hazarnis, A. (2012). “Intrusion detection system using log files and reinforce-
ment learning.” International Journal of Computer Applications 45(19), 28–35.
Dey, S., Janeja, V. P., and Gangopadhyay, A. (2009). “Temporal neighborhood dis-
covery through unequal depth binning.” IEEE International Conference on Data
Mining (ICDM’09), 110–119.
Dey, S., Janeja, V. P., and Gangopadhyay, A. (2014). “Discovery of temporal neigh-
borhoods through discretization methods.” Intelligent Data Analysis, 18(4),
609–636.
Ding, Y., Yan, E., Frazho, A., and Caverlee, J. (2009, November). “Pagerank for
ranking authors in co-citation networks.” Journal of the American Society for
Information Science and Technology, 60(11), 2229–2243.
Domo. (2017). “Data never sleeps.” www.domo.com/learn/data-never-sleeps-5?aid=
ogsm072517_1&sf100871281=1. Last accessed November 2021.
Drinkwater, D. (2016). “Does a data breach really affect your firm’s reputation?” www
.csoonline.com/article/3019283/data-breach/does-a-data-breach-really-affect-
your-firm-s-reputation.html. Last accessed June 2017.
Duchene, F., Garbayl, C., and Rialle, V. (2004). “Mining heterogeneous multivariate
time-series for learning meaningful patterns: application to home health telecare.”
arXiv preprint cs/0412003.
Duczmal, L. and Renato, A. (2004). “A simulated annealing strategy for the detection
of arbitrarily shaped spatial clusters.” Computational Statistics and Data Analysis,
45(2), 269–286.
Eberle, W., Graves, J., and Holder, L. (2010). “Insider threat detection using a graph-
based approach.” Journal of Applied Security Research, 6(1), 32–81.
ENISA (European Union Agency for Network And Information Security). (2016, Jan).
“Threat Taxonomy: a tool for structuring threat information.” https://library
.cyentia.com/report/report_001462.html. Last accessed November 2021.
Ester, M., Frommelt, A., Kriegel, H.-P., and Sander, J. (1998). “Algorithms for
characterization and trend detection in spatial databases.” Proceedings of the
Fourth International Conference on Knowledge Discovery and Data Mining
(KDD'98), 44–50.
Ester, M., Kriegel, H., and Sander, J. (1997). “Spatial data mining: a database
approach.” The 5th International Symposium on Advances in Spatial Databases,
Springer-Verlag, 47–66.
Ester, M., Kriegel, H. P., Sander, J., and Xu, X. (1996, August). “A density-based
algorithm for discovering clusters in large spatial databases with noise.” KDD, 96
(34), 226–231.
Estevez-Tapiador, J. M., Garcia-Teodoro, P., and Diaz-Verdejo, J. E. (2004). “Anomaly
detection methods in wired networks: a survey and taxonomy.” Computer
Communications, 27(16), 1569–1584.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 171

Fabrikant, A., Koutsoupias, E., and Papadimitriou, C. H. (2002). “Heuristically opti-

mized trade-offs: a new paradigm for power laws in the Internet.” International
Colloquium on Automata, Languages and Programming. Springer, 110–122.
Faloutsos, M., Faloutsos, P., and Faloutsos, C. (1999). “On power-law relationships
of the internet topology.” The Structure and Dynamics of Networks, M. Newman,
A.-L. Barabási, and D. J. Watts (Eds.). Princeton University Press, 195–206.
Famili, A., Shen, W. M., Weber, R., and Simoudis, E. (1997). “Data preprocessing and
intelligent data analysis.” Intelligent Data Analysis, 1(1–4), 3–23.
Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P. (1996). “From data mining to
knowledge discovery in databases.” AI Magazine 17(3), 37.
Feily, M., Shahrestani, A., and Ramadass, S. (2009, June). “A survey of botnet and
botnet detection.” Third International Conference on Emerging Security
Information, Systems and Technologies, 2009. SECURWARE’09. IEEE, 268–273.
Ferebee, D., Dasgupta, D., and Wu, Q. (2012, December). “A cyber-security storm
map.” 2012 International Conference on Cyber Security (CyberSecurity). IEEE,
93–102.
FireEye. (2019). “Mandiant Purple Team Assessment data sheet.” www.fireeye.com/
content/dam/fireeye-www/services/pdfs/pf/ms/ds-purple-team-assessment.pdf.
Last accessed November 2021.
Fischer, P., Lea, S. E., and Evans, K. M. (2013). “Why do individuals respond to
fraudulent scam communications and lose money? The psychological determin-
ants of scam compliance.” Journal of Applied Social Psychology, 43(10),
2060–2072.
Fodor, I. K. (2002). “A survey of dimension reduction techniques.” Center for Applied
Scientific Computing, Lawrence Livermore National Laboratory 9, 1–18.
Frank, E., and Witten, H. I. (1998). “Generating accurate rule sets without global
optimization.” Proceedings of the 15th International Conference on Machine
Learning (ICML’98), Madison, Wisconsin, E. Frank and I. H. Witten (Eds.).
Morgan Kaufmann, 144–151.
Frank, L., Greitzer, R., and Hohimer, E. (2011, Summer). “Modeling human behavior
to anticipate insider attacks.” Journal of Strategic Security: Strategic Security in
Cyber Age, 4(2), 25–48.
Freeman, L. C. (1978). “Centrality in social networks conceptual clarification.” Social
Networks, 1(3), 215–239.
Frei, S., May, M., Fiedler, U., and Plattner, B. (2006, September). “Large-scale
vulnerability analysis.” Proceedings of the 2006 SIGCOMM Workshop on
Large-Scale Attack Defense. ACM, 131–138.
Freund, Y. and Schapire, R. E. (1996). “Experiments with a new boosting algorithm.”
Proceedings of the 13th International Conference on Machine Learning. Morgan
Kaufmann, 148–146.
Gandhi, R., Sharma, A., Mahoney, W., et al. (2011). “Dimensions of cyber-attacks:
cultural, social, economic, and political.” IEEE Technology and Society Magazine,
30(1), 28–38.
Garcia, S., Luengo, J., Sáez, J. A., Lopez, V., and Herrera, F. (2013). “A survey
of discretization techniques: taxonomy and empirical analysis in supervised
learning.” IEEE Transactions on Knowledge and Data Engineering, 25(4),
734–750.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 172 References

Garcia-Teodoro, P., Diaz-Verdejo, J., Maciá-Fernández, G., and Vázquez, E. (2009).

“Anomaly-based network intrusion detection: techniques, systems and chal-
lenges.” Computers and Security, 28(1), 18–28.
Garg, A., Upadhyaya, S., and Kwiat, K. (2013). “A user behavior monitoring and
profiling scheme for masquerade detection.” Handbook of Statistics: Machine
Learning: Theory and Applications, 31, 353.
Geenens, P. (2020). “FireEye hack turns into a global supply chain attack.” https://
securityboulevard.com/2020/12/fireeye-hack-turns-into-a-global-supply-chain-
attack/. Last accessed November 2021.
Gennari, J. H., Langley, P., and Fisher, D. (1989). “Models of incremental concept
formation.” Artificial Intelligence, 40(1–3), 11–61.
Gesenhues, A. (2014). “Google Dorking: it’s all fun & games until the hackers show
up.” http://searchengineland.com/google-dorking-fun-games-hackers-show-
202191. Last accessed November 2021.
Glaz, J., Naus, J., and Wallenstein, S. (2001). Scan Statistics. Springer Verlag.
Goldberg, L. R. (1990). “An alternative ‘description of personality’: the big-five factor
structure.” Journal of Personality and Social Psychology, 59(6), 1216.
Golnabi, K., Min, R. K,. Khan, L., and Al-Shaer, E. (2006). “Analysis of firewall policy
rules using data mining techniques.” 2006 IEEE/IFIP Network Operations and
Management Symposium, NOMS 2006. IEEE/IFIP, 205–315.
Goodfellow, I. (2016). “NIPS 2016 tutorial: generative adversarial networks.”
arXiv:1701.00160.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. (2014). “Generative adversarial
nets.” Proceedings of the 27th International Conference on Neural Information
Processing Systems – Volume 2 (NIPS'14). MIT Press, 2672–2680.
Gopalani, S. and Arora, R. (2015). “Comparing Apache Spark and Map Reduce with
performance analysis using K-means.” International Journal of Computer
Applications, 113(1), 8–11.
Gormley, T., Reingold, N., Torng, E., and Westbrook, J. (2000). “Generating adversar-
ies for request-answer games.” Proceedings of the 11th ACM-SIAM Symposium on
Discrete Algorithms. ACM-SIAM, 564–565.
Grazioli, S. (2004). “Where did they go wrong? An analysis of the failure of know-
ledgeable internet consumers to detect deception over the internet.” Group
Decision and Negotiation, 13(2), 149–172.
Griffith, D. A. (1987). Spatial Autocorrelation: A Primer. Association of American
Geographers.
Grover, A., Gholap, J. Janeja, V. P., et al. (2015). “SQL-like big data environments:
case study in clinical trial analytics.” Proceedings of 2015 IEEE International
Conference on Big Data, 2680–2689.
Gu, G., Zhang, J., and Lee, W. (2008, February). “BotSniffer: detecting botnet com-
mand and control channels in network traffic.” Proceedings of the 15th Annual
Network and Distributed System Security Symposium, vol. 8, 1–18.
Guardian, The. (2016). “Norway, the country where you can see everyone’s tax
returns.” www.theguardian.com/money/blog/2016/apr/11/when-it-comes-to-tax-
transparency-norway-leads-the-field. Last accessed November 2021.
Guha, S., Rastogi, R., and Shim, K. (1998, June). “CURE: an efficient clustering
algorithm for large databases.” ACM Sigmod Record, 27(2), 73–84.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 173

