Long-Term Observation on Browser Fingerprinting: Users' Trackability and Perspective

Abstract

Browser fingerprinting as a tracking technique to recognize users based on their browsers' unique features or behavior has been known for more than a decade. We present the results of a 3-year online study on browser fingerprinting with more than 1,300 users. This is the first study with ground truth on user level, which allows the assessment of trackability based on fingerprints of multiple browsers and devices per user. Based on our longitudinal observations of 88,000 measurements with over 300 considered browser features, we optimized feature sets for mobile and desktop devices. Further, we conducted two user surveys to determine the representativeness of our user sample based on users' demographics and technical background, and to learn how users perceive browser fingerprinting and how they protect themselves.

Full text

Proceedings on Privacy Enhancing Technologies ; 2020 (2):558–577
Gaston Pugliese*, Christian Riess, Freya Gassmann, and Zinaida Benenson
Long-Term Observation on Browser
Fingerprinting: Users’ Trackability and
Perspective
Abstract: Browser fingerprinting as a tracking technique
to recognize users based on their browsers’ unique fea-
tures or behavior has been known for more than a
decade. We present the results of a 3-year online study
on browser fingerprinting with more than 1,300 users.
This is the first study with ground truth on user level,
which allows the assessment of trackability based on
fingerprints of multiple browsers and devices per user.
Based on our longitudinal observations of 88,000 mea-
surements with over 300 considered browser features,
we optimized feature sets for mobile and desktop de-
vices. Further, we conducted two user surveys to deter-
mine the representativeness of our user sample based on
users’ demographics and technical background, and to
learn how users perceive browser fingerprinting and how
they protect themselves.
Keywords: browser fingerprinting, tracking, privacy
DOI 10.2478/popets-2020-0041
Received 2019-08-31; revised 2019-12-15; accepted 2019-12-16.
1 Introduction
In 2020, the PETS paper How Unique Is Your Web
Browser? [1] by Peter Eckersley will celebrate its 10th
publication anniversary. It was the first paper describ-
ing a study on browser fingerprinting (Panopticlick)
that gained far-reaching attention and can thus be seen
as the origin of this research field. Back then, Eckersley
showed that a small set of eight browser characteristics,
including the user-agent string and the list of installed
*Corresponding Author: Gaston Pugliese: Friedrich-
Alexander University Erlangen-Nürnberg, E-mail address:
gaston.pugliese@cs.fau.de
Christian Riess: Friedrich-Alexander University Erlangen-
Nürnberg, E-mail address: christian.riess@fau.de
Freya Gassmann: Saarland University, E-mail address:
f.gassmann@mx.uni-saarland.de
Zinaida Benenson: Friedrich-Alexander University Erlangen-
Nürnberg, E-mail address: zinaida.benenson@cs.fau.de
plugins, were sufficient to uniquely recognize between
83.6% and 94.2% of browsers in his dataset.
Motivation. The collected browser fingerprints of
Panopticlick have not been published. Ten years later,
after many related studies [2–6], research on browser
fingerprinting still lacks available and appropriate data.
Depending on the investigated aspects of browser finger-
printing, the requirements for data collection are quite
extensive: long-term, large-scale, fine-grained, diversi-
fied, sometimes cross-browser and cross-device. These
requirements apply to research studying the evolution
of browser features over time, assessing users’ trackabil-
ity, or examining the effectiveness of countermeasures.
At the time of writing, we are aware of only two
available datasets of browser fingerprints. Tillmann col-
lected fingerprints in 2012 [2] and published a dataset
that does not contain raw values of the most discrimi-
nating features, like fonts or plugins, and the data col-
lection was performed over a short period of one month.
Vastel et al. published an unfiltered sample of their
data in 2018 of fingerprints collected via browser ex-
tensions [3]. Although the dataset contains fingerprints
from 1.5 years, it lacks mobile devices, other browser
families than Firefox and Chrome (after filtering) as well
as precise timestamps and frequencies of observations.
Furthermore, most studies relied on cookies to rec-
ognize recurring browsers [1, 2, 4, 6]. This, however,
is not a reliable way to establish ground truth for long-
term observations, especially for privacy studies with al-
legedly savvy users: cookies can be easily deleted, either
manually by the user or automatically by the browser,
and they do not help to recognize users if they switch
browsers or devices. Moreover, it diminishes the relia-
bility of ground truth on fingerprints being unique-by-
entity and trackable (Sec. 2.2).
In general, both the overall long-term trackability of
users across multiple browsers and devices, and users’
understanding of or actions against browser fingerprint-
ing have received little attention so far. Furthermore,
the representativeness of fingerprint datasets has only
been investigated with respect to technical aspects [1–
8], but not with respect to user demographics.

Long-Term Observation on Browser Fingerprinting 559
Finally, the impact of participation in the studies on
users’ perception of fingerprinting remains unknown.
Research Questions. Based on the considerations
above, we present a 3-year online study with the goals
to collect longitudinal fingerprint data and to investi-
gate the users’ perspective on browser fingerprinting
for the first time. Our data collection (Sec. 3) estab-
lishes ground truth on user level instead of browser or
device level. Thereby, we can link fingerprints to indi-
vidual study participants over longer periods of time
without relying on the persistence of client-side identi-
fiers and regardless of the number of devices or browsers
they used. Furthermore, we conducted two user surveys
to determine the demographic representativeness of our
user sample and to understand the users’ perception of
browser fingerprinting, the role of their study participa-
tion on this perception, and their protection measures.
We aim to answer the following research questions:
–RQ1: How trackable are users based on their
browser fingerprints regardless of the number of
browsers and devices in use?
–RQ2: How do different feature sets perform regard-
ing fingerprint stability and trackability of users?
–RQ3: Do demographics of users, as well as their
technical background, privacy concerns and privacy
behavior, correlate with their trackability?
–RQ4: How do users perceive browser fingerprinting,
what is the role of study participation in this per-
ception, and which countermeasures do they apply?
Hypotheses. For RQ3, we formulate the follow-
ing six hypotheses, here combined into one sentence for
brevity. H1-6: The following user characteristics are re-
lated to their trackability: (1) age, (2) gender, (3) educa-
tion level, (4) computer science background, (5) privacy
behavior, (6) privacy concerns.
Contributions. The main contributions of this pa-
per cover technical findings as well as insights in users’
perception of browser fingerprinting based on quantita-
tive and qualitative analyses:
1. We present a novel long-term study on browser
fingerprinting with ground truth on user level.
Between 2016–2019, we collected 88,088 measure-
ments of 305 browser features of 1,304 participants
(Sec. 3).
2. We present two online surveys with study partic-
ipants to assess their demographic characteristics
and thus the representativeness of our sample as
well as to study participants’ privacy behavior, com-
prehension of browser fingerprinting, and applied
countermeasures (Sec. 3.2, 5, and 6).
Collect
browserfeatures
Define
browserfeatures
tocollect
Preprocess
browserfeatures
Feature
stemming
Featureset
optimization
Compose
fingerprints
Fingerprint
linkability
Evaluate
fingerprint
metrics
Fig. 1. Systemized workflow for browser fingerprinting
3. We present a simple, yet effective approach for op-
timizing feature sets towards different metrics (e.g.,
stability of fingerprints) for desktop and mobile de-
vices. Further, we introduce feature stemming as a
way to improve feature stability, e.g., by stripping
off version substrings (Sec. 4).
4. We make a dataset of our long-term study available
for research purposes (Sec. 7).
2 Background
A fingerprint is “a set of information elements that
identifies a device or application instance” and finger-
printing is the “process of uniquely identifying” these
entities [9]. For browser fingerprinting, these informa-
tion elements (features) can be obtained passively from
the client HTTP headers (e.g., user-agent string or lan-
guage), and actively using a client-side script to col-
lect information like screen resolution or plugins. Un-
like cookies which are stateful identifiers stored on the
client side, fingerprinting is considered a stateless track-
ing technique [10].
In the following, we review concepts and termi-
nology of browser fingerprinting, and we provide an
overview of studies that collected browser fingerprint-
ing data since 2009 (Table 1).
2.1 Evaluating Browser Fingerprints
Figure 1 shows a workflow for browser fingerprinting:
(1) The browser features that shall be collected are de-
fined and the fingerprint script is implemented for the
client and server side. (2) The fingerprint script is de-
ployed to collect the clients’ browser features. If ap-

