# A Comprehensive Evaluation on the Benefits of Context Based Password Cracking for Digital Forensics

## Full Text

This Markdown is an automated text extraction for search and reading tools. The [hosted PDF](BenefitsOfContextBasedPasswordCracking.pdf) is the authoritative publication source.

Source PDF: [BenefitsOfContextBasedPasswordCracking.pdf](BenefitsOfContextBasedPasswordCracking.pdf)
Source SHA-256: `ab8296c3858c9c97ec0cae75501a6c55fbf2177ce561e1920f937a6f779da140`
Generated with `pdftotext -nopgbrk`; formatting and reading order may differ from the PDF.

---

Journal of Information Security and Applications 84 (2024) 103809

Contents lists available at ScienceDirect

Journal of Information Security and Applications
journal homepage: www.elsevier.com/locate/jisa

A comprehensive evaluation on the benefits of context based password
cracking for digital forensics
Aikaterini Kanta a,b , Iwen Coisel c , Mark Scanlon a ,∗
a Forensics and Security Research Group, School of Computer Science, University College Dublin, Ireland
b
c

University of Portsmouth, Portsmouth, United Kingdom
Europol, The Hague, The Netherlands

ARTICLE

INFO

Keywords:
Password
Dictionary
Contextual information
Password cracking
Wordlist

ABSTRACT
Password-based authentication systems have many weaknesses, yet they remain overwhelmingly used and
their announced disappearance is still undated. The system admin overcomes the imperfection by skilfully
enforcing a strong password policy and sane password management on the server side. But in the end, the
user behind the password is still responsible for the password’s strength. A poor choice can have dramatic
consequences for the user or even for the service behind, especially considering critical infrastructure. On the
other hand, law enforcement can benefit from a suspect’s weak decisions to recover digital content stored
in an encrypted format. Generic password cracking procedures can support law enforcement in this matter
— however, these approaches quickly demonstrate their limitations. This article proves that more targeted
approaches can be used in combination with traditional strategies to increase the likelihood of success when
contextual information is available and can be exploited.

1. Introduction
Password-based authentication is older than modern digital society
might realise. It is such an archaic system yet remains a crucial component of the security of most digital systems (albeit not necessarily the
only one). Human beings are often predictable, and malicious actors are
unfortunately able to exploit poorly chosen passwords to illegitimately
enter into the system. While the consequences on a personal level can
be severe, when a large service or critical infrastructure is targeted, the
scale can become dramatic. It is therefore imperative to undergo a risk
assessment when deploying a service requiring authentication and to
make sure that all the precautions are taken.
1.1. Why are we still using passwords?
With all the known weaknesses of password-based authentication
systems, one might wonder: why we are still using them? One explanation might be the public acceptance of this mechanism. Everybody
has already used password-based authentication. In fact, it is considered that on average, a human has between 70 and 150 online
password-protected accounts. But with that being said, what are the
alternatives? Single-sign-on (SSO) strategies, active-directory, and password managers offer substitutes or enhancements over simple password
authentication. Each of these solutions increases account security as

they often require an additional element for a malicious actor to access
a given system, e.g., having access to the key wallet protected by the
password manager. Nevertheless, they still rely on a password at one
stage or another and are unfortunately not yet widely adopted [1]. This
is also largely true with two-factor authentication, where one of the
factors often remains ‘‘something you know’’ – namely a password.
Lastly, some people might argue that they are no longer using
passwords to unlock their phones, make payments, etc., but instead
use a fingerprint or facial scan for identification. However, any service
relying on a fingerprint reader or facial scanner on their phone can
be bypassed by knowing the master code of the phone — this allows
anyone to define a new fingerprint and bypass this security feature.
Therefore, passwords are not dead and will most likely continue to
be used in one way or another for the foreseeable future. It is therefore of utmost importance to strengthen password-based authentication
systems. Of course, this includes safe storage of the password on the
server/device side. Furthermore, passwords selected by users should be
strong in the very first place to ensure the best level of security.
1.2. Law enforcement investigation
The other side of the coin is that security reinforcements and the
ready availability of strong encryption tools can also benefit criminal

∗ Corresponding author.

E-mail addresses: katerina.kanta@port.ac.uk (A. Kanta), iwen.coisel@europol.europa.eu (I. Coisel), mark.scanlon@ucd.ie (M. Scanlon).
https://doi.org/10.1016/j.jisa.2024.103809
2214-2126/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

parameters that were taken into account. Section 4 presents ten different experiments with targeted custom dictionaries compared to popular
baseline dictionaries. Finally, a discussion of the results and possible
avenues for future work are outlined in Section 6.

enterprise and hamper lawful investigation [2]. Law enforcement agencies (LEAs) are nowadays encountering digital evidence in almost all
investigations. An outstanding proportion of offenders, like any other
member of society, have at least a mobile phone and a personal computer. These devices follow the security trends of the manufacturers and
the content is most likely protected with a basic standard of protection
at a minimum. Offenders often take additional security precautions
if they are aware of the risks of investigation — as highlighted in
the latest Internet Organised Crime Threat Assessment (IOCTA) report
from Europol [3]. For example, they might employ additional levels of
encryption over what might be enabled by default, such as full disk
encryption or encrypted communication — again often protected by a
password.
As stated by Plunkett et al. [4], ‘‘the lack of passwords, particularly
during the execution of search warrants, has hindered investigations’’.
It can be crucial to get access to such content during an investigation
— necessitating the retrieval of the suspect’s password(s). Of course,
criminals are not always inclined to share their passwords with the
investigators. It is not always possible to compel the suspect to divulge
his/her passwords through a court order. For example, compelling password surrender could be considered as against the Fifth Amendment
in the USA protecting suspects from self-incrimination [5]. In some
other countries, it is considered a crime to not reveal a password under
court order, e.g., in the United Kingdom within Section 49 of the
Regulation of Investigatory Powers Act 2000. Nevertheless, the suspect
may well decide to not reveal the password if the sentence incurred
is lower than what might be expected should police gain access to the
device(s). In each of these cases, LEAs have no other choice than to
conduct password cracking processes to recover the suspect password
and examine the targeted content [6].
The approach followed by digital investigators is diverse from those
of malicious password crackers. The latter is predominantly interested
in getting one hit to enter into a system under any user’s account, or to
gain access to the maximum of entries in a given dataset. The former
are more interested in one specific user’s account — the one of the
targeted suspect. Therefore, it stands to reason that the cracking process
can benefit from a tailored approach using the available contextual
information of the suspect [7].

