2. user-centered research activities that can be used to inform the
design and
development of cyber situation awareness technology. One of
the most important
responsibilities of a cyber security analyst is to watch over and
protect his network
from harm. Maintaining situation awareness of the wide variety
of events that occur
and massive amounts of data generated is one of many analytic
challenges. Situation
awareness technology aims to reduce the data overload burden
placed on the analyst.
Good situation awareness technology requires good design, and
good design requires
a good understanding of the user and a focus on the user during
the design process.
In the case of a cyber security analyst, the practice of good
user-centered design is
focused on his security-related work processes on a large
computer network. One way
of understanding how a cyber security analyst accomplishes
situation awareness on a
large computer network is to study the questions he may ask
himself during the course
of a network event. Studying the relationships between these
questions will lead to a
better understanding of the analysts’ mental model of cyber
situation awareness. A
mental model of cyber situation awareness is a valuable tool in
the user-centered
design of cyber-related technology and to researchers who
cannot always study cyber
security analysts in the field.
* This article is a work of the U.S. Government, and per 17
3. U.S.C. §105, receives no copyright
protection within the U.S. In those countries that afford the
article copyright protection, the
U.S. Government puts the article into the public domain.
146 C.L. Paul and K. Whitley
2 Background
There is a growing body of work within the cyber security field
that is focused on
understanding the work processes of cyber security analysts.
For example, the results
of a series of interviews with a wide variety of cyber security
analysts by Werlinger et
al. [8] described three stages of computer network incident
response activities:
preparation; anomaly detection; and, anomaly analysis. Work by
Goodall et al. [5]
discussed the work process for network intrusion detection
analysts in four task
stages: monitoring the network for events; triaging an event;
analysis of an event; and,
response to an event. Thompson et al. [7] expanded Goodall et
al.’s work to include a
pre-processing stage before the monitoring stage that involves
intrusion detection
system preparation, as well as expanding the triage stage to
include activities for
determining the cause of an event and deciding if the event
should be escalated to
analysis. Although individually these studies describe different
phases of activities
4. within the cyber analytic work process, together they infer a
general cyber analysis
work process model: preparation, monitoring, detection,
analysis, and response to
network events.
Few researchers have specifically focused on situation
awareness during the cyber
analysis work process. Situation awareness is a state of
knowledge within the context
of a dynamic system, often with three stages: perception,
comprehension, and
projection [4]. Work by D’Amico et al. [2] examined the
analytic questions of
intrusion detection analysts to understand how they fused
complex data during
different stages of situation awareness. They developed a model
of situation
awareness that extended and overlapped with the model for the
cyber analysis work
process: event detection (monitoring and detection); situation
assessment (analysis);
and, threat assessment (response). Later research by D’Amico et
al. [3] added role-
based work processes that corresponded to their model of
situation awareness, such
as: triage analysis; escalation analysis; correlation analysis;
threat analysis; incident
response; and, forensic analysis.
However, there is still a general lack of information on cyber
security analysts,
their work processes, and how they establish and maintain
situation awareness.
Conducting empirical and ethnographic research with cyber
security analysts is often
5. difficult. There are a number of challenges to involving cyber
security analysts in
research, such as establishing contact with cyber security
analysts who have the time
to participate in research and are willing to share potentially
sensitive information
related to their jobs [1]. Additionally, the role of a cyber
security analyst is difficult to
define and ranges from a system administrator, intrusion
detection analyst, to an
incident responder. Cyber security analysts may take on the
same, different, or
overlapping responsibilities depending on the scope of the job
role or size of the
organization [3]. As computer networks become larger and more
complex,
understanding how cyber security analysts manage the large
amounts of information
generated by these networks and maintain awareness of the
increasing number of
events on these networks will be critical to future technology
design and
development.
A Taxonomy of Cyber Awareness Questions for the User-
Centered Design 147
3 Methodology
A combination of ethnographic research methods were used to
understand the mental
model of cyber security analysts responsible for a large
network. First, interviews and
observations were conducted to gain an understanding of
6. analysts’ work environment.
Then, a card sorting activity was conducted to understand
analysts’ conceptual models
of situation awareness on a network. Analysts in our study were
primarily responsible
for intrusion detection and not incident response.
3.1 Interviews and Observations
Interviews were conducted with six cyber security experts.
Participants had at least
one year of previous or current experience working in support
of a network operations
center as well as additional experience in cyber security. The
interviews were
open-ended with no structured topics except for the overall
purpose of the interview.
Participants were asked to discuss their experiences in cyber
security and within the
operations center. Participants were asked to talk freely, and
were only interrupted
with follow-up and clarification questions. If the topic did not
come up during the
initial discussion, participants were prompted to discuss their
experiences with cyber
situation awareness and the types of high level orientation
questions they ask
themselves during a new or ongoing event. Interviews lasted
between 45 minutes to
1.5 hours. To supplement the interviews, approximately 25
additional hours of
observations of a round-the-clock network operations center
were conducted. This
included general observations of analyst work during normal
operations, attending
operations center meetings, and observing two training
7. exercises. Participant
interruptions were minimal and participants were available to
answer questions and
discuss their activities. Observation sessions lasted between one
and four hours each.
3.2 Card Sorting Activity
Card sorting is a knowledge elicitation method that helps people
describe
relationships between and hierarchy among concepts [6]. An
open card sorting study
was conducted with 12 cyber security analysts using 44 cyber
situation awareness
questions. Participants had at least one year of previous or
current experience working
in a round-the-clock network operations center and were
primarily responsible for
network intrusion detection. Participants were not responsible
for incident response.
Cyber Awareness Questions. A list of questions was derived
from the interview and
observation data. These were questions analysts reported asking
themselves to
establish and maintain awareness of new and ongoing network
events. The
informality and similarity between questions was not edited to
preserve any nuance
that existed in question phrasing. Table 1 provides a list of the
cyber awareness
questions derived from interviews and observations of cyber
security analysts and
used in the card sorting study.
8. 148 C.L. Paul and K. Whitley
Procedure. Participants sorted the 44 cyber awareness questions
into groups that best
reflected their understanding of the questions. Once the
questions were sorted into
groups, participants labeled each group with a descriptive word
or phrase. At the end
of the activity the study moderator debriefed the participants’
work by asking them to
explain how they sorted the questions and why. Card sorting
sessions lasted between
45 minutes and one hour.
Table 1. Cyber awareness questions used in card sorting study
3.3 Analysis
An analysis of the top question pairs based on descriptive
statistics and question
co-occurrence provided insights to the most critical cyber
situation awareness
questions. Co-occurrence was calculated as the number of
participants who sorted two
questions together independent of the group the questions were
sorted into during the
card sorting activity. Graph visualization of question co-
occurrence was then used to
1. Are there more or less bad guys attacking
my network than normal?
2. Can I see the attack I know is happening?
3. Does the attack have a negative effect on
9. other business operations?
4. Does this attack matter?
5. Have I seen an attack like this before?
6. How did the bad guys get into my
network?
7. How is my network being attacked?
8. How is my network different from last
week?
9. How serious is the attack?
10. How successful was the attack?
11. Is anything different happening on my
network than normal?
12. Is anything interesting happening on my
network?
13. Is it a good day on the network?
14. Is my network configured correctly?
15. Is my network healthy?
16. Is something bad happening on the
network?
17. Is something happening on the network?
18. Is the event on my network good, bad, or
just different?
19. Is there more or less traffic on my network
than normal?
20. Is this a new attack I have not seen before?
21. What are the bad guys doing on my
network?
10. 22. What did the bad guys do?
23. What did the bad guys take?
24. What do I do about the attack?
25. What do I not see happening on my
network?
26. What does my network look like to the
bad guys?
27. What does my network look like?
28. What does the attack look like?
29. What does the event on my network
mean?
30. What happened on the network last
night?
31. What is different on my network from
last week?
32. What is happening on my network now?
33. What is happening with my network?
34. What is normal for my network?
35. What is not normal for my network?
36. What is the most important event
happening on my network?
37. What is the status of my network?
38. What malware have been detected on my
network?
39. What systems are up or down on my
network?
40. Where are the bad guys attacking from?
11. 41. Where on my network am I being
attacked?
42. Who is attacking my network?
43. Why is my network being attacked?
44. Why are computers on my network not
available?
A Taxonomy of Cyber Awareness Questions for the User-
Centered Design 149
analyze clusters of questions. Graph features such as, network
weight, clusters, and
bridges were used to identify topic areas. Content analysis of
the clusters provided
insight to topic areas that shape an analyst mental model for
cyber situation
awareness. Knowledge from the interviews and observations
provided additional
context and was integrated into the interpretation and
understanding of the results.
3.4 Limitations
Cyber security analysts are often difficult to involve in research
[1]. Only a limited
number of cyber security analysts were available to participate
in this study. Card
sorting studies can be run with a large number of participants
using quantitative
analysis methods or a small number of participants using
qualitative analysis methods
[6]. We chose to conduct a small qualitative card sorting study
12. because of the benefits
of in-depth qualitative analysis and the challenges recruiting
cyber security analysts.