Long-Term Observation on Browser Fingerprinting 560
plicable, these features are enriched with further state-
ful identifiers (e.g., cookies, session ID after authenti-
cation, personalized token in URL). Depending on the
type of ground truth, a fingerprint can be linked to ei-
ther an individual browser instance, device, or even user.
(3) The collected browser features are preprocessed
which may include normalization (e.g., screen resolu-
tion [8]) or derivation of additional information (e.g.,
user-agent parsing). Our work contributes to this step
of the workflow and proposes feature stemming and fea-
ture set optimization to improve the stability of features
and to compile an optimized feature set from collected
features (Sec. 4). (4) The actual fingerprints are com-
posed using a feature set. In practice, fingerprints can
be handled as MD5 hashes, or as vectors of raw feature
values. The latter enables further examinations like es-
tablishing linkability between evolved fingerprints [1, 3],
which we do not consider in this work. (5) Finally, the
collected fingerprints can be evaluated w.r.t. various
metrics, e.g., the anonymity set size [11] or stability.
2.2 Formal Concepts
Considering browser fingerprinting formally, we de-
note the feature set consisting of nfeatures as X=
{x1, . . . , xn}(n∈N). Further, we denote the feature
value domain of x∈ X , i.e., the set of all possible values
of x∈ X as V(x). As fingerprints are linked to entities
(e.g., individual browsers, devices, or users) and they
are observed at a point in time, we denote the set of
entities as E={e1, e2, . . .}, and the time domain as T.
We denote the fingerprint type of size kw.r.t. Xas:
T= (x1, . . . , xk) (xi∈ X ,1≤i≤k).(1)
We denote the fingerprint space Vof type Tw.r.t.
Xas the cross product of V(xi)(xi∈ X ):
V=V(x1)× V(x2)×. . . V(xk).(2)
The set of all fingerprints w.r.t. Tis denoted as F
and we define fingerprints (f∈ F) as follows:
f:E × T→ V w.r.t. T.(3)
Finally, for two timestamps t,t0, and without loss
of generality t<t0, we define the stability period s(e, t)
of fingerprint f(e, t)as
s(e, t) = max(t0−t)s. t.
∀t00, t ≤t00 ≤t0:f(e, t) = f(e, t00 ) = f(e, t0).(4)
Based on the previous equations, we derive following
basic metrics for fingerprints.
Def. 1. A fingerprint f(e, t)w.r.t. Tis unique-by-
entity if, and only if, it is linked to a single entity, i.e.,
∀e0∈ E,∀t0∈T, e06=e⇒f(e0, t0)6=f(e, t).(5)
Def. 2. A fingerprint f(e, t)w.r.t. Tis unique-by-
appearance if, and only if, it was observed once, i.e.,
∀e0∈ E,∀t0∈T , f(e0, t0) = f(e, t)⇒e0=e∧t0=t. (6)
Def. 3. A fingerprint f(e, t)is stable if, and only if,
its stability period is >0.
Def. 4. A fingerprint f(e, t)is trackable if, and only
if, f(e, t)is (i) unique-by-entity and (ii) stable.
Analogously to Def. 4, an entity is considered trackable,
if at least one trackable fingerprint is linked to it. In the
sequel, whenever we use the terms such as “stable” or
“trackable”, we refer to the definitions above.
In related work, the concept of uniqueness is often
vague as it is not stated whether feature values or fin-
gerprints are unique-by-appearance or unique-by-entity
w.r.t. to the corresponding ground truth.
2.3 Related Work
To our knowledge, Mayer [7] performed the first docu-
mented online study on browser fingerprinting. He col-
lected 1,298 fingerprints from 1,328 browser instances
and reported that 98.5% of the fingerprints were unique
and thus 96.23% of the browsers uniquely identifiable.
Eckersley [1] presented Panopticlick1and re-
ported between 83.6%–94.2% of 470,161 fingerprints
to be unique, depending on the availability of Flash
or Java. Moreover, he established linkability between
evolving fingerprints with an accuracy of 99.1%. Till-
mann [2] collected 23,709 fingerprints from 18,692
browsers within one month, and reported that 92.57%
of the fingerprints were unique.
Fifield and Egelman [6] uniquely identified 43% out
of 1,016 browsers using the dimension of rendered font
glyphs combined with the user-agent string.
Laperdrix et al. [4] collected 118,934 fingerprints on
AmIUnique2within three months, where 90% and 81%
of the fingerprints from desktop and mobile devices were
reported as unique, respectively.
1https://panopticlick.eff.org/
2https://amiunique.org/

Long-Term Observation on Browser Fingerprinting 561
Ground truth
Study Year Start End Features Browsers Users Fingerprints
User Demographics
Dataset Availability
Cookie
IP Address
Browser ID
User ID
Mayer [7] 2009 02/2009 02/2009 3 1,328 - 1,298   ○   
Eckersley [1] 2010 01/2010 02/2010 8 - - 470,161   ○ ○  
Tillmann [2] 2013 11/2012 12/2012 48 18,692 - 23,709  ○ ○   
Fifield and Egelman [6] 2015 - - 4311,016 - 1,016   ○   
Laperdrix et al. [4] 2016 11/2014 02/2015 17 - - 118,934   ○   
Cao et al. [8] 2017 - - 49 - 1,903 3,615   ○   ○
Vastel et al. [3] 2018 07/2015 08/2017 17 1,905 - 98,598  ○   ○ 
Gómez-Boix et al. [5] 2018 12/2016 06/2017 17 - - 2,067,942   ○   
This study 2019 02/2016 02/2019 3052- 1,304 88,088 ○ ○    ○
1Width and height of 43 Unicode code points, each rendered in six default CSS font families; 2including parsed and derived ones
Table 1. Overview on browser fingerprinting studies that collected datasets of fingerprints between 2009 and 2019 including this study;
indicating the publication year, start and end date of data collection as well as the number of considered features and fingerprints after
data cleansing and filtering, the number of distinct browsers or users observed (if explicitly specified), whether user demographics were
collected, the public availability of the collected data, and the ground truth for each dataset (○yes, no)
Cao et al. [8] investigated cross-browser fingerprint-
ing and collected 3,615 fingerprints from 1,903 users.
They proposed novel rendering tasks for canvas finger-
printing and reported an identification rate of 99.24%
on unique fingerprints.
Vastel et al. [3] used browser extensions for Fire-
fox and Chrome to collect fingerprints, which is a ma-
jor improvement towards establishing ground truth for
long-term observations. They proposed Fp-Stalker, an
algorithm to establish linkability between fingerprints
despite changing features.
Gómez-Boix et al. [5] collected 2,067,942 finger-
prints from a real-world population on a popular french
website using the same feature set as [3, 12]. They re-
ported only 33.6% and 18.5% of the fingerprints from
desktop and mobile devices to be unique. This notable
lower share of unique fingerprints indicates the need for
real-world samples of browser fingerprints for a thor-
ough assessment of the threat of browser fingerprinting
on individuals’ privacy.
Compared to our work, most studies above collected
the data only within six months or less, except [3]. Our
evaluations are based on data collected within three
years, which we believe can provide new insights on
users’ long-term trackability. In contrast to most prior
works, except for [8, 13], we did not rely on cookies to
recognize recurring users. We used personalized links to
establish ground truth on user level, like [8], and allowed
users to choose the browsers and devices they want to
use freely, unlike [8]. To our knowledge, our study is the
first to have ground truth on user level while allowing
multiple devices per user.
Previous studies did not collect demographics of
their participants. Yet, we assume that demograph-
ics are important to assess the representativeness of a
study: Studies with explicitly recruited users report a
much higher share of unique fingerprints than those with
real-world users [5]. Demographic characteristics may be
one reason for this discrepancy. Recruited user samples
are likely to have a higher share of students and pro-
fessionals with technical background as they are more
likely to be interested in browser fingerprinting. To our
knowledge, our study is the first to provide such data.
Furthermore, our study is the first to conduct user
surveys to investigate how users perceive browser fin-
gerprinting, how they protect themselves against it, and
how their participation in the study impacts their view
on browser fingerprinting and privacy.
Finally, while prior studies considered between 3
and 49 browser features (see Table 1), our study consid-
ers 305 features, including derived ones (e.g., via user-
agent parsing or feature stemming). Unlike other stud-
ies, we do not use a predefined feature set, but instead
investigate how different feature sets perform regarding
fingerprints’ stability and how they affect users’ track-
ability on the long run.

Long-Term Observation on Browser Fingerprinting 562
3 Method
Below we describe our study design: How we collected
browser fingerprints, how we conducted user surveys,
and how we compiled the final dataset for evaluation.
3.1 Study Design
We designed an online user study where participants
register themselves on our website: https://browser-
fingerprint.cs.fau.de/. In the following, we describe the
main components and considerations during the design
of our study.
3.1.1 Establishing Ground Truth
Participation in a study on browser fingerprinting might
encourage users to experiment with their browsers to be-
come undistinguishable, or to use multiple browsers and
devices out of curiosity. As we aimed for a long-term
observation of fingerprints from multiple devices and
browsers per user, we required a level of ground truth
that is more reliable than that provided using cook-
ies. We refrained from user accounts, apps, or browser
extensions to establish ground truth as it would have
raised the hurdle for participation (e.g., remembering
credentials), require an additional communication chan-
nel to remind the users to submit measurements, or limit
our sample to specific devices or browsers [8, 14].
We decided to require users to sign up with their
email addresses. Afterwards, they verified their email
addresses using a verification link sent to them. Thereby,
we establish ground truth on user level, enable users
to use as many browsers and devices as they wish, and
also control the reminders for periodic measurements by
sending weekly emails with personalized links to them.
3.1.2 Ethical Considerations
The study received an approval from the data protection
office at our institution. We attached great importance
to transparency towards participants about browser fin-
gerprinting, the purpose of our study, and the data we
collect. We fully disclosed our study method on our
website, including all conceptional and technical details
(e.g., feature collection, user surveys, data storage).
During the entire study, participants were free to
provide their fingerprints by visiting the weekly links we
sent them via email. At any given point in time, partic-
ipants were able to quit their participation by visiting
an unsubscription link we added to each of the weekly
emails. If participants decided to quit, we automatically
removed their email addresses from our database. All
data is stored on a server hosted at our institution, with
only project researchers having access to the data.
3.1.3 Fingerprinting and User Experience
After participants verified their email address, they re-
ceived a new personalized link once a week via email
which referred to a subpage of our study website where
their browser features were collected. Each weekly link
had a validity period of one week to reduce the poten-
tial effect of publicly shared links that could distort the
evaluation of individual participants retrospectively.
One of our goals was to investigate how users per-
ceive browser fingerprinting, and whether our study has
impact on this perception. Hence, we provided informa-
tion about browser fingerprinting and our study as well
as FAQs written in non-technical terms. The website,
surveys, and email texts were in English and German.
After collecting the browser features using a custom
script, we informed the participants about the unique-
ness and recognizability of their fingerprints and pro-
vided an overview on all of their features. We further
provided descriptive statistics about our study.
To provide another incentive for participation on
a regular basis, every four weeks we sent an individual
study report to each participant. From these reports, the
participants could learn for how long their three most
stable fingerprints were trackable, for desktop and mo-
bile devices separately, and how their trackability com-
pares to the results of other participants.
3.1.4 User Survey 1: Demographics and Background
To determine the representativeness of our user sam-
ple, and to answer the research question RQ3 about the
relation between various user characteristics and their
trackability, we designed a user study to be adminis-
tered shortly after the users registered for the study.
The survey was tested with seven experts on usable se-
curity and five private contacts with non-technical back-
ground, and iteratively improved during the tests.

