Evaluating the Influence of Multi-Factor Authentication and Recovery
Settings on the Security and Accessibility of User Accounts
Andre B
¨
uttner
a
and Nils Gruschka
b
University of Oslo, Gaustadall
´
een 23B, 0373 Oslo, Norway
Keywords:
Authentication, MFA, Security, Accessibility, Account Access Graphs.
Abstract:
Nowadays, most online services offer different authentication methods that users can set up for multi-factor
authentication but also as a recovery method. This configuration must be done thoroughly to prevent an ad-
versary’s access while ensuring the legitimate user does not lose access to their account. This is particularly
important for fundamental everyday services, where either failure would have severe consequences. Neverthe-
less, little research has been done on the authentication of actual users regarding security and the risk of being
locked out of their accounts. To foster research in this direction, this paper presents a study on the account
settings of Google and Apple users. Considering the multi-factor authentication configuration and recovery
options, we analyzed the account security and lock-out risks. Our results provide insights into the usage of
multi-factor authentication in practice, show significant security differences between Google and Apple ac-
counts, and reveal that many users would miss access to their accounts when losing a single authentication
device.
1 INTRODUCTION
Online services play an ever-increasingly important
role in our digital world. We use them in many as-
pects of our private and business lives, like finance,
transport, communication, entertainment, etc. Big
tech companies—like Google and Apple—provide us
with different services, including cloud storage, de-
vice management, payment, or single sign-on (SSO)
to other online services. We highly depend on these
services for more and more everyday tasks, and their
availability and security are more crucial than ever.
Most online services require an account, and the
user must authenticate to give proof of the owner-
ship of their account. Authentication is still done
mainly through passwords, although it is commonly
known that passwords are a relatively weak method
to protect a user account (Taneski et al., 2019). This
has been confirmed in many studies, highlighting the
problems with simple passwords and the reluctance
to use password managers (Hayashi and Hong, 2011;
Shen et al., 2016; Davis et al., 2022). Therefore,
multi-factor authentication (MFA) is usually offered
to increase security and make compromising an ac-
count by a malicious user significantly harder. With
a
https://orcid.org/0000-0002-0138-366X
b
https://orcid.org/0000-0001-7360-8314
MFA, the user needs to provide, in addition to the
password, one
1
or more authentication factors, such
as a one-time password (OTP) from an authenticator
app or sent via SMS or a security key. Despite the
apparent advantage, many users consider MFA incon-
venient and, therefore, refrain from activating it (Das
et al., 2019). However, an increasing number of ser-
vices make MFA mandatory and enforce configuring
a second authentication factor. In addition, there is a
movement towards passwordless authentication with
the use of FIDO2 passkeys (Google, 2022), but, at the
time of writing, it is not widely adopted yet.
Suppose the MFA methods are not accessible, or
users forget their password (or lose access to their
password storage). In that case, most online services
provide additional authentication factors, so-called
account recovery mechanisms. Typical examples of
these mechanisms are a link or code sent via email
or SMS. However, such account recovery can be ex-
ploited if they are less secure than the main authen-
tication, e.g., if a recovery email address is not pro-
tected sufficiently (Li et al., 2018). Also, a recovery
factor might be bound to the same device as the main
authentication. For example, the phone receiving the
recovery SMS might be the same as the phone used
1
In this case, also called two-factor authentication
(2FA). In this paper, we will generally use the term MFA.
Büttner, A. and Gruschka, N.
Evaluating the Influence of Multi-Factor Authentication and Recovery Settings on the Security and Accessibility of User Accounts.
DOI: 10.5220/0012319000003648
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 10th International Conference on Information Systems Security and Privacy (ICISSP 2024), pages 691-700
ISBN: 978-989-758-683-5; ISSN: 2184-4356
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
691
for generating the MFA OTPs, which is certainly not
an uncommon combination on a smartphone. Statis-
tics show that smartphones often get stolen, damaged,
or lost (Bitkom, 2021; Beatriz Henr
´
ıquez, 2022),
which can (in addition to the financial loss) make con-
nected accounts inaccessible. Many providers warn
users if the authentication is insecurely configured
(e.g., no MFA activated or weak password). Still, they
can not derive which device the configured factors are
bound to. Furthermore, little research was done on
this issue.
To tackle this research question, we conducted
an online survey with 185 test participants to study
the authentication configuration of their Google and
Apple accounts
2
. To analyze the participants’ re-
sponses, we extended and applied Account Access
Graphs (AAG) (Hammann et al., 2019; P
¨
ohn et al.,
2022), a new approach to evaluate MFA and recovery
configurations. The analysis revealed insights into the
account security and accessibility “in the wild”. The
main contributions of this paper are thereby as fol-
lows:
1. An improved accessibility scoring for AAG mod-
els with higher practical significance.