To compensate for a smaller study, we triangulated our results
with graph
visualization analysis and the results from observations and
interviews.
4 Results
4.1 Top Question Pairs
There were 144 card pairs with 50% (6/12 participants) co-
occurrence representing
98% (43/44) of the questions in the study. There were 21
question pairs with 75%
(9/12 participants) co-occurrence representing 52% (23/44) of
the questions in the
study. Overall, there was good representation of all the
questions in the study within
the highest co-occurrence pairs. Table 2 provides a list of the
cyber awareness
question pairs with 75% co-occurrence. Additionally, we found
three types of
relationships between the highest co-occurrence question pairs
that we define as:
question similarity, question sets, and question order. The
question types were derived
from qualitative analysis of the question relationships.
Similarity. The first type of question pair relationship was
based on similarity (A is
the same as B) in which two questions are asking the same
thing. These questions are
essentially the same, just asked differently depending on the
situation:
13. “How is my network different from last week?”
“What is different on my network from last week?” (9/12
participants)
Set. The second type of question pair relationship was a logical
set (A and B are the
same type) in which questions are distinctly different but
related in purpose or goal:
“Is anything different happening on my network than normal?”
“Is anything interesting happening on my network?” (10/12
participants)
150 C.L. Paul and K. Whitley
Table 2. Top 75% co-occurrence (9/12 or more participants)
cyber awareness question pairs
While the framework of this question pair is very similar, e.g.,
“Is anything … on my
network?”, the use of “different” and “interesting” make the
questions distinct. Based
on the knowledge gained from the interviews and observations,
“different” is not
always “interesting” but both are equally important and asked.
Order. The third type of question pair relationship was a logical
order (A comes
before B) in which a question was a logical follow-up or a
requirement to the previous
question. For example, the order of these questions implies an
14. analytic process,
including the priority or requirement to answer certain
questions before others:
“What does the attack look like?”
“Have I seen an attack like this before?” (10/12 participants)
CO Question Question
92% Is anything interesting happening on my
network?
Is something bad happening on the network?
92% What did the bad guys take? How successful was the
attack?
83% What happened on the network last night? What is
different on my network from last
week?
83% Is something happening on the network? Is anything
interesting happening on my
network?
83% Is anything interesting happening on my
network?
Is anything different happening on my network
than normal?
83% What does the attack look like? Have I seen an attack like
this before?
75% How is my network different from last
week?
15. What is different on my network from last
week?
75% Is anything different happening on my
network than normal?
Is something happening on the network?
75% Is anything different happening on my
network than normal?
Is something bad happening on the network?
75% Is anything different happening on my
network than normal?
What is happening with my network?
75% Is anything different happening on my
network than normal?
Is there more or less traffic on my network
than normal?
75% Is something happening on the network? Is something bad
happening on the network?
75% What is happening with my network? What do I not see
happening on my network?
75% Is it a good day on the network? Is my network healthy?
75% Is it a good day on the network? What is the status of my
network?
75% What is the status of my network? What is normal for my
network?
75% What is the status of my network? What systems are up or
down on my network?
75% What does the attack look like? Who is attacking my
16. network?
75% Have I seen an attack like this before? Who is attacking my
network?
75% Does this attack matter? How serious is the attack?
75% What did the bad guys do? What did the bad guys take?
A Taxonomy of Cyber Awareness Questions for the User-
Centered Design 151
Fig. 1. Graph visualizations of question co-occurrence with
potential topics. A) Base network
with 50% (6/12 participants) co-occurrence highlighted
revealing two topic areas; B) 50%
co-occurrence network with two main topic areas; C) 75% (9/12
participants) co-occurrence
network revealing six sub-topic areas.
4.2 Graph Visualization and Content Analysis
The most interesting graph visualizations were those that were
expressed by the
highest number of participants. Figure 1 shows graph
visualizations for all question
pairs (Fig.1-A), 50% co-occurrence (Fig.1-B) that represented
questions paired by at
least half of the participants in the study, and 75% co-
occurrence (Fig.1-C) that
represented questions paired by a majority of participants.
An overlay of the 50% co-occurrence question pairs on the base
network
visualization showed two question co-occurrence clusters,
potentially revealing two
main topic areas (Fig.1-A). A visualization of the 50% co-
17. occurrence question pairs
(Fig.1-B) showed the two clusters found in the base network
(Fig.1-A) as well as
graph features such as sub-clusters and bridges that identify
possible sub-clusters. A
visualization of the 75% co-occurrence question pairs (Fig.1-C)
showed six small
clusters that are a sub-set of the two 50% co-occurrence clusters
(Fig.1-B).
Content analysis of the questions in the two clusters (Fig.1-B)
revealed potential
topics in Event Detection (T1) and Event Orientation (T2).
Further content analysis of
the six clusters from the 75%+ co-occurrence question pairs
(Fig.1-C) revealed
potential sub-topics such as Network Baseline (T1.1), Change
Detection (T1.2),
Network Activity (T1.3), Event Identification (T2.1), Mission
Impact (T2.2), and
Damage Assessment (T2.3).
Further analysis of different levels of co-occurrence
visualization disambiguated
the relationships between question pairs that were not clearly
from one of the six
75%+ co-occurrence clusters. For example, several additional
questions can be
classified in one of the six topics by examining the visualization
for the 67% co-
occurrence question pairs (8/12 participants) and 58% co-
occurrence question pairs
(7/12 participants). These additional questions are included in
Table 3 taxonomy of
cyber awareness questions.
18. A) B) C)
152 C.L. Paul and K. Whitley
5 Taxonomy of Cyber Awareness Questions
The presented taxonomy of cyber awareness questions offers
insights into different
stages of cyber situation awareness (Table 3). The categories
were derived from the
results of the study and previous work in cyber situation
awareness. The questions
were organized into categories based on their co-occurrence
score from our study,
ranging from 58% (7/12 participants) to 92% (11/12
participants) co-occurrence.
Event Detection. This category contains questions that analysts
ask prior and during
initial awareness of a network event. In Event Detection, these
questions roughly
align with the perception phase of situation awareness.
Network Baseline. A network baseline is a model or snapshot of
the network when it
functioning in a “normal” state, in which “normal” is often the
best approximation of
healthy, acceptable operation. Comparison to their mental
baseline was a common
way analysts in this study articulated how their analytic needs
precede cyber events.
Change Detection. Change detection is the ability to compare
19. states of the network to
identify differences and trends. The concept differs only
slightly from network Base-
line in that, here, analysts focus on the comparison between two
network states.
Network Activity. Network activity reflects a shift from
“normal” to “not normal”
network activity that acts as a cue for the analyst to narrow his
attention for in-depth
analysis. These questions relate closest to the situation
awareness concept of percep-
tion as well as allude to the transition between Change
Detection and Event Identifi-
cation.
Event Orientation. This category contains questions that
analysts ask and are most
closely aligned with the comprehension stage of situation
analysis. In Event
Orientation, analysts are working to maximize insight into an
identified cyber event.
Identification. Identification is the recognition that a subset of
network activity
warrants analytic attention. This category is the detailed
analysis of an event to
identify who, what, when, where, and why and attack is
happening and to possibly
link the activity to familiar threats.
Mission Impact. Mission impact is analysis to prioritize the
importance of an
identified threat. Analysts must judge the severity of the threat
to business operations,
such as personnel necessary to respond to the threat, to help
20. determine how to
distribute limited resources for investigating and responding to
the threat.
Damage Assessment. Damage assessment is analysis to inform a
response to an
identified threat. These questions differ somewhat from Mission
Impact; here, the
goal is to understand the full effects of the attack on the internal
network.
A Taxonomy of Cyber Awareness Questions for the User-
Centered Design 153
Table 3. Taxonomy of Cyber Awareness Questions for Cyber
Situation Awareness
Event Detection Event Orientation
Network Baseline Event Identification
• Is it a good day on the network?
• Is my network configured correctly?
• Is my network healthy?
• What does my network look like?
• What is happening on the network now?
• What is normal for my network?
• What is not normal for my network?
• What is the status of my network?
• What systems are up or down on my
network?
21. • Have I seen an attack like this before?
• Is this a new attack I have not seen before?
• How is my network being attacked?
• What are the bad guys doing on my
network?
• What does the attack look like?
• Where on my network am I being attacked?
• Who is attacking my network?
• Where are they bad guys attacking from?
• Why is my network being attacked?
Change Detection Mission Impact
• How is my network different from last week?
• What happened on the network last night?
• What is different on my network from last
week?
• Does this attack matter?
• How serious is the attack?
• What do I do about the attack?
Network Activity Damage Assessment
• Is anything different happening on my
network than normal?
• Is anything interesting happening on my
network?
• Is something bad happening on the network?
• Is something happening on the network?
• Is the event on my network good, bad, or
22. just different?
• Is there more or less traffic on my network
than normal?
• What do I not see happening on my
network?
• What does the event on my network mean?
• What is happening with my network?
• What is the most important event happening
on my network?