Guha, S., Rastogi, R., and Shim, K. (2000). “ROCK: a robust clustering algorithm for
categorical attributes.” Information Systems, 25(5), 345–366.
Gupta, H., Sural, S., Atluri, V., and Vaidya, J. (2016). “Deciphering text from touchsc-
reen key taps.” IFIP Annual Conference on Data and Applications Security and
Privacy. Springer International Publishing, 3–18.
Guralnik, V. and Srivastava, J. (1999). “Event detection from time series data.”
KDD’99: Proceedings of the Fifth ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining. ACM, 33–42.
Haining, R. (2003). Spatial Data Analysis: Theory and Practice. Cambridge University Press.
Halevi, T., Lewis, J., and Memon, N. (2013, May). “A pilot study of cyber security and
privacy related behavior and personality traits.” Proceedings of the 22nd
International Conference on World Wide Web Companion. International World
Wide Web Conferences Steering Committee, 737–744.
Halliday, J. (2010, September 24). “Stuxnet worm is the ‘work of a national govern-
ment agency’.” www.guardian.co.uk/technology/2010/sep/24/stuxnet-worm-
national-agency. Last accessed November 2021.
Han, J., and Fu, Y. (1995, September). “Discovery of multiple-level association rules
from large databases.” VLDB’95, Proceedings of 21th International Conference
on Very Large Data Bases, U. Dayal, P. Gray, and S. Nishio (Eds.). Morgan
Kaufmann, 420–431.
Han, J., Pei, J., and Yin, Y. (2000, May). “Mining frequent patterns without candidate
generation.” ACM Sigmod Record, 29(2), 1–12.
Hellerstein, J. L., Ma, S., and Perng, C.-S. (2002). “Discovering actionable patterns in
event data.” IBM Systems Journal, 41(3), 475–493.
Heron, S. (2007). “The rise and rise of the keyloggers.” Network Security 2007(6), 4–6.
Hinneburg, A., and Keim, D. A. (1998, August). “An efficient approach to clustering in
large multimedia databases with noise.” KDD, 98, 58–65.
Hoffman, S. (2011). “Cyber attack forces internet shut down for DOE lab, on July 8.”
www.crn.com/news/security/231001261/cyber-attack-forces-internet-shut-down-
for-doe-lab.htm. Last accessed February 23, 2014.
Hong, J., Liu, C. C., and Govindarasu, M. (2014). “Integrated anomaly detection for cyber
security of the substations.” IEEE Transactions on Smart Grid, 5(4), 1643–1653.
Hu, M. and Liu, B. (2004, August). “Mining and summarizing customer reviews.”
Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining. ACM, 168–177.
Hu, Z., Baynard, C. W., Hu, H., and Fazio, M. (2015, June). “GIS mapping and spatial
analysis of cybersecurity attacks on a Florida university.” 2015 23rd International
Conference on Geoinformatics. IEEE, 1–5.
Hu, Z., Wang, H., Zhu, J., et al. (2014). “Discovery of rare sequential topic patterns in
document stream.” Proceedings of the 2014 SIAM International Conference on
Data Mining, 533–541.
Huang, Y., Pei, J., and Xiong, H. (2006). “Mining co-location patterns with rare events
from spatial data sets.” GeoInformatica, 10(3), 239–260.
Hussain, M., Al-Haiqi, A., Zaidan, A. A., Zaidan, B. B., Kiah, M. M., Anuar, N. B., and
Abdulnabi, M. (2016). “The rise of keyloggers on smartphones: A survey and
insight into motion-based tap inference attacks.” Pervasive and Mobile
Computing, 25, 1–25.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 174 References

Ingols, K., Lippmann, R., & Piwowarski, K. (2006, December). “Practical attack graph
generation for network defense.” Computer Security Applications Conference,
2006. ACSAC’06. 22nd Annual. IEEE, 121–130.
Iyengar, V. S. (2004). “On detecting space-time clusters.” KDD’04: Proceedings of the
Tenth ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining. ACM Press, 587–592.
Jain, A. K. (2010). “Data clustering: 50 years beyond K-means.” Pattern Recognition
Letters, 31(8), 651–666.
Jakobsson, M. (2007). “The human factor in phishing.” Privacy & Security of
Consumer Information, 7(1), 1–19.
Janeja, V. P. (2019). Do No Harm: An Ethical Data Life Cycle, Sci on the Fly. AAAS.
Janeja, V. P., Adam, N. R., Atluri, V., and Vaidya, J. (2010). “Spatial neighborhood
based anomaly detection in sensor datasets.” Data Mining and Knowledge
Discovery, 20(2), 221–258.
Janeja, V. P. and Atluri, V. (2005). “LS3: A linear semantic scan statistic technique for
detecting anomalous windows. Proceedings of the 2005 ACM Symposium on
Applied Computing, 493–497.
Janeja, V. P. and Atluri, V. (2005). “FS3: A random walk based free-form spatial scan
statistic for anomalous window detection.” Fifth IEEE International Conference
on Data Mining (ICDM’05). IEEE Computer Society, 661–664.
Janeja, V. P, and Atluri, V. (2008). “Random walks to identify anomalous free-form
spatial scan windows.” IEEE Transactions on Knowledge and Data Engineering,
20(10), 1378–1392.
Janeja, V. P. and Atluri, V. (2009). “Spatial outlier detection in heterogeneous neigh-
borhoods.” Intelligent Data Analysis, 13(1), 85–107.
Janeja, V. P., Azari, A., Namayanja, J. M., and Heilig, B. (2014, October).
“B-dids: Mining anomalies in a Big-distributed Intrusion Detection System.”
In 2014 IEEE International Conference on Big Data (Big Data) (pp. 32–34).
IEEE.
Jarvis, R. A. and Patrick, E. A. (1973). “Clustering using a similarity measure based
on shared near neighbors.” IEEE Transactions on Computers, 100(11),
1025–1034.
Jha, S., Sheyner, O., and Wing, J. (2002). “Two formal analyses of attack graphs.” Computer
Security Foundations Workshop, 2002. Proceedings. 15th IEEE. IEEE, 49–63.
Ji, Y., Zhang, X., Ji, S., Luo, X., and Wang, T. (2018, October). “Model-reuse attacks
on deep learning systems.” Proceedings of the 2018 ACM SIGSAC Conference on
Computer and Communications Security. ACM, 349–363.
Joachims, T. (2002, July). “Optimizing search engines using clickthrough data.”
Proceedings of the Eighth ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining. ACM, 133–142.
John, O. P. and Srivastava, S. (1999). “The big-five trait taxonomy: history,
measurement, and theoretical perspectives.” Handbook of Personality:
Theory and Research, vol. 2., L. A. Pervin and O. P. John (Eds.). Guilford
Press, 102–138.
Kang, I., Kim, T., and Li, K. (1997). “A spatial data mining method by Delaunay
triangulation.” Proceedings of the 5th ACM International Workshop on Advances
in Geographic Information Systems. ACM, 35–39.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 175

