Conference PaperPDF Available

Plenty of Phish in the Sea:Analyzing Potential Pre-Attack Surfaces


Abstract and Figures

Advanced Persistent Threats (APTs) are one of the main challenges in modern computer security. They are planned and performed by well-funded, highly-trained and often state-based actors. The first step of such an attack is the reconnaissance of the target. In this phase, the adversary tries to gather as much intelligence on the victim as possible to prepare further actions. An essential part of this initial data collection phase is the identification of possible gateways to intrude the target. In this paper, we aim to analyze the data that threat actors can use to plan their attacks. To do so, we analyze in a first step 93 APT reports and find that most (80%) of them begin by sending phishing emails to their victims. Based on this analysis, we measure the extent of data openly available of 30 entities to understand if and how much data they leak that can potentially be used by an adversary to craft sophisticated spear-phishing emails. We then use this data to quantify how many employees are potential targets for such attacks. We show that 83% of the analyzed entities leak several attributes of uses, which can all be used to craft sophisticated phishing emails.
Content may be subject to copyright.
Plenty of Phish in the Sea:
Analyzing Potential Pre-Attack Surfaces
Tobias Urban13 [0000000309080038], Matteo Große-Kampmann12 3, Dennis
Tatang3, Thorsten Holz3, and Norbert Pohlmann1
1Institute for Internet-Security, Westphalian University of Applied Sciences,
2Aware7 GmbH, Germany
3Ruhr-Universit¨at Bochum, Germany
Abstract. Advanced Persistent Threats (APTs) are one of the main
challenges in modern computer security. They are planned and performed
by well-funded, highly-trained and often state-based actors. The first step
of such an attack is the reconnaissance of the target. In this phase, the
adversary tries to gather as much intelligence on the victim as possible
to prepare further actions. An essential part of this initial data collection
phase is the identification of possible gateways to intrude the target.
In this paper, we aim to analyze the data that threat actors can use to
plan their attacks. To do so, we analyze in a first step 93 APT reports
and find that most (80 %) of them begin by sending phishing emails
to their victims. Based on this analysis, we measure the extent of data
openly available of 30 entities to understand if and how much data they
leak that can potentially be used by an adversary to craft sophisticated
spear phishing emails. We then use this data to quantify how many
employees are potential targets for such attacks. We show that 83% of
the analyzed entities leak several attributes of uses, which can all be used
to craft sophisticated phishing emails.
Keywords: advanced persistent threats ·phishing ·OSINT ·reconnais-
sance ·MITRE ·cyber kill chain ·measurement study.
1 Introduction
Today, advanced persistent threats (APTs) represent one of the most dangerous
types of attacks, as a malicious actor focuses a tremendous amount of resources
into an attack on a selected target. Often such attacks utilize social engineering
methods—especially spear phishing—to initially infect the system in the target’s
network (e.g., via an email attachment) [19]. For an attacker, one of the first
steps is to collect as much information as possible on the target to plan their
further steps (e. g., used technologies or intelligence on employees to craft spear-
phishing emails) [22]. This data collection mostly happens unnoticed since the
adversaries often rely on open-source intelligence (OSINT) data, which can be
accessed by anyone. The collection of such data cannot be measured, or at least
the crawling cannot be distinguished from benign traffic.
2 T. Urban et al.
In this paper, we aim to understand and measure which publicly avail-
able data malicious actors can potentially utilize to plan and conduct their
attacks with a strong emphasis on data an adversary can use to design so-
phisticated phishing campaigns. To the best of our knowledge, all previous work
exclusively aims to detect attacks while they happen, to investigate them after
the adversaries performed the attack, or to compare different APT campaigns
(e. g., [4, 10, 11, 16, 21, 25]). We aim to illuminate the data publicly available to
adversaries during their initial reconnaissance phase by analyzing a diverse set of
organizations (n= 30). In a first step, we analyze 93 APT reports with a strong
focus on the different approaches how actors get access to a company’s network
and which techniques they use to do so. We show that an overwhelming majority
of 80 % use targeted phishing emails to lure users to unknowingly infect their
system (e. g., clicking on a malicious email attachment). Based on this finding,
we crawled nearly 5 million websites, analyzed more than 250,000 documents,
and over 18,000 social media profiles regarding data that can be used to create
personalized phishing emails. We then quantify the magnitude of publicly avail-
able data companies (unknowingly) leak and show that 90% of them leak data
that adversaries can use for the desired task. Furthermore, we show that, on av-
erage, 71 % of the employees we identified leaked several attributes that can be
used for phishing attacks as we found several work-related information on them
that an adversary can use in a targeted phishing campaign (e. g., supervisors,
the focus of work, or the used software).
In summary, we make the following key contributions:
1. We analyze real-world APT campaigns and identify the most common tactics
adversaries use during an attack and map these tactics and techniques onto
the MITRE PRE-ATT&CK framework.
2. We measure the magnitude of data that companies (unknowingly) expose
that can be used by adversaries to craft spear phishing emails. To this end, we
crawl several publicly available data sources (e.g., social networks and openly
available information on data leaks) and the company’s infrastructure.
3. We analyze how many employees of a company leak enough attributes to
write highly sophisticated phishing mails. We find that over 83% of all an-
alyzed companies provide rich target for spear phishing attacks.
2 Background
Before describing our approach to determine the Internet-facing attack surface of
a company, we provide background information necessary to follow our method.
2.1 Advanced Persistent Threats
Advanced Persistent Threats (APTs) are attacks executed by sophisticated and
well-resourced, often state-sponsored, groups. In contrast to other adversaries,
the actions of these groups are often politically motivated, but they also aim to
Analyzing Potential Pre-Attack Surfaces 3
achieve an economic gain from their efforts. They target every business sector
and design their attacks in a way that remains undetected for a very long time,
in contrast to e. g., ransomware attacks. While common adversaries often choose
their target by chance, APT threat actors typically target a specific company
or business sector and invest a lot of time and energy until they eventually suc-
cessfully obtain access. To enable such attacks, these groups utilize traditional
attack vectors like social engineering (e. g., spear phishing), but also sometimes
collect information by physically infiltrating the target companies (e.g., dump-
ster diving).
Spear Phishing In computer security, phishing describes the act when an ad-
versary impersonates a trusted entity (e.g., a popular brand or bank) with the
intent to trick users into exposing personal data (e. g., credit card numbers or
credentials) or spreading malware via malicious attachments or links [27]. While
these attempts commonly target tens of thousands of users, spear phishing tar-
gets a limited group of people (e. g., few people within a company or one research
group at a university) or sometimes only a single person (e. g., the head of a de-
partment). As these phishing campaigns target specific individuals, adversaries
can craft emails in a way that they perfectly suit the audience (e. g., by per-
sonal salutations in emails) and are often successful [3]. Adversaries persistently
exploit phishing and spear-phishing because exploiting humans is often easier
compared to bypassing technical security measures [8].