• Why are computers on my network not
available?
• Does the attack have a negative effect on
other business operations?
• How successful was the attack?
• What did the bad guys do?
• What did the bad guys take?
154 C.L. Paul and K. Whitley
6 Conclusion
In this paper we utilized user-centered and ethnographic
research methods to explore
and understand the mental model of cyber security analysts
responsible for a large
network. Our results lead to the contribution of a taxonomy of
23. cyber awareness
questions that describes a set of questions analysts ask
themselves while they establish
and maintain situation awareness during a network event.
This taxonomy provides valuable information about the cyber
security analyst and
will support the user-centered design and development of cyber
situation awareness
technology. For example, the taxonomy could be used during
the design of cyber
situation awareness visualization. Good design is especially
important for large-scale
visualizations that display large amounts of data. This taxonomy
of cyber awareness
questions would help inform the design of visualizations that
would help analysts
better establish and maintain situation awareness of a large
computer network.
However, this study only addresses part of the picture. Our
taxonomy does not
include questions related to incident response while other
models of cyber situation
awareness do. The cyber security analysts in our study were
specialized and only
responsible for intrusion detection related activities as opposed
to other research that
studied generalists (e.g., [8]) or specific types of cyber security
analysts (e.g., [2, 5,
7]). This may explain the lack of incident response topic area
and questions in our
taxonomy. Additional work in this cyber situation awareness
will contribute
additional questions and topic areas to the taxonomy.
24. References
1. Botta, D., Werlinger, R., Gagné, A., Beznosov, K., Iverson,
L., Fels, S., Fisher, B.: Towards
Understanding IT Security Professionals and Their Tools. In:
ACM Symposium on Usable
Privacy and Security, pp. 100–111 (2007)
2. D’Amico, A., Whitley, K., Tesone, D., O’Brien, B., Roth, E.:
Achieving Cyber Defense
Situational Awareness: A Cognitive Task Analysis of
Information Assurance Analysts. In:
Human Factors and Ergonomics Society Annual Meeting, pp.
229–233 (2005)
3. D’Amico, A., Whitley, K.: The Real Work of Computer
Network Defense Analysts. In:
Symposium on Visualizations for Computer Security, pp. 19–37
(2007)
4. Endsley, M.R.: Toward a Theory of Situation Awareness in
Dynamic Systems. Human Fac-
tors 37(1), 32–64 (1995)
5. Goodall, J.R., Lutters, W.G., Komlodi, A.: Developing
expertise for network intrusion de-
tection. Information Technology & People 22(2), 92–108 (2009)
6. Hudson, W.: Card Sorting. In: Soegaard, M., Dam, R. (eds.)
The Encyclopedia of Human-
Computer Interaction, 2nd edn. The Interaction Design
Foundation, Aarhus (2013)
7. Thompson, R.S., Rantanen, E.M., Yurcik, W.: Network
Intrusion Detection Cognitive Task
Analysis: Textual and Visual Tool Usage and
25. Recommendations. In: Human Factors and
Ergonomics Society Annual Meeting, pp. 669–673 (2006)
8. Werlinger, R., Muldner, K., Kawkey, K., Beznosov, K.:
Preparation, detection, and analy-
sis: the diagnostic work of IT security incident response.
Information Management & Com-
puter Security 18(1), 26–42 (2010)
Objective: To determine the effects of an
adversary’s behavior on the defender’s accurate and
timely detection of network threats.
Background: Cyber attacks cause major work
disruption. It is important to understand how a defender’s
behavior (experience and tolerance to threats), as well
as adversarial behavior (attack strategy), might impact
the detection of threats. In this article, we use cognitive
modeling to make predictions regarding these factors.
Method: Different model types representing a
defender, based on Instance-Based Learning Theory
(IBLT), faced different adversarial behaviors. A defender’s
model was defined by experience of threats: threat-prone
(90% threats and 10% nonthreats) and nonthreat-prone
(10% threats and 90% nonthreats); and different tolerance
levels to threats: risk-averse (model declares a cyber attack
after perceiving one threat out of eight total) and risk-
seeking (model declares a cyber attack after perceiving
seven threats out of eight total). Adversarial behavior is
simulated by considering different attack strategies: patient
(threats occur late) and impatient (threats occur early).
26. Results: For an impatient strategy, risk-averse
models with threat-prone experiences show improved
detection compared with risk-seeking models with
nonthreat-prone experiences; however, the same is not
true for a patient strategy.
Conclusions: Based upon model predictions,
a defender’s prior threat experiences and his or her
tolerance to threats are likely to predict detection
accuracy; but considering the nature of adversarial
behavior is also important.
Application: Decision-support tools that consider
the role of a defender’s experience and tolerance to
threats along with the nature of adversarial behavior are
likely to improve a defender’s overall threat detection.
Keywords: cyber situation awareness, Instance-Based
Learning Theory, defender, adversarial behavior, experi-
ences, tolerance
Cyber attacks are the disruption in the nor-
mal functioning of computers and the loss of
private information in a network due to mali-
cious network events (threats), and they are
becoming widespread. In the United Kingdom,
organizers of the London 2012 Olympic Games
believe that there is an increased danger of
cyber attacks that could seriously undermine
the technical network supporting everything,
from recording world records to relaying results
to commentators at the Games (Gibson, 2011).
With the prevalence of “Anonymous” and
“LulzSec” hacking groups and other threats to
corporate and national security, guarding
against cyber attacks is becoming a significant
27. part of IT governance, especially because most
government agencies and private companies
have moved to online systems (Sideman, 2011).
Recently, President Barack Obama declared
that the “cyber threat is one of the most serious
economic and national security challenges we
face as a nation” (White House, 2011).
According to his office, the nation’s cyber-
security strategy is twofold: (1) to improve our
resilience to cyber incidents and (2) to reduce
the cyber threat. To meet these goals, the role of
the security analyst (called “defender”
onwards), a human decision maker who is in
charge of protecting the online infrastructure of
a corporate network from random or organized
cyber attacks, is indispensable (Jajodia, Liu,
Swarup, & Wang, 2010). The defender protects
a corporate network by identifying, as early and
accurately as possible, threats and nonthreats
during cyber attacks.
In this article, we derive predictions about
the influence of a simulated defender’s experi-
ence and his or her tolerance to threats on threat
detection accuracy for different simulated
adversarial behaviors using a computational
model. Adversarial behaviors are exhibited
through different simulated attack strategies
that differ in the timing of threat occurrence in a
sequence of network events. We simulate a
defender’s awareness process through a compu-
tational model of dynamic decision making
Address correspondence to Varun Dutt, School of Com-
puting and Electrical Engineering, School of Humanities and
Social Sciences, Indian Institute of Technology, Mandi, PWD
29. the attack; situation comprehension: under-
standing why and how the current situation is
caused and what is its impact; and situation pro-
jection: determining the expectations of a future
attack, its location, and its impact (Jajodia et al.,
2010; Tadda, Salerno, Boulware, Hinman, &
Gorton, 2006).
During a cyber attack, there could be both
malicious network events (threats) and benign
network events (nonthreats) occurring in a
sequence. Threats are generated by attackers,
while nonthreats are generated by friendly users
of the network. In order to accurately and timely
detect cyber attacks, a defender relies on highly
sophisticated technologies that aid in the detec-
tion of threats (Jajodia et al., 2010). One of
these cyber technologies is called an intrusion
detection system (IDS), a program that alerts
defenders of possible network threats. The IDS
is not a perfect technology, however, and its
predictions have both false positives and false
negatives (PSU, 2011). Although there is ample
current research on developing these technolo-
gies, and on evaluating and improving their
efficiency, the role of the defender behavior,
such as the defender’s experience and tolerance
to threats, is understudied in the cyber-security
literature (Gardner, 1987; Johnson-Laird, 2006;
PSU, 2011). In addition, it is likely that the
nature of adversarial behavior also influences
the defender’s cyberSA (Gonzalez, 2012). One
characteristic of adversarial behavior is the
attacker’s strategy regarding the timing of
threats during a cyber attack: An impatient
attacker might inject all threats in the beginning
30. of a sequence of network events; however, a
patient attacker is likely to delay this injection
to the very end of a sequence (Jajodia et al.,
2010). For both these strategies, there is pre-
vailing uncertainty in terms of exactly when
threats might appear in a cyber attack. Thus, it
is important for the defender to develop a timely
and accurate threat perception to be able to
detect a cyber attack. Thus, both the nature of
the defender’s and adversary’s behaviors may
greatly influence the defender’s cyberSA.