2. Background and related work
Human-chosen passwords have been analysed for more than a
decade since the first data breaches including users’ passwords have
occurred. The most famous is the ‘‘Rockyou’’ data breach that happened
in 2009. It is still today widely used in the literature for two main
reasons. Firstly, it was at the time the largest data breach, exposing more than 32 million accounts with approximately 14 million
unique passwords. Secondly, the passwords were not stored safely in
the database, i.e., they were stored in plain text. This presented a
significant advantage for password analysis, as the dataset also contains
examples of the strongest passwords. More recent data breaches have
exposed a significantly larger quantity of users. Plain text passwords
are not typically exposed in more recent leaks and when encrypted
passwords are leaked the strongest passwords in the leak are generally
not retrieved.
2.1. Reinforcing password strength and management
Collected passwords from the aforementioned data breaches are useful for many fields of research, e.g., designing new password cracking
processes, reinforcing password strength meters and models [8,9], etc.
In all cases, data breaches represent a risk to the individual safety of
each user concerned. Nowadays, passwords are almost always safely
stored on the server side — whether they are stored encrypted or
using a cryptographically secure hash function. Consequently, if the
data leaks for whatever reason, undesired users would not immediately
gain access to the password(s) of the user(s). The function used to
safely store the password plays a crucial role at this stage, as it will
directly impact the capacity of the malicious person conducting an
offline attack.
For example, using a modern gaming graphics card, e.g., an Nvidia
GeForce RTX 3090, in conjunction with hashcat1 password cracking
software, a malicious actor can try above 65 × 109 password candidates
per second for a relatively weak hashing function, e.g., MD5. This number drops to around 100,000 per second in the case of a more secure
password hashing function, e.g., Bcrypt (with 32 iterations). While the
latter case offers more resistance to an attacker, a weak password in the
Bcrypt scenario would still be found quicker than a strong one in the
MD5 scenario. It is worth noting that the figures presented above are
for a single graphics card and much faster performance can be achieved
leveraging large GPU clusters, such as those offered by cloud computing
providers, e.g., Amazon Web Services.2
Therefore, in terms of protection from password cracking and increasing security, passwords selected by users should be strong in the
first place. Raising awareness is generally the best method to ensure
that users understand the purpose and the importance of security and
good password selection and management [7]. However, the reality is
that users choose weak passwords and, even worse, often reuse their
passwords [10–12]. This renders a strong password weaker in case it
can be obtained from other sources by an attacker [13], e.g., keyloggers
or phishing attacks. There are solutions to encourage users to not reuse
the same passwords across different services, such as the ‘‘Have I Been
Pwned’’3 service or the automated detection of password reuse in some
Internet browsers.

1.3. Contribution of this work
The work outlined as part of this paper proves that context plays
a role during users’ generation of passwords and can therefore be
exploited by LEAs during their lawful criminal investigation. There is
no dataset available focusing on a single user. As a result, the analysis
outlined below is focused on a community level in order to extrapolate
how likely a contextual-based approach is to succeed. Nevertheless,
the bespoke, context-based approach outlined as part of this paper is
proven to find passwords exclusively recoverable using this technique,
i.e., those that were not found by currently used, generic approaches.
The contribution of this work includes:
• An overview of the experimental methodology used in this paper,
with a breakdown of the parameter definition process from dictionary creation parameters to password cracking tools selected.
• An extensive experimentation and results section, analysing the
approach’s performance across ten datasets of varying topics —
proving the impact of context in password cracking.
• A thorough discussion of the uses, benefits and limitations of the
contextual approach, as well as further optimisation steps in the
future.
The remainder of the paper is organised as follows: Section 2 offers
a brief literature review of related work in password cracking and
dictionary creation. Section 3 presents the experimental methodology
that was followed as part of this paper, with an analysis of the different

1

https://hashcat.net/.
aws.amazon.com.
3
https://haveibeenpwned.com/.
2

2

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

Password policies and/or password strength meters are often deployed to ensure a certain level of security. The most famous one
is most probably the initial password policy designed by NIST in
2013 [14]. It requires passwords to be of a minimum length of 8 and to
include lowercase, uppercase, special characters and digits. Such policies exist in many variations, but they tend to give a false impression
of security. End users also tend to be predictable. For example, if the
previous password was ‘‘password’’, they would simply capitalise the
first letter and add a digit or special character at the end to make it
compliant, namely transform it into ‘‘Password1!’’. Such a password
is compliant to the password policy, yet it is not secure as password
crackers have seen this pattern many times in data leaks of passwords
and have adapted their attacks to mimic popular behaviour.
Password strength meters have therefore evolved to more advanced
techniques, often analysing the inner structure of the password and detecting popular patterns. This is the case for example of the well-known
zxcvbn tool [8] but also of many other academic solutions [9].
Users therefore need to create more complex passwords. Ideally,
they should be long and random, but this makes them more difficult
to remember. As highlighted in the analysis of passwords from 3.9
billion accounts [15], users tend to use password fragments that can
potentially be linked to their context, such as a city or the name of a pet.
Such facts provide a hook to design a context-based password-cracking
attack if an individual is targeted, such as during a digital investigation.

Fortunately or not, humans are often predictable. Human password
selection is far from random and often follows the distribution of
natural language [18]. Password trends are frequently analysed from
leaked datasets and for years, the password appearing most often is
‘‘123456’’.5 A dictionary of the most commonly used passwords can
be designed from these predictable user tendencies. Typical trends,
such as adding a ‘!’ at the end or capitalising the first letter, can be
easily combined with such lists of popular password candidates. These
candidate modifications are referred to as ‘‘mangling rules’’ and can be
generated automatically from data breaches, e.g., using the PACK suite
of tools.6 The input for these tools can be a list of passwords obtained
from one or several data breaches or humanly designed on purpose.
More modern approaches rely on machine learning techniques
analysing a given input set to produce a list of password candidates
which can be combined with mangling rules, similar to dictionarybased approaches [19,20]. Those techniques include Markov-based
models [21], probabilistic context-free grammars [22], and neural
network based attacks [23]. One such example of a neural network
are Generative Adversarial Networks (GANs); where a neural network
is developed to create password candidates that fall as close to the
distribution of real passwords stemming from real-world password
leaks [24].