Long-Term Observation on Browser Fingerprinting 563
We collected year of birth, gender, country of resi-
dence, current occupation, and the highest level of ed-
ucation. We also asked whether they study or work in
computer science or a related discipline, and whether
they have ever heard of browser fingerprinting before.
We measured privacy concerns using the Westin in-
dex [15] and asked the following questions about the
privacy behavior on the Web:
– Have you ever used following privacy measures: Do
Not Track, Tor (e.g., Tor browser), private mode
in browsers (privacy or incognito mode), deleting
cookies, denying third-party cookies?
– Have you ever used following browser extensions:
Ad Blocker, NoScript, BetterPrivacy, HTTPS Ev-
erywhere, Privacy Badger, Ghostery, Disconnect,
uBlock Origin?
For each privacy behavior, users could answer with
“yes”, “no” or “don’t know”. We computed a privacy
behavior index based on the number of “yes” answers.
3.1.5 User Survey 2: Impact of Our Study
To answer RQ4 about users’ perception of browser fin-
gerprinting, applied countermeasures, and impact of our
study, we developed a second user study. The survey was
tested with twelve experts on usable security and five
laypersons, and iteratively improved during the tests.
Understanding, Concern, Protection. We
gathered quantitative data on users’ understanding and
perception of browser fingerprinting by asking them to
indicate their agreement or disagreement with the fol-
lowing statements on a 5-point Likert scale from strongly
disagree to strongly agree:(1) Browser fingerprinting
can be used by websites to recognize me. (2) I think that
I understand how browser fingerprinting works. (3) I am
concerned that websites and companies try to finger-
print my browser. (4) I think most websites on the Web
use browser fingerprinting. (5) I think websites that I
frequently visit use browser fingerprinting. (6) It is im-
portant for me to be protected from browser fingerprint-
ing. (7) I am capable of protecting myself from browser
fingerprinting. (8) Protecting myself from browser fin-
gerprinting requires a lot of effort.
The items were presented to each user in a random-
ized order to counter the influence of the previous items
on the answers to the subsequent ones. In the above
order, items 1 and 2 measure users’ (perceived) under-
standing of fingerprinting, items 3-5 refer to the concern
about fingerprinting and perception of its spread, and
items 6-8 ask users’ opinion about protection measures.
Countermeasures. To identify applied protection
measures, we asked the following question, where par-
ticipants could enter their countermeasures into prede-
fined boxes (number of boxes was not fixed): Did you try
to protect yourself from browser fingerprinting? If yes,
what did you try (at least occasionally)? Please specify
as many measures as you can think of.
Study impact. We asked the following questions,
where the users first indicated “yes” or “no”, and then
could explain their reasons (free text): (1) Did you ex-
perience or learn something new by participating in our
study? (2) Did your study participation change any of
your thoughts/feelings regarding the Web? (3) Did your
study participation change your behavior on the Web?
3.2 Data Collection
Recruiting and Initial Data Sample. Our goal was
to recruit as many participants as possible from the gen-
eral public. Thus, we used mailing lists of international
universities, press releases, and private contacts of re-
searchers. Moreover, participants could share our study
via email or via privacy-respecting Twitter, Facebook
and Google+ buttons with a predefined text. We also
used announcements at computer science and security
conferences, mailing lists, and forums.
Within 1,111 days between February 9, 2016 and
February 23, 2019, we collected a total of 124,046 mea-
surements from 2,315 participants. Each measurement
consists of 305 features (App. D). We do not call the
data fingerprints yet, but measurements, because the
number of distinct fingerprints is depended on the fea-
ture set used to compose the fingerprints (Sec. 2, 4).
Data Cleansing. As we aim to assess users’ track-
ability on the long run, we had to filter our raw dataset
to remove the noise induced by exploitation of our study
design and by users’ curiosity. Using a keyed-hash stored
additionally for each email address, we merged 42 par-
ticipants that re-subscribed with the same email address
after unsubscribing. We discarded 2,421 measurements
resulted from testing and debugging with 11 of our own
subscriptions to the study. We removed 11,294 first-
week measurements of all participants to reduce the ef-
fect of their initial curiosity and experimentation during
the first week on later evaluations. This resulted in 266
of the participants having no measurements left. To en-
sure a minimum observation time of at least four weeks
with four measurements, we had to remove 676 par-
ticipants (2,173 measurements). Further, we removed
19,647 measurements from the upper 1% of partici-

Long-Term Observation on Browser Fingerprinting 564
Fig. 2. Growth of filtered dataset during course of study
Fig. 3. Collected data per participation week in filtered dataset
pants with the most measurements (16) as their amount
and frequency of measurements appeared dubious (e.g.,
due to sharing of personalized links, or automated re-
quests). Finally, we discarded 423 measurements from
clients that were neither desktop nor mobile devices
(e.g., crawler, command-line tools).
Dataset for Evaluation. Our final dataset con-
sists of 88,088 measurements from 1,304 participants
and is used in the sequel. On average, participants pro-
vided 67.6 measurements (σ=78.4) over a period of 63.2
weeks (σ=47.7). Figure 2 shows the growth of the num-
ber of participants and measurements in our filtered
dataset, and Figure 3 shows the total number of par-
ticipants and measurements per participation week.
Sample Characteristics. In total, 1,275 study
participants (97.8%) answered the first user survey.
79.1% of participants were from Germany (1,008); the
remaining shares within the top 5 countries were: 3.8%
U.S. (48), 2.2% Netherlands (28), 1.6% UK (20), and
1.0% Switzerland (13). The majority of participants
were male (76.5%), between 30 and 49 years old (40.2%),
have an academic background (64.8%), are employed
(47.7%), study or work in computer science or a related
field (57.5%), and already knew browser fingerprinting
(68.5%). A full overview of our participants’ demograph-
ics is shown in Appendix A.
Based on the Westin index [15], 60.9% of the partici-
pants were categorized as privacy fundamentalist, 36.5%
n%
Browser
Delete cookies 1,200 94.1
Use of private Mode 1,080 84.7
Denying third-party cookies 1,036 81.3
Set Do-not-track flag 916 71.8
Tor / Tor browser 605 47.5
Extensions
AdBlock 1,058 83.0
NoScript 666 52.2
Ghostery 526 41.3
HTTPS Everywhere 479 37.6
uBlock 340 26.7
Privacy Badger 238 18.7
Better Privacy 234 18.4
Disconnect 195 15.3
Table 2. Participants’ privacy behavior (N=1,275)
as pragmatic, and 2.6% as unaware. Regarding users’
privacy behavior, Table 2 shows the results on whether
the 13 privacy measures we asked for have ever been
used by the participants. Counting the “yes” answers,
the mean privacy behavior index is 6.7.
4 Trackability of Participants
In this section, we present a data-driven approach to se-
lect suitable tracking features. We first create stabilized
versions of features, and subsequently perform an auto-
mated feature selection. The feature selection greedily
maximizes an objective function that can be adjusted to
the target application and for different device types. We
propose objective functions to maximize the number of
trackable users, and to maximize feature stability.
Our results outperform the hand-crafted feature
sets of Panopticlick [1] and AmIUnique [4]. More-
over, we believe that automated feature selection can
provide a more realistic view on how the trackability of
users might be improved by privacy-violating websites
even without using linkability techniques [1, 3].
4.1 Feature Stemming
Several browser features change regularly. This involves,
e.g., changing version strings after system updates, or
varying screen resolution upon connecting the device
to an external monitor. Vastel et al. distinguish changes
like version strings as automatic evolutions, changes like
varying screen resolution as context-dependent evolu-

Long-Term Observation on Browser Fingerprinting 565
tions, and changes like disabling cookies or enabling
“Do-not-track” as user-triggered evolutions [3].
Particularly automatic evolution can negatively im-
pact the stability of browser fingerprints. Hence, we
manually selected features and removed their variable
elements (such as version strings via regular expres-
sions). We denote this normalization as feature stem-
ming in reference to linguistic processing to reduce
“words with the same root [...] to a common form [...] by
stripping each word of its derivational and inflectional
suffixes” [16]. Appendix B shows examples for the era-
sure of version substrings, assignment of IDs to plug-
ins and MIME types, and sorting of list-like features in
alphabetical order to gain robustness against targeted
order randomization.
As shown in Appendix C, stemmed versions of fea-
tures have been selected by the data-driven feature se-
lection presented in the next section.
4.2 Feature Set Optimization
We propose two objective functions for feature set op-
timization, namely the number of trackable users and
the average stability period. Both objective functions
might be suitable choices for real-world tracking tasks.
Thus, we consider them equally suitable for privacy as-
sessments of users’ trackability.
Number of Trackable Users (Ju). To maximize
the number of trackable users, we define a mapping
function φ:F 7→ R that reduces the set of all fin-
gerprints Fto the set of trackable fingerprints R. The
resulting set Rdepends on a specific choice of a fin-
gerprint type T. Therefore, if we seek to maximize the
number of trackable users, the goal is to find via the
objective function Juan optimum fingerprint type T∗,
Ju:T∗=arg max
T
|ψ(φ(T))|,(7)
where the mapping ψ: (e, t)7→ (e),(e, t)∈ R extracts
the entity from a trackable fingerprint.
Stability of Trackable Fingerprints (Js). To
maximize the stability period of trackable fingerprints
(Def. 4), we propose the objective function Jsto seek
Js:T∗=arg max
T
Ps(φ(T))
|E| ,(8)
where the function s: (e, t)7→ (tmax)calculates the
stability period of a fingerprint as defined in Eqn. 4.
Note that |E| is constant across fingerprint types Tin
Eqn. 8, and can hence be omitted during optimization.
Greedy Sequential Search. In practice, using
the objective functions Eqn. 7 or Eqn. 8 in a brute-
force search through all subsets of features is infeasi-
ble. However, we can perform a search for a sufficiently
good type. This is referred to as feature selection in
the pattern recognition literature, with various exist-
ing approaches that vary in computational effort and
reliability of the results [17]. We are not aware of an
efficient search strategy that guarantees convergence to
the global maximum of Eqn. 7 or Eqn. 8. Hence, we re-
sort to a greedy sequential search. A greedy search is
only guaranteed to converge to a local optimum. Yet,
the presented empirical results outperform manual fea-
ture selection strategies, which illustrates the benefit of
a data-driven optimization over hand-crafted solutions.
In detail, the search iteratively enlarges the cardi-
nality of the fingerprint type by adding that feature that
yields the largest improvement to the objective func-
tion. Thus, it starts with an empty fingerprint type T(0),
|T (0)|= 0. In iteration i, the type expansion is
T(i)=T(i−1) ∪xopt ,where (9)
xopt /∈ T (i−1),(10)
xopt =arg max
x
Jfor J∈ {Ju, Js},(11)
and Juor Jschosen as in Eqn. 7 and Eqn. 8, respectively.
The iteration is terminated when there is no feature
found that further improves the chosen objective func-
tion. If multiple features offer identical improvement in
one iteration, we choose the lexicographically first.
4.2.1 Evaluation on Our Dataset
The dataset is split in non-overlapping training and test
sets by odd and even user IDs with 652 users in both
sets. The splits contain 45,063 and 43,025 measurements
over a period of 157.6 and 157.3 weeks, respectively.
The test set measurements consist of 29,989 desktop
and 13,036 mobile measurements of 621 and 432 users.
Both sets are statistically similar, with shares of mobile
devices 29.3% and 30.3%, JavaScript enabled 86% and
85.7%, Flash enabled 13.7% and 13%, and on average
69 and 66 measurements per participant, respectively.
The training set is used to select device-dependent
(desktop TD, mobile TM) and device-independent (to-
tal TT) feature sets. For each device, we select one fea-
ture set to maximize the number of trackable partici-
pants using Eqn. 7 (denoted as TD
u,TM
u,TT
u), and one
feature set to maximize stability using Eqn. 8 (denoted
as TD
s,TM
s,TT
s). Also on the test set, we compare to the