Due to the high demand for defenders, their
lack of availability for laboratory experiments,
and the difficulty of studying real-world cyber-
security events, an important alternative used to
study cyberSA is computational cognitive mod-
eling. A cognitive model is a representation of
the cognitive processes and mechanisms
involved in performing a task. An important
advantage of computational cognitive modeling
is to generate predictions about human behavior
in different tasks without first spending time
and resources in running large laboratory exper-
iments involving participants. Of course, the
model used would need to be validated against
human data to generate accurate predictions
with high confidence. The cognitive model of
cyberSA that we present here relies on the
Instance-Based Learning Theory, a theory of
decisions from experience in dynamic environ-
ments that has demonstrated to be comprehen-
sive and robust across multiple decision making
tasks (Gonzalez & Dutt, 2011; Gonzalez, Dutt,
& Lejarraga, 2011; Gonzalez et al., 2003;
31. Lejarraga, Dutt, & Gonzalez, 2010).
In the next section, we explain the role that a
defender’s experience and tolerance to threats
and the role that the nature of adversarial behav-
ior might play in determining a defender’s
cyberSA. Then, we present IBLT and the par-
ticular implementation of a cognitive model of
cyberSA. This presentation is followed by a
simulation experiment where the model’s expe-
rience, tolerance, and the exposure to different
adversarial behaviors were manipulated. Using
the model, we predict two measures of cyberSA:
timing and accuracy. Finally, we conclude with
the discussion of the results and their implica-
tions to the design of training and decision-
support tools for improving a defender’s detec-
tion of cyber attacks.
Cyber SA, AttACker, And defender 607
ROLE OF EXPERIENCE, TOLERANCE,
AND ADVERSARIAL BEHAVIOR
A defender’s cyberSA is likely to be influ-
enced by at least three factors: the mix of threat
and nonthreat experiences stored in memory;
the defender’s tolerance to threats, namely, how
many network events a defender perceives as
threats before deciding that these events repre-
sent a cyber attack; and the adversarial behavior
(i.e., an attacker’s strategy). The adversarial
behavior is different from the first two factors.
First, actions from the attacker are external or
32. outside of the defender’s control. Second, pre-
viously encountered adversarial behaviors
might influence the defender’s current experi-
ences and tolerance to threats.
Prior research indicates that the defender’s
cyberSA is likely a function of experience with
cyber attacks (Dutt, Ahn, & Gonzalez, 2011;
Jajodia et al., 2010) and tolerance to threats
(Dutt & Gonzalez, in press; McCumber, 2004;
Salter, Saydjari, Schneier, & Wallner, 1998).
For example, Dutt, Ahn, et al. (2011) and Dutt
and Gonzalez (in press) have provided initial
predictions about a simulated defender’s cyber
SA according to its experience and tolerance.
Dutt and Gonzalez (in press) created a cognitive
model of a defender’s cyberSA based upon
IBLT and populated the model’s memory with
threat and nonthreat experiences. The model’s
tolerance was determined by the number of
events perceived as threats before it declared
the sequence of network events to be a cyber
attack. Accordingly, a model with a greater pro-
portion of threat experiences is likely to be
more accurate and timely in detecting threats
compared with one with a smaller proportion of
such experiences. That is because, according to
IBLT, possessing recent and frequent past expe-
riences of network threats also increases the
model’s opportunity to remember and recall
these threats with ease in novel situations.
However, it is still not clear how results in Dutt
and Gonzalez (in press) were impacted by dif-
ferent adversarial behaviors.
Furthermore, recent research in judgment
33. and decision making has discussed how experi-
encing outcomes, gained by sampling alterna-
tives in a decision environment, determines our
real decision choices after sampling (Gonzalez
& Dutt, 2011; Hertwig, Barron, Weber, & Erev,
2004; Lejarraga et al., 2010). For example, hav-
ing a greater proportion of negative experiences
or outcomes in memory for an activity (e.g.,
about threats) makes a decision maker (e.g.,
defender) cautious about said activity (e.g., cau-
tious about threats) (Hertwig et al., 2004;
Lejarraga et al., 2010).
Prior research has also predicted that a
defender’s tolerance to threats is likely to influ-
ence his or her cyberSA. For example, Salter,
Saydjari, Schneier, and Wallner (1998) high-
lighted the importance of understanding both
the attacker and the defender’s tolerance, and
according to Dutt and Gonzalez (in press), a
defender is likely to be more accurate when his
or her tolerance is low rather than high. That is
because possessing a low tolerance is likely to
cause the defender to declare cyber attacks very
early on, which may make a defender more
timely and accurate in situations actually
involving early threats. Although possessing
low tolerance might be perceived as costly to an
organization’s productivity, it is expected to
boost the organization’s productivity.
The studies discussed previously provided
interesting predictions about a simulated
defender’s experience and tolerance. However,
these studies did not consider the role of adver-
34. sarial behavior (i.e., attacker’s strategies) and
the interactions between a defender’s behavior
and an adversary’s behavior. Depending upon
adversarial behavior, threats within a network
might occur at different times and their timing
is likely to be uncertain (Jajodia et al., 2010).
For example, an impatient attacker could exe-
cute all threats very early on in a cyber attack,
whereas a patient attacker might decide to delay
the attack and thus threats would appear late in
the sequence of network events. There is evi-
dence that the recency of information in an
observed sequence influences people’s deci-
sions when they encounter this information at
different times (early or late) in the sequence
(Dutt, Yu, & Gonzalez, 2011; Hogarth &
Einhorn, 1992). The influence of recency is
likely to be driven by people’s limited working
memory capacity (Cowan, 2001), especially
when making decisions from experience in
emergency situations (Dutt, Cassenti, &
608 June 2013 - Human Factors
Gonzalez, 2011; Dutt, Ahn, et al., 2011;
Gonzalez, 2012). Given the influence of
recency, we expect a defender with a greater
proportion of threat experiences and a low tol-
erance to be more accurate and timely against
an impatient attack strategy compared with a
defender with fewer threat experiences and a
high tolerance. However, we do not expect that
to be the case for a patient attack strategy. That
is because according to IBLT, a model (repre-
35. senting a defender) will make detection deci-
sions by recalling similar experiences from
memory. When the model has a greater propor-
tion of threat experiences in memory, it is more
likely to recall these experiences early on, mak-
ing it accurate if threats occur early in an attack
(i.e., generated by an impatient strategy). The
activated threat experiences would be recalled
faster from memory and would also increase the
likelihood that the accumulation of evidence for
threats exceeds the model’s low tolerance. By
the same logic, when threats occur late in cyber
attacks (i.e., generated by a patient strategy),
the situation becomes detrimental to the accu-
racy and timeliness of a model that possesses
many threat experiences in memory and has a
low tolerance. In summary, a model’s experi-
ence of threats, its tolerance to threats, and an
attack strategy may limit or enhance the mod-
el’s cyberSA. These model predictions generate
insights for the expected behavior of a defender
in such situations.
CYBER INFRASTRUCTURE
AND CYBER ATTACKS
The cyber infrastructure in a corporate net-
work may consist of different types of servers
and multiple layers of firewalls. We used a
simplified network configuration consisting of
a webserver, a fileserver, and two firewalls (Ou,
Boyer, & McQueen, 2006; Xie, Li, Ou, Liu, &
Levy, 2010). An external firewall (“firewall 1”
in Figure 1) controls the traffic between the
Internet and the Demilitarized Zone (DMZ; a
36. subnetwork that separates the Internet from the
company’s internal LAN network). Another
firewall (“firewall 2” in Figure 1) controls the
flow of traffic between the webserver and the
fileserver (i.e., the company’s internal LAN
network). The webserver resides behind the
first firewall in the DMZ (see Figure 1). It
handles outside customer interactions on a
company’s Web site. The fileserver resides
behind the second firewall and serves as repos-
itory accessed by workstations used by corpo-
rate employees (internal users) to do their daily
operations. These operations are made possible
by enabling workstations to run executable files
from the fileserver.
Generally, an attacker is identified as a com-
puter on the Internet that is trying to gain access
to the internal corporate servers. For this cyber-
infrastructure, attackers follow a pattern of
“island-hopping” attack (Jajodia et al., 2010, p.
30), where the webserver is compromised first,
and then it is used to originate attacks on the
fileserver and other company workstations.
A model of the defender, based upon IBLT, is
exposed to different island-hopping attack
sequences (depending upon the two adversarial
timing strategies). Each attack sequence is com-
posed of 25 network events (a combination of
both threats and nonthreats), whose nature
(threat or nonthreat) is not known to the model.
However, the model is able to observe alerts
that correspond to some network events (that
Figure 1. A typical cyber-infrastructure in a
37. corporate network. The attacker uses a computer on
the Internet and tries to gain access to the company’s
workstations through the company’s webserver and
fileserver.
Source. Adapted from Xie, Li, Ou, Liu, and Levy
(2010).
Cyber SA, AttACker, And defender 609
are regarded as threats) generated from the
intrusion-detection system (Jajodia et al., 2010).
Out of 25 events, there are 8 predefined threats
that are initiated by an attacker (the rest of the
events are initiated by benign users). The model
does not know which events are generated by
the attacker and which are generated by corpo-
rate employees. By perceiving network events
in a sequence as threats or nonthreats, the model
needs to identify, as early and accurately as pos-
sible, whether the sequence constitutes a cyber
attack. In this cyber-infrastructure, we repre-
sented adversarial behavior by presenting event
sequences with different timings for the 8
threats: an impatient strategy, where the 8
threats occur at the beginning of the sequence,
and a patient strategy, where the 8 threats occur
at the end of the sequence.