Cyber Kill Chain The term cyber kill chain, coined by Lockheed Martin [22], is
referring to the military term “kill chain”, and both terms describe the structure
of an attack. However, the cyber kill chain is often used defensively in incidence
response or digital forensics to model the attack performed by an adversary [32].
The chain maps each attack to seven phases that can be grouped into two sec-
tions, based on the stage of the attack. First, the attacker profiles the target
(1: “Reconnaissance”), then she builds the malware used to infiltrate the tar-
get (2: “Weaponize”), which is then transferred to the target (3: “Delivery”).
Afterwards, the attacker triggers the payload (4: “Exploitation”), and installs a
backdoor and establishes a persistent bridgehead into the target’s network (5:
“Installation”). Finally, she builds a C&C infrastructure to communicate with
the infected hosts (6: “Command and Control”) and performs the malicious ac-
tions of desire (7: “Act on Objective”). Phases one, two, and sometimes three are
referred to as the pre-attack stage, while the remaining stages are referred to as
attack stage. In this work, we only focus on the pre-attack phase and specifically
on the reconnaissance phase.
2.2 MITRE Framework
The MITRE cooperation created and still maintains the PRE-ATT&CK [31]
and ATT&CK (“Adversarial Tactics, Techniques, and Common Knowledge”)
[30] frameworks. The platform collects and systematizes techniques and tactics
of real-world adversaries which were obtained from several attacks with the goal
4 T. Urban et al.
that companies can learn from those attacks and improve their security con-
cepts. All collected events are organized in different categories based on their
appearance in the cyber kill chain [32]. The framework assigns a unique four-
digit identifier to each category and technique so that it can be referenced easier
(e. g., T1189 in TA0001).
The PRE-ATT&CK framework is designed to focus on the stages that usually
occur before the attack is performed. For example, this includes choosing a
victim, collecting data on the victim, or setting up the infrastructure needed
to perform the attack (e. g., implementing the needed malware or setting up
the C&C infrastructure). While the ATT&CK framework often contains very
technical and specific information, the PRE-ATT&CK framework is often more
general as it is by nature not as easy to determine which actions the actor
performed. For example, if an adversary used a specific type of malware, one
can analyze it and draw conclusions based on the sample. However, one cannot
undoubtedly determine why a specific employee was phished based on technical
data. Appendix A provides an overview of the pre-attack techniques and tactics
of the framework that are relevant for our work.
3 Advance Persistent Threat Analysis
In this section, we provide the results of an analysis of 93 real-world APT reports
we studied. More specifically, we perform a technical mapping of these reports
onto the MITRE PRE-ATT&CK framework.
3.1 APT Report Analysis
As noted above, the cyber kill chain describes the multiple stages of an attack.
To the best of our knowledge, no systematic research went into the analysis of
the early steps of this process in which adversaries collect data on their victims
to plan and initiate their campaigns. To close this gap and to gain further in-
sights into the methods adversaries use, we manually analyze 93 openly available
reports and technical blogs on APT campaigns with a strong emphasis on these
steps (i. e., the reconnaissance phase). We use these reports as security compa-
nies, in contrast to academic researchers, often have unique insights into these
APTs, especially in terms of incident response. In total, 40 different companies
provide the reports of the APTs (e. g., Symantec,Kaspersky Lab or Palo Alto
Overall, the analysis of the APT reports in our dataset attributed them to
66 different malicious actors. In 32 cases, the report does neither identify nor
disclose the actor. We argue that this broad distribution of actors allows us to
draw a more generalizable conclusion on the methods used by them. According
to the analyzed reports, the attacks in our dataset happened between 2011 and
2020. Figure 1a provides a detailed overview of the number of analyzed APTs
each year. Two reports do not report on the year in which the APT happened.
Nearly all reports lack information about the reconnaissance phase (91 %). This
Analyzing Potential Pre-Attack Surfaces 5
(a) Reported incidents by
(b) Reported delivery vec-
(c) Reported exploits.
Fig. 1: Overview of the analyzed APT reports.
knowledge gap probably roots in the fact that this cannot easily be analyzed,
especially from an incident response point of view. However, in 41% of the cases,
the target group (e. g., business sector or company type) could be identified. In
rare cases where data on the reconnaissance phase is present, the actor used data
publicly available (“Open-Source Intelligence” or OSINT) to identify promising
targets for further steps. Reportedly, an overwhelming majority (88 %) of all
APTs used social engineering techniques to deploy their attack tools (e. g., mal-
ware) in companies’ infrastructure. Furthermore, email seems to be the most
popular way to get in touch with the victim (80%). Other means of communi-
cation with victims include social media (3 %), phishing websites (4 %), or SMS
(1 %). In the cases where the malicious actor did not rely on social engineering,
the attackers abused vulnerabilities collected from public data on the compa-
nies infrastructure (4 %), data collected from other services (3 %) or the reports
only hold vague or inconclusive data on the delivery phase (e. g., “banks in Rus-
sia”). Figure 1b provides an overview of the reported delivery vectors. In the
exploitation phase, the actors mostly used Microsoft Office documents that con-
tained malicious macros (69 %). In the remaining cases, the adversaries either
used case-specific malware or exploits they tailored for a product the company
uses, as shown in Figure 1c.
In summary, there is a lack of knowledge of how attackers collect the data
on their victims. However, in the early stages of an attack, social engineering
is the most common attack vector. Most malicious actors use email (e.g., spear
phishing) as a primary channel to get in touch with their targets. In these emails,
they utilized office documents that contain malicious macros to infect the user’s
system. While the analysis of the APT campaigns yielded the most common
ways of how adversaries try to infiltrate companies, it is unclear which data they
used to perform these attacks, or how and where they acquired it.
6 T. Urban et al.
3.2 MITRE PRE-ATT&CK Analysis
The MITRE PRE-ATT&CK framework consists of 15 groups that describe dif-
ferent stages of the pre-attack phase [31]. In this work, we focus on data that
can be publicly accessed by an adversary that provides her insights on the target
company, the used infrastructure, and employees of the company. As previously
described, an adversary can use this data to perform sophisticated social engi-
neering attacks, like spear phishing.
We analyzed the framework to test which of the listed tactics can be ana-
lyzed using publicly measurable data using only non-offensive collection meth-
ods, which we used as a basis to design our measurement. Adversaries probably
also use offensive tools (e. g., vulnerability scanners or buying information leaks
online) to collect information, but due to ethical considerations, we renounce
to use such tactics. Three computer security experts with a strong background
in online measurements or threat intelligence (i.e., the first three authors of
this paper) analyzed the framework. The experts were instructed to analyze all
techniques and tactics in the framework and assessed whether they are publicly
measurable using only non-offensive collection methods. The final inter-rater
agreement whether a technique is measurable in our setting or not shows sub-
stantial agreement (Fleiss’ Kappa: κ= 0.73; agreement >90 %). In the rare
cases of discrepancies, the option that got the majority was selected to resolve
such matters.