Kang, U., Tsourakakis, C., Appel, A., Faloutsos, C., and Leskovec, J. (2010). “Radius
plots for mining tera-byte scale graphs: algorithms, patterns, and observations.”
Proceedings of the 2010 SIAM International Conference on Data Mining (SDM),
548–558.
Kang, U., Tsourakakis, C., and Faloutsos, C. (2009). “Pegasus: a peta-scale graph
mining system – implementation and observations.” 2009 Ninth IEEE
International Conference on Data Mining, 229–238.
Karypis, G., Han, E. H., and Kumar, V. (1999). “Chameleon: hierarchical clustering
using dynamic modeling.” Computer, 32(8), 68–75.
Kaspersky. (2020). “Cyberthreat real-time map.” https://cybermap.kaspersky.com/.
Last accessed November 2020.
Kath, O., Schreiner, R., and Favaro, J. (2009, September). “Safety, security, and
software reuse: a model-based approach.” Proceedings of the Fourth
International Workshop in Software Reuse and Safety. www.researchgate.net/
publication/228709911_Safety_Security_and_Software_Reuse_A_Model-Based_
Approach. Last accessed November 2021.
Kato, K. and Klyuev, V. (2017, August). “Development of a network intrusion detec-
tion system using Apache Hadoop and Spark.” 2017 IEEE Conference on
Dependable and Secure Computing. IEEE, 416–423.
Katsini, C., Abdrabou, Y., Raptis, G. E., Khamis, M., and Alt, F. (2020, April). “The
role of eye gaze in security and privacy applications: survey and future HCI
research directions.” Proceedings of the 2020 CHI Conference on Human
Factors in Computing Systems, 1–21.
Kaufman, L. and Rousseeuw, P. (1987). Clustering by Means of Medoids. North-
Holland.
Kaufman, L. and Rousseeuw, P. J. (1990). Finding Groups in Data: An Introduction to
Cluster Analysis. John Wiley & Sons.
Keim, D. A., Mansmann, F., Panse, C., Schneidewind, J., and Sips, M. (2005).
“Mail explorer – spatial and temporal exploration of electronic mail.”
Proceedings of the Seventh Joint Eurographics/IEEE VGTC Conference on
Visualization, 247–254.
Keim, D. A., Mansmann, F., and Schreck, T. (2005). “Analyzing electronic mail using
temporal, spatial, and content-based visualization techniques.” Informatik 2005–
Informatik Live!, vol. 67, 434–438.
Keogh. E., Lin. J., and Fu. A. (2005). “Hot sax: efficiently finding the most unusual
time series subsequence.” Fifth IEEE International Conference on Data Mining
(ICDM'05), doi: 10.1109/ICDM.2005.79.
Kianmehr, K. and Koochakzadeh, N. (2012). “Learning from socio-economic charac-
teristics of IP geo-locations for cybercrime prediction.” International Journal of
Business Intelligence and Data Mining, 7(1/2), 21–39
Kim, S., Edmonds, W., and Nwanze, N. 2014. “On GPU accelerated tuning for a
payload anomaly-based network intrusion detection scheme.” Proceedings of the
9th Annual Cyber and Information Security Research Conference (CISR ‘14).
ACM, 1–4. doi: 10.1145/2602087.2602093.
Kim Zetter Security. (2013). “Someone’s been siphoning data through a huge security
hole in the internet.” www.wired.com/2013/12/bgp-hijacking-belarus-iceland/.
Last accessed December 2016.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 176 References

Knorr, E. M., Ng, R. T., and Tucakov, V. (2000). “Distance-based outliers: algorithms
and applications.” VLDB Journal – The International Journal on Very Large Data
Bases, 8(3–4), 237–253.
Koh, Y. S. and Ravana, S. D. (2016). “Unsupervised rare pattern mining: a survey.”
ACM Transactions on Knowledge Discovery from Data (TKDD), 10(4), 45.
Koh, Y. S. and Rountree, N. (2005). “Finding sporadic rules using apriori-inverse.”
PAKDD (Lecture Notes in Computer Science), vol. 3518. T. B. Ho, D. Cheung,
and H. Liu (Eds.). Springer, 97–106.
Koike, H. and Ohno, K. (2004). “SnortView: visualization system of snort logs.”
Proceedings of the 2004 ACM Workshop on Visualization and Data Mining for
Computer Security (VizSEC/DMSEC '04). ACM, 143–147. doi: 10.1145/
1029208.1029232.
Kosner, A. W. (2012). “Cyber security fails as 3.6 million social security numbers
breached in South Carolina.” www.forbes.com/sites/anthonykosner/2012/10/27/
cyber-security-fails-as-3-6-million-social-security-numbers-breached-in-south-car
olina/?sh=5f3637784e9e. Last accessed March 2021.
Kotsiantis, S. and Pintelas, P. (2004). “Recent advances in clustering: a brief
survey.” WSEAS Transactions on Information Science and Applications, 1(1),
73–81.
Kotsiantis, S. B., Zaharakis, I. D., and Pintelas, P. E. (2006). “Machine learning: a
review of classification and combining techniques.” Artificial Intelligence Review,
26(3), 159–190.
Kulldorff, M. (1997). “A spatial scan statistic.” Communications of Statistics – Theory
Meth., 26(6), 1481–1496.
Kulldorff, M., Athas, W., Feuer, E., Miller, B., and Key, C. (1998). “Evaluating cluster
alarms: a space-time scan statistic and brain cancer in Los Alamos.” American
Journal of Public Health, 88(9), 1377–1380.
Kurgan, L. A. and Musilek, P. (2006, March). “A survey of knowledge discovery and
data mining process models.” Knowledge Engineering Review, 21(1) 1–24. doi: 10
.1017/S0269888906000737.
L24. (2016). “First national cyber security exercise Cyber Shield.” http://l24.lt/en/
society/item/150489-first-national-cyber-security-exercise-cyber-shield-2016-will-
be-held. Last accessed April 12, 2017.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). “Deep learning.” Nature, 521(7553),
436.
Lee, G., Yun, U., Ryang, H., and Kim, D. (2015). “Multiple minimum support-based
rare graph pattern mining considering symmetry feature-based growth technique
and the differing importance of graph elements.” Symmetry, 7(3), 1151.
Leskovec, J. (2008). “Dynamics of large networks.” Dissertation. ProQuest
Dissertations Publishing.
Leskovec, J., Chakrabarti, D., Kleinberg, J., and Faloutsos, C. (2005). “Realistic,
mathematically tractable graph generation and evolution, using Kronecker multi-
plication.” European Conference on Principles and Practice of Knowledge
Discovery in Databases: PKDD 2005, A. M. Jorge, L. Torgo, P. Brazdil, R.
Camacho, and J. Gama (Eds.). Lecture Notes in Computer Science, vol. 3721.
Springer. doi: 10.1007/11564126_17.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 177

Leskovec, J. and Faloutsos, C. (2007). “Scalable modeling of real graphs using

kronecker multiplication.” International Conference on Machine Learning
(ICML ‘07). ACM, 497–504. doi: 10.1145/1273496.1273559.
Leskovec, J., Kleinberg, J., and Faloutsos, C. (2005, August). “Graphs over time:
densification laws, shrinking diameters and possible explanations.” Proceedings
of the Eleventh ACM SIGKDD International Conference on Knowledge Discovery
in Data Mining. ACM, 177–187.
Leskovec, J., Kleinberg, J., and Faloutsos, C. (2007). “Graph evolution: densification
and shrinking diameters.” ACM Transactions on Knowledge Discovery from Data
(TKDD), 1, 2-es. doi: 10.1145/1217299.1217301.
Lewis, D. M. and Janeja, V. P. (2011). “An empirical evaluation of similarity coeffi-
cients for binary valued data.” International Journal of Data Warehousing and
Mining (IJDWM), 7(2), 44–66. doi: 10.4018/jdwm.2011040103.
Lewis, J. A. (2005). “Computer espionage, Titan Rain and China.” http://csis.org/files/
media/csis/pubs/051214_china_titan_rain.pdf. Last accessed March 2020.
Leyden, J. (2012, March 29). “NSA’s top spook blames China for RSA hack.” www.theregister
.co.uk/2012/03/29/nsa_blames_china_rsa_hack/. Last accessed November 2021.
Li, J., Dou, D., Wu, Z., Kim, S., and Agarwal, V. (2005). “An Internet routing forensics
framework for discovering rules of abnormal BGP events.” ACM SIGCOMM
Computer Communication Review, 35(5), 55–66.
Li, X., Wang, L., & Sung, E. (2008). AdaBoost with SVM-based component classifiers.
Engineering Applications of Artificial Intelligence, 21(5), 785–795.
Lin, J., Keogh, E., Lonardi, S., and Chiu, B. (2003). “A symbolic representation of time
series, with implications for streaming algorithms.” DMKD’03: Proceedings of the
8th ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge
Discovery. ACM, 2–11.
Lin, Q., Adepu, S., Verwer, S., and Mathur, A. (2018, May). “TABOR: a graphical
model-based approach for anomaly detection in industrial control
systems.” Proceedings of the 2018 on Asia Conference on Computer and
Communications Security (ASIACCS). ACM, 525–536. doi: 10.1145/
3196494.3196546.
Liu, B., Hsu, W., and Ma, Y. (1999a). “Mining association rules with multiple
minimum supports.” Proceedings of the 5th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM, 337–341.
Liu, H., Hussain, F., Tan, C. L., and Dash, M. (2002). “Discretization: an enabling
technique.” Data Mining and Knowledge Discovery, 6(4), 393–423.
Liu, Z., Wang, C., and Chen, S. (2008). “Correlating multi-step attack and constructing
attack scenarios based on attack pattern modeling.” International Conference on
Information Security and Assurance, 2008. ISA 2008. IEEE, 214–219.
Limmer, T. and Dressler, F. (2010). “Dialog-based payload aggregation for intrusion detec-
tion.” Proceedings of the 17th ACM Conference on Computer and Communications
Security (CCS ‘10). ACM, 708–710. doi: 10.1145/1866307.1866405.
Lockheed Martin. (2015). “Gaining the advantage, applying Cyber Kill Chain® meth-
odology to network defense 2015.” Technical report. www.lockheedmartin.com/
content/dam/lockheed-martin/rms/documents/cyber/Gaining_the_Advantage_Cyber_
Kill_Chain.pdf. Last accessed November 2021.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 178 References