INSTANCE-BASED LEARNING
MODEL OF DEFENDER’S CYBER SA
IBLT is a theory of how people make deci-
sions from experience in dynamic environments
(Gonzalez et al., 2003). Computational models
38. based on IBLT have been shown to generate
accurate predictions of human behavior in many
dynamic decision-making situations similar to
those faced by defenders (Dutt, Ahn, et al., 2011;
Dutt, Cassenti, et al., 2011; Dutt & Gonzalez, in
press; Gonzalez & Dutt, 2011; Gonzalez et al.,
2011). IBLT proposes that every decision situa-
tion is represented as an experience called an
instance that is stored in memory. Each instance
in memory is composed of two parts: situation
(S) (the knowledge of attributes that describe an
event), a decision (D) (the action taken in such
situation), and utility (U) (a measure of expected
result of a decision that is to be made for an
event). For a situation involving securing a net-
work from threats, the situation attributes are
those that can discriminate between threat and
nonthreat events: the IP address of a computer
(webserver, fileserver, or workstation, etc.)
where the event occurred, the directory location
in which the event occurred, whether the IDS
raised an alert corresponding to the event, and
whether the operation carried out as part of the
event (e.g., a file execution) by a user of the net-
work (which could be an attacker) succeeded or
failed. In the IBL model of a defender, an
instance’s S part refers to the situation attributes
defined previously, and the U slot refers to the
expectation in memory that a network event is a
threat or not. For example, an instance could be
defined as [webserver, c:, malicious code, suc-
cess; threat], where “webserver,” “c:,” “mali-
cious code,” and “success” constitute the
instance’s S part and “threat” is the instance’s U
part (the decision being binary: threat or not, is
39. not included in this model).
IBLT proposes that a decision maker’s men-
tal process is composed of five mental phases:
recognition, judgment, choice, execution, and
feedback. These five decision phases represent
a complete learning cycle where the theory
explains how knowledge in instances is
acquired, reused, and reinforced by human
decision makers. Because the focus of this arti-
cle is on cyberSA rather than on decision mak-
ing, we only focus on the recognition, judgment,
and choice phases in IBLT (not the execution
and feedback phases). Among these, the IBLT’s
recognition and judgment phases accomplish
the recognition and comprehension stages in
cyberSA, and IBLT’s choice phase is used to
make a decision in different situations after rec-
ognition and comprehension has occurred. To
calculate the decision, the theory relies on
memory mechanisms such as frequency and
recency. The formulation of these mechanisms
has been taken from a popular cognitive archi-
tecture, ACT-R (Anderson & Lebiere, 1998,
2003) (the model reported here uses a simpli-
fied version of the activation equation in
ACT-R).
Figure 2 shows an example of the processing
of network events through the three phases in
IBLT. The IBLT’s process starts with the recog-
nition phase in search for decision alternatives to
classify a sequence of network events as a cyber
attack or not. During recognition, an instance
with the highest activation and closest similarity
to the network event is retrieved from memory
40. and is used to make this classification. For exam-
ple, the first instance in memory matches the first
network event because it is most similar to the
event, and thus, it is retrieved from memory for
further processing. Next, in the judgment phase,
the retrieved instance is used to evaluate whether
the network event currently being evaluated is
610 June 2013 - Human Factors
perceived as a threat or not. As seen in Figure 2,
this evaluation is based upon the U part of the
retrieved instance (as described previously, the
instance’s U part indicates the expectation
whether the network event is a “threat” or “non-
threat”). Based upon the U part, a “threat-evi-
dence” counter is incremented by one unit if the
network event is classified as a threat; otherwise
not. The threat-evidence counter represents the
accumulation of evidence for threats and in a
new network scenario, it starts at 0. For the first
network event in Figure 2, the retrieved instance’s
U part indicates a nonthreat, so the threat-evi-
dence counter is not incremented and remains at
0. For the second network event, however, the
retrieved instance’s U part indicates a threat, so
the counter is incremented by 1.
In the choice phase, the model decides
whether to classify a set of previously evaluated
network events in the sequence as part of a cyber
attack or to keep accumulating more evidence by
further observing network events. In IBLT, this
classification is determined by the “necessity
41. level,” which represents a satisficing mechanism
used to stop search of the environment and be
“satisfied” with the current evidence (e.g., the
satisficing strategy; Simon & March, 1958). This
necessity level is the mechanism used to simulate
defenders of different tolerance levels. Tolerance
is a free parameter that represents the number of
network events perceived as threats before the
model classifies the sequence as a cyber attack.
Therefore, when the threat-evidence counter
becomes equal to the tolerance parameter’s
value, the model classifies the sequence as a
cyber attack. For example, for the first network
event in Figure 2, the threat-evidence counter
(=0) is less than the tolerance (=1) and the model
continues to evaluate the next network event. For
the second network event, however, the threat-
evidence counter (=1) becomes equal to the tol-
erance (=1) and the model stops and classifies
the entire sequence of events as a cyber attack.
Translated to actual network environments, the
convergence of the threat-evidence counter with
tolerance would mean stopping online operations
in the company. Otherwise, the model will keep
observing more events and let the online opera-
tions continue undisrupted if it has not classified
a sequence as a cyber attack.
Figure 2. The processing of network events by the Instance-
Based Learning model. The model uses
recognition, judgment, and choice phases for each observed
event in the network, and it decides to stop
when the threat-evidence counter equals the tolerance
parameter.
42. Cyber SA, AttACker, And defender 611
An instance is retrieved in the recognition
phase from memory according to an activation
mechanism (Gonzalez et al., 2003; Lejarraga
et al., 2010). The activation of an instance i in
memory is defined using a simplified version of
ACT-R’s activation equation:
A
i
= B
i
+ Sim
i
+ ε
i
,
where i refers to the ith instance that is pre-pop-
ulated in memory, and i = 1, 2, . . . constitutes
the total number of pre-populated instances in
memory; B
i
is the base-level learning mecha-
nism and reflects both the recency and fre-
quency of use for the ith instance since the time
it was created; and ε
43. i
is the noise value that is
computed and added to an instance i’s activa-
tion at the time of its retrieval attempt from
memory.
The B
i
equation is given by
B t ti i
d
t ti
= −
−
∈ −
∑ln ( )
{ ,..., }1 1
.
44. In this equation, the frequency effect is pro-
vided by t – 1, the number of retrievals of the ith
instance from memory in the past. The recency
effect is provided by t – t
i
, the time since the tth
past retrieval of the ith instance (in Equation 2,
t denotes the current event number in the sce-
nario). The d is the decay parameter and has a
default value of 0.5 in the ACT-R architecture,
and this is the value we assume for the IBL
model.
Sim
i
refers to the similarity between the attri-
butes of the situation and the attributes of the ith
instance. Sim
i
is defined as
Sim P * Ml li
k
i l
=
=∑ 1
The P Ml lil
k
45. *
=∑ 1 is the similarity compo-
nent and represents the mismatch between a
situation’s attributes and the situation (S) part of
an instance i in memory. The k is the total num-
ber of attributes for a situation event that are
used to retrieve the instance i from memory.
The value of k = 4 as there are four attributes
that characterize a situation in the network. As
mentioned previously, these attributes are IP,
directory, alert, and operation in an event. The
match scale (P
l
) reflects the amount of weight-
ing given to the similarity between an instance
i’s situation part l and the corresponding situa-
tion event’s attribute. P
l
is generally a negative
integer with a common value of –1.0 for all situ-
ation slots k of an instance i, and we assume this
value for the P
l
. The M
li
or match similarities
represents the similarity between the value l of
46. a situation event’s attribute and the value in the
corresponding situation part of an instance i in
memory. Typically, M
li
is defined using a
squared distance between the situation event’s
attributes and the corresponding instance’s situ-
ation slots (Shepard, 1962). Thus, M
li
is equal to
the sum of squared differences between a situa-
tion event’s attributes and the corresponding
instance’s S part. In order to find the sum of
these squared differences, the situation events’
attributes and the values in the corresponding S
part of instances in memory were coded using
numeric codes. Table 1 shows the codes
assigned to the S part of instances and the situa-
tion events’ attributes.
The noise value ε
i
(Anderson & Lebiere,
1998; Lejarraga et al., 2010) is defined as
ε
η
η
i
47. i
i
s= ×
1
,
where η
i
is a random draw from a uniform dis-
tribution bounded in [0, 1] for an instance i in
memory. We set the parameter s in an IBL
model to make it a part of the activation equa-
tion (Equation 1). The s parameter has a default
value of 0.25 in the ACT-R architecture, and we
assume this default value in the IBL model. We
use default values of d and s parameters.