Long-Term Observation on Browser Fingerprinting 566
Device type Desktop Mobile Total
Feature set [1] [4] TD
uTD
s[1] [4] TM
uTM
s[1] [4] TT
uTT
s
User (%)
w/ unique FPs (appear.) 95.7 97.1 87.3 91.0 80.3 88 88.7 77.1 95.7 97.4 96.8 91.4
w/ unique FPs (entity) 98.1 99.4 98.6 98.7 89.1 94.4 94.2 88.4 98.2 99.5 99.4 98.3
w/ trackable FPs 84.4 85.7 89.2 89.2 67.6 72.9 72.7 64.6 91.3 93.1 94.5 94.5
Stability trackable FPs (weeks)
Mean of means p. user 3.1 3.4 9.6 11.6 3.3 3.1 3.1 10.7 3.2 3.3 3.7 11.9
Std. of means p. user 3.3 3.9 9.3 10.5 6.4 5.1 5.1 11.8 4.1 3.8 4.6 10.8
q25 of means p. user 1.3 1.5 4.1 4.5 1.0 1.1 1.1 2.1 1.4 1.5 1.8 4.6
q50 of means p. user 2.2 2.4 7.9 9.0 2.2 2.3 2.3 7.0 2.3 2.4 2.9 9.4
q75 of means p. user 3.8 3.9 12.2 16.0 3.5 3.3 3.4 14.4 3.8 3.8 4.2 15.7
Mean of maxima p. user 8.2 8.6 21.1 25.7 6.8 6.8 6.8 20.2 9.2 9.5 10.0 27.2
Std. of max. p. user 12.3 12.7 19.5 23.7 12.1 10.9 10.9 24.2 13.9 13.8 14.3 25.0
q25 of max. p. user 2.1 2.4 6.0 7.0 1.9 2.0 2.0 3.0 2.4 2.9 3.2 8.0
q50 of max. p. user 4.8 5.0 15.2 18.1 4.0 4.1 4.2 10.0 5.0 5.5 6.0 19.1
q75 of max. p. user 8.1 8.3 29.3 40.0 7.0 8.0 8.0 27.8 9.0 9.0 10.0 42.0
FPs (%)
Distinct FPs (n) 12,330 13,801 7,473 8,029 3,935 4,793 4,825 4,308 16,265 18,594 16,541 9,822
Unique FPs (appear.) 57.0 60.7 49.2 53.2 47.9 50.9 51.4 70.1 54.8 58.2 53.9 49.1
Unique FPs (entity) 94.4 97.7 97.4 97.8 90.1 91.8 91.6 94.1 93.4 96.2 95.8 95.4
Trackable FPs 37.1 36.7 48.0 44.4 41.8 40.6 39.8 23.7 38.2 37.7 41.6 46.1
Table 3. Feature set optimization towards “users with trackable fingerprints” (Ju) and “stability of trackable fingerprints” (Js). Fea-
ture sets of Panopticlick [1], AmIUnique [4], and computed feature sets (TD,TM,TT) applied on the test split of our dataset. Bold
values indicate best result(s) p. row. Highlighted rows indicate the optimization criterion for TD,TM,TTw.r.t. Juand Js.
prior works with preset feature sets Panopticlick [1]
and AmIUnique [4]. For the latter, the feature on the
order of HTTP headers is omitted, as our data does not
provide this information (discussed in Section 7).
Results. Our feature selection chose from the 305
features with Juand Json desktop data 9 and 15 fea-
tures, on mobile data 9 and 15 features, and on the full
dataset 11 and 24 features (Appendix C).
The results of our feature selection are shown in Ta-
ble 3. Three results are shown in vertical order: results
on the number of tracked participants, results on the
stability of the fingerprints, and statistics on the num-
ber of fingerprints. We discuss these results below.
The first gray row is the central result on the num-
ber of tracked users. Here, both proposed feature selec-
tion methods achieve with 89.2% by some margin the
best result for desktop devices. On mobile devices, AmI-
Unique is with 72.9% slightly better than our feature
set TM
uwith 72.7%. On the whole dataset (“Total”),
both proposed methods achieve the best results (94.5%).
The second gray row is the central result on the sta-
bility of fingerprints. As a performance metric, we calcu-
late the mean over the average stability periods of each
user. When optimizing for the number of tracked users
via Ju, the stability is roughly comparable to Panop-
ticlick and AmIUnique. However, this changes com-
pletely when optimizing for the stability via Js: here, the
achieved stabilities range between 10.7 weeks for mobile
devices and 11.9 weeks for the whole dataset, which out-
performs the related methods by a factor of about 3.
The reported standard deviations are also considerably
larger, since the fingerprints of some users are remark-
ably stable fingerprints. This is further illustrated in the
quartiles in the lines below. For the mean over the max-
imally stable fingerprints per user, the third quartile of
42 weeks on the whole dataset shows that some users
have extremely stable fingerprints.
The fingerprint statistics in the third part of Table 3
show that our method extracts on a considerably lower
number of distinct fingerprints on desktop devices, and
on mobile devices a slightly larger number. High per-
centages in the following row “unique-by-appearance”
indicate that a larger number of fingerprints is useless
for tracking as they only appear once and thus have
no stability. By tendency higher percentages here coin-
cide with lower number of tracked users and reduced
stability. Conversely, a larger percentage in “unique-by-
entity” and “trackable” fingerprints improves the track-
ing result. Thus, it is interesting to note that the pro-
posed feature selection apparently runs into a local opti-
mum when working with mobile device data, by produc-
ing a large share of unique-by-appearance fingerprints
and as a consequence a reduced percentage of trackable
fingerprints. On the other hand, the feature selection