2. A security and accessibility analysis of the au-
thentication setups by actual Apple and Google
users.
The remainder of this paper is structured as fol-
lows. Section 2 presents related work on MFA and
AAGs. In Section 3, the methodology of AAGs is de-
scribed, as well as our security and accessibility scor-
ing. Afterwards, we describe our study design in Sec-
tion 4. In Section 5, we present our study results. Sec-
tion 6 discusses the limitations, our new accessibility
scoring, and the implications of our results on MFA.
The paper is finally concluded in Section 7.
2 RELATED WORK
A study from 2015 analyzed how widely MFA au-
thentication is adopted among Gmail users (Petsas
et al., 2015). Within their test set, 6.39% of all users
had an MFA method enabled, about 62% had config-
ured a phone for account recovery, and 17% had pro-
vided a recovery email address. This gives a good pic-
ture of the adoption of MFA at the time of the study,
but it might have changed significantly throughout the
years. Their study is similar to ours regarding the col-
lected data, i.e., enabled MFA and recovery methods.
2
We selected Google and Apple, as they are two of the
largest online services and offer a large variety of authenti-
cation configuration options.
However, they obtained the data by using the Google
Password Reminder feature with leaked email lists
while we conducted an anonymous survey with the
participants’ consent.
Much research was done on the authentication and
account recovery procedures offered by online ser-
vices. A study by Gavazzi et al. (Gavazzi et al., 2023)
investigated how popular websites support MFA and
risk-based authentication (RBA). They found that
among the 208 websites they tested, about 42% pro-
vided MFA, and about 22% applied RBA. Amft et al.,
for example, evaluated the account recovery of 1303
websites based on their documentation and tested the
actual recovery procedure on 71 of these websites. By
this, they discovered many insecure procedures, even-
tually allowing them to bypass MFA authentication
through recovery and significant deviations between
the documentation and the actual procedure (Amft
et al., 2023). Another study analyzed usability flaws
in account recovery when MFA is enabled on 78 pop-
ular websites (Gerlitz et al., 2023). Our paper, in con-
trast, focuses on which authentication and recovery
methods provided by a service are adopted by users.
Other work has been done to compare different
authentication methods based on security, privacy,
and usability properties. The conclusion is that it
is not feasible to find authentication schemes that
perfectly satisfy all these properties (Ometov et al.,
2018). More recent research has conducted a formal
analysis of multi-factor authentication methods pro-
vided by Google, revealing some weaknesses in their
protocols (Jacomme and Kremer, 2021).
Furthermore, different studies exist that compare
the usability of MFA methods. For example, Reese
et al. have looked at the perceived usability of five
different MFA methods in terms of usability and their
efficiency (Reese et al., 2019). In another study, the
time used on MFA methods has been compared in two
universities, showing that this can take up a significant
amount of working hours over a long time (Reynolds
et al., 2020).
Account Access Graphs (Hammann et al., 2019;
P
¨
ohn et al., 2022) is a methodology to analyze authen-
tication systems systematically. It allows the mod-
eling of authentication and recovery methods of on-
line services and the dependencies between different
user accounts. Further, it is possible to define ded-
icated scoring schemes to evaluate various aspects
of an account, such as security and accessibility. It
was shown that this could even become a practical
tool for users to increase awareness about their ac-
count use and to support them in making it more se-
cure (P
¨
ohn et al., 2022). Moreover, AAGs were used
in a lab study to observe patterns in account setups
ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy
692
Table 1: Security scores assigned to different example au-
thentication methods.
Score Category Authentication Methods
High Hardware-based Security Key, Smart Card
Medium Software-based SMS Code, OTP Apps
Low Knowledge-based Password, PIN
(Hammann et al., 2022). The methodology used in
this work mostly leans on the foundation provided in
(P
¨
ohn et al., 2022).
In this paper, we have enhanced the methodol-
ogy of AAGs with an improved accessibility scoring
scheme and applied it to actual user accounts. By au-
tomatically creating graphs of users, we show that an-
alyzing user accounts with AAGs can be scalable for
research on larger populations.
3 ACCOUNT ACCESS GRAPHS
In an Account Access Graph as presented in (P
¨
ohn
et al., 2022), authentication methods and accounts are
modeled as nodes in a graph. Account nodes are il-
lustrated as rectangles, and authentication nodes as
rounded rectangles. These nodes can be connected
through intermediary nodes representing logical op-
erators, including conjunctions ’&’ and disjunctions
|’, depicted as circles. For instance, MFA can be
modeled as a conjunction of two or more authentica-
tion methods. Recovery or alternative authentication
mechanisms are modeled as a disjunction. Graphs can
be created for the entire authentication system of an
online service or the account setup of a specific user
and can be used for qualitative and quantitative anal-
ysis.