EXECUTION AND RESULTS OF
THE IBL MODEL
The IBL model representing a simulated
defender was created using Matlab. The model
runs over different network event sequences that
48. represent the different timing strategies of attack
as described previously. All sequences contained
25 network events. The model’s memory was
prepopulated with instances representing defend-
(1)
(2)
(3)
(4)
612 June 2013 - Human Factors
ers with different experiences, and the model used
different levels of tolerance. The IBL model used
Equations 1, 2, and 3 to retrieve an instance with
the highest activation and made a decision about
whether an event is a threat or nonthreat. The
proportion of threat and nonthreat instances in the
model’s prepopulated memory classified it into
two kinds: threat-prone model, whose memory
consisted of 90% of threat instances and 10% of
nonthreat instances for each network event in the
sequence, and nonthreat-prone model, whose
memory consisted of 10% of threat instances and
90% of nonthreat instances for each situation
event in the sequence. Although we assumed that
the 90% and 10% classification for threat-prone
and nonthreat-prone models as extreme values,
one could readily change this assumption to other
intermediate values (between 90% and 10%) in
the model.
49. After an instance was retrieved from memory,
a decision was made to classify a sequence as a
cyber attack or not depending upon the tolerance
and the value of the threat-evidence counter. The
tolerance level would classify the model into two
kinds: risk-averse (model declares a cyber attack
after perceiving one threat) and risk-seeking
(model declares a cyber attack after perceiving
seven threats). Based upon the aforementioned
manipulations, we created four simulated model
types: nonthreat-prone and risk-seeking, non-
threat-prone and risk-averse, threat-prone and
risk-seeking, and threat-prone and risk-averse.
Furthermore, adversarial behavior was simulated
by considering different attack strategies about the
timing of threats: patient (the last eight events in
an event sequence were actual threats) and impa-
tient (the first eight events in an event sequence
were actual threats).
We had initially assumed 1,500 simulations
in the model; however, we found that by reduc-
ing the number of simulations to one third
(=500) of that, there was a minuscule change in
values of our dependent variables. Thus, we
decided to use only 500 simulations of the
model as they were sufficient for generating
stable model results. We ran 500 simulations
(each simulation consisting of 25 network
events), and the model’s cyberSA was evalu-
ated using its accuracy and detection timing in
eight groups defined by: experience (threat-
prone and nonthreat-prone), tolerance (risk-
averse and risk-seeking), and attacker’s strategy
50. (impatient and patient). Accuracy was evalu-
ated by computing the d′ (= Z[hit rate] – Z[false-
alarm rate]), hit rate (= hits/[hits + misses]), and
false-alarm rate (= false alarms/[false alarms +
correct rejections]) (Wickens, 2001) over the
course of 25 network events and averaged
across the 500 simulations. The model’s deci-
sion for each network event was marked as a hit
if an instance with its U slot indicating a threat
was retrieved from memory for an actual threat
event in the sequence. Similarly, the model’s
decision was marked as a false alarm if an
instance with its U slot indicating a threat was
retrieved from memory for an actual nonthreat
event in the sequence. Hits and false alarms
were calculated for all events that the model
observed before it declared a cyber attack and
stopped or when all the 25 events had occurred
(whichever came first). In addition to the hit
rate, false-alarm rate, and d′, we also calculated
the model’s accuracy of stopping in different
simulations of the scenario. Across the 500 sim-
ulations of the scenario, the simulation accu-
racy was defined as Number of scenarios with a
hit/(Number of scenarios with a hit + Number
of scenarios with a miss). The model’s decision
for each simulated scenario (out of 500) was
marked as a “scenario with a hit” if the model
stopped before observing all the 25 events; oth-
erwise, if the model observed all the 25 events
in the scenario, its decision was marked as a
TABLE 1: The Coded Values in the S Part of
Instances in Memory and Attributes of a Situation
Event
51. Attributes Values Codes
IP Webserver 1
Fileserver 2
Workstation 3
Directory Missing value –100a
File X 1
Alert Present 1
Absent 0
Operation Successful 1
Unsuccessful 0
aWhen the value of an attribute was missing, then the
attribute was not included in the calculation of similarity.
Cyber SA, AttACker, And defender 613
“scenario with a miss.” As both the patient and
impatient scenarios were only attack scenarios
where an attacker attacked the network, we
were only able to compute the simulation accu-
racy in terms of scenarios with a hit or a miss.
Furthermore, detection timing was calculated in
each simulation as the “proportion of attack
steps,” defined as the percentage of threat
events out of a total 8 that have occurred after
which the model classifies the event sequence
as a cyber attack and stops. Therefore, higher
percentages of attacks steps would indicate the
model to be less timely in detecting cyber
attacks. Again, we expected a threat-prone and
risk-averse model to be more accurate and
timely against an impatient attack strategy com-
52. pared with a nonthreat-prone and risk-seeking
model; however, we don’t expect that to be the
case for a patient attack strategy.
RESULTS
Accuracy
As expected, the attack strategy interacted
with the model’s type (experience and tolerance)
to influence its accuracy. This interaction is illus-
trated in Figure 3, which shows averages of d′,
hit rate, and false-alarm rate, across the 500
simulated participants in each of the eight groups.
For an impatient strategy, the d′ was higher for
threat-prone models than the nonthreat-prone
models, regardless of the risk tolerance (threat-
prone risk-seeking: M = 2.26, SE = .05; threat-
prone risk-averse: M = 2.71, SE = .05;
nonthreat-prone risk-seeking: M = −0.06, SE =
.05; nonthreat-prone risk-averse: M = 0.33, SE =
.05). However, for the patient strategy the d′ was
higher for the nonthreat-prone models than for
the threat-prone models, again regardless of the
risk tolerance (threat-prone risk-seeking: M =
−2.63, SE = .05; threat-prone risk-averse: M =
−2.63, SE = .05; nonthreat-prone risk-seeking: M
= −0.29, SE = .05; nonthreat-prone risk-averse:
M = −0.35, SE = .05). These results suggest that
the nonthreat-prone model is unable to recognize
threats from nonthreats for both patient and
impatient attack strategies. In all cases of the
nonthreat-prone models, the hit rates and false-
alarm rates are very low. Similarly, in the patient
strategy, the accuracy (d′) is very low. The mod-
els show very high false-alarm rates and very
low hit rates. Also, as expected it is only when
53. the attack strategy is impatient and the model has
a threat-prone and risk-averse disposition that
the d′ is the highest.
Furthermore, we compared the effect of
model type and attack strategy on the model’s
simulation accuracy of stopping. The simula-
tion accuracy was higher for the impatient strat-
egy (M = 60.70%, SE = .01) compared with the
patient strategy (M = 50.60%, SE = .01). Also,
the simulation accuracy was higher was for
threat-prone models (96.95%) compared to
nonthreat-prone models (16.75%) and risk-
averse models (59.55%) compared to risk-seek-
ing models (54.15%). However, the attack
strategy did not interact with the model type to
influence the simulation accuracy. Thus, irre-
spective of the attack strategy (patient or impa-
tient), the threat-prone and risk-averse models
performed more accurately compared to non-
threat-prone and risk-seeking models.
Timing
Again as expected, the attack strategy inter-
acted with the model type (experience and toler-
ance) to influence the proportion of attack steps.
Figure 4 shows the nature of this interaction
across 500 simulated participants in each of the
eight groups. For the impatient strategy, it mat-
tered whether models were threat- or nonthreat-
prone, as well as whether they were risk-averse
or risk-seeking, whereas for the patient strategy,
it only mattered whether the models were threat-
or nonthreat-prone, irrespective of whether they
54. were risk-seeking or risk-averse. For the impa-
tient strategy, the proportions of attack steps
needed by threat-prone models (53.15%) were
much less than those needed by nonthreat-prone
models (92.60%). Also, for the impatient strat-
egy, the proportions of attack steps needed by
risk-averse models (50.90%) were much less
compared with risk-seeking models (94.85%).
For the patient strategy, however, although the
proportion of attack steps needed by threat-prone
models (10.10%) were much less than nonthreat-
prone models (88.80%), there were no differ-
ences in the proportion of attack steps between
risk-averse (48.80%) and risk-seeking (50.10%)
models. In general, as expected, the threat-prone
and risk-averse model used the least proportion
of attack steps irrespective of the attack strategy,
patient or impatient.
614 June 2013 - Human Factors
DISCUSSION
Cyber attacks are becoming increasingly
common and they might cause major disruption
of work and the loss of important information.
Therefore, it is important to investigate defender
and adversarial behaviors that influence the
accurate and timely detection of network
threats. In this endeavor, Instance-Based
Figure 3. The influence of model type and the attack strategy on
model’s accuracy.
55. Cyber SA, AttACker, And defender 615
Learning Theory predicts that both defender
and adversary behaviors are likely to influence
the defender’s accurate and timely detection of
threats (i.e., cyberSA). Results from an IBL
model predict that defender’s cognitive abili-
ties, namely, experience and tolerance, and the
attacker’s strategy about timing of threats
together impact a defender’s cyberSA.
First, we found that the model’s accuracy (d′)
is positive only when the attack strategy is
impatient and the model has a threat-prone dis-
position regardless of the model’s tolerance.