The results show that one cannot measure several techniques using data that
is publicly available. As a result, we only consider four of the 15 groups (i. e.,
Technical Information Gathering, People Information Gathering,Organizational
Information Gathering, and Technical Weakness Identification) in our analysis
(see Table A in Appendix A). The remaining groups are not measurable without
internal insights of the adversary. For example, an analyst could measure the
Target Selection or Adversary OPSEC phase if she infiltrates the adversary’s
internal infrastructure and monitors all events. We consider this to be out-of-
scope as (1) we want to identify protection mechanisms for companies, and only
highly specialized experts can perform such infiltration and (2) such penetration
is likely in a legal gray (if not black) area. The techniques that we exclude are
often either (1) described too general in the framework (e. g., “Conduct active
scanning”), (2) out-of-scope because we refrain from using offensive technologies
(e. g., “Conduct social engineering ”), or (3) can be done reliably in an automated
fashion (e. g., “Identify supply chains”).
Summary The analysis of the APT campaigns revealed that social engineering
enables most of them, commonly conducted via spear-phishing emails. However,
the reports could only rarely reconstruct which data attackers used to write the
emails. The MITRE PRE-ATT&CK lists several techniques adversaries can use
to collect such data. However, several of these described techniques are very
broad, cannot be measured straightforwardly, and are sometimes not under the
control of the company. Therefore, the question arises to what extend companies
(unknowingly) expose such data.
Analyzing Potential Pre-Attack Surfaces 7
4 Measuring Data Collection Opportunities
Based on the analysis of the framework presented in the previous section, we
developed tools to collect the data types that a malicious actor can use to craft
sophisticated spear-phishing emails as they are the most prevalent intrusion vec-
tor. We used two different crawling approaches to collect data for each company:
(1) Analyzing sources directly maintained by the companies (e.g., websites) and
(2) information present on third-party websites but that the company directly
or indirectly provides (e. g., job postings or social media profiles).
4.1 Data Description
In our analysis, we perform an in-depth analysis of 30 entities (27 companies, two
government agencies, and one non-profit organization). For the sake of simplicity,
we use the term company for these 30 entities in the following. To choose these
companies, we used a list of large, international companies and chose 27 from
this list, with an emphasis on banking and e-payment companies. We focused on
one sector as the described malicious actors tend to attack financial institutions
or large organizations. However, the chosen companies are active in a variety
of industry sectors and are of different sizes regarding revenue and number of
employees. On average, the revenue of the analyzed companies is 60 billion USD
(min: 27 million USD; max: 790 billion USD), and they employ 55,484 people
(min: 49; max: 375,000). We took these numbers from the official figures the
companies provided for 2018. Ten of the companies are active in the “banking”
or “digital payment” sector (37%), while the others are distributed over eleven
sectors (e. g., “Food” or “Aeronautics”).
For ethical reasons, we refrain from naming any of the companies and will use
pseudonyms for all companies in the remainder of this work (i.e., Comp. #X ).
In our measurements, we used no legal or ethical questionable tools and only
accessed data that is publicly available. More specifically, we use in this study
three different types of data sources to measure the pre-attack surface of a com-
pany: (1) data the company (unknowingly) provides, (2) data publicly available
through social media sites, and (3) data leaked in known data breaches. An
extended ethical discussion is presented in Section 7.
4.2 Data Collection
As previously mentioned, we rely for our analysis on “Open Source Intelligence”
(OSINT) data, i. e., data sources that are publicly available. In the following, we
describe the used data sources in more detail.
Company Controlled Entities To crawl each companies’ infrastructure, we
built a crawler that we initialize with 1 to ndomains owned by a company (seed
domains). If possible, we read the TLS certificate present on these domains and
8 T. Urban et al.
try to identify further domains that can be protected by this certificate (i. e., Sub-
ject Alternative Name (SAN) and multi-domain SSL certificates). Furthermore,
we perform a DNS enumeration to discover further domains and infrastructure
operated by the company. After identifying the “landing pages” of all domains
associated with a company, we visit each page and recursively all first-party links
occurring on each website to a certain depth (n= 6). Hence, we try to visit ev-
ery single webpage publicly linked by a company. Using this approach, we miss
resources that are only available if the user has a specific link to the resource.
Analyzing Metadata Most popular file types offer proprietary options to store
additional information regarding the file (“metadata”). Such metadata, for ex-
ample, includes authors of the document, the software used to create the docu-
ment (e. g., pdfTeX-1.40.17), email addresses of the author, or its title. From
an adversary’s point of view, this information may provide specific insights into
the victim. For example, the authors of a document, in combination with its
title/content, can be used to craft specific phishing emails for a single or small
group of users. With a given type of software, the adversary might also be
able to attach a file that exploits a specific bug in that software to infect the
user’s system. We only used email addresses whose domain part’s effective top-
level domain (eTLD) +1 fit the eTLD+1 of the seed domain(s) we analyzed.
For example, if our crawler scanned and extracted the email address in one file, we dropped the file. Aside from metadata analysis,
we identified emails by analyzing the content of websites and documents. For
our study, we download all files that we find during the crawling process and
extract the metadata. Overall, we analyze 36 different file types. These files in-
cludes .pdf files, office documents (e. g., *.docx or *.odt), and various image
types (e. g., *.png or *.jpeg). If a document contains an author or other per-
sonally identifiable information (e. g., email addresses or names), we map them
to other properties (e. g., used software). More specifically, we create relations
between users, the software they use, and possible topics on which they work.
For example, if we identified a Microsoft Office v1.0 document written by two
authors (Alice and Bob) with the title World Peace—Status Quo and Outlook,
we can conclude that both worked on “World Peace” using Microsoft Office.
Company Infrastructure We mainly focus on the vulnerability of companies to-
wards social engineering attacks, especially spear phishing. Thus, we describe
our measurements regarding parts of the companies infrastructure that might
be abused by an adversary for this specific kind of attack. Adversaries might
use so-called homoglyph domains (e. g., changing ‘l’ to 1) to trick employees into
visiting them with the belief to navigate on the secure infrastructure of the com-
pany (but an adversary, of course, controls this infrastructure). We perform a
simple cybersquatting detection by creating a list—based on the seed domains—
of URLs that “look” similar to humans by applying techniques like homoglyphs,
simple permutations, or by using different eTLDs. Afterward, we test if any of
these URLs exist and try to assess who registered them. We use whois requests
and data from SSL certificates to identify the registering organization.