Long-Term Observation on Browser Fingerprinting 567
works remarkably well on desktop data and the whole
dataset with a somewhat lower number of distinct fin-
gerprints, but these fingerprints are more precise with
respect to the number of tracked users and fingerprint
stability.
4.2.2 Evaluation on the FP-STALKER Dataset
We also evaluate our feature selection on the Fp-
Stalker dataset3which is the only available large fin-
gerprint dataset at the time of writing (see Table 1).
As shown in Figure 1, linkability should be applied af-
ter the preprocessing (which includes feature stemming
and feature set optimization) and fingerprint composi-
tion, and is thus not considered in this work. However,
the Fp-Stalker dataset is suitable for the evaluation
of feature set optimization, as we explain below.
Dataset. The fingerprints of 1,819 browser in-
stances were collected using browser extensions for
Chrome and Firefox. We discard all other browser in-
stances, and also browsers with more than a single
browser and operating system family, as this is a strong
indicator for spoofing. We further remove fingerprints
that are inconsistent according to Fp-Stalker, that are
unique-by-appearance, that have tiny screen resolutions
(e.g., 8x8), and the upper 1% of browsers with more
than four canvas fingerprints. This left 1,198 browsers
and 4,816 distinct fingerprints for evaluation.
Since the dataset is pre-grouped into fingerprints
using a given feature set, we use the timestamps to
split the fingerprints into individual measurements us-
ing each fingerprint’s time of first appearance, update,
and last appearance, resulting in 14,887 measurements.
We split the data into training and test, each with
599 browser instances with 7,025 and 7,862 measure-
ments from within 72.7 weeks, respectively. We also de-
rive features via user-agent parsing, screen resolution
normalization [8], and feature stemming, and split the
WebGL fingerprint into individual features.
Results. As shown in Table 4, optimizing towards
Juprovides almost identical stability as the initial Fp-
Stalker feature set (baseline). In both cases, the av-
erage stability of trackable fingerprints per browser is
1.8 weeks. In the upper quartile of average and maxima
stabilities, Juoutperforms Fp-Stalker by 0.1 weeks,
where one browser less is trackable for Ju. The share
of trackable fingerprints is equal to the baseline (95%).
3https://github.com/Spirals-Team/FPStalker
Criterion Baseline JuJs
Features (n) 18 8 13
Brow. (%)
w/ unique FPs (appear.) 16.5 16.4 11.5
w/ unique FPs (entity) 97.8 97.5 95.7
w/ trackable FPs 91.0 90.8 89.1
Stability trackable FPs (weeks)
Mean of means p. browser 1.8 1.8 3.7
Std. of means p. browser 4.1 4.1 5.6
q25 of means p. browser 0.2 0.2 0.2
q50 of means p. browser 0.6 0.5 1.5
q75 of means p. browser 1.4 1.5 4.4
Mean of maxima p. browser 2.9 2.9 6.3
Std. of maxima p. browser 5.1 5.0 8.3
q25 of maxima p. browser 0.2 0.2 0.3
q50 of maxima p. browser 1.3 1.3 2.6
q75 of maxima p. browser 3.1 3.2 9.4
FPs (%)
Distinct FPs (n) 2,513 2,501 1,648
Unique FPs (appear.) 4.6 4.6 4.6
Unique FPs (entity) 99.6 99.6 98.5
Trackable FPs 95.0 95.0 93.9
Table 4. Feature set optimization on Fp-Stalker dataset com-
pared to its initial feature set (baseline) [3]. Two feature sets
optimized towards criteria “entities with trackable fingerprints”
(Ju) and “stability of trackable fingerprints” (Js) were computed
on training set and applied on the test split. Bold values indicate
best result(s) per row. Highlighted rows indicate the optimiza-
tion criteria.
Jucomposes 2,501 distinct fingerprints using 8 features,
while the baseline yields 2,513 using 18 features.
Optimizing directly towards the stability of track-
able fingerprints (Js) yields an average stability of 3.7
weeks, hence outperforms the baseline by a factor of 2.
On q75,Jsachieves 4.4 weeks on the average stabilities
and 9.4 weeks on the maximum stabilities, which is 3
times higher than the baseline. The number of distinct
fingerprints of the baseline is 2,513, which is consider-
ably larger than the 1,648 distinct fingerprints that Js
composes out of 7,862 measurements on the test set.
The trade-off for this improvement in stability is that
11 browsers are less trackable with Js, and the share of
trackable fingerprints is slightly lower (93.9% vs 95%).
The results show that a data-driven feature selec-
tion can considerably improve a given figure of merit,
in this case stability.
5 Trackability Factors
We conducted quantile regressions [18] to test Hypothe-
ses H1-6 (Sec. 1) about the relations between user char-
acteristics and trackability.

Long-Term Observation on Browser Fingerprinting 568
Model 1 Model 2 Model 3
q25 q50 q75 q25 q50 q75 q25 q50 q75
Age 0.00964+0.0147∗∗∗ 0.0245∗∗∗ 0.0130∗0.0157∗0.0171+0.0128∗0.0189∗∗ 0.0272∗∗∗
(1.93) (3.72) (3.55) (2.19) (2.55) (1.83) (2.40) (3.07) (3.62)
Gender: male1-0.246 -0.349+-0.691∗∗ -0.396 -0.220 -0.396∗-0.478+-0.511∗-0.764∗∗
(-1.20) (-1.87) (-2.76) (-1.54) (-1.07) (-1.97) (-1.86) (-2.49) (-2.77)
Education2
Doctorate 0.702∗∗∗ 0.295+-0.145 0.619∗∗ 0.139 0.0422 0.614∗0.241 -0.237
(3.57) (1.67) (-0.48) (2.70) (0.72) (0.12) (2.55) (1.03) (-0.82)
High school 0.139 0.0702 0.0376 0.0455 -0.00544 -0.147 0.203 0.180 -0.0297
(0.58) (0.50) (0.14) (0.21) (-0.04) (-0.50) (0.80) (1.03) (-0.08)
<High school 0.00201 -0.232 -0.0986 0.0447 -0.194 -0.264 0.0653 -0.0885 -0.214
(0.01) (-0.91) (-0.31) (0.22) (-0.84) (-1.45) (0.27) (-0.31) (-0.56)
Other -0.113 -0.0766 -0.184 -0.365 -0.407 -0.0393 -0.202 -0.115 -0.130
(-0.30) (-0.15) (-0.23) (-0.61) (-0.70) (-0.03) (-0.39) (-0.25) (-0.12)
CS background -0.390∗∗ -0.482∗∗∗ -0.362∗∗
(-2.91) (-3.94) (-2.88)
Privacy behavior index3-0.0308 -0.0858∗∗ -0.102∗∗∗
(-1.01) (-3.06) (-4.82)
Westin index 0.152 0.00252 -0.110
(1.43) (0.03) (-0.56)
Constant 1.466∗∗∗ 2.456∗∗∗ 3.507∗∗∗ 1.438∗∗∗ 2.634∗∗∗ 4.079∗∗∗ 1.061∗∗∗ 2.073∗∗∗ 3.486∗∗∗
(6.06) (11.22) (9.51) (3.89) (9.28) (9.65) (3.54) (7.24) (6.93)
Observations 1,031 1,031 1,021
Pseudo R20.0220 0.0237 0.0259 0.0189 0.0236 0.0303 0.0179 0.0173 0.0239
tstatistics in parentheses; +p < 0.10, ∗p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001; 1(ref. female); 2(ref. university degree); 3range 0-13
Table 5. Simultaneous quantile regression on the median of weeks each user was trackable using feature set TT
u
Compared to a standard linear regression, a quan-
tile regression provides a “more complete picture [. . . ]
about the relationship between the outcome y [dependent
variable] and the regressors x [independent variables] at
different points in the conditional distribution of y” [18].
Thus, it is possible to figure out if there are different ef-
fects of the independent variables at different stages of
the dependent variable, more precisely: the first quar-
tile (25%), the second quartile (50%, median), and the
third quartile (75%). The simultaneous quantile regres-
sion estimates three regressions simultaneously (one per
quartile) and calculates the standard errors via boot-
strapping. Standard errors are used to estimate the t-
test for each coefficient and decide whether relations are
significant. “The bootstrap generates multiple samples by
resampling from the current sample” [18] which is nec-
essary to account for multiple testing due to the three
simultaneous regressions to ensure correct standard er-
rors and thus correct t-tests.
To analyze the effects of the regressors on the three
stages, regression coefficients are compared. Coefficients
can be interpreted as changes in the dependent vari-
able, when the independent variable increases by one.
As a measure of fit, Pseudo R2values are provided. The
range of Pseudo R2is from 0 to 1, with high Pseudo-R2
indicating a good fit of the regression model.
Table 5 presents results of three regression models.
All models include age, gender and education level, in
addition Model 1 contains the variable computer science
(CS) background,Model 2 the privacy behavior index
and Model 3 the Westin index. These three variables
were tested in different models because of the correlation
between them. For example, users with a CS background
have, on average, a 1.7 points higher privacy behavior
index than those without CS background.
Model 1. According to H1, there is a relation be-
tween age and the number of weeks that users were
trackable. The positive coefficients of age indicate that
older users are longer trackable than younger ones. For
example, the coefficient of 0.0147 in the second quartile
means that with each year a person is older, their me-
dian trackability is 0.0147 weeks higher, so older users
are longer trackable than younger ones.
The t-tests show that the age effects are only sig-
nificant for the second and third quartile. The age coef-
ficient of 0.0245 means that with each year a person is
older, the third quartile of trackability is 0.0245 weeks
higher. The slightly higher coefficient indicates that in

Long-Term Observation on Browser Fingerprinting 569
the group of long trackability duration, older users are
longer trackable.
For gender (H2), a negative significant effect can
be found for the third quartile. Thus, in the group of
the longest trackable users, men are shorter trackable
than women with a difference of 0.691 weeks. Significant
effects for education (H3) can only be found between
users with a PhD and a university degree for the first
quartile. The first quartile of trackability is 0.7 weeks
lower for users with a PhD than for those with a uni-
versity degree.
According to the three significant negative coeffi-
cients in Model 1, users with a CS background (H4) are
shorter trackable than those without: The duration dif-
fers from 0.39 to 0.482 and 0.362 weeks, concerning the
first, second and third quartile.
Model 2. Here, the same variables are contained
as in Model 1 except for CS background, and the pri-
vacy behaviour index (H5) is considered additionally.
Age, gender and education show the same effects as in
Model 1. The coefficients of privacy behaviour index in-
dicate that users differ in the second and third quartiles
where those who exhibit more privacy-protecting be-
haviour are shorter trackable. The difference is 0.086 in
the second and 0.102 in the third quartile.
Model 3. Besides age, gender and education, the
Westin index (H6) is contained. There is no significant
relation between the Westin index and trackability.
Summary. There are significant relations between
trackability and age, gender, education level, CS back-
ground and privacy behaviour index. However, these ef-
fects are rather low: The independent variables influence
the trackability duration only by some days. Also, the
Pseudo R2values as the measures of fit are quite low.
To sum up, we could find some significant effects, but
they only have low impact on the trackability duration.
6 Users’ Perspective
Between December 27, 2017 and January 24, 2018,
while the study was running for almost two years, we
conducted a second user survey (Sec. 3.1.5) to an-
swer the research question RQ4 about users’ perception
of browser fingerprinting, applied countermeasures and
impact of our study.
We recruited participants by sending an email to
760 study participants who were subscribed at that
point in time, and 243 (32%) responded.
Fig. 4. Agreement w/ statements on browser fingerprinting
(N=234)
We performed a qualitative content analysis [19] of
the free-text answers about the study impact. For each
of the three questions, two researchers read the first
50 answers independently and identified categories for
those answers. They then discussed their categories and
compiled a codebook together before starting their ini-
tial coding for the first 50 answers of each question.
Afterwards, they calculated Cohen’s Kappa κ[20] and
discussed their initial coding results and disagreements.
Then, they agreed on clearer category descriptions and
coded all answers of each question. The achieved inter-
coder agreement was excellent (κ > 0.75) for 16 out of
22 categories, and good (κ > 0.4) for the remaining 6
categories [21]. Finally, remaining disagreements were
discussed, so that full agreement could be reached.
6.1 Understanding, Concern, Protection
234 participants indicated their agreement or disagree-
ment with 8 items (Fig. 4). 99.1% agreed that browser
fingerprinting can be used by websites to recognize
users. 79.5% indicated that they think they understand
how browser fingerprinting works.
85.5% were concerned that websites and companies
try to fingerprint their browser. 62.4% thought that
most websites use browser fingerprinting, and 74.8%
thought that the websites they frequently visit use it.