This result is explained given the influence of
recency of information on decisions in the
model (Dutt, Ahn, et al., 2011; Gonzalez &
Dutt, 2011). That is because an impatient strat-
egy’s early threats would increase the activation
of threat instances in the threat-prone model’s
memory early on, and the early threats are also
likely to increase the chances that the accumu-
lation of evidence for threats would exceed the
model’s tolerance level, irrespective of whether
it is risk-seeking or risk-averse. Therefore, both
factors are likely to make the model perform
more accurately against an impatient strategy
when its cognitive disposition is threat-prone,
irrespective of its risk tolerance.
Second, we found that the proportion of attack
steps was influenced by both memory and toler-
56. ance against the impatient strategy, whereas only
the memory seemed to influence the proportion of
attack steps against the patient strategy. We believe
the likely reason for this observation is the fact
that when threats occur early, the accumulation of
evidence builds up toward the tolerance level and
it influences the timing of detection; however,
when threats occur late, the accumulation of evi-
dence might have already reached the tolerance
level causing the model to stop much before
encountering these late occurring threats. This
observation is supported by increased number of
false alarms, as well as the model needing lesser
proportion of attack steps against the patient strat-
egy (i.e., when the threats occurred late).
Also, we found an interaction between dif-
ferent attack strategies and the model’s type:
For an impatient attack strategy, possessing
threat-prone experiences helped the model’s
accuracy (due to high hit rates), whereas pos-
sessing threat-prone experiences hurt the mod-
el’s accuracy against a patient strategy (due to
high false-alarm rates). This result is expected
given that when threats occur early, possessing
a majority of threat instances in the model
increases the likelihood of detecting these
threats early on. Moreover, based on the same
reasoning, increasing the likelihood of detect-
ing threats causes the model to detect these
threats earlier, which hurts the accuracy when
these threats actually occur late in the attack.
Another important observation is that the
model possessing nonthreat-prone experiences
57. seemed to show a close to zero d′ irrespective of
the attack strategy. A probable reason for this
Figure 4. The influence of model type and the attack strategy on
proportion of attack steps.
616 June 2013 - Human Factors
observation is the following: Having lesser pro-
portion of threat experiences in memory would
make it difficult for the model to retrieve these
experiences whether the attack occurs early or
late. Thus, the overall effect would be a decrease
in ability to detect threats from nonthreats when
the proportion of threat experiences in memory
is low. Indeed we find that the hit rate in the
model possessing nonthreat-prone experiences
is very low.
Our results are clear in the one way to
improve defenders’ performance: It is important
to train defenders with cases involving multiple
threats that would results in a threat-prone
memory and prepare them for impatient attack-
ers. Furthermore, it is important to determine
the defender’s tolerance to risk, as that will
determine how timely the defenders address an
attack from impatient attackers. Unfortunately,
as our results show, these two requirements
would not be sufficient to prepare defenders for
patient attacker strategies. Because the model’s
predicted accuracy (d′) is low in all cases in
which the attacker follows a patient strategy, we
would need to determine better ways to improve
58. accuracy in these cases. Being trained with a
threat-prone memory would not be enough in
this case, given the high number of false alarms
produced in this type of training, although for-
tunately only a small number of steps would be
needed to determine an attack in these cases.
Although in our experimental manipulations
we have simulated defenders with characteristics
of memory and tolerance that varied at two oppo-
site ends of the spectrum of several possibilities,
one could easily modify our defender characteris-
tics in the model to intermediate values. Thus, for
example, the threat-prone and nonthreat-prone
defenders could each have a 60%–40% and 40%–
60% mix of threat and nonthreat instances in
memory rather than the currently assumed 90%–
10% and 10%–90% mix. Even if one changes this
mix to intermediate values, we believe the direc-
tion of results obtained would agree with our cur-
rent results. Second, recent research in cyber
security has led to develop methods for correlat-
ing alerts generated by IDS sensors into attack
scenarios and these methods seem to greatly sim-
plify security defenders’ job functions (Albanese,
Jajodia, Pugliese, & Subrahmanian, 2011; Debar
& Wespi, 2001; Ning, Cui, & Reeves, 2002). In
future work, we plan to consider analyzing a
defender’s behavior with respect to these newer
tools. Third, given the low d′ values we could
attempt to improve the model’s performance by
using feedback for the decisions made. Feedback
was not provided because defenders in the real
world do not get this feedback during a real-time
cyber attack (and might only learn about the
59. attack after it has occurred). However, we do use
a squared similarity assumption in the model and
this assumption enables the model to observe the
different attributes of network events. We believe
that this similarity mechanism allows the model
to produce some differences between hit and
false-alarm rates on account of the memory and
tolerance manipulations.
If our model’s predictions on defender
behavior (experiences and tolerance) are correct
and the model is able to represent the cyberSA
of human defenders, then it would have signifi-
cant potential to contribute toward the design of
training and decision-support tools for analysts.
Based on our model predictions, it might be bet-
ter to devise training and decision-support tools
that prime analysts to experience more threats
in a network. Moreover, our model’s cyberSA
was also impacted by how risk-seeking it was to
the perception of threats. Therefore, companies
recruiting analysts for network-monitoring
operations would benefit by evaluating the
defender’s risk-seeking/risk-aversion tenden-
cies by using risk measures like BART (Lejuez
et al., 2002) or DOSPERT (Blais & Weber,
2006). Furthermore, although risk-orientation
may be a person’s characteristic (like personal-
ity), there might be training manipulations that
could make defenders conscious of their risk-
orientation or alter it in some ways.
At present, we know that predictions gener-
ated from the model in this article need to be
validated against real human data; however, it is
difficult to study real-world cyber-attack events
60. because these occurrences are uncertain, and
many attacks occur on proprietary networks
where getting the data after they have occurred
raises ownership issues (Dutt, Ahn, et al., 2011).
Yet, as part of future research, we plan to run
simulated laboratory studies assessing human
behavior in situations involving different adver-
sarial strategies that differ in the timing of
threats. An experimental approach involving
Cyber SA, AttACker, And defender 617
human participants (even if not real defenders)
will allow us to validate our model’s predictions
and improve its relevance and the default
assumptions made with its free parameters. In
these studies, we believe that some of the inter-
esting factors to manipulate would include the
threat/nonthreat experiences stored in memory.
One method is to provide training to partici-
pants on scenarios that present them with a
greater or smaller proportion of threats before
they actually participate in detecting threats in
island-hopping attacks (i.e., priming memory of
participants with more or less threat instances
as we did in the model). Also, we plan to record
the participants’ risk-seeking and risk-averse
behavior using popular measures involving
gambles to control for their tolerance level (typ-
ically a risk-seeking person is more tolerant to
risks compared with a risk-averse person).
Thus, our next goal in this research program
will be to validate our model’s predictions.
61. ACKNOWLEDGMENTS
This research was supported by the
Multidisciplinary University Research Initiative
Award on Cyber Situation Awareness (MURI;
#W911NF-09-1-0525) from Army Research Office
to Cleotilde Gonzalez. The authors are thankful to
Noam Ben Asher and Hau-yu Wong, Dynamic
Decision Making Laboratory, Carnegie Mellon
University, for providing insightful comments on an
earlier version of the article.
KEY POINTS
•• Due to most corporate operations becoming
online, the threat of cyber attacks is growing; a
key element in keeping online operations safe
is the cyber security awareness (cyberSA) of a
defender, who is in charge of monitoring online
operations.
• The defender’s cyberSA is measured by his or
her accurate and timely detection of cyber attacks
before they affect online operations. It is likely
influenced by the defender’s behavior (experi-
ence and tolerance level) and adversary’s behav-
ior (strategies about different timing of threats).
• Defenders who are risk-averse and possess prior
threat experiences are likely to improve their
detection performance in situations involving
impatient attackers; however, not in situations
involving patient attackers.
• A cognitive model based on the Instance-Based
Learning Theory (IBLT) represents a simulated
62. defender. The model is simulated 500 times each
for the different combination of the adversary’s
and defender’s behaviors. This experiment gener-
ates predictions about the effects of those behav-
iors on the defender’s cyberSA.
• Application of our results include the design of
training tools that increase defenders’ compe-
tency and the development of decision-support
tools that improve their on-job performance in
detecting cyber attacks.
REFERENCES
Albanese, M., Jajodia, S., Pugliese, A., & Subrahmanian, V. S.
(2011). Scalable analysis of attack scenarios. In Proceedings of
the 16th European Conference on Research in Computer Secu-
rity (pp. 416–433). Leuven, Belgium: Springer-Verlag Berlin.
doi:10.1007/978-3-642-23822-2_23
Anderson, J. R., & Lebiere, C. (1998). The atomic components
of
thought. Hillsdale, NJ: Lawrence Erlbaum Associates.
Anderson, J. R., & Lebiere, C. (2003). The Newell test for a
the-
ory of mind. Behavioral and Brain Sciences, 26(5), 587–639.
doi:10.1017/S0140525X03000128
Blais, A.-R., & Weber, E. U. (2006). A Domain-Specific Risk-
Taking (DOSPERT) scale for adult populations. Judgment and
Decision Making, 1(1), 33–47.