2.2. Password cracking techniques

Whether considering dictionary-based approaches or machine learning ones, an input dictionary is required. This dictionary can be
straightforwardly a single data breach or a concatenation of several
data breaches. Amassing the largest amount of leaked datasets might
seem like a good idea to detect password reuse. However, it is not
optimal if one considers using mangling rules in combination with such
input. Indeed, those breaches contain entries that, once modified, could
be interpreted as junk. Maintaining a good balance between the size
and the quality of the entries is a valuable activity that will increase
the success rate and especially reduce the time before success. There
are automated tools that help users to automatically sanitise wordlists,
such as the demeuk tool7 from the Netherlands Forensic Institute.
A more fine-grained analysis of the breaches can provide detailed
information about users’ tendencies and draft specific rules to mimic
observed creation patterns. For example, when asked to create a password with lowercase and uppercase letters, users are most likely to
capitalise the first letter of their password [15]. When asked to include
numbers and/or special characters in their passwords, they are very
likely to use number sequences such as ‘123’, number repetitions such
as ‘111’, meaningful numbers such as ‘314’, or use letter substitutions
such as ‘@’ for ‘a’ and ‘1’ for ‘i’ [15]. Automated tools, such as PACK8 or
Pipal9 can analyse datasets to produce a set of realistic mangling rules
to be combined with a dictionary at a later stage.
General trends can be drawn from the analysis of large passwords
datasets. Culture, education, social origin and religion have an impact
during a user’s password selection process. For example, AlSabah et al.
[25] highlighted that Arabic users were three times more likely to
include their mobile phone number in their password, while users
from India and Pakistan were more prone to use names. According
to another study, more than half of Chinese users use passwords only
made of digits [26] and Chinese passwords have a different letter
distribution, structure and semantic patterns compared to their English
counterparts [27]. In a study of the linguistic and cultural impact on
password cracking of three different language spheres, it was shown

2.3. Password statistics

If users generated their passwords following a totally random distribution, no cracking strategy would have an advantage over an exhaustive, brute-force search — where all combinations of allowable
characters of any accepted length are tested. This approach is guaranteed to work; the only unknown variable is how long it will take. Consider the aforementioned gaming card example testing 100,000 Bcrypt
candidates per second, it would take approximately three months to
try all candidates’ passwords made up of digits, special, lowercase
and uppercase characters up to length 6. This raises exponentially to
22 years for length 7, 2 millennia for length 8, etc. As a consequence,
this approach is not preferred and will only be used as a last resort
or if the investigator knows that the length is limited, allowing a full
exploration in a reasonable time.
A time-memory trade-off approach relying on the principle of the
Hellman table [16], Rainbow Tables [17], focused on mitigating the
time required to explore a given space. For the price of some storage
capacities and the pre-computation of the space to be explored, a
password belonging to this given predefined space can be retrieved
in a negligible time compared to the pre-computation step. This is
highly valuable if one knows that several passwords are meant to be
encountered in practice. Many projects have worked collaboratively
to generate such rainbow tables, e.g., the rainbowcrack project4 for
the functions MD5, NTLM and SHA1. While efforts are still made to
improve the performance for generating such rainbow tables, it is
rendered almost useless in the field of password cracking due to the
popular usage of a salt in the storage of passwords. A salt is a random
string concatenated to the password before using it as the input to the
hashing function. Theoretically, rainbow tables could still be built for
salted passwords, but the defined space to explore must incorporate
this salt. There is no fixed length for a salt, but it is generally long
enough, e.g., 32 bits or more, to render rainbow tables no longer
feasible in practice. Using salted passwords has the additional benefit
that two identical passwords should have two different salts (as they
are randomly generated) and will therefore have two different hashes
in the database.

4

5

https://nordpass.com/most-common-passwords-list/.
https://github.com/iphelix/pack.
7
https://github.com/NetherlandsForensicInstitute/demeuk.
8
https://github.com/iphelix/pack.
9
https://digi.ninja/projects/pipal.php.
6

http://project-rainbowcrack.com/table.html.
3

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

Table 1
The ten datasets involved in the experiments.

that when the attackers leveraged the knowledge of each language,
the number of candidates needed to crack 10% of the passwords was
significantly lower [28].
This contextual information, while being personal, can often be
accessed using side sources of information [6,29]. This is particularly
true in the law enforcement scenario; where the investigators often
have some information about the suspect to hand. The target always
remains human and as such, generic approaches should always be tried
first to grab the low-hanging fruit. In a second stage, a more targeted
approach could be considered. Yet, there is no automated manner to
design a contextualised dictionary targeting a user or a community of
users. Below, a method is outlined to generate contextual dictionaries,
and it is demonstrated that it can be useful to crack passwords that
could be missed by generic approaches.

Dataset

Size

Date of breach

AxeMusic
JeepForum
Minecraft
MangaTraders
Wattpad
Battlefield
Wanelo
EverydayRecipes
Zynga
DoSportsEasy

252,752
239.347
143,248
618,237
23,531,304
419,940
2,130,060
25,271
42,908,386
46,113

N/A
November 2017
June 2015
June 2014
June 2020
June 2011
December 2018
N/A
August 2019
N/A

community of users around a specific interest/topic, ten datasets were
chosen from different online community leaks. These datasets come
from hashes.org and their use for the purpose of this research has
been approved y the Office of Research Ethics of [redacted for blind
review]. As can be seen from Table 1, the datasets have been picked to
represent a variable sample of topics and interests. These include data
breaches from forums focused on music, cars, video games, recipes,
and shopping. The datasets are all in English and contain only unique
passwords, there are no repetitions. They are also of various length; the
smallest being approximately 25,000 and the largest being 43 million
— to encompass as big a variance as possible.