Long-Term Observation on Browser Fingerprinting 570
The majority (78.5%) indicated that being pro-
tected from browser fingerprinting is important to them.
However, 68.3% agreed that protecting themselves from
browser fingerprinting requires a lot of effort, and 55.1%
stated they were not capable of protecting themselves.
6.2 New Insights and Experiences
193 participants answered the question whether they
experienced or learned something new by participating
in our study on browser fingerprinting (see Table 6). 169
of them (87.6%) answered in the affirmative.
“Yes” answers. 75 users (44.4% of 169) stated
technical insights. For example, P127, learned “how easy
it is to collect user data”, and P151 about “IP address
leakage over WebRTC”, which was one of the features
we collected. 64 users (37.9%) experienced their individ-
ual trackability. Thus, P7 explained: “It seems I’m easily
recognizable when I browse the Internet”, and P29: “I’m
much more recognisable than I thought”.
33 users (19.5%) expressed awareness, e.g., P111
learned “that browser fingerprints exist” or P91
who learned: “what fingerprinting actually means and
whether it is good or negative”. Finally, 24 (14.2%) of the
answers were about countermeasures. P38 said: “[...]Ap-
ple’s unification [of browser features] results in a worse
recognition by websites”, and P67 realized: “Without
blocking JavaScript it is hard to stop fingerprinting”.
“No” answers. 24 participants stated that they
did not learn or experienced anything new. 13 of them
did not receive enough information. For instance, P12
‘[...] knew browser fingerprinting already”, and P98
would have liked “[...] hints on how to protect myself”.
10 users indicated that they did not put enough atten-
tion or effort into the study, such as P3 who “only
looked at the [fingerprinting] results”. 10 answers con-
cerned lack of background knowledge to understand in-
formation that was provided during the study, e.g., P84:
“my knowledge of computers [...] is too low”.
6.3 Participants’ Feelings
Did participation in our study changed participants’
thoughts or feelings regarding the Web? 166 participants
answered this question (see Table 7).
“Yes” answers. 78 participants answered in the
affirmative. Most stated that they realized the conse-
quences of browser fingerprinting, and thus became dis-
illusioned or disappointed about privacy. Thus, P133
said: “I did not think that things with user tracking and
all this stuff can be so bad. The era of digital human
husbandry is here”. Participants also expressed their in-
security or distrust, such as P219 who feels “more per-
secuted” or P100 who thinks “there is real concern that
this could be used en masse in the future to track people”.
Some answers referred to protection, i.e., wanting
countermeasures, or how countermeasures do not exist
or perform badly. For instance, P144 expressed: “Pri-
vacy is a real concern nowadays, and I don’t think most
people can take appropriate measures to protect them-
selves”. The users also said that they became more vig-
ilant, such as P7 who tries to stay offline more often.
“No” answers. 68 of 88 users whose thoughts and
feelings did not change were already aware of data col-
lection, browser fingerprinting, or tracking, and thus ex-
perienced no surprises. Thus, P71 was aware of the “sad
state of internet tracking” and P80 says that “compa-
nies have always done everything to the disadvantage
of their customers”. Some participants expressed their
helplessness because they cannot escape data collection,
whereas other users stated that they have no problem
with tracking, because “Advertising finances many free
services and I don’t want to pay for everything” (P102).
6.4 Participants’ Behavior
We also asked participants whether they changed their
behavior on the Web (see Table 8).
“Yes” answers. Out of 155 users, 53 said “yes”.
Most participants either applied specific countermea-
sures or changed their browsing behavior (conscious
browsing). For example, P33 uses multiple profiles, P45
“started to use Firefox for social media/websites that do
a lot of tracking, alongside Google Chrome”; P117 avoids
“unnecessary browsing”, and P4 started to “block more
[content] to browse more safely”.
Some users started looking for protection, i.e., ei-
ther doing research or thinking about it, but not apply-
ing countermeasures yet. Others became more cautious,
such as P143: “I’m much more cautious with my data
and closed my Facebook account after using it for 10
years”. Some users were unspecific about their behav-
ior, just stating that they protect themselves.
“No” answers. Most respondents (102 out of 155)
did not change their behavior. The main reason was the
lack of protection: “I believe my efforts are in vain and it
is really up to the Browser Vendors and Website Hosters
to do something.” (P162). Other users think that they
are already protected, e.g., they use only “trustworthy”

Long-Term Observation on Browser Fingerprinting 571
Category n%κDescription
Yes
(N=169)
Technical insights 75 44.4 0.68 Learned/experienced how fingerprinting works (details, statistics)
Individual trackability 64 37.9 0.81 Affected by fingerprinting (e.g., uniqueness, recognizability, stability)
Awareness 33 19.5 0.73 Gained/raised awareness on fingerprinting
Countermeasures 24 14.2 0.84 Existence, effectiveness, or importance of countermeasures
No
(N=24)
Not enough information 13 54.2 1.00 Wished or expected more information (e.g., on fingerprinting details)
Not enough attention/effort 10 41.7 1.00 Did not pay enough attention to study details
Not enough knowledge 10 41.7 1.00 Lack of background knowledge to understand provided information
Some answers were assigned to multiple categories
Table 6. Answer categories: Did you experience or learn something new by participating in our study? (N= 193)
Category n%κDescription
Yes
(N=78)
Realized consequences 47 60.3 0.76 Disappointment about privacy (e.g., realized magnitude of tracking)
Insecurity & distrust 19 24.4 0.77 Feeling insecure, worried, distrustful, persecuted, or uncertain
Protection 14 17.9 0.83 Countermeasures do not exist or perform badly
Vigilance 12 15.4 0.79 Became (more) cautious while browsing the Web
No
(N=88)
No surprises 68 77.3 0.88 Already aware of data collection, browser fingerprinting, or tracking
Helplessness 7 8.0 0.82 Helplessness towards data collection and tracking
No problem with tracking 7 8.0 0.71 Tracking has beneficial aspects, or is not important
Some answers were assigned to multiple categories
Table 7. Answer categories: Did your study participation change any of your thoughts or feelings regarding the Web? (N= 166)
websites (P16), or do not click on “dangerous links”
(P141). Finally, some users did not change their behav-
ior due to the cost of protection. P219 stated that “the
most effective ways to protect myself from fingerprint-
ing are also quite invasive.” and P94 is “not sure how I
could make my browser less unique while still continuing
to use my browser for the things I want to do”.
6.5 Applied Countermeasures
We asked participants whether they tried to protect
themselves from browser fingerprinting, and if yes,
which countermeasures they did apply (N= 118). Be-
low, we discuss categories of countermeasures that were
applied by more than 5% of the users. We note that
many users reported measures that do not protect from
fingerprinting, although they are generally good security
and privacy practices (see discussion in Sec. 7).
Browser extensions. 38 respondents reported rea-
sonable browser extensions such as NoScript or uBlock
to reduce their fingerprintable features by blocking
scripts on specific domains or in general. 10 respondents
reported partly effective browser extensions such as Pri-
vacy Badger which detects and blocks canvas finger-
printing from third-party domains. 26 respondents re-
ported extensions that provide no fingerprinting protec-
tion, such as HTTPS Everywhere or CookieAutoDelete.
Browser settings. Two participants hardened
their browser by customizing its configuration (e.g.,
privacy.resistFingerprinting in about:config),
and 19 participants reported browser settings that do
not prevent fingerprinting, such as Do-not-track, or
disabling third-party cookies.
Disabling JavaScript. Nine respondents disabled
JavaScript either completely or partly for some websites
and thus limited fingerprinting to HTTP features.
Private browsing modes. Eight participants used
the incognito or private browsing mode, which does not
protect from fingerprinting.
Spoofing. Six respondents spoofed their user-agent
to hide their genuine browser and operating system. Un-
fortunately, spoofing can be detected and introduces in-
consistencies that might make users distinguishable [14].
Tor browser. 22 respondents used the Tor browser,
which is reportedly the currently best countermea-
sure [22] due to its unification approach and hardening.
Multiple browsers and devices. 22 respondents
used different browsers/devices for specific tasks. Al-
though this may separate identities exposed to websites,
it does not prevent browser fingerprinting per se.
Virtual private network. 10 respondents used
VPNs for browser fingerprinting protection. However,
this only hides the client’s IP address, but is ineffective
against browser fingerprinting.