In this paper, we use AAGs to evaluate the security
and accessibility of actual user accounts. The respec-
tive scoring schemes are described below.
3.1 Security Scoring
As suggested in (Hammann et al., 2019), there are
different possibilities to assess the security in an
AAG. We adopted a scoring scheme inspired by the
level of assurance as found in standards like NIST
(Grassi et al., 2020) or the EU electronic identifica-
tion and trust services (eIDAS) (European Commis-
sion, 2023), which is a well-established method to
evaluate authentication methods. Consequently, we
chose to use an ordinal scale including low, medium
and high. An overview of how security values are as-
signed to different authentication methods is given in
Table 1.
H
&
M
|
M
Account
L
Password
M
SMS
H
Security Key
Figure 1: Example graph for showing how security scores
are calculated. The scores are indicated as L (low), M
(medium), and H (high).
The security scores of parent nodes are calculated
as follows. For the logical conjunction, the score
is calculated using the maximum score of the child
nodes. This is because all child nodes must be ac-
cessed to access this node, so the highest score defines
the security level. The score for a logical disjunction
is calculated by the minimum score of the children,
as accessing the weakest of the children is sufficient
to access a node. Figure 1 shows an example for cal-
culating security scores.
Some nodes in AAGs might represent other ac-
counts, such as recovery email addresses. Ideally,
the scores for these nodes should be derived from the
respective AAGs specifically for that email account.
However, since we could not consider all possible
email accounts used, we opted for a worst-case calcu-
lation and assigned the value low to those leaf nodes.
3.2 Accessibility Scoring
Users might lose their credentials or a device required
for MFA authentication and be locked out of their
accounts. Account accessibility scores indicate how
many options users have to access their accounts.
Consequently, a low accessibility score implies a high
risk and a high score means a low risk of losing access
to an account.
The accessibility scoring, as proposed in (P
¨
ohn
et al., 2022), provides a first estimation of whether the
accessibility of a user account is stronger or weaker
compared to others. However, it is not precise enough
to give practical conclusions. We, therefore, propose
an improved accessibility scoring.
The basic idea behind our improved scoring re-
mains the same. All authentication methods have a
physical “representation”, e.g., passwords might be
stored on a computer or remembered in memory, and
Evaluating the Influence of Multi-Factor Authentication and Recovery Settings on the Security and Accessibility of User Accounts
693
SMS
OTPPassword
&
|
Account
Memory Tablet Phone
Figure 2: Example graph with access methods.
SMS retrieval requires a phone to which it is sent.
These elements are represented as access method in
the AAG graph and are connected to the respective
authentication node.
In contrast to the other nodes in an AAG, access
methods may have more than one parent node (as
shown in Figure 2). If multiple different access meth-
ods can access an authentication method, it means that
either one can access it. As a consequence, these ac-
cess methods have a disjunctive relation. To assess
the accessibility of a specific AAG, we use the logical
formula based on the nodes in the graph.
(Memory (Tablet Phone)) Phone (1)
(Memory Tablet) (Memory Phone) Phone (2)
(Memory Tablet) Phone (3)
As an illustrating example, the equations above
show the boolean terms describing the accessibility
of the instance in Figure 2. First, we derive the term
from the leaf nodes of the graph and replace them with
the respective access methods (1). We then transform
the term (2) to get a complete list of possible combi-
nations of access methods for accessing that account.
Finally, we reduce the term so only those access meth-
ods remain that are at least required for accessing the
account (3). This means that an account cannot be ac-
cessed if all of the device combinations in this term
are inaccessible.
We derive a numerical score based on this reduced
term for a simple comparison at a larger scale. This
is done by counting the number of occurrences n
i
of
each access method i within the reduced term and
assigning them the value s
i
= 1/n
i
. The total ac-
cessibility score is calculated using the minimum of
the conjunctive terms and the sum of the disjunctive
terms. The resulting score s
acc
gives us the lower
bound number of access methods one must lose to
lose access completely. At the same time, s
acc
1
can be interpreted as the upper bound number of ac-
cess methods that can be lost without any risk of being
locked out of an account. For the given example, the
accessibility score is calculated as follows:
s
acc
= min(1, 1) + 1 = 2
The accessibility score is 2, and losing one access
method is therefore not critical. However, losing two
methods, e.g., the phone and tablet, can make the ac-
count inaccessible to the user.
4 STUDY DESIGN
This section presents our study design. In particular,
the research questions to be answered by this study
are formulated. Furthermore, we describe the study
procedure, the AAG models created for Apple and
Google, the user account survey, and some ethical
considerations.