Luengo, J., García, S., and Herrera, F. (2012). “On the choice of the best imputation
methods for missing values considering three groups of classification methods.”
Knowledge and Information Systems, 32(1), 77–108.
Ma, S. and Hellerstein, J. L. (2001a). “Mining mutually dependent patterns.”
Proceedings of the 2001 International Conference on Data Mining (ICDM’01),
San Jose, CA, November 2001. IEEE, 409–416.
Ma, S. and Hellerstein, J. L. (2001b). “Mining partially periodic event patterns
with unknown periods.” Proceedings of the 2001 International Conference
on Data Engineering (ICDE’01), Heidelberg, Germany, April 2001. IEEE,
205–214.
MacQueen, J. (1967, June). “Some methods for classification and analysis of
multivariate observations.” Proceedings of the Fifth Berkeley Symposium on
Mathematical Statistics and Probability. University of California Press, vol. 1,
no. 14, 281–297.
Mandiant. (2013). “APT1: exposing one of China’s cyber espionage units.” www
.mandiant.com/resources/apt1-exposing-one-of-chinas-cyber-espionage-units.
Last accessed November 2021.
Manyika, J., Chui, M., Brown, B., et al. (2011). Big data: The next frontier for
innovation, competition, and productivity. McKinsey Global Institute.
Maron, M. and Kuhns, J. (1960). “On relevance, probabilistic indexing, and infor-
mation retrieval.” Journal of the Association for Computing Machinery 7,
216–244.
Massicotte, F., Whalen, T., and Bilodeau, C. (2003). “Network mapping tool for real-
time security analysis.” RTO IST Symposium on Real Time Intrusion Detection,
12-1–12-10.
McAfee. (2010). “Protecting your critical assets.” www.wired.com/images_blogs/
threatlevel/2010/03/operationaurora_wp_0310_fnl.pdf. Last accessed November
2021.
McAfee. (2011). “Global energy cyberattacks: “Night Dragon.” www.heartland.org/
publications-resources/publications/global-energy-cyberattacks-night-dragon. Last
accessed November 2021.
McAfee. (2018). “The economic impact of cybercrime – no slowing down.” McAfee,
Center for Strategic and International Studies (CSIS). www.mcafee.com/enter
prise/en-us/solutions/lp/economics-cybercrime.html.
McBride, M., Carter, L., and Warkentin, M. (2012). “Exploring the role of
individual employee characteristics and personality on employee compliance with
cybersecurity policies.” RTI International Institute for Homeland Security
Solutions, 5(1), 1.
McGuire, M. P., Janeja, V.P., and Gangopadhyay, A. (2008, August). “Spatiotemporal
neighborhood discovery for sensor data.” International Workshop on Knowledge
Discovery from Sensor Data. Springer, 203–225.
McGuire, M. P., Janeja, V. P., and Gangopadhyay, A. (2012). “Mining sensor datasets
with spatio-temporal neighborhoods.” Journal of Spatial Information Science
(JOSIS), 2013(6), 1–42.
Meng, X., Bradley, J., Yavuz, B., et al. (2016). “Mllib: machine learning in Apache
Spark.” Journal of Machine Learning Research, 17(1), 1235–1241.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 179

Mezzour, G. (2015). “Assessing the global cyber and biological threat.” Thesis,
Carnegie Mellon University. doi: 10.1184/R1/6714857.v1.
Miller, H. J. (2004). “Tobler’s first law and spatial analysis.” Annals of the Association
of American Geographers, 94(2), 284–289.
Miller, W. B. (2014). “Classifying and cataloging cyber-security incidents within
cyber-physical systems.” Doctoral dissertation, Brigham Young University.
Misal, V., Janeja, V. P., Pallaprolu, S. C., Yesha, Y., and Chintalapati, R. (2016,
December). “Iterative unified clustering in big data.” 2016 IEEE International
Conference on Big Data (Big Data). IEEE, 3412–3421.
Mitra, B., Sural, S., Vaidya, J., and Atluri, V. (2016). “A survey of role mining.” ACM
Computing Surveys (CSUR), 48(4), 1–37.
MITRE ATT&CK. (2020). ATT&CK Matrix for Enterprise. https://attack.mitre.org/.
Last accessed November 2021.
Molina, L. C., Belanche, L., and Nebot, À. (2002). “Feature selection algorithms: a
survey and experimental evaluation.” 2002 IEEE International Conference on
Data Mining, 2002. ICDM 2003. Proceedings. IEEE, 306–313.
Namayanja, J. M. and Janeja, V. P. (2014, October). “Change detection in temporally
evolving computer networks: a big data framework.” 2014 IEEE International
Conference on Big Data. IEEE, 54–61.
Namayanja, J. M. and Janeja, V. P. (2015, May). “Change detection in evolving computer
networks: changes in densification and diameter over time.” 2015 IEEE International
Conference on Intelligence and Security Informatics (ISI). IEEE, 185–187.
Namayanja, J. M. and Janeja, V. P. (2017). “Characterization of evolving networks for
cybersecurity.” Information Fusion for Cyber-Security Analytics, I. Alsmadi, G.
Karabatis, and A. Aleroud (Eds.). Studies in Computational Intelligence, vol. 691.
Springer International Publishing, 111–127. doi: 10.1007/978-3-319-44257-0_5.
Namayanja, J. M. and Janeja, V. P. (2019). “Change detection in large evolving
networks.” International Journal of Data Warehousing and Mining, 15(2), 62–79.
Naseer, S., Saleem, Y., Khalid, S., et al. (2018). “Enhanced network anomaly detection
based on deep neural networks.” IEEE Access, 6, 48231–48246.
Naus, J. (1965). “The distribution of the size of the maximum cluster of points on the
line.” Journal of the American Statistical Association, 60, 532–538.
Neill, D., Moore, A., Pereira, F., and Mitchell, T. (2005). “Detecting significant
multidimensional spatial clusters.” Advances in Neural Information Processing
Systems 17, MIT Press, 969–976.
Netscout. (2020). “A global threat visualization.” www.netscout.com/global-threat-
intelligence. Last accessed November 2020.
Ng, R. T. and Han, J. (1994, September). “Efficient and effective clustering methods for
spatial data mining.” Proceedings of the 20th International Conference on Very
Large Data Bases (VLDB ’94). Morgan Kaufmann , 144–155.
Ng, R. T., Lakshmanan, L. V. S., Han, J., and Pang, A. 1998. “Exploratory mining and
pruning optimizations of constrained associations rules.” Proceedings of the
1998 ACM SIGMOD International Conference on Management of Data
(SIGMOD ’98). ACM,13–24.
Nguyen, T. T. and Janapa Reddi, V. (2019). “Deep reinforcement learning for cyber
security.” arXiv:1906.05799.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 180 References

Nicosia, V., Tang, J., Mascolo, C., et al. (2013). “Graph metrics for temporal net-
works.” Temporal Networks: Understanding Complex Systems, P. Holme and J.
Saramäki (Eds.). Springer, 15–40. doi: 10.1007/978-3-642-36461-7_2.
NIST (National Institute of Standards and Technology). (2015). NIST Big Data
Interoperability Framework (NBDIF), V1.0. https://bigdatawg.nist.gov/V1_
output_docs.php. Last accessed April 12, 2017.
NIST. (2017). National vulnerability database. http://nvd.nist.gov/. Last accessed
September 2017.
Nunes, D. S., Zhang, P., and Silva, J. S. (2015). “A survey on human-in-the-loop
applications towards an internet of all.” IEEE Communications Surveys and
Tutorials, 17(2), 944–965.
O’Gorman, G. and McDonald, G. (2012). “The Elderwood Project.” www.infopoint-
security.de/medien/the-elderwood-project.pdf. Last accessed November 2021.
Ohm, M., Sykosch, A., and Meier, M. (2020, August). “Towards detection of software
supply chain attacks by forensic artifacts.” Proceedings of the 15th International
Conference on Availability, Reliability and Security (ARES ’20). ACM, 1–6. doi:
10.1145/3407023.3409183.
Openshaw, S. (1987). “A mark 1 geographical analysis machine for the
automated analysis of point data sets.” International Journal of GIS, 1(4),
335–358.
OSQuery. (2016). https://osquery.io/. Last accessed March 2020.
Otoum, S., Kantarci, B., and Mouftah, H. (2019, May). “Empowering reinforcement
learning on big sensed data for intrusion detection.” 2019–2019 IEEE
International Conference on Communications (ICC). IEEE, 1–7.
Paganini, P. (2014). “Turkish government is hijacking the IP for popular DNS pro-
viders.” http://securityaffairs.co/wordpress/23565/intelligence/turkish-govern
ment-hijacking-dns.html. Last accessed June 2017.
Parekh, J. J., Wang, K., and Stolfo, S. J. (2006). “Privacy-preserving payload-based
correlation for accurate malicious traffic detection.” Proceedings of the
2006 SIGCOMM Workshop on Large-Scale Attack Defense (LSAD ‘06). ACM,
99–106. doi: 10.1145/1162666.1162667.
Patcha, A. and Park, J. M. (2007). “An overview of anomaly detection techniques:
existing solutions and latest technological trends.” Computer Networks, 51(12),
3448–3470.
Paul, R. J. and Taylor, S. J. E. (2002). “Improving the model development process:
what use is model reuse: is there a crook at the end of the rainbow?” Proceedings
of the 34th Conference on Winter Simulation: Exploring New Frontiers (WSC
‘02). Winter Simulation Conference, 648–652.
Pei, J. and Han, J. (2000). “Can we push more constraints into frequent pattern
mining?” Proceedings of the Sixth ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining. ACM Press, 350–354.
Peña, J. M., Lozano, J. A., and Larrañaga, P. (2002). “Learning recursive Bayesian
multinets for data clustering by means of constructive induction.” Machine
Learning, 47(1), 63–89.
Pfeiffer, T., Theuerling, H., and Kauer, M. (2013). “Click me if you can! How do users
decide whether to follow a call to action in an online message?” Human Aspects of
Information Security, Privacy, and Trust. HAS 2013, L. Marinos and I.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 181