Long-Term Observation on Browser Fingerprinting 572
Category n%κDescription
Yes
(N=53)
Device-bound protection 20 37.7 0.96 Specific countermeasures (e.g., multiple devices/browsers/profiles)
Conscious browsing 14 26.4 0.74 Changed browsing behavior (e.g., not visiting certain websites)
Looking for protection 9 17.0 0.71 Looking for or thinking about countermeasures
More cautious 7 13.2 0.85 Being more cautious wo/ mentioning specific behavior
Unspecific protection 7 13.2 0.57 Applying countermeasures or changing behavior but wo/ any details
No
(N=102)
Lack of protection 37 36.3 0.87 Not knowing how to protect oneself / no (proper) defense
Already protected 27 26.5 0.92 Already cautious enough, or already applying countermeasures
Cost of protection 16 15.7 0.89 Not changing behaviour due to cost-benefit imbalance
Some answers were assigned to multiple categories
Table 8. Answer categories: Did your study participation change your behavior on the Web? (N= 155)
7 Discussion
Formal Concepts. Based on the definitions we pro-
posed (Sec. 2.2), a fingerprint is considered trackable if
it is unique-by-entity and stable. An entity is considered
trackable if it has at least one trackable fingerprint. Al-
though the focus of this paper is on fingerprinting, which
is a stateless tracking technique [10], these definitions
can also apply on stateful techniques. Tracking cook-
ies, for example, must be unique-by-entity and stable to
track individual entities over time.
In practice, adversaries may combine stateless and
stateful tracking techniques when they need a certain
level of ground truth. Furthermore, a unique stateful
identifier can make up for a fingerprint that is (tem-
porarily) not unique. If the goal of an adversary, how-
ever, is to distinguish different types of users, such as
users with a specific operating system, specific browser
language, or users that have visited a specific website at
least once, a fingerprint or an identifier does not have to
be unique-by-entity. Even if users might not be track-
able individually due to being in an anonymity set of
size >1 [11], they remain distinguishable from others by
being in this particular anonymity set.
Users’ Trackability. Since our study enabled par-
ticipants to use multiple browser and devices, we as-
sessed their trackability using the average and maxi-
mum stabilities of their fingerprints (Sec. 4).
Using feature sets from prior works [1, 4] on our test
set and considering splits for measurements of desktop
and mobile devices as well as all measurements, 67.6%–
93.1% of the users were trackable. The mean stability
of trackable fingerprints per user averaged to 3.1–3.4
weeks, and the maximum stability to 6.8–8.6 weeks.
Utilizing feature set optimization and crafting
device-dependent and device-independent feature sets
aiming for stability, we increased the mean stability
per user averaged to 10.7–11.9 weeks, and to 20.2–
27.2 weeks for the most stable fingerprint per user with
64.6%–94.5% of the users being trackable.
Likewise, we confirmed the applicability of data-
driven feature selection on the dataset of Fp-Stalker.
Although the level of ground truth on our dataset is
on user level, participants’ alternating use of, for exam-
ple, desktop and mobile browsers should have no quali-
tative impact on our results since such fingerprints are
distinguishable even after stemming features, such as
the user-agent string, and thus unlikely to be merged.
Our feature selection increases users’ trackability,
which has similar negative impact on privacy as the
linking of fingerprints that change over time [1, 3]. We
believe that both data-driven feature selection and fin-
gerprint linkability should be considered in privacy as-
sessments, as their combination provides a more realistic
view on users’ trackability.
Although we found significant relations between the
trackability of users and their demographics and privacy
behavior, the effect sizes were quite low. We did not find
any correlations between users’ trackability and their
use of countermeasures. As our user sample is biased
towards German tech-savvy, well educated male users,
these relations should be re-examined on user samples
that better represent the Internet population.
Device-dependent Fingerprinting. As shown in
Appendix C, some selected features (e.g., audio sam-
ple rate, accept-language) are similar for both desktop
and mobile devices. Other features, e.g., based on Flash,
were only picked for desktop devices due to lack of sup-
port on mobile devices. We argue that device-dependent
fingerprinting is a reasonable strategy for trackers. In
practice, such device-dependent feature sets require de-
vice type detection. This can only to a limited extend be
counteracted by spoofing browser features, as spoofing
may yield characteristic inconsistencies [14].

Long-Term Observation on Browser Fingerprinting 573
Feature Set Optimization. We showed on two
datasets the applicability of feature selection (Sec. 4.2).
However, our approach does not yield a one-fits-all fea-
ture set for every dataset. Its greedy algorithm may
yield suboptimal results. Yet, we believe that the re-
sults enable new insights to the importance of features
towards certain metrics on a given sample, such as the
stability of fingerprints. The greediness of our method
can be relaxed by not only yielding the first-best fea-
ture candidate for propagation, but all candidates of
the same quality while maximizing (or minimizing) to-
wards a given criterion and propagating them tree-wise.
Exhaustive, non-greedy propagation could help to assess
the (un)importance of features, and pre-filtering feature
candidates supports the selection.
Feature Coverage. When we designed our study
in 2015/16, the order of HTTP headers [4] had not been
discussed in related work yet and could thus not be con-
sidered in our evaluation in Sec. 4.2.1. We did not add
new features during the ongoing study to have a consis-
tent superset of features. However, we argue that this
feature is unlikely to change the results significantly due
to its low entropy [4] and low feature importance [3].
Data Collection. Most studies on browser finger-
printing are not performed in a real-world scenario with-
out users’ knowledge; thus, users behave differently [5].
Nonetheless, we have thoroughly filtered our data to re-
duce possible effects on our evaluations induced by par-
ticipants’ curiosity on becoming non-distinguishable.
The availability of data is essential for privacy re-
search on browser fingerprints. Our participants could
opt-in to contribute their measurements to a dataset for
scientific purposes. This data, however, will be restricted
to academia only (upon request) due to the abuse po-
tential by parties not committed to users’ privacy.
Users’ Perspective. Participants in our study
understood that browser fingerprinting can be used
to track them, and the majority indicated that they
understand how it works. Nevertheless, although we
asked participants explicitly about protection against
browser fingerprinting (Sec. 3.1.5), a noticeable num-
ber of reported protection measures were ineffective
against fingerprinting. This might indicate misconcep-
tions about fingerprinting. However, there is also a rea-
sonable chance that participants interpreted this ques-
tion more broadly in terms of protecting their online
behavior in general, and thus reported countermeasures
against other forms of threats, such as stateful tracking.
The majority of users were concerned about browser
fingerprinting, but they overestimated its prevalence on
websites. They expressed that protection is important
to them, but that it requires a lot of effort and they do
not feel capable of protecting themselves.
The most reported impact of our study on users
was awareness of how browser fingerprinting works and
their individual trackability. Our study had no impact
on most users’ behavioral changes, as they reported the
lack of protection, already being cautious, or the cost-
benefit imbalance of protection. On whether our study
affected how users now feel or think about the Web,
the respondents were divided: Most who affirmed it are
now disappointment about the prevalence of tracking,
and most who negated it were already aware about it.
8 Conclusion
In this paper, we presented the results of a 3-year on-
line study on browser fingerprinting with 1,304 partic-
ipants. Our study is the first to establish ground truth
on user level for long-term observations with multiple
devices per user, and to provide information on the de-
mographic representativeness of the dataset. Based on
two user surveys, we studied users’ privacy behavior,
their perception of browser fingerprinting, the counter-
measures they apply, and the impact of our study. We
investigated the trackability of users, and proposed fea-
ture stemming and feature set optimization to assess
fingerprint stability. Compiling device-dependent and
device-independent feature sets, our method increased
the mean stability of trackable fingerprints per user by
factor 3 compared to existing feature sets [1, 4] on our
data, and by factor 2 on a public dataset [3]. Ten years
after Panopticlick raised awareness, the situation has
not changed: Browser fingerprinting remains a threat to
users’ privacy. Hopefully, the twentieth anniversary will
be more enjoyable.
Acknowledgements
We thank Felix Freiling for simplifying the formalization
in Section 2.2. We further thank the anonymous review-
ers and our shepherd, Paul Syverson, for their thorough
and valuable comments. This research received no spe-
cific grant from any funding agency in the public, com-
mercial, or not-for-profit sectors.