3. Experimental methodology
In order to prove the role contextual information can play in password cracking, experimentation and analysis is needed. This section
outlines the methodology used and each of the parameters used as part
of the experimentation, and describes how/why they were chosen.
3.1. Baseline selection
Most dictionary attacks use wordlists that originate from one or
more different data breaches. The passwords in these data breaches
are sometimes leaked in their plaintext form, but more often they are
hashed (and salted). This means that in order to make use of these lists
in dictionary and/or other password cracking attacks, the passwords
need to be cracked first. That is not always possible for 100% of the
leaked data — meaning that it can be argued that some of these cracked
lists do not contain the ‘‘harder to crack’’ passwords.
One list that does not fall in this category is called ‘‘RockYou’’
and contains 32 million passwords that were leaked in 2009. Because
RockYou was leaked in plain text, it is a very popular wordlist that
has been used by many researchers as a way to extract insights on
users’ password habits [30,31] or as a baseline to compare other
attacks against [32]. While the plain text passwords offer a significant advantage compared to an incomplete list from hashed leaks,
one drawback of RockYou is that it was leaked in 2009. Since then,
password policies around the globe have changed and stricter measures
have been adopted — with passwords often needing to be longer
and containing more than one upper/lowercase, number, and symbol
characters. Therefore, in terms of adopting a baseline to compare the
dictionary approach proposed as part of this paper against, RockYou
was evaluated against a more modern dictionary, Ignis-10M.10 Ignis10M contains passwords from various leaks, and statistical analysis
comparing its makeup against RockYou can be found on the project’s
GitHub. Besides Ignis-10M, smaller versions of the Ignis wordlists were
tested, but the 10M version achieved the best results. In terms of
performance, RockYou and Ignis-10M had comparable results, with
Ignis-10M performing slightly better. Therefore, Ignis-10M was chosen
as the baseline for the rest of the experiments outlined below.

3.3. Password cracking methodology
In order to test the effect of context with these ten datasets, a
pipelined approach to context based dictionary creation is adopted, as
can be seen in Fig. 1.
According to this approach, a starting seed word is chosen that
corresponds to a specific article on Wikipedia. For the purpose of this
research, a structured version of Wikipedia, called DBPedia11 is used.
From the starting article in DBPedia, every link in the text of the article
is visited and saved as a new entry in the dictionary list. Initially,
the abstract and/or full article were also scanned and keywords were
extracted, but for the most part these keywords coincided with the
links within the article. For this reason, the decision was made to only
consider the links, as the extraction of keywords from the article did
not result in significant added value.
In the next step, these new DBPedia articles that are saved in
the dictionary list, are then visited in turn and the same process is
performed on them until a given depth from the original link is reached.
For efficiency reasons, each of the explored links are stored in a set to
avoid exploring the same input repeatedly. Subsequently, the links will
be used to generate the dictionary. Some generic sub-strings inserted by
Wikipedia are removed from the entry, e.g., ‘‘List of’’ or ‘‘Category:’’.
If the entry is composed of several words, the entry is saved as such
but also a version without the stop words, if there are any contained
therein.
These dictionary lists are then used as input with the password
cracking tool of choice in order to crack the passwords of the data
sets listed in Table 1. The password candidate creator tool that was
chosen for this set of experiments was Prince,12 as according to previous
work by [32], it had the best performance. Prince is a password
candidate generator that takes a dictionary list as input and outputs
various combinations of the words in the dictionary list. This list of
password candidates is then fed into John The Ripper13 – the tool
that performs the password cracking. To automate this process, the
Password Guessing Framework14 was used.

3.2. Dataset selection
Ideally, when talking about context-based decryption in a digital
forensic setting, experimentation would be conducted on real cases
by a digital investigator focusing on one specific target. But as this
constitutes privileged/sensitive information, it is not possible to do this
in a research context. Therefore, the focus is shifted in this paper to
what is referred to as ‘‘the community approach’’. In order to prove
the importance of contextual information in password cracking using a

11

https://www.dbpedia.org/.
https://github.com/hashcat/princeprocessor.
13
https://www.openwall.com/john/.
14
https://github.com/RUB-SysSec/Password-Guessing-Framework.
12

10

https://github.com/ignis-sec/Pwdb-Public.
4

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

Fig. 1. Pipelined context based dictionary creation.
Table 2
The ten dictionaries produced by DBPedia.

3.4. Parameter optimisation
For the purpose of this paper, ten seed words were chosen in order
to correspond thematically to each leaked dataset, as shown in Table 1.
These seed words can be found in Table 2. It should be noted here
that the seed word for Battlefield is Battlefield_(video_game_series) in
order to represent the video game Wikipedia article, but will be referred
as simply Battlefield for the remainder of the paper. The seeds words
were chosen to be as thematically close to the topic as possible. For
example, for the Zynga leak, the world ‘‘Zynga’’ was also chosen as the
starting point for creating the dictionary. For Wanelo, a leak from a
website about shopping, the word ‘‘shopping’’ was used. As observed
in the table, half of the seed words were chosen to be the same word
as the target dataset, such as ‘‘Minecraft’’ and ‘‘Battlefield’’, while the
other five were chosen to be a generic one-word description/category
of the purpose of the website, such as ‘‘Cooking’’ for EverydayRecipes
and ‘‘Sports’’ for DoSportsEasy. The rationale behind this was to study
whether an identical seed word to the target leak over a generalised
topic would be significant.

Dataset

Seed word

Size

Axemusic
Jeepforum
Minecraft
Mangatraders
Wattpad
Battlefield
Wanelo
EverydayRecipes
Zynga
DoSportsEasy

Music
Car
Minecraft
Manga
Fanfiction
Battlefield
Shopping
Cooking
Zynga
Sports

1,001,173
853,825
243,803
180,641
641,007
415,311
627,487
524,269
443,443
31,918

one. Therefore, selecting a depth of 3 layers is the optimal choice. When
keeping the number of guesses constant across the experimentation, it
is important for the list to be long and detailed enough, but not too
long as to include words that are too thematically distant from the seed
word.
Finally, as can be seen in Table 2, even though each of these datasets
are of depth 3, their size varies according to how many links are
contained in each Wikipedia/DBPedia page visited.