Askoxylakis (Eds.). Lecture Notes in Computer Science, vol. 8030. Springer,

155–166. doi: 10.1007/978-3-642-39345-7_17.
Phillips, G., Shenker, S., and Tangmunarunkit, H. (1999). “Scaling of multicast trees:
comments on the Chuang–Sirbu scaling law.” Proceedings of the Conference on
Applications, Technologies, Architectures, and Protocols for Computer
Communication (SIGCOMM ’99). ACM, 41–51. doi: 10.1145/316188.316205.
Picard, R. W. (2003). “Affective computing: challenges.” International Journal of
Human-Computer Studies, 59(1), 55–64.
Press, S. J. and Wilson, S. (1978). “Choosing between logistic regression and discrimin-
ant analysis.” Journal of the American Statistical Association, 73(364), 699–705.
Qiu, J., Gao, L., Ranjan, S., and Nucci, A. (2007, September). “Detecting bogus BGP
route information: Going beyond prefix hijacking.” In 2007 Third International
Conference on Security and Privacy in Communications Networks and the
Workshops-SecureComm 2007 (pp. 381–390). IEEE.
Quader, F., and Janeja, V. (2014). Computational Models to Capture Human Behavior
In Cybersecurity Attacks. Academy of Science and Engineering (ASE).
Quader, F., and Janeja, V. P. (2021). Insights into Organizational Security Readiness:
Lessons Learned from Cyber-Attack Case Studies. Journal of Cybersecurity and
Privacy, 1(4), 638–659.
Quader, F., Janeja, V., and Stauffer, J. (2015, May). “Persistent threat pattern discov-
ery.” 2015 IEEE International Conference on Intelligence and Security
Informatics (ISI). IEEE, 179–181.
Quinlan, J. R. (1979). Discovering Rules by Induction from Large Collections of
Examples: Expert Systems in the Micro Electronic Age. Edinburgh University
Press.
Quinlan, J. R. (1993). C4.5: Programs for Machine Learning. Morgan Kaufmann.
Ramaswamy, S., Rastogi, R., and Shim, K. (2000, May). “Efficient algorithms for
mining outliers from large data sets.” ACM Sigmod Record, 29(2), 427–438).
Rashid, F. Y. (2013, April 4). “DHS: spear phishing campaign targeted 11 energy
sector firms.” www.securityweek.com/dhs-spear-phishing-campaign-targeted-11-
energy-sector-firms. Last accessed February 23, 2014.
Raveh, A. and Tapiero, C. S. (1980). “Periodicity, constancy, heterogeneity and the
categories of qualitative time series.” Ecology, 61(3), 715–719.
Revow, M., Williams, C. K., and Hinton, G. E. (1996). “Using generative models for
handwritten digit recognition.” IEEE Transactions on Pattern Analysis and
Machine Intelligence, 18(6), 592–606.
Risk Based Security. (2014). “A breakdown and analysis of the December, 2014 Sony
hack.” www.riskbasedsecurity.com/2014/12/05/a-breakdown-and-analysis-of-the-
december-2014-sony-hack/. Last accessed November 2021.
Rivest, R. L. and Vuillemin, J. (1976). “On recognizing graph properties from adja-
cency matrices.” Theoretical Computer Science, 3(3), 371–384.
Roddick, J. F. and Hornsby, K., (Eds.). (2001). Temporal, Spatial, and Spatio-
Temporal Data Mining, First International Workshop TSDM 2000 Lyon,
France, September 12, 2000, Revised Papers. Lecture Notes in Computer
Science, vol. 2007. Springer-Verlag.
Roddick, J. F., Hornsby, K., and Spiliopoulou, M. (2001). “An updated bibliography of
temporal, spatial, and spatio-temporal data mining research.” Temporal, Spatial,

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 182 References

and Spatio-Temporal Data Mining, First International Workshop TSDM 2000

Lyon, France, September 12, 2000, Revised Papers. Springer-Verlag, 147–164.
Roddick, J. F. and Spiliopoulou, M. (1999). “A bibliography of temporal, spatial and
spatio-temporal data mining research.” SIGKDD Explorations Newsletter, 1(1),
34–38.
Roman, J. The Hadoop Ecosystem Table. https://hadoopecosystemtable.github.io/. Last
accessed April 13, 2017.
Rousseeuw, P. J. (1987). “Silhouettes: a graphical aid to the interpretation and
validation of cluster analysis.” Computational and Applied Mathematics. 20,
53–65.
Sadhwani, H. (2020). “Introduction to threat hunting.” https://medium.com/@hirensadh
wani2619/introduction-to-threat-hunting-8dff62ba52ca. Last accessed November
2021.
Sainani, H. (2018). “IP reputation scoring – a perspective on clustering with meta-
features augmentation.” Thesis.
Sainani, H., Namayanja, J. M., Sharma, G., Misal, V., and Janeja, V. P. (2020). “IP
reputation scoring with geo-contextual feature augmentation.” ACM Transactions
on Management Information Systems (TMIS), 11(4), 26:1–26:29.
Samuel, A. W. (2004). “Hactivism and future of political participation.” www.alexan-
drasamuel.com/dissertation/pdfs/Samuel-Hacktivism-entire.pdf. Last accessed
November 2021.
Sander, J., Ester, M., Kriegel, H. P., and Xu, X. (1998). “Density-based clustering in
spatial databases: the algorithm GDBSCAN and its applications.” Data Mining
and Knowledge Discovery, 2(2), 169–194.
Sandoval, G. (2008). “YouTube blames Pakistan network for 2-hour outage.” www
.cnet.com/news/youtube-blames-pakistan-network-for-2-hour-outage/. Last
accessed June 2017.
Schlamp, J., Carle, G., and Biersack, E. W. (2013). “A forensic case study on as
hijacking: the attacker’s perspective.” ACM SIGCOMM Computer
Communication Review, 43(2), 5–12.
Schlosser, A. E., White, T. B., and Lloyd, S. M. (2006). “Converting web site visitors
into buyers: how web site investment increases consumer trusting beliefs and
online purchase intentions.” Journal of Marketing, 70(2), 133–148.
SecDev Group. (2009). “Tracking GhostNet: investigating a cyber espionage network.”
www.nartv.org/mirror/ghostnet.pdf. Last accessed November 2021.
Shashanka, M., Shen, M., and Wang, J. (2016). “User and entity behavior analytics for
enterprise security.” 2016 IEEE International Conference on Big Data (Big Data).
IEEEE. 1867–1874. doi: 10.1109/BigData.2016.7840805.
Shearer, C. (2000). “The CRISP-DM model: the new blueprint for data mining.”
Journal of Data Warehousing, 5, 13–22.
Shekhar, S., Lu, C., and Zhang, P. (2001). “Detecting graph-based spatial outliers:
algorithms and applications.” Proceedings of the 7th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, 371–376.
Sheng, S., Holbrook, M., Kumaraguru, P., Cranor, L. F., and Downs, J. (2010). “Who
falls for phish? A demographic analysis of phishing susceptibility and effective-
ness of interventions.” Proceedings of the SIGCHI Conference on Human Factors
in Computing Systems. ACM, 373–382.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 183