Cowan, N. (2001). The magical number 4 in short-term memory:
A reconsideration of mental storage capacity. Behavioral and
Brain Sciences, 24, 87–185.
63. Debar, H., & Wespi, A. (2001). Aggregation and correlation of
intrusion-detection alerts. In Proceedings of the 4th Interna-
tional Symposium on Recent Advances in Intrusion Detection
(pp. 85–103). Davis, CA: Springer-Verlag. doi:10.1007/3-540-
45474-8_6
Dutt, V., Ahn, Y. S., & Gonzalez, C. (2011). Cyber situation
aware-
ness: Modeling the security analyst in a cyber-attack scenario
through Instance-Based Learning. Lecture Notes in Computer
Science, 6818, 280–292. doi:10.1007/978-3-642-22348-8_24.
Dutt, V., Cassenti, D. N., & Gonzalez, C. (2011). Modeling a
robot-
ics operator manager in a tactical battlefield. In Proceedings
of the IEEE Conference on Cognitive Methods in Situation
Awareness and Decision Support (pp. 82–87). Miami Beach,
FL: IEEE. doi:10.1109/COGSIMA.2011.5753758
Dutt, V., & Gonzalez, C. (in press). Cyber situation awareness:
Modeling the security analyst in a cyber attack scenario through
Instance-Based Learning. In C. Onwubiko & T. Owens (Eds.),
Situational awareness in computer network defense: Princi-
ples, methods and applications. Hershey, PA: IGI Global.
Dutt, V., Yu, M., & Gonzalez, C. (2011). Deciding when to
escape
a mine emergency: Modeling accumulation of evidence about
emergencies through Instance-based Learning. Proceedings of
the Human Factors and Ergonomics Society Annual Meeting,
55(1), 841–845. doi:10.1177/1071181311551175
Endsley, M. R. (1995). Toward a theory of situation aware-
ness in dynamic systems. Human Factor, 37(1), 32–64.
doi:10.1518/001872095779049543
64. Gardner, H. (1987). The mind’s new science: A history of the
cogni-
tive revolution. New York, NY: Basic Books.
Gibson, O. (2011, January 19). London 2012 Olympics faces
increased
cyber attack threat. The Guardian. Retrieved from http://www
618 June 2013 - Human Factors
.guardian.co.uk/uk/2011/jan/19/london-2012-olympics-cyber-
attack
Gonzalez, C. (2012). Training decisions from experience with
decision making games. In P. Durlach & A. M. Lesgold (Eds.),
Adaptive technologies for training and education (pp. 167–
178). New York, NY: Cambridge University Press.
Gonzalez, C., & Dutt, V. (2011). Instance-based learning: Inte-
grating decisions from experience in sampling and repeated
choice paradigms. Psychological Review, 118(4), 523–551.
doi:10.1037/a0024558
Gonzalez, C., Dutt, V., & Lejarraga, T. (2011). A loser can be a
winner:
Comparison of two instance-based learning models in a market
entry competition. Games, 2(1), 136–162.
doi:10.3390/g2010136
Gonzalez, C., Lerch, J. F., & Lebiere, C. (2003). Instance-based
learning in dynamic decision making. Cognitive Science,
27(4), 591–635. doi:10.1016/S0364-0213(03)00031-4
65. Hertwig, R., Barron, G., Weber, E. U., & Erev, I. (2004).
Decisions
from experience and the effect of rare events in risky choice.
Psychological Science, 15(8), 534–539. doi:10.1111/j.0956-
7976.2004.00715.x
Hogarth, R. M., & Einhorn, H. J. (1992). Order effects in belief
updating: The belief-adjustment model. Cognitive Psychology,
24, 1–55. doi:10.1016/0010-0285(92)90002-J
Jajodia, S., Liu, P., Swarup, V., & Wang, C. (2010). Cyber situ-
ational awareness. New York, NY: Springer.
Johnson-Laird, P. (2006). How we reason. London, UK: Oxford
University Press.
Lejarraga, T., Dutt, V., & Gonzalez, C. (2010). Instance-based
learning: A general model of decisions from experience in
repeated binary choice. Journal of Behavioral Decision Mak-
ing, 23, 1–11. doi:10.1002/bdm.722
Lejuez, C. W., Read, J. P., Kahler, C. W., Richards, J. B.,
Ramsey,
S. E., Stuart, G. L., & . . . Brown, R. A. (2002). Evaluation of a
behavioral measure of risk-taking: The Balloon Analogue Risk
Task (BART). Journal of Experimental Psychology: Applied,
8(2), 75–84. doi:10.1037/1076-898X.8.2.75
McCumber, J. (2004). Assessing and managing security risk in
IT
systems: A structured methodology. Boca Raton, FL: Auerbach
Publications.
Ning, P., Cui, Y., & Reeves, D. S. (2002). Constructing attack
sce-
narios through correlation of intrusion alerts. In Proceedings
66. of the 9th ACM Conference on Computer & Communications
Security (CCS ‘02) (pp. 245–254). Washington, DC: ACM.
doi:1-58113-612-9/02/0011.
Ou, X., Boyer, W. F., & McQueen, M. A. (2006). A scal-
able approach to attack graph generation. In Proceed-
ings of the 13th ACM Conference on Computer and
Communications Security (pp. 336–345). Alexandria, VA:
ACM. doi:10.1145/1180405.1180446
PSU. (2011). Center for cyber-security, information privacy,
and
trust. Retrieved from http://cybersecurity.ist.psu.edu/research.
php
Salter, C., Saydjari, O., Schneier, B., & Wallner, J. (1998).
Toward
a secure system engineering methodology. In Proceedings of
New Security Paradigms Workshop (pp. 2–10). Charlottesville,
VA: ACM. doi:10.1145/310889.310900
Shepard, R. N. (1962). The analysis of proximities: Multidi-
mensional scaling with an unknown distance function. Psy-
chometrika, 2, 125–140. doi:10.1007/BF02289630
Sideman, A. (2011). Agencies must determine computer security
teams in face of potential federal shutdown. Retrieved from
http://fcw.com/articles/2011/02/23/agencies-must-determine-
computer-security-teams-in-face-of-shutdown.aspx
Simon, H. A., & March, J. G. (1958). Organizations. New
York,
NY: Wiley.
Tadda, G., Salerno, J. J., Boulware, D., Hinman, M., & Gorton,
S.
67. (2006). Realizing situation awareness within a cyber environ-
ment. In Proceedings of SPIE Vol. 6242 (pp. 624204). Orlando,
FL: SPIE. doi:10.1117/12.665763
White House, Office of the Press Secretary. (2011). Remarks by
the President on securing our nation’s cyber infrastructure.
Retrieved from http://www.whitehouse.gov/the_press_office/
Remarks-by-the-President-on-Securing-Our-Nations-Cyber-
Infrastructure/
Wickens, T. D. (2001). Elementary signal detection theory. New
York, NY: Oxford University Press.
Xie, P., Li, J. H., Ou, X., Liu, P., & Levy, R. (2010). Using
Bayes-
ian networks for cyber security analysis. In Proceedings of the
2010 IEEE/IFIP International Conference on Dependable Sys-
tems and Networks (DSN) (pp. 211–220). Hong Kong, China:
IEEE Press. doi:10.1109/DSN.2010.5544924
Varun Dutt received his PhD in engineering and
public policy from Carnegie Mellon University in
2011. He is an assistant professor at the School of
Computing and Electrical Engineering and School of
Humanities and Social Sciences, Indian Institute of
Technology, Mandi, India. Prior to this assignment,
he worked as a postdoctoral fellow at the Dynamic
Decision Making Laboratory, Carnegie Mellon
University. He is also the knowledge editor of the
English daily Financial Chronicle. His current
research interests are in dynamic decision making
and modeling human behavior.
Young-Suk Ahn is an MS (software engineering)
student in the School of Computer Science at
Carnegie Mellon University. He is also the CEO of
68. Altenia Corporation, Panama. His current research
interests focus on dynamic decision making and
software engineering.
Cleotilde Gonzalez received her PhD in manage-
ment information systems from Texas Tech
University in 1996. She is an associate research
professor and director of the Dynamic Decision
Making Laboratory, Department of Social and
Decision Sciences, Carnegie Mellon University. She
is affiliated faculty at HCII, CCBI, and CNBC. She
is on the editorial board of Human Factors and is
associate editor of the Journal of Cognitive
Engineering and Decision Making.
Date received: October 31, 2011
Date accepted: August 19, 2012
Homework 9
Due a week after the first class at 11:59 pm
Read the assigned articles in D2L. Answer the questions below.
The answers must demonstrate that you have substantively
engaged with the material and you haven’t simply goggled the
question and copy/pasted the answer.
1. Automation of a wide variety of behaviors continues to
proliferate. Name a cognitive task that has become automated in
your lifetime. Do you use the automated platforms for this task?
Are there conditions under which you would change your
behavior (either use the automation if you do not, or not use the
automation if you currently do)?