3.4.1. Generated dictionary level depth
The seed words mentioned above were subsequently used with the
methodology outlined above in order to create custom dictionary lists.
One parameter that needs to be defined at this stage, is the depth of
these datasets, i.e., how many layers down from the seed word should
be explored during dictionary creation. For this purpose, multiple
dictionary lists for each seed word were generated; ranging from one
layer to three layers, and in some cases four layers. Of course, the time
to generate these dictionaries depends on the number of links in each
level. The latency of the Internet connection has a significant impact
on the speed to gather the pages from the online version of DBPedia.
As indicative durations, Layer 1 is almost instantaneous, Layer 2 took
∼20 s, Layer 3 took ∼30 m and Layer 4 took approximately ∼1 day.
The performance of these varying layer depths was assessed for
a selection of the aforementioned datasets and it was found that the
dictionary lists produced by only 1 or 2 layers achieved lacklustre
performance. For example, with the experimentation using the Wattpad
leak, the custom 2 layer dictionary cracked 1.6% of the total passwords,
while the 3 layer dictionary to cracked 42.1%.
A 4 layer dictionary was produced with the ‘‘Manga’’ seed word.
This was used with the leak from Mangatraders, and it was still found
the 3 layer dictionary performed better than the 4 layer one. More
specifically, the 3 layer dictionary found 57.2% of the passwords,
while the 4 layer one found 34.4%. This is due to smaller dictionaries
facilitating more mangling for a fixed number of guesses than a larger

3.4.2. Password mangling rules
As mentioned in Section 2, password mangling rules are set during
password cracking processes in order to imitate real users’ password
habits. For example, adding numbers or symbols at the end of a
chosen password when the corresponding password policy requires
them. These are generally useful and should be tailored according to the
target. For the experiments outlined as part of this paper, the default
mangling rules of John the Ripper were used on both the contextual
dictionaries and the baseline dictionary.
3.4.3. Number of guessing attempts
When it comes to password cracking, the time taken to explore
the password search space defined is directly related to the number of
attempts permitted during the cracking phase’s execution. Despite the
brute-force cracking mantra of every password being crackable given
enough time, this is realistically impractical in real-world scenarios.
With a reduced search space and using a non-brute-force technique,
more attempts will crack more passwords and/or have a higher likelihood of cracking a specific password — but at the expense of time and
resources. As a result, password cracking typically requires a reasonable
limit for the number of attempts to be decided upon.
5

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

Table 3
Total passwords cracked and improvement of the combination approach. The L_3 Excl.
column contains passwords found only by L_3 dictionaries, while the L_3 Imp. column
contains the improvement over Ignis-10M provided by the L_3 dictionaries.

In order to decide on the number of attempts to limit each experiment presented as part of this paper, a number of options were
evaluated. To overcome the difference in dictionary sizes generated
for a specific topic and/or generated dictionary level, a fixed size of
guessing attempts was selected after experimentation and this was 10
billion. A lower number of guessing attempts produced worse results
for both the baseline dictionary and the contextual dictionaries. On
the other hand, more guessing attempts did result in more cracked
passwords, but the trade-off between the additionally found passwords
and the running time of the cracking process was deemed inefficient
for the purposes of this paper.
4. Results
As mentioned in Section 3, ten datasets across various topics were
chosen. Related contextual dictionaries with the help of Wikipedia/
DBPedia were created. The depth of the dictionaries was selected as
3 and the number of guesses as 10 billion, as defined in the previous
section. The cracking process over time for these ten datasets, with both
the baseline dictionary (Ignis-10M) and the contextual dictionary, can
be seen in Fig. 2.
From Fig. 2, it can be observed that for all ten datasets, Ignis-10M is
the best performing dictionary. This is to be expected, as Ignis-10M is
a compilation of different data leaks and contains some of the most
popular passwords used by real-world users. Ignis-10M is also a 10
million entry dictionary, while the contextual dictionaries, as seen in
Table 2 range from 1 million to 30 thousand candidates. It is therefore
expected that Ignis-10M will perform better in comparison, and it will
crack the most passwords across all different datasets as it is the most
varied dictionary.
Focusing a little more into the varying results of the ten different
contextual dictionaries (denoted as L_3s), it can be seen in Fig. 2
that Music_3 and Car_3 had some of the best performances, while
Sports_3 had the worst. This can also be explained by the size of these
dictionaries, with Music_3 being 1 million while Sports_3 is only 30
thousand. A dictionary of 30k candidates, even with the permutations
allowed by 10 million guesses, cannot produce enough variance. This
serves to highlight the importance of picking the correct seed word
for generating a dictionary. The layer 3 dictionary Sport_3 was also
generated, and it contained 1,068,758 candidates, which is a very
significant increase over Sports_3. If ‘‘Sport’’ was used as a seed word
instead of ‘‘Sports’’, better results would be achieved when cracking
DoSportsEasy, but the decision was made to use ‘‘Sports’’ as part of the
experimentation outlined in this paper, to demonstrate the pitfalls of
picking a bad seed word.
One interesting metric when it comes to the performance of these
contextual dictionaries is how well they would do ‘‘stacked’’, i.e., in
a combination attack. To this end, Table 3 shows the percentage
of unique passwords cracked by Ignis-10M and the L_3 contextual
dictionaries. Column 3 of the table also presents the percentage of
passwords that were only cracked with the L_3 dictionaries for each of
the ten cases, i.e., the exclusively cracked passwords. Finally, Column
4 presents the improvement over Ignis-10M if it is combined with the
contextual approach.
As can be seen in Table 3, in most cases the contextual dictionary has found approximately half the passwords found by Ignis-10M.
Although in some cases, like JeepForum and EverydayRecipes, this
number is even higher. Considering that Ignis-10M is compiled by
a number of different data leaks and therefore contains actual used
passwords across a range of services, the results of the L_3 dictionaries
that are only dictionary words without any extra modification, is
quite impressive. Once again, the only outlier is Sports_3, but this is
somewhat expected since the input dictionary that was created from
DBPedia contained only 30 thousand candidates. The passwords found
exclusively by the contextual dictionaries offer on average an additional 2% of passwords, which in some cases represents a significant
improvement over what was found by Ignis-10M alone.