Shi, L. and Janeja, V. P. (2009, June). “Anomalous window discovery through scan
statistics for linear intersecting paths (SSLIP).” Proceedings of the 15th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining.
ACM, 767–776.
Shropshire, J., Warkentin, M., Johnston, A., and Schmidt, M. (2006). “Personality and
IT security: an application of the five-factor model.” AMCIS 2006 Proceedings.
Association for Information Systems AIS Electronic Library (AISeL), 415.
Simmon, E., Sowe, S. K., and Zettsu, K. (2015). “Designing a cyber-physical cloud
computing architecture.” IT Professional, (3), 40–45.
Simmons, C., Ellis, C., Shiva, S., Dasgupta, D., and Wu, Q. (2009). AVOIDIT: A cyber
attack taxonomy. 9th Annual Symposium on Information Assurance, 2–12. https://
nsarchive.gwu.edu/sites/default/files/documents/4530310/Chris-Simmons-Charles-
Ellis-Sajjan-Shiva.pdf. Last accessed November 2021.
Skariachan, D. and Finkle, J. (2014). “Target shares recover after reassurance on
data breach impact.” www.reuters.com/article/us-target-results/target-shares-
recover-after-reassurance-on-data-breach-impact-idUSBREA1P0WC20140226. Last
accessed March 2020.
Sklower, K. (1991, Winter). A Tree-Based Packet Routing Table for Berkeley UNIX.
USENIX.
SMR Foundation. (2021). NodeXL. www.smrfoundation.org/nodexl/. Last accessed
November 2021.
Snare. (2020). Snare. www.snaresolutions.com/central-83/. Last accessed March 2020.
Snort. (2020). Snort Rules Infographic. https://snort-org-site.s3.amazonaws.com/pro
duction/document_files/files/000/000/116/original/Snort_rule_infographic.pdf?X-
Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAIXACIED2
SPMSC7GA%2F20210316%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date
=20210316T191343Z&X-Amz-Expires=172800&X-Amz-SignedHeaders=host&
X-Amz-Signature=bcfc7d75d223ab40badd8bd9e89ded29cc98ed3896f0140f5532
0ee9bcdf1383. Last accessed March 2020.
Spitzner, L. (2003). Honeypots: Tracking Hackers, vol. 1. Addison-Wesley.
Srikant, R. and Agrawal, R. (1996). “Mining quantitative association rules in large
relational tables.” Proceedings of the 1996 ACM SIGMOD international
Conference on Management of Data. ACM Press, 1–12.
Statista. (2020). eCommerce report 2020 Statista Digital Market Outlook. www.statista
.com/study/42335/ecommerce-report/. Last accessed November 2021.
Statista. (2021, March). Digital population worldwide. www.statista.com/statistics/
617136/digital-population-worldwide/. Last accessed November 2021.
Statista. “IoT market – forecasts.” www.statista.com/statistics/1101442/iot-number-of-
connected-devices-worldwide/. Last accessed November 2021.
Stauffer, J. and Janeja, V. (2017). “A survey of advanced persistent threats and the
characteristics. Technical report.
Stephens, G. D. and Maloof, M. A. (2014, April 22) “Insider threat detection.” U.S.
Patent No. 8,707,431.
Stouffer, K., Falco, J., and Scarfone, K. (2009). “Guide to industrial control systems
(ICS) security.” Technical report, National Institute of Standards and Technology.
Stubbs, J., Satter, R., and Menn, J. (2020). “U.S. Homeland Security, thousands of
businesses scramble after suspected Russian hack.” www.reuters.com/article/

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 184 References

global-cyber/global-security-teams-assess-impact-of-suspected-russian-cyber-
attack-idUKKBN28O1KN. Last accessed November 2021.
Stutz, J. and Cheeseman, P. (1996). “AutoClass – a Bayesian approach to classifica-
tion.” Maximum Entropy and Bayesian Methods. Springer, 117–126.
Sugiura, O. and Ogden, R. T. (1994). “Testing change-points with linear trend.”
Communications in Statistics – Simulation and Computation, 23(2), 287–322.
doi: 10.1080/03610919408813172.
Sugiyama, M. and Borgwardt, K. (2013). “Rapid distance-based outlier detection via
sampling.” Proceedings of the 26th International Conference on Neural
Information Processing Systems – Volume 1 (NIPS'13). Curran Associates,
467–475.
Tan, Y., Vuran, M. C., and Goddard, S. (2009, June). “Spatio-temporal event model for
cyber-physical systems.” 29th IEEE International Conference on Distributed
Computing Systems Workshops, 2009. ICDCS Workshops’ 09. IEEE, 44–50.
Tang, X., Eftelioglu, E., Oliver, D., and Shekhar, S. (2017). “Significant linear hotspot
discovery.” IEEE Transactions on Big Data, 3(2), 140–153.
Tango, T. and Takahashi, K. (2005). “A flexibly shaped spatial scan statistic for
detecting clusters.” International Journal of Health Geographics, 4(11). doi:
10.1186/1476-072X-4-11.
Tao, F., Murtagh, F., and Farid, M. (2003). “Weighted association rule mining using
weighted support and significance framework.” Proceedings of the Ninth ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining,
KDD’03. ACM Press, 661–666.
Tartakovsky, A. G., Polunchenko, A. S., and Sokolov, G. (2013). “Efficient computer
network anomaly detection by changepoint detection methods.” IEEE Journal of
Selected Topics in Signal Processing, 7(1), 4–11.
Ten, C. W., Hong, J., and Liu, C. C. (2011). “Anomaly detection for cybersecurity of
the substations.” IEEE Transactions on Smart Grid, 2(4), 865–873.
Thakur, V. (2011, December 8). “The Sykipot Attacks.” www.symantec.com/connect/
blogs/sykipot-attacks. Last accessed November 2021.
Tim, O., Firoiu, L., and Cohen, P. (1999). “Clustering time series with hidden markov
models and dynamic time warping.” Presented at IJCAI-99 Workshop on
Sequence Learning.
Tobler, W. R. (1970). “A computer model simulation of urban growth in the Detroit
region.” Economic Geography, 46(2), 234–240.
Townsend, M., Rupp, L., and Green, J. (2014). “Target CEO ouster shows new board
focus on cyber attacks.” www.bloomberg.com/news/2014-05-05/target-ceo-ouster-
shows-new-board-focus-on-cyber-attacks.html. Last accessed November 2021.
Trend Micro Incorporated. (2012a). ”Detecting APT activity with network
traffic analysis.” www.trendmicro.com/cloud-content/us/pdfs/security-intelli
gence/white-papers/wp-detecting-apt-activity-with-network-traffic-analysis.pdf.
Last accessed November 2021.
Trend Micro Incorporated. (2012b), “Spear-phishing email: most favored APT attack
bait.”
Trinius, P., Holz, T., Göbel, J., and Freiling, F. C. (2009, October). “Visual analysis of
malware behavior using treemaps and thread graphs.” 6th International Workshop
on Visualization for Cyber Security, 2009. VizSec 2009. IEEE, 33–38.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 185

Tsuchiya, P. F. (1988). “The landmark hierarchy: a new hierarchy for routing in very
large networks.” Symposium Proceedings on Communications Architectures and
Protocols (SIGCOMM '88). ACM, 35–42.
Vaarandi, R. and Podiņš, K. (2010). “Network IDs alert classification with frequent
itemset mining and data clustering.” 2010 International Conference on Network
and Service Management. IEEE, 451–456.
Vaidya, J., Atluri, V., and Guo, Qi. (2007). “The role mining problem: finding a
minimal descriptive set of roles.” Proceedings of the 12th ACM Symposium on
Access Control Models and Technologies. ACM, 175–184.
Van Mieghem, V. (2016). “Detecting malicious behaviour using system calls.”
Master’s thesis, Delft University.
Venkatasubramanian, K., Nabar, S., Gupta, S. K. S., and Poovendran, R. (2011).
“Cyber physical security solutions for pervasive health monitoring systems.”
E-Healthcare Systems and Wireless Communications: Current and Future
Challenges, M. Watfa (Ed.). IGI Global, 143–162.
Verizon Wireless. (2017). “Data breach digest.” https://enterprise.verizon.com/
resources/articles/2017-data-breach-digest-half-year-anniversary/. Last accessed
November 2021.
Verma, J. P. and Patel, A. (2016, March–September). “Comparison of MapReduce and
Spark programming frameworks for big data analytics on HDFS.” International
Journal of Computer Science and Communication, 7(2), 80–84.
Versprille, A. (2015). “Researchers hack into driverless car system, take control of
vehicle.” www.nationaldefensemagazine.org/articles/2015/5/1/2015may-research
ers-hack-into-driverless-car-system-take-control-of-vehicle. Last accessed
November 20201.
Villeneuve, N. and Sancho, D. (2011). “The ‘lurid’ downloader.” www.trendmicro.com/
cloud-content/us/pdfs/security-intelligence/white-papers/wp_dissecting-lurid-apt.pdf.
Last accessed November 2021.
Vishwanath, A., Herath, T., Chen, R., Wang, J., and Rao, H. R. (2011). “Why do
people get phished? Testing individual differences in phishing vulnerability within
an integrated, information processing model.” Decision Support Systems, 51(3),
576–586.
Wall, M. E., Rechtsteiner, A., and Rocha, L. M. (2003). “Singular value decomposition
and principal component analysis.” A Practical Approach to Microarray Data
Analysis, D. P. Berrar, W. Dubitzky, and M. Granzow (Eds.). Springer, 91–109.
doi: 10.1007/0-306-47815-3_5.
Wang, H., Wu, B., Yang, S., Wang, B., and Liu, Y. (2014). “Research of decision tree
on YARN using MapReduce and Spark.” World Congress in Computer Science,
Computer Engineering, and Applied Computing. American Council on Science
and Education, 21–24.
Wang, K., He, Y., and Han, J. (2003). “Pushing support constraints into association
rules mining.” IEEE Transactions Knowledge Data Engineering, 15(3),
642–658.
Wang, K. and Stolfo, S. J. 2004. “Anomalous payload-based network intrusion detec-
tion.” Recent Advances in Intrusion Detection. RAID 2004, E. Jonsson, A. Valdes,
and M. Almgren (Eds.). Lecture Notes in Computer Science, vol. 3224. Springer,
89–96.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 186 References