Analyzing Potential Pre-Attack Surfaces 9
Furthermore, we aim to identify isolated components in a company’s infras-
tructure that is not connected to any other entity of the company’s infrastruc-
ture. Examples for connections between the components are hyperlinks or shared
IP addresses. Such isolated components could be legacy systems running with-
out the direct knowledge of the responsible administrators, might be used as
test systems, or in case of domains, might be run by an adversary in preparation
of an attack. For all domains registered by a company (excluding homoglyph
domains), we tested if the websites use trustworthy SSL certificates (e.g., not
expired ones). If companies use certificates that are not trustworthy, adversaries
might be able to intercept or eavesdrop the connection, which allows them to
collect sensitive data. Finally, we check whether companies register domains with
names similar to their original domain. Domain parking can be used to register
domains up front before a service is run on the domain. Furthermore, a service
provider can use this practice to avoid “domain drop catching”. Domain drop
catching is a (malicious) practice to registers a domain right after it expired and
then to use it for different purposes [18,26]. As users usually do not know when
and if domains expire, they will still visit the domain and might be exposed to
malicious content.
Social Media Employment-oriented social media platforms, like LinkedIn, are
commonly used by millions of people [20]. As these platforms are supposed to
maintain business relationships, they can also be abused by adversaries to collect
intelligence on a company [1]. This data might provide several details about the
internal workings of a company, and its employees and their careers, contacts, or
supervisors. Furthermore, companies do not have real control over which data is
shared and posted on such platforms, and adversaries might use these sites to
get in touch with the employees, undetected by any security mechanism of the
In this work, we use data obtained from various sources (e. g., different APIs).
Some of these APIs are deprecated as of July 2020 but were still available when
we collected our dataset. One example was the LinkedIn API that allowed to
crawl user data based on an email address (i. e.,
gmail/profile/viewByEmail/mailaddress. The malicious actor could use this end-
point to determine whether an identified email address had a corresponding pro-
file. To mimic the potential workflow of an adversary, we utilized search engines
to perform site-specific searches (e. g., <COMPANYNAME>).
To further enrich our dataset, we utilized publicly available tools that automate
the crawling process of social media sites (e. g., CrossLinked [24]).
Data Leaks Finally, adversaries may utilize data from previous data breaches
to prepare their attack. In this work, we use the Have I Been Pwned API [15] to
test if a company ever leaked data that can be used in another attack on that
company. The API exposes data leaks from over 400 websites and over 110.000
“pastes”. In this case, pastes are indications of data leaks in which the adversaries
provides examples of the acquired data to prove that she actually got access to
10 T. Urban et al.
Fig. 2: Overview of our data collection approach.
sensitive information. The API does not directly provide any of the breached
data but returns categories of data that the leak contained. For example, if one
provides an email address, the API will return data types that were leaked along
with the address in different data breaches (e.g., results in dates
of birth, employers, and job titles). Figure 2 provides an overview of the
three types of data sources we analyzed.
4.3 Identified Data
Data Crawled from Companies’ Infrastructure In total, we scanned 30
entities and identified 492 domains (eTLD+1 and suffix) operated by them.
Furthermore, we identified 18,873 employees, of which 8,994 appeared in data
leaks, or they provided valuable data in public social media profiles.
Metadata Analysis During the measurement, we visited 4,912,938 distinct web-
pages and extracted metadata from 271,124 documents. Table 1a provides an
overview of the identified data types identified based on the metadata of files
we found on the crawled webpages. The min,max, and mean value describe how
many instances were obtained for each company (e. g., we identified the names of
634 employees of a company). Ninety percent of the analyzed companies leak the
names of their employees. Overall, we identified 22,361 email addresses, of which
6,335 were exclusively exposed via metadata (intentionally or unintentionally).
As we extracted them from metadata, this might also provide insights to the
adversary on which projects they work on (e. g., based on the file’s content).
Aside from names, the email address is essential as actors can use them to get
in touch with potential victims. Almost three-quarters of all companies in our
dataset leaked an employee’s email address.
Furthermore, once the malicious actor understands the structure of a com-
pany’s email addresses (e. g.,, she can presumably make ed-
ucated guesses on the local parts of further addresses if she knows the employees’
names. In our dataset, the amount of identified email addresses would increase,
Analyzing Potential Pre-Attack Surfaces 11
on average, by 52 % for each company. 81 % of the companies exposed third
parties they work with (i. e., collaborating partners that created a document).
The three named data types can be used to craft user/team specific spear-
phishing campaigns. For example, an adversary could impersonate a partner the
employee worked with. Aside from personal data, the metadata of a file might
expose intelligence on the inner workings of a company. In our dataset, 90% of
the companies leaked the software they used to create a document, and almost
two-thirds leaked data paths they use in the company to store documents (e. g.,
Z:\Project X\Results). An attacker can use this information when preparing
for the attack (e. g., zero-day exploits for the used software).
Company Domains Table 1b presents potential information on the infrastructure
that an actor can collect and later use for an attack. Furthermore, it provides
hints that adversaries already actively make use of homoglyph domains. The
most troubling finding of this measurement is that for 18 (60 %) of the analyzed
companies, an adversary actively abuse a homoglyph domain, at the time of
our crawl. Note that we only counted domains for which we find a substantial
string similarity of more than 95 %, and therefore, our results can be seen as
a lower bound. For one company, we found twelve active domains of this type
(avg: 4.5). The presence of such domains indicates that adversaries are likely
already actively trying to misguide users or employees of such services (e. g.,
password phishing). However, we also observed that some companies are aware
of this endangerment and acquire some of these domains and “park” them for
brand protection purposes as a kind of proactive defense.
Often websites or other services are connected by various mechanisms (e.g.,
hyperlinks or services that share the same IP address). In our dataset, half of the
companies operate services that have no connection to others. These domains
might pose a problem if the companies no longer maintain them and, therefore,
could be less protected (e. g., legacy interfaces). On the other hand, these services
might not pose a problem at all because the companies are fully aware of them.
We found that eight entities (26%) operated domains that use an invalid or
outdated certificate. An adversary might abuse these by intercepting the TLS
encryption to such domains to collect more data on users or employees, whoever
primarily uses these services. All of these companies operated at least one isolated
domain (avg: 5) that uses an expired or otherwise untrusted certificate, which
reinforces the assumption that the companies no longer maintain them.
Data Available in Data Leaks and in Social Media Profiles In addition to
the analysis of the companies’ infrastructure and data they expose via metadata,
we analyzed if the business accounts of employees (e. g., email addresses) occur
in publicly known data breaches (see Section 4.2). For each company in our
dataset, we found data on at least three of the identified employees (max: 1,102).
In absolute numbers, 11 companies (46 %) leaked data of less than 30 employees,
and only four (12 %) did not leak data on any employee that we identified. In
relative numbers, two-thirds (20) of the analyzed companies leaked data of more
12 T. Urban et al.
Table 1: Overview of the data extracted from the company’s own infrastructure.
(a) Overview of the identified information.
The min and mean values exclude companies
that did not provide the type of data.