4.1 Research Questions
We designed a study to learn how users set up their
Google and Apple accounts and how much they de-
pend on their devices. One goal was to analyze
how users access their passwords. Nowadays, pass-
word managers are becoming increasingly popular
and are often even integrated into operating systems
or browsers. This might affect the choice of pass-
words, i.e., whether they are easy to remember or
rather created randomly and how accessible the pass-
word is. Furthermore, we wanted to check how users
adopt MFA and recovery methods compared to earlier
research (Petsas et al., 2015). In addition, we were in-
terested in the general security of the users’ accounts
by considering the primary authentication and recov-
ery methods. Finally, we wanted to assess how many
devices users depend on. From this, the following re-
search questions were derived:
RQ1. How do the users access their passwords?
RQ2. Which MFA and recovery methods did the
users enable?
RQ3. How secure are the account setups?
RQ4. How many access methods do the user ac-
counts depend on?
4.2 Methodology
For this study, we first created the models that de-
scribe the authentication systems for Google and Ap-
ple. We then created an online survey based on the au-
thentication systems in which actual users were asked
ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy
694
EmailPhone
|Sign in by phone
Security keySMS / VoiceBackup codesAuthenticator appGoogle prompts
|Password
&
|
Google
Figure 3: AAG for Google.
which authentication and recovery methods they use
for their Google or Apple accounts and what means or
devices they use to access them. After that, we gen-
erated user-specific AAG models based on the survey
responses and calculated the respective security and
accessibility scores. The general survey responses, as
well as the security and accessibility scores, were fi-
nally analyzed concerning the research questions.
A research tool and some scripts have been used to
facilitate the creation and analysis of the AAGs
3
. The
web application tool can create and visualize AAGs
and calculate security and accessibility scores. More-
over, we created Python scripts to automatically con-
vert the survey responses of the test participants into
actual AAGs. These AAGs were then imported into
the research tool and subsequently examined.
4.3 General AAG Models
The general AAG models for Google and Apple were
created in January 2023. It was ensured that the
AAGs were valid when the survey was conducted.
This is important to consider as authentication sys-
tems change over time. For instance, since Spring
2023, Google has started to support FIDO2 passkeys
as a passwordless alternative to its other authentica-
tion methods (Brand, 2023), which could not be cov-
ered in our study. The two AAG models used in this
study are described in the following paragraphs.
4.3.1 Google
To create an AAG model for Google accounts, the se-
curity settings have been examined. This is where
3
Our tools, survey questionnaires, anonymous partici-
pant data, and AAG files can be accessed at https://github
.com/Digital-Security-Lab/user-account-study-icissp2024
(Last accessed: 2023-12-06)
users can change their passwords or enable other
types of authentication. Note that figuring out all au-
thentication methods requires experimenting because
specific options are interdependent. For Google ac-
counts, the default primary authentication method is
the password. In addition to that, it provides the
option sign-in by phone and several MFA methods.
Google users can optionally set up a recovery email
address or phone number for recovery. The respec-
tive AAG model is shown in Figure 3. In practice,
Google is applying risk based authentication (RBA)
(Wiefling et al., 2019), which means that it might
request different authentication factors depending on
certain features of the user’s client. This, however,
can not be modeled by the AAGs because they cur-
rently only represent static models.
4.3.2 Apple
The AAG model for Apple has been created based
on the Apple ID security settings. As with Google,
the primary authentication method here is a password.
However, users have fewer options to configure MFA
in their Apple accounts. Secondary authentication
factors can be either trusted phone numbers that are
manually added by the user or trusted devices that
are automatically configured when an Apple device is
signed in to that Apple ID account. There is also the
possibility of recovering accounts through customer
service. This can not easily be modeled or evaluated
using AAGs. Also, there is an option to set a recov-
ery contact. These two options were excluded from
the model. However, an Apple account can also be
recovered using a trusted device. This option is im-
plicitly enabled. Therefore, this has been added to the
model. Furthermore, a user can explicitly configure a
printable recovery key. If this is selected, this is the
only possible way to recover an Apple account. The
Evaluating the Influence of Multi-Factor Authentication and Recovery Settings on the Security and Accessibility of User Accounts
695
Recovery KeyDevice
|
SMSDevice
|Password
&
|
Apple
Figure 4: AAG for Apple.
final AAG model for the Apple account is shown in
Figure 4.
4.4 User Account Survey
The surveys were created as two separate online
forms, each adapted to Google or Apple account set-
tings respectively. Both surveys consist of two main
parts.
First, the test participants had to make an enumer-
ated list of the devices that they actively use. They
should assign them to the categories of (1) phone, (2)
computer/laptop, (3) tablet, (4) smart watch and (5)
security key. These devices were referred to in subse-
quent questions.
For the second part, the participants were sup-
posed to log in to their Google or Apple accounts.
Afterward, they had to specify by which means they
could access their password, i.e., whether they could
remember it, whether it was stored in a password
manager, in a browser, on a device, or whether they
wrote it down on paper. They could choose multiple
options if applicable. After that, they were guided to
specific account settings and asked about them.