Dataset

Ignis-10M

L_3

L_3 Excl.

L_3 Imp.

Axemusic
Jeepforum
Minecraft
Mangatraders
Wattpad
Battlefield
Wanelo
EverydayRecipes
Zynga
DoSportsEasy

41.3%
68%
38.4%
57.2%
39.7%
60.6%
42.1%
64.4%
37.9%
41.7%

20.5%
39.2%
11.2%
28.2%
15.2%
29%
19.3%
36.7%
15.7%
1%

2.47%
2.32%
0.76%
2.61%
0.69%
2.21%
2.38%
2.24%
1.22%
0.06%

5.97%
5.19%
3.88%
4.56%
17.86%
3.64%
5.64%
3.47%
10.61%
0.15%

Table 4
Class 3 passwords. The L_3 Excl. column contains passwords found only by L_3
dictionaries, while the L_3 Imp. column contains the improvement over Ignis-10M
provided by the L_3 dictionaries.
Dataset

Ignis-10M

L_3

L_3 Excl.

L_3 Imp.

Axemusic
Jeepforum
Minecraft
Mangatraders
Wattpad
Battlefield
Wanelo
EverydayRecipes
Zynga
DoSportsEasy

4504
3770
3247
24,524
223,567
17,330
47,604
254
417,404
934

581
276
80
1906
16,854
755
3855
41
33,752
6

286
118
46
942
7758
281
1709
24
15,735
1

6.3%
3.1%
1.5%
3.8%
3.5%
1.6%
3.6%
9.4%
3.8%
0.1%

For example, with the Wattpad leak, while the passwords found exclusively by Fanfiction_3, represent a 0.69% increase, this translates to
a 17.86% improvement over Ignis-10M. The reason for this is that while
Ignis-10M finds more passwords, these are passwords that are repeated
many times in the leak, while the passwords found by Fanfiction_3 do
not have as many repetitions. This could indicate that the passwords
found by Fanfiction_3 are less frequently chosen by users and therefore
less encountered.
It can also be observed that the choice of either generalising the
seed word or keeping it the same as the target dataset did not influence
the results in any significant manner. Nonetheless, from the five best
performing dictionaries, four were from the ‘‘generalised seed word’’
category.
Looking at the case of a single law enforcement officer wanting to
gain access to an encrypted device, the number of popular passwords
cracked from one data leak is not the optimal way to judge the
effectiveness of a dictionary. In fact, if a suspect is hiding behind an
encrypted device, it is reasonable that they are more tech-savvy, and
it follows that there is a good chance that their password would be
stronger than those found on the most popular password lists.
It is therefore important to also look at the quality of passwords
cracked by the baseline dictionary and the contextual approach, i.e., the
strength of these cracked passwords. To this end, Fig. 3 shows the
breakdown of the found passwords by both approaches, classified into
five classes of strength. The password strength meter used for this classification is zxcvbn and the five classes range from 0 to 4, with Class
0 being the weakest passwords and Class 4 containing the strongest.
As can be seen in Fig. 3, Class 1 and Class 2 passwords are those
most commonly found, mostly from both Ignis-10M and the contextual
dictionaries. This is because the passwords in these categories are easier
to crack and would most likely be found by various approaches, as
confirmed by [15]. It is therefore the Class 3 and Class 4 passwords
that are the most interesting.
Tables 4 and 5 outline the passwords found by Ignis-10M, L_3, the
exclusively retrieved by L_3, and the improvement percentage for Class
3 and Class 4 passwords cracked. By examining these two tables, it is
6

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

Fig. 2. Passwords cracked by Ignis-10M and bespoke layer 3 dictionaries (Seed word for bespoke dictionary in parentheses).

5. Discussion

notable that on average, approximately half the passwords found by
L_3, are not found by Ignis-10M, cementing the importance of the proposed contextual dictionaries further. Furthermore, it can be observed
that although the numbers are smaller compared to Class 3, Class 4
contains the strongest passwords and the percentage improvement of
using L_3 on top of Ignis-10M is higher for Class 4. Notable examples
are Fanfiction and Music achieving a 4.9% and 7.7% improvement
respectively, In addition, for EverydayRecipes the percentage improvement is 42.8%, while acknowledging that the absolute numbers of
recovered passwords are quite low for both.

As the experiments of the previous section demonstrated, the added
value of considering context in password cracking is evident. In a
community-based approach, a bespoke, targeted dictionary can provide
a significant increase in the number of found passwords and can be
adopted ahead or alongside the baseline in the password cracking
pipeline. Of course, one scenario that cannot be tested is that of a
single suspect and their seized encrypted device(s). In such a case, a
bespoke dictionary whose parameters can be tweaked and tailored to
7

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

Fig. 3. zxcvbn classification of passwords cracked by Ignis-10M and bespoke layer 3 dictionaries (Seed word for bespoke dictionary in parentheses).

6. Conclusion and future work

the suspect can be created with ease using the proposed methodology
and procedure as described in this paper. This would result in the
investigator easily having the means to produce a custom dictionary,
or dictionaries, for a specific case. When racing against the clock or
when an encrypted device presents the largest roadblock in an ongoing
case, contextual dictionaries tailored to the suspect at hand could prove
invaluable to progressing an investigation.

This paper offers an extensive experiment using ten different data
sets of ten varying topics to provide a definite proof of the value
of considering contextual information in password cracking. Humans
are creatures of habit, and that is no different in their password
selection process — where they often choose familiar words that are
8

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.
Table 5
Class 4 passwords. The L_3 Excl. column contains passwords found only by L_3
dictionaries, while the L_3 Imp. column contains the improvement over Ignis-10M
provided by the L_3 dictionaries.
Dataset

Ignis-10M

L_3

L_3 Excl.

L_3 Imp.