Data type aff. comp. min max mean
Names 90 % 7 634 227
Mail addresses 71 % 1 96 19
Third Parties 81 % 1 53 15
Software 90 % 5 205 71
Path 65 % 1 30 7
(b) Overview of the identified infrastruc-
ture information.
Type min max mean
Homoglyph Domains 1 12 4.5
Parked Domains 5 379 41.6
Isolated Domains 7 89 27.3
Untrusted Certificates 1 20 5.9
than one-third of the identified employees (max: 88 %). We found no statistical
significance between the amount of identified emails and the amount of leaked
data (ANOVA-Test p-value 0.03). Hence, companies that expose more emails
are not automatically likely to be present in more data leaks. As this might seem
to be counter-intuitive, it hints that some companies have policies in place to
reduce the potential of such data leaks (e. g., awareness campaigns). Overall, the
analyzed data leaks include 65 different data categories. The categories range
from personal data (e. g., credit status information,government issued IDs, or
device usage tracking data) over data directly tied to the employee’s professional
live (e. g., job titles,employers, or occupations) to other data an adversary could
use to plan an attack (e. g., instant messenger identities or password hints).
The category of a data leak shows a statistical correlation with the number of
instances that this data is leaked (ANOVA-Test p-value <0.0001). Hence, some
data types leak more often than others.
Figure 3 shows the type of leaked data for each company. The heatmap
highlights the ratio of identified leaks with the email addresses that we could
identify. The figure only lists the top 15 categories, which account for 89 % of
all leaking instances. It shows that some companies leak excessively more data
than others (ANOVA-Test p-value <0.0001) but that there is no dominating
data type that is leaked. In our dataset, the top leaked types are passwords
(10 %), phone numbers(8 %), and geolocations (7 %), excluding the name and
email addresses of the users that the adversary needs to identify an employee.
The biggest challenge with data actors collect from data leaks is that companies
have virtually no measure to delete it. Furthermore, in none of the cases, it
was the company itself that leaked the data but other platforms on which the
employees registered to use the service, using their business email address. Hence,
one solution could be to raise awareness with employees only to use the work
email if necessary and to provide as little information as possible when using the
respective services.
Summary In this section, we demonstrated that companies excessively leak data
that provides insights into their inner workings or on the employees of the com-
Analyzing Potential Pre-Attack Surfaces 13
Fig. 3: Overview of the data extracted from other data sources.
panies. However, it is not clear whether an adversary can meaningfully combine
this data to plan further steps link designing successful phishing campaigns.
5 Assessing Potential Phishing Targets
Based on the insights of our study, we now introduce a metric to asses the
likeliness that an employee serves as a good spear phishing target. The presence
of data that we identify in this work does not necessarily pose a security problem
per se. Each data point on its own is properly not problematic if obtained by
an adversary, but taken together, they reveal intelligence that can be used, for
example, to craft personalized phishing emails. Therefore, it is important to
analyze and interpret the collected data.
5.1 Identifying Potential Phishing Targets
We now numerically analyze whether users are promising targets for spear phish-
ing attacks from an adversary’s point of view. In this work, we only analyze
technical aspects and not the personal experience of each person, which is out-
of-scope of this work but an essential aspect if someone falls for a phishing
attempt [17]. Previous work that analyzed the effectiveness of spear phishing
found that sources that impersonate an individual from the victim’s company
(e. g., from the human resources department) are quite effective [3,12]. The work
shows that 34 %–60 % of all participants clicked on a link in such email. There-
fore, we assume that if we could identify other persons working in the company
and especially if they are working together (e. g., co-worker, supervisor, or team
member), a phishing attempt might be more effective. Furthermore, if the adver-
sary knows the software used by the victim, she can craft and append an exploit
specifically for the used software to the email, which increases the chances of
a successful compromise. Thus, we also consider the used technology of each
employee as an essential aspect.
14 T. Urban et al.
Fig. 4: Number of leaked attributes for each user in our dataset.
5.2 Spear Phishing Targets in the Wild
Figure 4 shows the number of attributes leaked for each employee in our dataset.
Most employees only leak two categories of information (i. e., their name and
email address), which we use to identify users of a company. For those indi-
viduals, it is not possible to craft targeted phishing mails (at least using our
data). However, we identified 5,910 (62%) employees that leak between seven
and 15 attributes. Those employees are employed at 25 different companies
(83 %). 878 (9 %) employees leak between 24 and 28 attributes and are em-
ployed at 18 companies (60%). From both employee groups, an adversary can
potentially pick several attributes and craft highly specific spear phishing mails.
Only four (13 %) companies in our dataset do not leak any additional data on
their employees, aside the name and email address). For all of these compa-
nies, we identified relatively few employees. One reason for this is that most of
these companies are active in business fields with no (public) customer interac-
tion (e. g., investment banking). The results show that almost all companies in
our dataset provide a considerable pre-attack surface to motivated attackers. In
absolute numbers, companies that leaked more data also provide more targets
for an attacker (ANOVA-Test p-value 0.001). However, taking the relation
between the amount of identified emails and leaked data into account (similar
to Figure 3), we did not find a correlation. Hence, leaking more data does not
necessarily mean that the adversary can identify more spear phishing targets.
Our results show that the OSINT data sources that we utilized provide a rich
data source from which threat actors can profit.
6 Related Work
Previously research on effective APT detection or prevention mostly focused on
detecting them at and/or after the “Delivery” phase in the cyber kill chain [32].
APT detection is highly complicated as information from many sources (e. g.,
human behavior, intrusion detection systems, or system logs) have to be com-
bined to make an informed decision. Machine learning approaches were studied
to process this enormous amount of data to detect APTs [2, 10, 16, 21]. Fur-
thermore, more heuristic solutions like correlating events [25], defining detection
Analyzing Potential Pre-Attack Surfaces 15
rule sets [33], detecting misuse on application level [23], or annotating security
events [11] have been proposed. Similar to our APT analysis, Lemay et al. [19]
analyzed APT reports. In their work, they provide a summary of 40 analyzed
APT reports. A large number of different works focus on the technical detection
of spear-phishing emails, content analysis of phishing websites, and the detection
of such websites based on their URLs (e.g., [13,14]). Furthermore, various studies
analyzed human aspects to understand why spear-phishing attacks are successful
(e. g., [3,12]). Finally, several papers systematize the extensive research that was
conducted in this area (e. g., [5–7]). Our work differs from these approaches as
we solely focus on the very first steps adversaries take when they plan their ma-
licious actions, the reconnaissance phase. To the best of our knowledge, we are
the first ones to show the variety and magnitude of information companies (un-
knowingly) provide that can be abused by adversaries to perform spear-phishing
attacks. Furthermore, we do not aim to understand the effectiveness of specific
phishing campaigns, but provide insights on how companies expose data.