Long-Term Observation on Browser Fingerprinting 574
References
[1] P. Eckersley, “How unique is your web browser?,” in Privacy
Enhancing Technologies, 10th International Symposium,
PETS 2010, Berlin, Germany, July 21-23, 2010. Proceed-
ings, pp. 1–18, 2010.
[2] H. Tillmann, “Browser fingerprinting - tracking ohne spuren
zu hinterlassen,” Master’s thesis, 2013.
[3] A. Vastel, P. Laperdrix, W. Rudametkin, and R. Rouvoy,
“FP-STALKER: Tracking Browser Fingerprint Evolutions,”
in 2018 IEEE Symposium on Security and Privacy, SP 2018,
Proceedings, 21-23 May 2018, San Francisco, California,
USA, pp. 728–741, 2018.
[4] P. Laperdrix, W. Rudametkin, and B. Baudry, “Beauty and
the Beast: Diverting Modern Web Browsers to Build Unique
Browser Fingerprints,” in IEEE Symposium on Security and
Privacy, SP 2016, San Jose, CA, USA, May 22-26, 2016,
pp. 878–894, 2016.
[5] A. Gómez-Boix, P. Laperdrix, and B. Baudry, “Hiding in the
Crowd: An Analysis of the Effectiveness of Browser Finger-
printing at Large Scale,” in Proceedings of the 2018 World
Wide Web Conference on World Wide Web, WWW 2018,
Lyon, France, April 23-27, 2018, pp. 309–318, 2018.
[6] D. Fifield and S. Egelman, “Fingerprinting Web Users
Through Font Metrics,” in Financial Cryptography and
Data Security - 19th International Conference, FC 2015, San
Juan, Puerto Rico, January 26-30, 2015, Revised Selected
Papers, pp. 107–124, 2015.
[7] J. R. Mayer, “Any person... a pamphleteer: Internet
Anonymity in the Age of Web 2.0,” 2009. Bachelor’s the-
sis: https://jonathanmayer.org/publications/thesis09.pdf,
accessed on August 5, 2019.
[8] Y. Cao, S. Li, and E. Wijmans, “(Cross-)Browser Finger-
printing via OS and Hardware Level Features,” in 24th
Annual Network and Distributed System Security Sympo-
sium, NDSS 2017, San Diego, California, USA, February 26
- March 1, 2017, 2017.
[9] A. Cooper, H. Tschofenig, B. Aboba, J. Peterson, J. Morris,
M. Hansen and R. Smith, “RFC 6973: Privacy Considera-
tions for Internet Protocols,” 2013. https://tools.ietf.org/
html/rfc6973, accessed on August 7, 2019.
[10] J. R. Mayer and J. C. Mitchell, “Third-Party Web Tracking:
Policy and Technology,” in IEEE Symposium on Security
and Privacy, SP 2012, 21-23 May 2012, San Francisco, Cali-
fornia, USA, pp. 413–427, 2012.
[11] C. Díaz, S. Seys, J. Claessens, and B. Preneel, “Towards
Measuring Anonymity,” in Privacy Enhancing Technologies,
Second International Workshop, PET 2002, San Francisco,
CA, USA, April 14-15, 2002, Revised Papers, pp. 54–68,
2002.
[12] P. Laperdrix, B. Baudry, and V. Mishra, “FPRandom: Ran-
domizing Core Browser Objects to Break Advanced Device
Fingerprinting Techniques,” in Engineering Secure Software
and Systems - 9th International Symposium, ESSoS 2017,
Bonn, Germany, July 3-5, 2017, Proceedings, pp. 97–114,
2017.
[13] A. Vastel, W. Rudametkin, and R. Rouvoy, “FP-TESTER:
Automated Testing of Browser Fingerprint Resilience,” in
2018 IEEE European Symposium on Security and Privacy
Workshops, EuroS&P Workshops 2018, London, United
Kingdom, April 23-27, 2018, pp. 103–107, 2018.
[14] A. Vastel, P. Laperdrix, W. Rudametkin, and R. Rouvoy,
“FP-SCANNER: The Privacy Implications of Browser Fin-
gerprint Inconsistencies,” in 27th USENIX Security Sympo-
sium, USENIX Security 2018, Baltimore, MD, USA, August
15-17, 2018., pp. 135–150, 2018.
[15] P. Kumaraguru and L. F. Cranor, Privacy Indexes: A Survey
of Westin’s Studies. 2005.
[16] J. B. Lovins, “Development of a Stemming Algorithm,”
Mech. Translat. & Comp. Linguistics, vol. 11, no. 1-2,
pp. 22–31, 1968.
[17] L. Molina, L. Belanche, and A. Nebot, “Feature Selection
Algorithms: A Survey and Experimental Evaluation,” in
IEEE International Conference on Data Mining, pp. 306–
313, 2002.
[18] A. C. Cameron and P. K. Trivedi, “Microeconometrics using
Stata, revised edition,” StataCorp LP, 2010.
[19] M. Schreier, Qualitative Content Analysis in Practice. Sage
Publications, 2012.
[20] J. Cohen, “A Coefficient of Agreement for Nominal Scales,”
Educational and psychological measurement, vol. 20, no. 1,
pp. 37–46, 1960.
[21] M. Banerjee, M. Capozzoli, L. McSweeney, and D. Sinha,
“Beyond Kappa: A Review of Interrater Agreement Mea-
sures,” Canadian journal of statistics, vol. 27, no. 1, pp. 3–
23, 1999.
[22] A. Datta, J. Lu, and M. C. Tschantz, “Evaluating Anti-
Fingerprinting Privacy Enhancing Technologies,” in The
World Wide Web Conference, WWW 2019, San Francisco,
CA, USA, May 13-17, 2019, pp. 351–362, 2019.
A Demographic Profile
Table 9 provides an overview on the demographics of
our study participants within our evaluated dataset.
B Feature Stemming
Table 10 shows examples for feature stemming (see Sub-
section 4.1) where both original feature values and their
stemmed versions are shown.
Stemming includes deletion of version substrings,
sorting of list-like values in alphabetical order, and as-
signment of IDs for plugins and MIME types using low-
ercase concatenations of their properties name,file and
suffix,type,desc, respectively, while removing non-
alphabetical characters.

Long-Term Observation on Browser Fingerprinting 575
total male female n/a
n%nrel. % n rel. % nrel. %
Data collection
participants 1,304 100.0 998 76.5 249 19.1 57 4.4
measurements 88,088 100.0 68,968 78.3 16,445 18.7 2,675 3.0
Age
Under 18 4 0.3 3 0.3 0 0.0 1 1.8
18-29 324 24.8 261 26.2 57 22.9 6 10.5
30-49 524 40.2 412 41.3 108 43.4 4 7.0
50-64 277 21.2 214 21.4 62 24.9 1 1.8
65 and over 109 8.4 90 9.0 19 7.6 0 0.0
n/a 66 5.1 18 1.8 3 1.2 45 78.9
1,304 100.0 998 100.0 249 100.0 57 100.0
Education
Less than high school 124 9.5 82 8.2 40 16.1 2 3.5
High school 130 10.0 106 10.6 23 9.2 1 1.8
University degree 731 56.1 580 58.1 144 57.8 7 12.3
Doctorate 114 8.7 86 8.6 26 10.4 2 3.5
Other 34 2.6 24 2.4 10 4.0 0 0.0
n/a 171 13.1 120 12.0 6 2.4 45 78.9
1,304 100.0 998 100.0 249 100.0 57 100.0
Occupation
pupil 23 1.8 21 2.1 0 0.0 2 3.5
student 329 25.2 259 26.0 65 26.1 5 8.8
selfemployed 120 9.2 105 10.5 14 5.6 1 1.8
employee 622 47.7 481 48.2 138 55.4 3 5.3
homemaker 12 0.9 4 0.4 8 3.2 0 0.0
pensioner 73 5.6 61 6.1 12 4.8 0 0.0
unemployed 12 0.9 12 1.2 0 0.0 0 0.0
other 30 2.3 23 2.3 6 2.4 1 1.8
n/a 83 6.4 32 3.2 6 2.4 45 78.9
1,304 100.0 998 100.0 249 100.0 57 100.0
Background
Computer science 750 57.5 637 63.8 94 37.8 19 33.3
Non-CS 513 39.3 351 35.2 153 61.4 9 15.8
n/a 41 3.1 10 1.0 2 0.8 29 50.9
1,304 100.0 998 100.0 249 100.0 57 100.0
Browser Fingerprinting
Knew before 893 68.5 749 75.1 119 47.8 25 43.9
Did not knew before 382 29.3 249 24.9 130 52.2 3 5.3
n/a 29 2.2 0 0.0 0 0.0 29 50.9
1,304 100.0 998 100.0 249 100.0 57 100.0
Table 9. Demographic profile of participants in our final dataset
C Feature Set Optimization
The feature sets crafted using data-driven feature se-
lection (see Section 4.2) is shown in Table 11: device-
dependent and device-independent feature sets, each op-
timized towards the number of trackable participants,
and towards the avg. stability of trackable fingerprints.
D Features
Due to the sheer amount of 305 browser features
that were either collected or derived from exist-
ing ones, we report them online: https://browser-
fingerprint.cs.fau.de/paper/pets-2020/artifacts/.

Long-Term Observation on Browser Fingerprinting 576
Feature Raw value Stemmed value
User-Agent (HTTP) Mozilla/5.0 (Linux; Android 7.0; SM-G920F
Build/NRD90M) AppleWebKit/537.36 (KHTML,
like Gecko) Chrome/59.0.3071.125 Mobile
Safari/537.36
Mozilla (Linux; Android; SM-G920F Build)
AppleWebKit (KHTML, like Gecko) Chrome
Mobile Safari
navigator.languages (JS) de,en-US,en de,en, en-US
navigator.plugins (JS) [{’file’: ’npIntelWebAPIUpdater.dll’,
’name’: ’Intel®Identity Protection
Technology’, ’desc’: ’Intel web components
updater - Installs and updates the Intel
web components’, ’ver’: ’5.0.9.0’}, ...]
[’intelidentityprotectiontechnology/
npintelwebapiupdater.dll’, ...]
navigator.mimeTypes (JS) [{’desc’: ’OpenXPS document’, ’suffixes’:
’oxps’, ’type’: ’application/oxps’},
{’desc’ :’XPSdocument’ ,’suffixes’ :’xps’
,’ type’:’ application/vnd.ms-xpsdocument’}
...]
[’oxps/applicationoxps/openxpsdocument’ ,
’xps/applicationvndmsxpsdocument/
xpsdocument’, ...]
Table 10. Examples for feature stemming

Long-Term Observation on Browser Fingerprinting 577
Optimization criterion Stability of trackable FPs Participants w/ trackable FPs
Device type Desktop Mobile Total Desktop Mobile Total
Feature set TD
sTM
sTT
sTD
uTM
uTT
u
flash_avhardware_disabled   ○   
flash_language ○     
flash_os_stemmed   ○   
flash_screen_resolution_stemmed_ratio   ○   
flash_type_touchscreen ○  ○   
flash_version      ○
fonts_count  ○    ○
fonts_flash ○   ○  
fonts_js ○  ○ ○  ○
http_accept_language     ○ 
http_accept_language_stemmed ○ ○ ○ ○  ○
http_donottrack ○ ○ ○   
http_useragent     ○ ○
http_useragent_stemmed  ○ ○ ○  
js_adblocker_enabled ○ ○ ○  ○ 
js_app_version      ○
js_audio_channeltype_moz_supported   ○   
js_audio_samplerate ○ ○ ○   
js_battery_get_supported  ○    
js_battery_level  ○    
js_canvas_2d_base64     ○ 
js_canvas_3d_base64      ○
js_cookies_enabled   ○   
js_donottrack     ○ 
js_donottrack_amiunique      ○
js_donottrack_navigator  ○ ○   
js_is_mobile ○     
js_language_browser   ○   
js_language_system   ○   
js_mimetypes    ○  
js_oscpu  ○    
js_oscpu_stemmed   ○   
js_pdf_reader      ○
js_platform    ○  
js_screen_devicepixelratio  ○ ○  ○ ○
js_screen_height     ○ 
js_screen_height_available     ○ ○
js_screen_resolution_avail_wh  ○  ○  
js_screen_resolution_avail_whc   ○   
js_screen_resolution_stemmed_avail_whc ○     
js_screen_resolution_stemmed_wh ○     
js_screen_resolution_ratio   ○   
js_storage_opendb_windows_enabled   ○   
js_storage_session_enabled  ○    
js_timezone     ○ 
js_vibrate_supported ○  ○   
js_webgl_version_stemmed  ○ ○   
uap_http_browser_family ○     
uap_http_os_family ○     
uap_http_os_major ○  ○ ○  
uap_http_os_minor  ○  ○  
uap_js_device_branding   ○   
Table 11. Results of greedy data-driven feature selection on our data