Axemusic
Jeepforum
Minecraft
Mangatraders
Wattpad
Battlefield
Wanelo
EverydayRecipes
Zynga
DoSportsEasy

351
96
667
4554
15,022
2487
2953
7
28,211
60

42
9
5
152
1095
51
199
3
1403
0

27
5
3
90
673
25
100
3
849
0

7.7%
5.2%
0.4%
1.9%
4.9%
1.0%
3.4%
42.8%
3.0%
0%

reinvigorate the progress when relevancy declines. For example, during
the exploration of layer 3, some candidates might be already too
thematically distant from the seed word. If these candidates can be
disregarded at this stage, they could give way to an exploration of
deeper relevant layers rather than shallower irrelevant entries. As a
result, the end dictionary may be the same length as layer 3, but would
contain more relevant entries.
Finally, two other avenues to be explored for optimising the results
of the candidate generation process are searching backwards during
the link exploration process and ranking the dictionary entries. For
example, if the seed word is ‘‘Manga’’, backward searching would
include Layer -1, that is, all the pages on Wikipedia that link to Manga.
This could provide a substantial addition of very relevant password
candidates. As part of future work, a system for keeping track of how
many times each entry was found could help indicate how relevant it
is to the seed word and could be used to rank the resultant list.

more easily remembered. This information can be leveraged in an
investigation, and the ability to exploit it could prove invaluable during
an investigation.
Of course, a contextual dictionary based around a single seed word
cannot compete on equal footing with a 20 to 300 times larger and
more well-rounded dictionary like Ignis-10M when the objective is
to crack as many passwords as possible. This means that when no
information is known about the target or the goal is to gain access
to a system by cracking the password of any user and not a specific
one, using a dictionary like Ignis-10M would provide a higher chance
of success.
If the usage scenario surrounds a single case and/or a single password and information can be determined about its owner and their
interested, then the contextual approach can be utilised. This would
make even more sense, considering that in digital cases, suspects might
be more likely to ‘‘try’’ harder to conceal their tracks and therefore
would choose their password with more prudence.
The most notable improvement when it comes to the results of the
contextual dictionaries is that for Class 3, and more importantly, Class
4, the number of extra passwords cracked with the L_3 dictionaries
offers a significant improvement over the baseline. The extra passwords
that were cracked not only lend credit to a combination approach, but
also showcase further that a smaller dictionary built around one seed
word related to the target data leak can indeed boost the number of
cracked passwords significantly.
Something else to note, regarding the poor performance of Sports_3
compared to both Ignis-10M and the other L_3 dictionaries, is that
when the size of the produced dictionary is too small, more layers or a
different seed word (one with a more detailed Wikipedia page, hence
more links) should be considered.

CRediT authorship contribution statement
Aikaterini Kanta: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization,
Writing – original draft. Iwen Coisel: Conceptualization, Methodology,
Resources, Supervision, Validation, Writing – review & editing. Mark
Scanlon: Methodology, Project administration, Resources, Supervision,
Validation, Writing – review & editing.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Data availability
The data that has been used is confidential.
References
[1] Kennison SM, Chan-Tin DE. Predicting the adoption of password managers: A
tale of two samples. Technol Mind Behav 2021.
[2] Sayakkara A, Le-Khac N-A, Scanlon M. Electromagnetic side-channel attacks:
Potential for progressing hindered digital forensic analysis. In: Companion
Proceedings for the ISSTA/ECOOP 2018 Workshops. ISSTA ’18, New York, NY,
USA: Association for Computing Machinery; 2018, p. 138–43. http://dx.doi.org/
10.1145/3236454.3236512.
[3] Europol. Internet Organised Crime Threat Assessment (IOCTA) 2023. Tech. rep.,
Publications Office of the European Union, Luxembourg; 2023.
[4] Plunkett J, Le-Khac N-A, Kechadi T. Digital forensic investigations in the cloud:
A proposed approach for Irish law enforcement. Tech. rep., Science Foundation
Ireland; 2015.
[5] Atwood JR. The encryption problem: Why the courts and technology are creating
a mess for law enforcement. In: Saint Louis University Public Law Review. Vol.
34, 2015.
[6] Kanta A, Coisel I, Scanlon M. Harder, better, faster, stronger: Optimising the performance of context-based password cracking dictionaries. Forensic Sci Int: Digit Investig 2023;44:301507. http://dx.doi.org/10.
1016/j.fsidi.2023.301507, URL: https://www.sciencedirect.com/science/article/
pii/S2666281723000082. Selected papers of the Tenth Annual DFRWS EU
Conference.
[7] Kanta A, Coisel I, Scanlon M. A survey exploring open source intelligence for
smarter password cracking. Forensic Sci Int: Digit Investig 2020;35:301075.
http://dx.doi.org/10.1016/j.fsidi.2020.301075.
[8] Wheeler DL. Zxcvbn: Low-budget password strength estimation. In: 25th USENIX
Security Symposium. USENIX security 16, 2016, p. 157–73.
[9] Galbally J, Coisel I, Sanchez I. A new multimodal approach for password
strength estimation—Part I: Theory and algorithms. IEEE Trans Inf Forensics
Secur 2017;12(12):2829–44. http://dx.doi.org/10.1109/TIFS.2016.2636092.
[10] Bonneau J, Herley C, Oorschot PCv, Stajano F. The quest to replace passwords: A
framework for comparative evaluation of web authentication schemes. In: 2012
IEEE Symposium on Security and Privacy. 2012, p. 553–67. http://dx.doi.org/
10.1109/SP.2012.44.

6.1. Future work
As part of future work, more consideration will be given to refining
the generated dictionaries. Although 3 layer dictionaries provided the
best results in 9 out of 10 experiments, this was not the case for the
10th. In a scenario with the bespoke dictionary is too short, additional
seed words could be provided (or the wordlist expanded using a large
language model [33]), and the resultant dictionaries could be merged,
additional layers could be used, or a combination thereof. Moreover,
these dictionaries are exactly that, dictionaries, whereas Ignis-10M
contains real-world passwords. It is therefore important to look into
ways of transforming the dictionaries into password candidates, with
the help of well-refined mangling rules, that could better imitate the
behaviour of users when they choose their password.
Another crucial step in that regard is to fine-tune the sanitisation
process for the dictionary words. For example, a link could contain
more than one word. Therefore, tweaking the manner in which these
are combined could result in better password candidates. In this vein,
trimming some branches during the dictionary generation process can
9