In the case of Google, they were asked whether
MFA methods were enabled and, if so, which ones.
In addition, they had to choose the devices by which
these methods could be accessed. For instance, if
Google prompts were selected, they had to indicate
which phones or tablets were configured for this
method. If MFA was disabled, they were asked if
sign-in by phone was enabled and for which of the
phones it was enabled. Last, they had to specify
which recovery options were enabled, including email
and phone. If a phone was selected, they had to
choose the phone from their list of devices with the
corresponding phone number.
For Apple, participants had to indicate which de-
vices were registered as trusted devices and which
used a trusted phone number. Finally, the participants
should state which recovery options were enabled in
their Apple accounts.
4.5 Ethical Considerations
In the survey, the test participants were asked about
what means they use to authenticate themselves to
their Google or Apple accounts. This could generally
be misused to exploit weak account configurations.
However, the survey was completely anonymous, and
no personal data was collected that could identify any
of the participants by any means. Also, no other sen-
sitive information, such as passwords or other secrets,
was collected. The study was conducted in compli-
ance with our university’s research ethics guidelines.
The test participants were thoroughly informed about
the procedure and fairly compensated for their time.
They could stop participating in the study at any time.
5 STUDY RESULTS
After several preliminary test runs, the survey was
conducted in mid-January 2023, with test participants
acquired using Prolific (Prolific, 2023). There were
94 submissions for Google (age 18-70, MV 34.89, SD
10.25) and 91 for Apple (age 19-67, MV 32.77, SD
9.7). In both cases, about half of the participants were
residents of the USA and the other half of Germany.
Below, we describe the study results concerning the
research questions.
5.1 Password Access (RQ1)
To answer RQ1, we assessed how people access their
passwords. A summary of the responses from our sur-
vey is given in Table 2. Most participants still tend to
use a password they can remember. This applies to
both Google and Apple. Yet, many also use a pass-
word manager or the password storage functionality
of a browser or device. According to our results,
less than 10% of the participants have written it down
on paper. It can be further observed that about one-
third of the Apple test participants use more than one
method to store their password, which applies to al-
most 48% of the Google participants.
5.2 Adoption of MFA and Recovery
Methods (RQ2)
In RQ2, we wanted to analyze to what extent the
users have adopted MFA and recovery methods. Ap-
ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy
696
Table 2: Responses of test participants about their password
usage.
Password access Apple
(n = 91)
Google
(n = 94)
I can remember it 64 64
Password manager 28 37
Stored by browser/device 28 46
I wrote it down on paper 9 7
Table 3: Frequency of authentication and recovery methods
used in the participants’ Google accounts (n = 94).
Authentication method MFA Frequency
Password-only # 22
Sign-in by phone # 8
Google prompts 26
Authenticator app 15
Backup codes 8
Voice or text message 38
Security key 2
Recovery phone n/a 64
Recovery email n/a 76
ple and Google are very different in their approach
to MFA. Apple devices are implicitly configured as
MFA methods for accounts they are signed in. There-
fore, Apple users most likely have MFA enabled au-
tomatically, so we did not examine this for Apple in
more detail.
Google users have several different authentication
and recovery methods that need to be configured man-
ually. The usage of each method is shown in Table 3.
According to our results, 68% of the Google test par-
ticipants have at least one MFA method enabled. In
the previously mentioned study from 2015, only less
than 7% had MFA enabled (Petsas et al., 2015). Fur-
thermore, in 2018, a Google researcher stated that by
that time, only 10% of Gmail users had set up MFA
(Milka, 2018). Compared to that, there is a significant
increase in our test results. One must keep in mind
that our sample size of 94 Google test participants is
relatively small. Yet, an increase was expected since
Google started to enroll MFA automatically for its
users as announced in 2021 (Risher, 2021). The most
common secondary authentication method is voice or
text message. Many of the participants also appear to
use Google prompts and authenticator apps. Backup
codes and security keys, in contrast, are used only
rarely.
Regarding recovery methods, there also seems to
be a wider adoption of a recovery email address (80%)
and a slightly increased use of the phone as a recovery
method (68%) compared to the results in (Petsas et al.,
2015).
5.3 Account Security (RQ3)
RQ3 was addressed by assessing the security scores
of the AAGs created for all test participants. In gen-
eral, it was observed that the recovery authentication
methods of Google often decrease the security of an
account, even if a user intended to use more secure
authentication methods like the security key. This
was particularly the case because most test partici-
pants used a recovery email. Nevertheless, it must
be kept in mind that the security score for email was
set to low because the accurate score could not be as-
sessed within this study. In reality, the score might be
higher depending on the email provider and account
setup. However, as mentioned before, it was shown
that recovery emails are often a weakness in an ac-
count setup (Li et al., 2018). The frequency of each
security score for Google and Apple is summarized
in Figure 5.