7 Ethical Consideration
For this study, we gathered and analyzed sensitive information on companies and
employees, thus individual persons. Our research institution does not require ap-
proval for this type of study, nor does it provide an Institutional Review Board
(IRB). Nevertheless, we took strict ethical considerations into account. Addi-
tionally, we followed the research community’s standard guidelines to protect
those whose data was collected and the infrastructure of the services we use. A
recent court ruling, according to the Electronic Frontier Foundation, found that
automated scraping of publicly available data is unlikely to violate the Computer
Fraud and Abuse Act (CFAA)” [9]. As a general rule, the collection of personal
information requires user consent; however, there are exceptions for cases where
this is not practical. In our case, publicly available sources are the basis of the
collection of information. By nature of our analysis, we cannot preempt to pro-
cess personal data. We also want to highlight that none of our collections tools
use any questionable tools to identify systems or persons. We do not perform any
kind of penetration testing to collect data and send all requests at a courteous
rate. The gathered data was collected for scientific purposes only, and we only
disclosed it to the involved companies. To protect the collected personal data,
we took additional safety measures: We encrypt the raw files for storage and
delete unused samples and data.
8 Discussion and Conclusion
Our approach comes with limitations that need further clarification. The most
decisive one, from a researchers’ perspective, is that there exists no ground truth
for our collected data. Hence, it implies that we do not know if adversaries profit
from the data sources that we utilize to plan their actions or if all companies are
unaware of the leakage of such data. However, the analyzed APT reports and
16 T. Urban et al.
several other sources (e. g., [22,28,29]) indicate that adversaries make excessive
use of OSINT data, and even if companies are aware of the leakage of data, it
might still be used by the adversaries. There is very little raw data available
on incidents especially how the attackers infiltrated their victims. Furthermore,
to build a ground truth for our research, one would need to impersonate the
malicious actor while she plans her attack, which is ethically not tenable. With
a company’s consent, we could perform an awareness phishing campaign using
the identified data. However, previous work already performed similar studies
and demonstrated that they are often successful (see Section 6). With our work,
we do not aim to determine the exact data used by adversaries in each attack,
which is probably impossible in an automated fashion, but we demonstrate the
sheer scale of data leaked by companies. Our results highlight that all analyzed
companies provide a large attack surface to adversaries that is not monitored or
protected by state-of-the-art security solutions. Furthermore, this data leakage
is not always under the control of the companies, nor is it always possible to
revert the leakage. Therefore, there is no clear path how to circumvent this type
of leakage or straightforward countermeasure. It is quite hard to successfully
prevent attacks on third party providers or reduce attack surfaces and therefore
to apply countermeasures. One way to decrease the potential damage by these
data leaks is to raise awareness with employees that this kind of data is regu-
larly abused by adversaries and that the principle of “data economy” should be
followed. Actionable tools to counter misuse of our analyzed data sources can
be to wipe the metadata from all uploaded files, to continually monitor data
leaks if they include passwords or other personal data of employees, or to in-
crease awareness in a way that empowers employees not to provide too much
work-related information on social media platforms.
This work was partially supported by the Ministry of Culture and Science of
North Rhine-Westphalia (MKW grant 005-1703-0021 “MEwM”), the federal
Ministry of Research and Education (BMBF grant 16KIS1016 “AWARE7”), and
the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) un-
der Germany’s Excellence Strategy – EXC-2092 CaSa – 390781972. We would
like to thank Sweepatic NV—a cybersecurity company which maps, monitors
and manages attack surfaces—for their support and access to their technology.
1. Balduzzi, M., Platzer, C., Holz, T., Kirda, E., Balzarotti, D., Kruegel, C.: Abusing
Social Networks for Automated User Profiling. In: Symposium on Recent Advances
in Intrusion Detection. (RAID’10) (2010)
2. Barre, M., Gehani, A., Yegneswaran, V.: Mining Data Provenance to Detect Ad-
vanced Persistent Threats. In: Proceedings of the 11th International Workshop on
Theory and Practice of Provenance. TaPP’ 19, USENIX Association, Berkeley ,
CA, USA (2019)
Analyzing Potential Pre-Attack Surfaces 17
3. Caputo, D., Pfleeger, S., Freeman, J., Johnson, M.: Going Spear Phishing: Explor-
ing Embedded Training and Awareness. IEEE Security & Privacy 12(1), 28–38
(Jan 2014).
4. Chen, P., Desmet, L., Huygens, C.: A Study on Advanced Persistent Threats. In:
Proceedings of the 15th International Conference on Communications and Mul-
timedia Security. pp. 63–72. CMS ’14, Springer International Publishing, Cham
(2014). 5
5. Chiew, K., Yong, K., Tan, C.: A survey of phishing attacks: Their types, vectors
and technical approaches. Expert Systems with Applications 106, 1–20 (2018).
6. Das, A., Baki, S., El Aassal, A., Verma, R., Dunbar, A.: SoK: A Comprehensive
Reexamination of Phishing Research from the Security Perspective. IEEE Commu-
nications Surveys Tutorials (2019).
7. Dou, Z., Khalil, I., Khreishah, A., Al-Fuqaha, A., Guizani, M.: SoK:
A Systematic Review of Software-Based Web Phishing Detection.
IEEE Communications Surveys Tutorials 19(4), 2797–2819 (2017).
8. Ferreira, A., Vieira-Marques, P.: Phishing Through Time: A Ten Year Story based
on Abstracts. In: Proceedings of the 4th International Conference on Information
Systems Security and Privacy. pp. 225–232. ICISSP ’18, INSTICC, SciTePress,
Set´ubal, Portugal (2018).
9. Fischer, C., Crocker, A.: Victory! Ruling in hiQ v. Linkedin Protects Scrap-
ing of Public Data, v-
linkedin-protects-scraping- public-data
10. Ghafir, I., Hammoudeh, M., Prenosil, V., Han, L., Hegarty, R., Rabie, K., Aparicio-
Navarro, F.J.: Detection of Advanced Persistent Threat Using Machine-Learning
Correlation Analysis. Future Generation Computer Systems 89, 349–359 (2018).
11. Gianvecchio, S., Burkhalter, C., Lan, H., Sillers, A., Smith, K.: Closing the Gap
with APTs Through Semantic Clusters and Automated Cybergames. In: Proceed-
ings of the 15th International Conference on Security and Privacy in Communi-
cation Systems. pp. 235–254. SecureComm ’19, Springer International Publishing,
Cham (2019). 12
12. Halevi, T., Memon, N., Nov, O.: Spear-Phishing in the Wild: A Real-World Study
of Personality, Phishing Self-Efficacy and Vulnerability to Spear-Phishing Attacks.
SSRN Electronic Journal (01 2015).
13. Han, Y., Shen, Y.: Accurate Spear Phishing Campaign Attribution and Early
Detection. In: Proceedings of the 31st ACM Symposium on Applied Com-
puting. pp. 2079—-2086. SAC ’16, ACM Press, New York, NY, USA (2016).