Wang, K., Yu, H., and Cheung, D. W. 2001. “Mining confident rules without support
requirement.” Proceedings of the Tenth International Conference on Information
and Knowledge Management. ACM Press, 89–96.
Wang, L., Singhal, A., and Jajodia, S. (2007a, July). “Measuring the overall
security of network configurations using attack graphs.” IFIP Annual
Conference on Data and Applications Security and Privacy. Springer, 98–112.
Wang, L., Singhal, A., and Jajodia, S. (2007b, October). “Toward measuring network
security using attack graphs.” Proceedings of the 2007 ACM Workshop on Quality
of Protection. ACM, 49–54.
Wang, P. A. (2011). “Online phishing in the eyes of online shoppers.” IAENG
International Journal of Computer Science, 38(4), 378–383.
Wang, Q. H. and Kim, S. H. (2009). Cyber Attacks: Cross-Country Interdependence
and Enforcement. WEIS.
Wang, W., Yang, J., and Muntz, R. (1997, August). “STING: A statistical information
grid approach to spatial data mining.” VLDB, 97, 186–195.
Ward, J. S. and Barker, A. (2013). “Undefined by data: a survey of big data definitions.”
arXiv:1309.5821.
Washington Post. (2013). “Target data flood stolen-card market.” www.washingtonpost
.com/business/economy/target-cyberattack-by-overseas-hackers-may-have-comprom
ised-up-to-40-million-cards/2013/12/20/2c2943cc-69b5-11e3-a0b9-249bbb34602c_
story.html?utm_term=.42a8cd8b6c0e. Last accessed November 2016.
Websense. (2011). “Advanced persistent threat and advanced attacks: threat analysis
and defense strategies for SMB, mid-size, and enterprise organizations Rev 2.”
Technical report.
Wei, L., Keogh, E., and Xi, X. (2006). “SAXually explicit images: finding unusual
shapes.” ICDM’06: Proceedings of the Sixth International Conference on Data
Mining. IEEE Computer Society, 711–720.
Weimerskirch, A. (2018). Derrick Dominic Assessing Risk: Identifying and Analyzing
Cybersecurity Threats to Automated Vehicles. University of Michigan,
Wilson, R. J. (1986). Introduction to Graph Theory. John Wiley & Sons.
Wireshark. (2021). www.wireshark.org. Last accessed November 2021.
Wright, R., Chakraborty, S., Basoglu, A., and Marett, K. (2010). “Where did they go
right? Understanding the deception in phishing communications.” Group Decision
and Negotiation, 19(4), 391–416.
Wright, R. T. and Marett, K. (2010). The influence of experiential and dispositional
factors in phishing: an empirical investigation of the deceived. Journal of
Management Information Systems, 27(1), 27.
Wu, M., Miller, R. C., and Garfinkel, S. L. (2006). “Do security toolbars actually
prevent phishing attacks?” Proceedings of the SIGCHI Conference on Human
Factors in Computing Systems (CHI ‘06). ACM, 601–610.
Wu, X., Kumar, V., Quinlan, J. R., et al. (2008). “Top 10 algorithms in data mining.”
Knowledge and Information Systems, 14(1), 1–37.
Wübbeling, M., Elsner, T., and Meier, M. (2014, June). “Inter-AS routing anomalies:
improved detection and classification.” 6th International Conference on Cyber
Conflict (CyCon 2014), 2014. IEEE, 223–238.
Wybourne, M. N., Austin, M. F., and Palmer, C. C. (2009). National Cyber Security.
Research and Development Challenges. Related to Economics, Physical

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 References 187

Infrastructure and Human Behavior. I3P: Institute for Information Infrastructure

Protection.
Xu, X., Ester, M., Kriegel, H. P., and Sander, J. (1998, February). “A distribution-based
clustering algorithm for mining in large spatial databases.” Proceedings, 14th
International Conference on Data Engineering, 1998. IEEE, 324–331.
Xu, X., Jäger, J., and Kriegel, H. P. (1999). “A fast parallel clustering algorithm for
large spatial databases.” High Performance Data Mining., Y. Guo and R.
Grossman (Eds.). Springer US, 263–290.
Yamanishi, K. and Takeuchi, J. (2002). “A unifying framework for detecting outliers
and change points from non-stationary time series data.” KDD’02: Proceedings of
the Eighth ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining. ACM, 676–681.
Yang, S.-C., and Wang, Y.-L. (2011, May). “System dynamics based insider threat
modeling.” International Journal of Network Security and Its Applications 3(3),
doi: 10.1109/ICDM.2007.61.
Yang, X., Kong, L., Liu, Z., et al. (2018). “Machine learning and deep learning methods
for cybersecurity.” IEEE Access, 6, 35365–35381.
Yankov, D., Keogh, E., and Rebbapragada, U. (2007). “Disk aware discord discovery:
finding unusual time series in terabyte sized datasets.” Seventh IEEE International
Conference on Data Mining (ICDM 2007), 381–390. doi: 10.1109/
ICDM.2007.61.
Yıldırım, M. and Mackie, I. (2019). “Encouraging users to improve password security
and memorability.” International Journal of Information Security, 18(6),
741–759.
Yinka-Banjo, C. and Ugot, O.-A. (2019). “A review of generative adversarial networks
and its application in cybersecurity.” Artificial Intelligence Review, 53,
1721–1736. doi: 10.1007/s10462-019-09717-4.
Yun, H., Ha, D., Hwang, B., and Ryu, K. H. (2003, September 15). “Mining association
rules on significant rare data using relative support.” Journal of Systems and
Software, 67(3) 181–191.
Yanan, S., Janeja, V. P., McGuire, M. P, .and Gangopadhyay, A. (2012). “Tnet: tensor-
based neighborhood discovery in traffic networks.” 2012 IEEE 28th International
Conference on Data Engineering Workshops, 331–336. doi: 10.1109/
ICDEW.2012.72.
Zetter, K. (2011, April 20). “Top federal lab hacked in spear-phishing attack.” www
.wired.com/threatlevel/2011/04/oak-ridge-lab-hack/. Last accessed February 23,
2014.
Zhang, K., Hutter, M., and Jin, H. (2009). “A new local distance-based outlier detection
approach for scattered real-world data.” Advances in Knowledge Discovery and
Data Mining. PAKDD 2009, T. Theeramunkong, B. Kijsirikul, N. Cercone, and
T. B. Ho (Eds.). Lecture Notes in Computer Science, vol. 5476. Springer,
813–822. doi: 10.1007/978-3-642-01307-2_84.
Zhang, P., Huang,, Y., Shekhar, S., and Kumar, V. (2003). “Correlation analysis of
spatial time series datasets: a filter-and-refine approach.” Advances in Knowledge
Discovery and Data Mining. PAKDD 2003, K. Y. Whang., J. Jeon, K. Shim, and
J. Srivastava (Eds.). Lecture Notes in Computer Science, Vol. 2637 Springer. doi:
10.1007/3-540-36175-8_53.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 188 References

Zhang, T., Ramakrishnan, R., and Livny, M. (1996, June). “BIRCH: an efficient data
clustering method for very large databases.” ACM Sigmod Record, 25(2),
103–114.
Zhao, Q., and Bhowmick, S. S. (2003). “Sequential pattern mining: a survey.”
Technical report, CAIS Nayang Technological University Singapore, 1–26.
Zhou, B., Cheung, D. W., and Kao, B. (1999, April). “A fast algorithm for density-
based clustering in large database.” Methodologies for Knowledge Discovery and
Data Mining. PAKDD 1999, N. Zhong and L. Zhou (Eds.). Lecture Notes in
Computer Science, Vol. 1574. Springer Berlin Heidelberg. doi: 10.1007/3-540-
48912-6_45.
Zimmermann, A., Lorenz, A., and Oppermann, R. (2007, August). “An operational
definition of context.” Modeling and Using Context. CONTEXT 2007, B.
Kokinov, D. C. Richardson, T. R. Roth-Berghofer, and L. Vieu (Eds.). Lecture
Notes in Computer Science, Vol. 4635. Springer Berlin Heidelberg. doi: 10.1007/
978-3-540-74255-5_42.
Zimmermann, V. and Renaud, K. (2021). “The nudge puzzle: matching nudge inter-
ventions to cybersecurity decisions.” ACM Transactions on Computer–Human
Interaction (TOCHI), 28(1), 1–45.

https://doi.org/10.1017/9781108231954.013 Published online by Cambridge University Press