One can observe that there is no Google account
with a total security score high. The only way to
achieve this would be to disable all recovery meth-
ods and only enable a security key as a second au-
thentication factor. Two test participants have not
enabled recovery methods but used Google prompts
as the second authentication factor, with a score of
medium within our scoring scheme. Most Google ac-
counts have a total score of low. This is mainly be-
cause most test participants have enabled email as a
recovery method. The lesson learned is that Google
account security in practice often depends on the se-
curity of recovery email accounts.
Two Apple users are rated with the score high.
This is because they have only set up a trusted de-
vice but no trusted phone number. Most Apple ac-
counts have a score of medium. The main reason
here is likely that most Apple users use their account
in connection with at least one Apple device, which,
as mentioned before, automatically becomes a second
authentication factor.
5.4 Account Accessibility (RQ4)
We answered RQ4 by analyzing the accessibility
scores for each test participant’s Google and Apple
accounts. The scores were derived from the respec-
tive AAGs, including the access methods used by the
test participants. In Figure 6, it can be observed that
the range of accessibility scores is between 1 and 5
for Apple and between 1 and 6 for Google.
As described in Section 3.2, an accessibility score
equal to 1 indicates that an account can become inac-
cessible if one of the access methods is unavailable,
and we, therefore, analyzed these setups in more de-
Evaluating the Influence of Multi-Factor Authentication and Recovery Settings on the Security and Accessibility of User Accounts
697
Low
Medium High
0
20
40
60
80
100
1
88
2
Security score
Frequency
Apple
Low
Medium High
0
20
40
60
80
100
80
14
0
Security score
Frequency
Google
Figure 5: Histogram over security scores of the participants’ Apple and Google accounts.
tail. Within the test sample, ten of the Google ac-
counts have an accessibility score equal to 1. For Ap-
ple, this applies to seventeen accounts. After manual
analysis, it was found that in sixteen cases for Apple
and all ten cases for Google, the primary phone was
the key access method that, if lost, could cause an ac-
count lockout. This is reasonable since smartphones
play a crucial role in our lives today. The Google ac-
count of one test participant was even dependent on
the availability of both a memorized password and a
phone for Google prompts.
Beyond that, most test participants had accessibil-
ity scores of 2 or higher. This means that the major-
ity’s risk of being locked out of their account is rela-
tively low.
6 DISCUSSION
In this section, we discuss the limitations of the con-
ducted study, reflect on the new accessibility scoring,
and, finally, discuss what conclusions can be drawn
by our research for MFA in practice and how our
study might contribute to future improvements.
6.1 Limitations
First of all, the study was conducted with a relatively
low sample size of 91 Apple users and 94 Google
users. Hence, the study is rather an initial study, and
the results must be interpreted with that in mind. Yet,
the results give us an idea of how many possible ac-
count setups are used in practice and what potential
accessibility risks exist. Also, the results only reflect
a snapshot of when the study was conducted, as au-
thentication systems change with time.
Another limitation is the simplification of the
Google model, where email accounts have been as-
signed a low security score. In practice, this will dif-
fer for each account. However, it requires a rather so-
phisticated survey covering all possible email account
setups.
Moreover, as mentioned earlier, RBA has not yet
been considered in AAGs because the models are cur-
rently static. This might influence both security and
accessibility. For that, a better understanding of RBA
is required, and future work should look into how dy-
namic AAGs may be modeled and evaluated.
6.2 Accessibility Scoring
An essential contribution of this work is the new ac-
cessibility scoring described in Section 3.2. In con-
trast to the previously proposed method (P
¨
ohn et al.,
2022), where no transformation or reduction was con-
ducted after assessing the first Boolean term, our ap-
proach has a more practical meaning since it shows
how many access methods a user depends on. For
the example shown in Figure 2, the score as in (P
¨
ohn
et al., 2022) would result in an accessibility score of
1.5, while our method resulted in a score of 2.
A numerical score does not provide context about
the devices. Therefore, it could be beneficial to parse
the logical formula into a human-readable description
with more context about the actual devices. For in-
stance, the final term from the example in Figure 2
can be compiled into a description like Access to
Account might be lost when losing both Phone and
Tablet, or losing your Phone and forgetting your pass-
word”. In practice, this could help inform online users
more effectively of the consequences of losing a par-
ticular access method and motivate them to set up an
alternative access method for authentication or recov-
ery.