14. Ho, G., Sharma, A., Javed, M., Paxson, V., Wagner, D.: Detecting Credential
Spearphishing in Enterprise Settings. In: Proceedings of the 26th USENIX Security
Symposium. pp. 469–485. USENIX Sec ’17, USENIX Association, Berkeley , CA,
USA (2017)
15. Hunt, T.: Have I Been Pwned: API v3 (2020),
v3, accessed: 2020-04-15
16. Kumar, G.R., Mangathayaru, N., Narsimha, G., Cheruvu, A.: Feature Clustering
for Anomaly Detection Using Improved Fuzzy Membership Function. In: Proceed-
ings of the 4th International Conference on Engineering & MIS. ICEMIS ’18, ACM
Press, New York, NY, USA (2018).
18 T. Urban et al.
17. Kumaraguru, P., Rhee, Y., Acquisti, A., Cranor, L.F., Hong, J., Nunge, E.: Protect-
ing People from Phishing: The Design and Evaluation of an Embedded Training
Email System. In: Proceedings of the 25thACM SIGCHI Conference on Human
Factors in Computing Systems. pp. 905––914. CHI ’07, ACM Press, New York,
NY, USA (2007).
18. Lauinger, T., Chaabane, A., Buyukkayhan, A.S., Onarlioglu, K., Robertson, W.:
Game of Registrars: An Empirical Analysis of Post-Expiration Domain Name
Takeovers. In: USENIX Security Symposium (2017)
19. Lemay, A., Calvet, J., Menet, F., Fernandez, J.M.: Survey of Publicly Available
Reports on Advanced Persistent Threat Actors. Computers & Security 72, 26–59
20. LinkedIn Corporation: Statistics (2020),
statistics, accessed: 2020-04-15
21. Liu, F., Wen, Y., Zhang, D., Jiang, X., Xing, X., Meng, D.: Log2vec: A Hetero-
geneous Graph Embedding Based Approach for Detecting Cyber Threats within
Enterprise. In: Proceedings of the 26thACM Conference on Computer and Com-
munications Security. pp. 1777—-1794. CCS ’19, ACM Press, New York, NY, USA
22. Lockheed Martin Corporation: Gaining the Advantage—Applying Cy-
ber Kill Chain Methodology to Network Defense (2014), https://www.
Gaining the Advantage Cyber Kill Chain.pdf, accessed: 2020-04-15
23. M. Milajerdi, S., Eshete, B., Gjomemo, R., Venkatakrishnan, V.N.: ProPatrol:
Attack Investigation via Extracted High-Level Tasks. In: Proceedings of the 4th
International Conference on Information Systems Security. pp. 107–126. ICISS ’18,
Springer International Publishing, Cham (2018).
030-05171-6 6
24. m8r0wn: CrossLinked (2020),, accessed:
25. Milajerdi, S., Gjomemo, R., Eshete, B., Sekar, R., Venkatakrishnan, V.:
HOLMES: Real-Time APT Detection through Correlation of Suspicious Informa-
tion Flows. In: Proceedings of the IEEE Symposium on Security and Privacy.
pp. 1137–1152. S&P ’19, IEEE Computer Society, Washington, DC, USA (2019).
26. Miramirkhani, N., Barron, T., Ferdman, M., Nikiforakis, N.: Panning for
Understanding the dynamics of domain dropcatching. In: International Conference
on World Wide Web (2018)
27. Parsons, K., McCormac, A., Pattinson, M., Butavicius, M., Jerram, C.:
The Design of Phishing Studies: The design of phishing studies: Chal-
lenges for researchers. Computers & Security 52(C), 194––206 (2015).
28. Paterson, A., Chappell, J.: The Impact of Open Source Intelligence
on Cybersecurity, pp. 44–62. Palgrave Macmillan UK, London (2014). 4
29. RSA Research: Reconnaissance—A Walkthrough of the “APT” Intelligence Gath-
ering Process (2015),
apt-intelligence.html, accessed: 2020-04-15
30. The MITRE Corporation: Mitre att&ck matrix for enterprise (2019), https://, accessed: 2020-04-15
31. The MITRE Corporation: MITRE PRE-ATT&CK Matrix (2019), https://attack., accessed: 2020-04-15
Analyzing Potential Pre-Attack Surfaces 19
32. Yadav, T., Rao, A.M.: Technical Aspects of Cyber Kill Chain. In: Proceedings of
the 2015 Security in Computing and Communications. pp. 438–452. SSCC’ 2015,
Springer International Publishing, Cham (2015).
319-22915-7 40
33. Yu, H., Li, A., Jiang, R.: Needle in a Haystack: Attack Detection from
Large-Scale System Audit. In: Proceedings of the 19th International Con-
ference on Communication Technology. pp. 1418–1426. ICCT ’19 (2019).
A Analyzed MITRE PRE-ATT&CK Techniques
Table A lists the groups analyzed in this work. For each group, the techniques
and tactics are shown and we indicate whether we analyzed it (“Meas.”), if we
collected the needed information on third-party websites (“3rd or from first-
party resources (“1st) ”, and how we collected them (“How obtained”). If we
did not collect data on a technique, the column “How obtained” provides a brief
explanation why.
How obtained
Technical Information Gathering (TA0015)