Journal of Information Security and Applications 84 (2024) 103809

A. Kanta et al.

[23] Melicher W, Ur B, Segreti SM, Komanduri S, Bauer L, Christin N, Cranor LF.
Fast, lean, and accurate: Modeling password guessability using neural networks.
In: Proceedings of the 25th USENIX Security Symposium. SEC ’16, USA: USENIX
Association; 2016, p. 175–91.
[24] Hitaj B, Gasti P, Ateniese G, Perez-Cruz F. PassGAN: A deep learning approach
for password guessing. In: Applied Cryptography and Network Security. Springer;
2019, p. 217–37.
[25] AlSabah M, Oligeri G, Riley R. Your culture is in your password: An analysis
of a demographically-diverse password dataset. Comput Secur 2018;77:427–41.
http://dx.doi.org/10.1016/j.cose.2018.03.014.
[26] Liu Z, Hong Y, Pi D. A large-scale study of web password habits of Chinese
network users. J Softw 2014;9:293–7.
[27] Wang D, Wang P, He D, Tian Y. Birthday, name and bifacial-security: Understanding passwords of Chinese web users. In: 28th USENIX Security Symposium.
USENIX Security 19, 2019, p. 1537–55.
[28] Mori K, Watanabe T, Zhou Y, Hasegawa AA, Akiyama M, Mori T. Comparative analysis of three language spheres: Are linguistic and cultural
differences reflected in password selection habits? IEICE Trans Inf Syst
2020;103(7):1541–55.
[29] Keküllüoğlu D, Magdy W, Vaniea K. From an authentication question to a
public social event: characterizing birthday sharing on twitter. Proceedings of
the International AAAI Conference on Web and Social Media 2022;16(1):488–99.
http://dx.doi.org/10.1609/icwsm.v16i1.19309, URL: https://ojs.aaai.org/index.
php/ICWSM/article/view/19309.
[30] Bonneau J. The science of guessing: Analyzing an anonymized corpus of 70
million passwords. In: 2012 IEEE Symposium on Security and Privacy. 2012, p.
538–52. http://dx.doi.org/10.1109/SP.2012.49.
[31] Kelley PG, Komanduri S, Mazurek ML, Shay R, Vidas T, Bauer L, Christin N,
Cranor LF, Lopez J. Guess again (and again and again): Measuring password
strength by simulating password-cracking algorithms. In: 2012 IEEE Symposium
on Security and Privacy. 2012, p. 523–37. http://dx.doi.org/10.1109/SP.2012.
38.
[32] Kanta A, Coisel I, Scanlon M. PCWQ: A framework for evaluating password
cracking wordlist quality. In: The 12th EAI International Conference on Digital
Forensics and Cyber Crime. ICDF2C ’21, New York, NY, USA: Springer; 2021.
[33] Scanlon M, Breitinger F, Hargreaves C, Hilgert J-N, Sheppard J. ChatGPT for digital forensic investigation: The good, the bad, and the unknown. Forensic Sci Int:
Digit Investig 2023;46:301609. http://dx.doi.org/10.1016/j.fsidi.2023.301609,
URL: https://www.sciencedirect.com/science/article/pii/S266628172300121X.

[11] Florencio D, Herley C. A large-scale study of web password habits. In: Proceedings of the 16th International Conference on World Wide Web. WWW ’07,
New York, NY, USA: Association for Computing Machinery; 2007, p. 657–66.
http://dx.doi.org/10.1145/1242572.1242661.
[12] Stobert E, Biddle R. The password life cycle: User behaviour in managing
passwords. In: 10th Symposium on Usable Privacy and Security. SOUPS 2014,
Menlo Park, CA: USENIX Association; 2014, p. 243–55.
[13] Wash R, Rader E, Berman R, Wellmer Z. Understanding password choices: How
frequently entered passwords are re-used across websites. In: Twelfth Symposium
on Usable Privacy and Security. SOUPS 2016, Denver, CO: USENIX Association;
2016, p. 175–88.
[14] Burr WE, Dodson DF, Polk WT. NIST Special Publication 800-63 - Electronic
Authentication Guideline. Tech. rep., National Institute for Standards and
Technology; 2004.
[15] Kanta A, Coray S, Coisel I, Scanlon M. How viable is password cracking in digital
forensic investigation? Analyzing the guessability of over 3.9 billion real-world
accounts. Forensic Sci Int: Digit Investig 2021.
[16] Hellman M. A cryptanalytic time-memory trade-off. IEEE Trans Inform Theory
1980;26(4):401–6. http://dx.doi.org/10.1109/TIT.1980.1056220.
[17] Oechslin P. Making a faster cryptanalytic time-memory trade-off. In: Boneh D,
editor. Advances in Cryptology - CRYPTO 2003. Berlin, Heidelberg: Springer
Berlin Heidelberg; 2003, p. 617–30.
[18] Bonneau J, Shutova E. Linguistic properties of multi-word passphrases. In:
Proceedings of the 16th International Conference on Financial Cryptography
and Data Security. FC ’12, Berlin, Heidelberg: Springer-Verlag; 2012, p. 1–12.
http://dx.doi.org/10.1007/978-3-642-34638-5_1.
[19] Du X, Hargreaves C, Sheppard J, Anda F, Sayakkara A, Le-Khac N-A, Scanlon M.
SoK: Exploring the state of the art and the future potential of artificial intelligence in digital forensic investigation. In: Proceedings of the 15th International
Conference on Availability, Reliability and Security. Association of Computing
Machinery; 2020, http://dx.doi.org/10.1145/3407023.3407068.
[20] Kanta A, Coisel I, Scanlon M. A novel dictionary generation methodology for
contextual-based password cracking. IEEE Access 2022;10:59178–88. http://dx.
doi.org/10.1109/ACCESS.2022.3179701.
[21] Narayanan A, Shmatikov V. Fast dictionary attacks on passwords using timespace tradeoff. In: CCS 2005 - Proceedings of the 12th ACM Conference on
Computer and Communications Security. 2005, p. 364–72. http://dx.doi.org/10.
1145/1102120.1102168.
[22] Weir M, Aggarwal S, Medeiros Bd, Glodek B. Password cracking using probabilistic context-free grammars. In: 2009 30th IEEE Symposium on Security and
Privacy. 2009, p. 391–405. http://dx.doi.org/10.1109/SP.2009.8.

10