6.3 MFA in Practice
Our study shows that several users depend solely on
their primary phones. For online services, it is chal-
lenging to assess the risk of their users being locked
out of their account because the services can not know
if, e.g., an authenticator app is installed on the same
ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy
698
1 2 3 4 5 6
0
20
40
10
20
30
40
50
17
38
26
9
1
0
Accessibility score
Frequency
Apple
1 2 3 4 5 6
0
20
40
10
20
30
40
50
10
45
29
4
5
1
Accessibility score
Frequency
Google
Figure 6: Histogram over accessibility scores of the participants’ Apple and Google accounts.
device that is used for SMS as a second authentication
factor, or that has stored a user’s password. We found
that, especially for Apple, seventeen test participants
might lose access to their accounts if they lose their
phones. For Google, this applies to ten users. Given
that, this also means that having access to the users’
phones is enough to access an account. From a user
perspective, one could argue that the term multi-factor
authentication does not always apply, even though
users have enabled multiple authentication methods in
their accounts. The likelihood of a phone being com-
promised is relatively low, as it usually requires local
authentication through a PIN or biometrics or rather
sophisticated malware. It is still a scenario that must
be considered, especially by users with high-security
requirements, e.g., due to their profession.
Also, we found that MFA and recovery are often
linked to the same device. Beyond this study, we ob-
served that, e.g., Facebook prohibits using the same
phone number for MFA and account recovery, which
is a step in the right direction. Surprisingly, other
applications, such as LinkedIn, automatically enable
both options when setting up a phone number. Web
applications should encourage users to use different
devices for different authentication methods. There-
fore, it might be helpful to have a tool for users to
keep an overview of their account and device depen-
dencies, as suggested in (P
¨
ohn et al., 2022).
It is generally not very obvious how accessible an
account is. This depends on how people set up their
accounts and how devices and password access are
linked. AAGs can help assess this and encourage ser-
vice providers and users themselves to improve the
security and accessibility of online accounts. There-
fore, we suggest the integration of AAGs into con-
sumer tools and online services. Likewise, AAGs can
be a powerful tool in an enterprise context for an ad-
ministrator to get an overview of the account security
and accessibility of employees.
7 CONCLUSION
In this paper, we have extended the methodology of
Account Access Graphs with a new accessibility scor-
ing scheme. Moreover, we have developed and con-
ducted a study in which we analyzed several aspects
of Apple and Google user accounts, such as their
MFA adoption and their security and accessibility.
Within our security scoring scheme, Apple accounts
turned out to be more secure, likely due to the na-
ture of Apple devices being implicitly configured for
multi-factor authentication. For Google, we found
that several users still have not enabled any second
authentication factor. One of the most important find-
ings concerning accessibility was that several Google
test participants and even more Apple test participants
entirely depended on their phones. Even though this
was not the majority, several users risk losing access
to their user accounts. All in all, our study has demon-
strated that AAGs can contribute to a deeper under-
standing of authentication methods.
We will pursue this research direction in future
work. Similar studies may be conducted for other
web services to see how secure and accessible their
users’ account setups are. Furthermore, it can be in-
vestigated how service providers can use the knowl-
edge gained to improve their authentication systems
concerning security and accessibility. Finally, dy-
namic AAG models may be developed for analyzing
accounts of online services that apply risk-based au-
thentication.
REFERENCES
Amft, S., H
¨
oltervennhoff, S., Huaman, N., Krause, A.,
Simko, L., Acar, Y., and Fahl, S. (2023). ”We’ve Dis-
abled MFA for You”: An Evaluation of the Security
and Usability of Multi-Factor Authentication Recov-
Evaluating the Influence of Multi-Factor Authentication and Recovery Settings on the Security and Accessibility of User Accounts
699
ery Deployments. In Proceedings of the 2023 ACM
SIGSAC Conference on Computer and Communica-
tions Security, pages 3138–3152.
Beatriz Henr
´
ıquez (2022). Mobile Theft and Loss Report -
2020/2021 Edition. https://preyproject.com/blog/mob
ile-theft-and-loss-report-2020-2021-edition.
Bitkom (2021). Gestohlen oder verloren: Vier von zehn
Personen ist schon mal das Handy abhandengekom-
men. https://www.bitkom.org/Presse/Presseinformat
ion/Gestohlen-oder-verloren-Vier-von-zehn-Persone
n-ist-schon-mal-das-Handy-abhandengekommen.
Brand, Christiaan, K. S. (2023). The beginning of the end of
the password. https://blog.google/technology/safety-s
ecurity/the-beginning-of-the-end-of-the-password/.
Das, S., Wang, B., and Camp, L. J. (2019). MFA is a Waste
of Time! Understanding Negative Connotation To-
wards MFA Applications via User Generated Content.
arXiv preprint arXiv:1908.05902.
Davis, D. K., Chowdhury, M. M., and Rifat, N. (2022).
Password Security: What Are We Doing Wrong? In
2022 IEEE International Conference on Electro Infor-
mation Technology (eIT), pages 562–567. IEEE.