Acquire OSINT data sets and information 7— — Too general
Conduct active scanning 7— — Too general
Conduct passive scanning 7— — out of scope
Conduct social engineering 7— — out of scope
Determine 3rd party infrastructure services 33 3Shodan and IP addresses
Determine domain and IP address space 33 3log addresses during crawls
Determine external network trust
dependencies 3 3 7 log 3rd party usage
Determine firmware version 33 3during crawl & metadata
Discover target logon/email address format 3 3 7 extract from metadata
Enumerate client configurations 3 3 7 during crawl & metadata
Enumerate externally facing entities 3 3 7 during crawl & metadata
Identify job postings and needs/gaps 3 7 7 No API present
Identify security defensive capabilities 7— — out of scope
Identify supply chains 7— — very hard automatically
Identify technology usage patterns 3 3 7 logged during crawls
Continued on next column
20 T. Urban et al.
Continued from previous column
How obtained
Identify web defensive services 3 3 7 analyzing 3rd party usage
Map network topology 33 3based on identified data
Mine technical blogs/forums 7 7 7 out of scope
Obtain domain/IP registration information 33 3whois queries
Spearphishing for Information 7— — out of scope
People Information Gathering (TA0016)
Acquire OSINT data sets and information 7— — Too general
Aggregate individual’s digital footprint 7— — very hard automatically
Conduct social engineering 7— — out of scope
Identify business relationships 7— — out of scope
Identify groups/roles 3 7 3 Based on social media data
Identify job postings and needs/gaps 3 7 3 No API present
Identify people of interest 3 7 7 based on collected data
Identify personnel with an authority/privilege 3 7 7 based on collected data
Identify sensitive personnel information 7— — out of scope
Identify supply chains 7— — out of scope
Mine social media 3 7 3 APIs of platforms
Organizational Information Gathering (TA0017)
Acquire OSINT data sets and information 7— — Too general
Conduct social engineering 7— — out of scope
Determine 3rd party infrastructure services 3 3 7 extracted during crawl
Determine centralization of IT management 7— — very hard automatically
Determine physical locations 3 3 7 extracted during crawl
Dumpster dive 7— — out of scope
Identify business processes/tempo 7— — out of scope
Identify business relationships 7— — social media data
Identify job postings and needs/gaps 3 7 3 No API present
Identify supply chains 7— — out of scope
Obtain templates/branding materials 3 3 7 extracted during crawl
... Spoofing is especially critical since most conversations on the internet nowadays happen via mail or chat. Social engineering and phishing are some of the most prevalent threats to organizations and individuals of all sizes and shapes [23]. This threat is multiple magnitudes greater if messages are not squatted, but the identity or the identity-creating characteristics can be spoofed. ...
Full-text available
We discover and evaluate a systemic threat to business chat applications that emphasizes the human layer of security called chishing-Phishing via business chat tools. The threat originates at the beginning of the connected world, and the underlying problem is called spoofing in its most basic form. It is possible to successfully spoof display names, profile pictures, and further personal information (profiles) in four out of six business chat applications that we analyzed in this work. This behavior poses a threat to an organization as it enables insider threats to flourish and, under certain circumstances, also invites external sources to pose as internal employees.
Defenders spend significant time interpreting low-level events while attackers, especially Advanced Persistent Threats (APTs), think and plan their activities at a higher strategic level. In this paper, we close this semantic gap by making the attackers’ strategy an explicit machine-readable component of intrusion detection. We introduce the concept of semantic clusters, which combine high-level technique and tactic annotations with a set of events providing evidence for those annotations. We then use a fully automated cybergaming environment, in which a red team is programmed to emulate APT behavior, to assess and improve defensive posture. Semantic clusters both provide the basis of scoring these cybergames and highlight promising defensive improvements. In a set of experiments, we demonstrate effective defensive adjustments which can be made using this higher-level information about adversarial strategy.
Phishing and spear phishing are typical examples of masquerade attacks since trust is built up through impersonation for the attack to succeed. Given the prevalence of these attacks, considerable research has been conducted on these problems along multiple dimensions. We reexamine the existing research on phishing and spear phishing from the perspective of the unique needs of the security domain, which we call security challenges: real-time detection, active attacker, dataset quality and base-rate fallacy. We explain these challenges and then survey the existing phishing/spear phishing solutions in their light. This viewpoint consolidates the literature and illuminates several opportunities for improving existing solutions. We organize the existing literature based on detection techniques for different attack vectors (e.g., URLs, websites, emails) along with studies on user awareness. For detection techniques we examine properties of the dataset, feature extraction, detection algorithms used, and performance evaluation metrics. This work can help guide the development of more effective defenses for phishing, spear phishing and email masquerade attacks of the future, as well as provide a framework for a thorough evaluation and comparison.
Conference Paper
Conventional attacks of insider employees and emerging APT are both major threats for the organizational information system. Existing detections mainly concentrate on users' behavior and usually analyze logs recording their operations in an information system. In general, most of these methods consider sequential relationship among log entries and model users' sequential behavior. However, they ignore other relationships, inevitably leading to an unsatisfactory performance on various attack scenarios. We propose log2vec, a heterogeneous graph embedding based modularized method. First, it involves a heuristic approach that converts log entries into a heterogeneous graph in the light of diverse relationships among them. Next, it utilizes an improved graph embedding appropriate to the above heterogeneous graph, which can automatically represent each log entry into a low-dimension vector. The third component of log2vec is a practical detection algorithm capable of separating malicious and benign log entries into different clusters and identifying malicious ones. We implement a prototype of log2vec. Our evaluation demonstrates that log2vec remarkably outperforms state-of-the-art approaches, such as deep learning and hidden markov model (HMM). Besides, log2vec shows its capability to detect malicious events in various attack scenarios.
Conference Paper
Earlier research focus towards anomaly detection has been towards using classifiers such as kNN, SVM and using existing distance measures to perform classification. Traditionally IDSs (Intrusion detection systems) have been developed by applying machine learning techniques and adopted single learning mechanism. This is later extended by developing Intrusion Detection Systems by adopting multiple learning mechanisms. Such systems have addressed better detection rates compared to single learning Intrusion Detection Systems. Dimensionality is one more serious concern which affects the performance of classification algorithms. Approaches such as "Feature selection" have been studied and adopted which selects a subset features from the feature set. However, the feature extraction approach for dimensionality reduction has proved to be better compared to feature selection and achieved better classification and detection rates. In this research, we address "Feature extraction" using "Evolutionary feature clustering" by proposing a "Novel fuzzy membership function" which addresses Dimensionality Reduction (DR). The idea is to transform the initial connection representation so that its equivalent representation has reduced noise affect and achieves better classification or detection rates. Experimental results on KDD datasets with 19 and 41 attributes, prove that the proposed approach has improved detection rates for R2L and U2R attack classes when compared to CANN, CLAPP, and SVM approaches. CANN approach recorded lower detection rates w.r.t U2R and R2L attacks. This failure is addressed in our earlier studies through proposing, CLAPP which proved comparatively better accuracy rates to CANN. The fuzzy membership function proposed in this paper, recorded better classification and detection rates in experiments conducted
Conference Paper
An event that is rarely considered by technical users and laymen alike is that of a domain name expiration. The massive growth in the registration of domain names is matched by massive numbers of domain expirations, after which domains are made available for registration again. While the vast majority of expiring domains are of no value, among the hundreds of thousands of daily expirations, there exist domains that are clearly valuable, either because of their lexical composition, or because of their residual trust. In this paper, we investigate the dynamics of domain dropcatching where companies, on behalf of users, compete to register the most desirable domains as soon as they are made available and then auction them off to the highest bidder. Using a data-driven approach, we monitor the expiration of 28 million domains over the period of nine months, collecting domain features, WHOIS records, and crawling the registered domains on a regular basis to uncover the purpose for which they were re-registered (caught). Among others, we find that on average, only 10% of the expired (dropped) domains are caught with the vast majority of the re-registrations happening on the day they are released. We investigate the features that make some domains more likely to be caught than others and discover that a domain that was malicious at the time of its expiration is twice as likely to be caught than the average domain. Moreover, previously-malicious domains are significantly more likely to be reused for malicious purposes than previously benign domains. We identify three types of users who are interested in purchasing dropped domains, ranging from freelancers who purchase one or two domains to professionals who invest more than $115K purchasing dropped domains in only three months. Finally, we observe that less than 11% were used to host web content with the remaining domains used either by speculators, or by malicious actors.