 Index

access control, 12, 16, 24, 28, 49 centroid, 34, 43, 110–111
accommodation, 99 CLARANS, 44–45
accuracy, 29, 40–41, 50 classification, 29–30, 40–41, 49–51, 55, 75, 90,
adjacency matrix, 130, 135 101, 152, 157
advanced persistent threat, 5, 69, 84–85, 139 client server, 14
anomalous window, 118 closeness, 131
anomaly, 18, 27–28, 34, 40, 42, 49, 81, 91–93, cloud, 2–3, 6, 15–16, 76
95–101, 104, 107, 110–112, 116–118, clustering, 11, 25, 34, 37, 40, 42–45, 48–49,
120–123, 129, 147, 158 53, 75, 92, 101, 105, 108–110, 117–118,
anomaly detection, 27, 34, 36, 91, 93, 100, 155
104, 116–117, 140 common vulnerabilities and exposures, 25
application security, 6 common vulnerability scoring system, 25, See
Apriori algorithm, 57–58 CVSS
APT, 5, 79, 84–89, 139 communication, 21
assets, 1–3, 14, 16, 24, 79 computational data model, 81–82, 84
association rule mining, 18–19, 36, 41, 59 computer networks, 3, 7–8, 11, 13, 16, 18, 27,
asymmetric coefficient, 39 67, 124, 148, 154
attack maps, 3, 113–114 conditional probability, 42, 55, 57
attack vector, 85 confidence, 30, 41, 57
autocorrelation, 67, 113–115, 117–118, 120, consistent, 22, 28, 71–72, 89, 134
122 consistent nodes, 71–72
autonomous systems, 95 context, 93
contextual anomalies, 91, 94
Bayesian, 24 contingency matrix, 38
behavioral, 11 CPS, 8, 67–68, 147, 149
behavioral data, 23 critical infrastructure, 3, 80
betweenness, 131 curse of dimensionality, 96–97, 109
BGP anomalies, 95 cyberattacks, 5, 24–26, 62, 68–70, 138–139,
big data, 9, 20, 60–61, 63, 65–66, 69–70, 72, 144, 156
74–76 cyber espionage, 5
blackhole attack, 79 cyberphysical, 7, 16, 147
Border Gateway Protocol, 95 cyberphysical systems, 8, 16, 67, 147
cybersecurity, 1–2, 6, 14–15, 27, 60–61, 112,
categorical data, 32 123, 128–129, 133–134, 147
central nodes, 70, 72, 132–133 cybersecurity data, 28
centrality, 21, 131, 133 cyber threats, 1

189

https://doi.org/10.1017/9781108231954.014 Published online by Cambridge University Press

 190 Index

data analytics, 7–9, 12, 60, 70, 137, 147 firewall, 14, 19
data cleaning, 33 first law of geography, 113
data Mining, 8, 29, 36 F-measure, 40–41
data preprocessing, 31–32 FP tree, 58
data reduction, 35 frequent itemset, 58
data storage, 3, 6, 76 frequent pattern, 55–56, 89
data transformation, 34
decision tree, 50–52, 54 Geary’s, 113
deep learning, 157–158, 161 generative adversarial network, 160–161
degrees, 22, 71, 132 geocontextual features, 152
denial of service, 4, 79, 125 geographic, 11
densification, 133 geolocations, 19, 27, 94, 115, 126, 141
density based, 42, 46–48 geological, 3
detection, 1, 13, 16–18, 20, 23, 25, 27, 34, 42, georeferenced, 68, 113
46, 59, 65, 67, 70–71, 73–74, 81, 84, 88, geospatial data, 152–153
91, 93–97, 99–101, 103–105, 108–109, graph data, 21
111–112, 116–118, 120–123, 129, 139, graphs, 21, 69, 129, 133–134
141, 147, 154, 158 ground truth, 40
diameter, 21, 69, 71, 127–128, 134
discord, 91, 121 hacktivism, 5
discordancy, 34, 102–104, 108–109 Hausdorff distance, 38
dissimilarity, 37, 119 header, 20
distance, 34, 37–38, 42–43, 46–49, 93, 101, heartbleed, 5
105, 107–109, 111, 113, 116, 122, 124, heterogeneous, 9, 30–31, 37, 60, 67, 73, 120, 151
131, 134 hierarchical, 42
distance based, 105, 108 high dimensional, 36, 47, 67, 96, 109
distance based outlier detection, 34 honeypots, 100
distributed control systems, 147 human behavior, 18
distributed denial of service (DDoS), 4, 125
distributed intrusion detection system, 73 IDS, 17
domain name system, 14 IDS logs, 11
imbalanced learning, 53
e-commerce, 16 inconsistent nodes, 71
edges, 21, 45, 72, 122, 124, 128, 130–131, 133, industrial control systems, 3, 147–148
148, 151 insider threat, 2, 24, 138, 140–142
eigenvector, 131 Intellectual property, 4
electronic assets, 1–2 intercluster, 42
ensemble, 50, 53–54, 75, 162 interface design, 143
entropy, 52–53 Internet of Things. See IoT
equal frequency binning, 34 interquartile range, 34, 104
Euclidean distance, 37, 109, 118 intracluster, 42
evaluation, 27, 29–30, 37, 40, 72, 111, 139, intrusion detection systems. See IDS
142, 149 IoT, 8, 149–150
exception, 91 IP reputation scoring, 152
explainability, 159 IQR, 34, 104
exploits, 86
eye tracker, 24 Jaccard coefficient, 39, 135
joint count, 113
false negative, 40
false negative rate, 41 K nearest neighbor, See K-NN
false positive, 40 key logging, 17

https://doi.org/10.1017/9781108231954.014 Published online by Cambridge University Press

 Index 191

K-means, 43–44, 66 outlier detection, 34, 67, 101, 105, 108–109, 117
K-NN, 47, 109 outliers, 34, 37, 46, 67, 91, 96–97, 99–105,
107–109, 111, 117, 121
lift, 30, 41, 84
linear semantic-based scan, 118 page rank, 21
links, 16, 47, 69, 124, 128, 137, 139, 142, parametric, 101–102
148–149 partitioning, 42–44
logical errors, 5 password attack, 79
password theft, 81–83
macroneighborhoods, 120 payload, 19
malfunctions, 8, 150 peculiarity, 91
malicious intent, 2, 81, 95, 142 personal assets, 2
malware, 19–20, 24, 28, 78, 81, 86, phishing, 5, 10, 25, 78, 80, 84–85, 124, 126,
124, 152 138–139, 144, 154
malware detection, 24 physical security, 2, 11, 27, 68, 142
Manhattan distance, 37 power grid, 2, 124
Markov model, 122 precision, 40–41, 151
Markov stationary distribution, 123 predict, 14
masking, 104 prevention, 1, 13, 17, 129
measures of similarity, 37 probabilistic label, 55
medoids, 43–45 probe, 81, 145
microneighborhood, 119 psychological profiling, 140
MILES, 53–54 psychology, 15
Minkowski distance, 37 public assets, 2
missing data, 33
missing data values, 33 rare patterns, 56, 58–59
misuse, 6, 11, 27, 142 recall, 40
mobile ad hoc, 79 recover, 1
mobile devices, 3, 61 rejection, 100
Moran’s, 113 replay attack, 79
multilevel mining, 73 respond, 1, 6, 137, 141
multipronged attack, 9 router, 14
multivariate, 34
scan statistic, 118
naive Bayes, 55 security indicators, 25
National Vulnerability Database, 25 security perimeter, 9
neighborhood, 46, 119–120, 124 security policies, 12, 19, 25, 138, 142, 152
network security, 7 semantic relationships, 120
network topology, 23, 134 semisupervised, 91
network traffic, 21 sensor networks, 3, 8, 16, 67, 124, 154
nodes, 22–23, 28, 69–71, 76, 122–123, 125, sequential, 103
127–128, 131–133, 135, 148, 159 silhouette coefficient, 42, 45
NodeXL, 21 similarity, 32, 34, 38–40, 43, 104, 119–120,
noisy data, 34 122, 129, 135, 156, 161
normal distribution, 101–102 simple matching coefficient, 39
normalization, 35 social engineering, 5, 78, 80
numeric data attributes, 32 social media, 16
NVD, 25 software vulnerability, 25
spam, 126, 152
OPTICS, 48–49, 105, 118 spatial, 11, 27, 32, 46, 67–68, 94, 108–109,
ordinal, 32 112–114, 116–118, 120–122, 124–125, 151

https://doi.org/10.1017/9781108231954.014 Published online by Cambridge University Press

 192 Index

spatial autocorrelation, 113–114 Trojan, 87, 100

spatial data, 67, 112 true negative, 40
spatial data mining, 67 true negative rate, 41
spatial heterogeneity, 113 true positive, 40
spatial neighborhoods, 32, 119, 124 true positive rate, 41
spatial relationships, 117, 119, 125
spatiotemporal scan statistic, 118 unauthorized access, 1–2, 4–6, 19, 70, 81, 154
spoofed, 121 univariate data, 34
spoofing, 19 unsupervised, 34, 91
SQL injection, 79, 86 unusual window, 118
standard deviation, 34, 83, 105 user interaction, 143
Stuxnet, 124 user’s perception, 139
sum of squared error, 30, 42
supervised, 91 variety, 9
supervisory control and data acquisition, 147 velocity, 9
support, 12, 41, 55, 57–59, 94, 132, 140 veracity, 9
support vector machines, 55 vertices, 21, 124, 128, 130–131
swamping, 104 volume, 9
vulnerability, 9, 25, 38–39, 69, 80, 82, 86, 125,
temporal, 11, 25, 27–28, 67, 69, 88, 94, 112, 148, 150
115, 118, 121, 123, 125–126, 133, 145, vulnerable, 2, 84–85, 136–137, 143, 162
149, 151, 155
threat hunting, 7 walk, 130
threat taxonomy, 78 watering hole, 85
threats, 1, 6–7, 9, 12, 14–15, 17–18, 20, 23, 25, weighted degree, 72
28, 63, 70, 79, 85, 88–90, 112, 123, 129, white listing, 129
136–137, 139, 141, 146–147, 150–151, wireless networks, 8
154, 162 Wireshark, 20
time series discords, 121
Titan Rain attack, 5 zero day, 86, 129, 154
topology analysis, 140 z-score, 35

https://doi.org/10.1017/9781108231954.014 Published online by Cambridge University Press