European Commission (2023). eIDAS Levels of Assurance.
https://ec.europa.eu/digital-building-blocks/wikis/di
splay/DIGITAL/eIDAS+Levels+of+Assurance.
Gavazzi, A., Williams, R., Kirda, E., Lu, L., King, A.,
Davis, A., and Leek, T. (2023). A Study of Multi-
Factor and Risk-Based Authentication Availability. In
32nd USENIX Security Symposium (USENIX Secu-
rity’23). USENIX Association, Anaheim, CA, USA.
Gerlitz, E., H
¨
aring, M., M
¨
adler, C. T., Smith, M., and Tiefe-
nau, C. (2023). Adventures in recovery land: Test-
ing the account recovery of popular websites when the
second factor is lost. In Nineteenth Symposium on Us-
able Privacy and Security (SOUPS 2023), pages 227–
243.
Google (2022). Passwordless login with passkeys. https:
//developers.google.com/identity/passkeys.
Grassi, P., Fenton, J., Newton, E., Perlner, R., Regenscheid,
A., Burr, W., Richer, J., Lefkovitz, N., Danker, J.,
Choong, Y.-Y., Greene, K., and Theofanos, M. (2020).
Digital Identity Guidelines: Authentication and Life-
cycle Management.
Hammann, S., Crabb, M., Radomirovic, S., Sasse, R., and
Basin, D. (2022). I’m Surprised So Much Is Con-
nected. In CHI Conference on Human Factors in
Computing Systems, pages 1–13.
Hammann, S., Radomirovi
´
c, S., Sasse, R., and Basin, D.
(2019). User account access graphs. In Proceedings
of the 2019 ACM SIGSAC Conference on Computer
and Communications Security, pages 1405–1422.
Hayashi, E. and Hong, J. (2011). A diary study of password
usage in daily life. In Proceedings of the SIGCHI
Conference on Human Factors in Computing Systems,
pages 2627–2630.
Jacomme, C. and Kremer, S. (2021). An extensive for-
mal analysis of multi-factor authentication protocols.
ACM Transactions on Privacy and Security (TOPS),
24(2):1–34.
Li, Y., Wang, H., and Sun, K. (2018). Email as a master key:
Analyzing account recovery in the wild. In IEEE IN-
FOCOM 2018-IEEE Conference on Computer Com-
munications, pages 1646–1654. IEEE.
Milka, G. (2018). Anatomy of Account Takeover. In
Enigma 2018 (Enigma 2018), Santa Clara, CA.
USENIX Association.
Ometov, A., Bezzateev, S., M
¨
akitalo, N., Andreev, S.,
Mikkonen, T., and Koucheryavy, Y. (2018). Multi-
factor authentication: A survey. Cryptography, 2(1):1.
Petsas, T., Tsirantonakis, G., Athanasopoulos, E., and Ioan-
nidis, S. (2015). Two-factor authentication: is the
world ready? Quantifying 2FA adoption. In Proceed-
ings of the eighth european workshop on system secu-
rity, pages 1–7.
P
¨
ohn, D., Gruschka, N., and Ziegler, L. (2022). Multi-
Account Dashboard for Authentication Dependency
Analysis. In Proceedings of the 17th International
Conference on Availability, Reliability and Security,
pages 1–13.
Prolific (2023). Prolific · Quickly find research participants
you can trust. https://www.prolific.com.
Reese, K., Smith, T., Dutson, J., Armknecht, J., Cameron,
J., and Seamons, K. (2019). A usability study of five
two-factor authentication methods. In Proceedings of
the Fifteenth Symposium on Usable Privacy and Secu-
rity.
Reynolds, J., Samarin, N., Barnes, J., Judd, T., Mason, J.,
Bailey, M., and Egelman, S. (2020). Empirical mea-
surement of systemic 2fa usability. In Proceedings of
the USENIX Conference.
Risher, M. (2021). A simpler and safer future without
passwords. https://blog.google/technology/safety-sec
urity/a-simpler-and-safer-future-without-passwords/.
Shen, C., Yu, T., Xu, H., Yang, G., and Guan, X. (2016).
User practice in password security: An empirical
study of real-life passwords in the wild. Computers
& Security, 61:130–141.
Taneski, V., Heri
ˇ
cko, M., and Brumen, B. (2019). System-
atic overview of password security problems. Acta
Polytechnica Hungarica, 16(3):143–165.
Wiefling, S., Lo Iacono, L., and D
¨
urmuth, M. (2019). Is
this really you? An empirical study on risk-based au-
thentication applied in the wild. In ICT Systems Se-
curity and Privacy Protection: 34th IFIP TC 11 In-
ternational Conference, SEC 2019, Lisbon, Portugal,
June 25-27, 2019, Proceedings 34, pages 134–148.
Springer.
ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy
700