Complete Aromatherapy Catalog 2013

12 Pages·2013·0.23 MB·English

The Anatomy of Smartphone Unlocking
A Field Study of Android Lock Screens
MarianHarbach1,AlexanderDeLuca2,SergeEgelman1,3
1InternationalComputerScienceInstitute,Berkeley,CA
2Google,Zurich,Switzerland
3UniversityofCalifornia,Berkeley,CA
[email protected],[email protected],[email protected]
ABSTRACT on improving smartphone authentication, including improv-
Topreventunauthorizedpartiesfromaccessingdatastoredon ingadoption. Researchersproposenewlockingmechanisms
theirsmartphones, usershavetheoptionofenablinga“lock onaregularbasis.
screen”thatrequiresasecretcode(e.g.,PIN,drawingapat-
Asweshowinthispaper,theresearchonthesesystemsoper-
tern,orbiometric)togainaccesstotheirdevices. Wepresent
atesonassumptionsaboutsmartphoneauthenticationthatare
a detailed analysis of the smartphone locking mechanisms
overly optimistic or simply not true. For instance, a system
currentlyavailabletobillionsofsmartphoneusersworldwide.
thattakes10secondstoauthenticatemightbefineinalabora-
Through a month-long field study, we logged events from a
torysetting,butitsusabilityintherealworldisquestionable
panel of users with instrumented smartphones (N = 134).
whenauserhastounlockherphoneupto50timesperday.
We are able to show how existing lock screen mechanisms
Thus,inordertoimprovetheusabilityandadoptionofsmart-
provide users with distinct tradeoffs between usability (un-
phoneauthenticationmechanisms,anunderstandingofusers’
lockingspeedvs. unlockingfrequency)andsecurity. Wefind
insitubehaviorsoverlongperiodsoftimeisnecessary.
that PIN users take longer to enter their codes, but commit
fewererrorsthanpatternusers, whounlockmorefrequently Recent research has taken steps to better understand smart-
and are very prone to errors. Overall, PIN and pattern users phonelockinginthewild,suchasbyidentifyingbasicprefer-
spentthesameamountoftimeunlockingtheirdevicesonav- ences[23],usagepatternsandperformance[12],andattitudes
erage. Additionally, unlock performance seemed unaffected towardsunlocking[10]. However,theperformancedatafrom
forusersenablingthestealthmodeforpatterns. Basedonour theselaboratoryandfieldstudiesisstillnotverydetailedand
results, we identify areas where device locking mechanisms weareunawareofanystudythathaslookedatexactauthen-
canbeimprovedtoresultinfewerhumanerrors–increasing ticationtimesanderrorratesinarealworldsetup. Harbach
usability–whilealsomaintainingsecurity. et al. performed the only prior field study of which we are
aware [12], and their instrumentation did not allow for the
captureofunlockingerrorsnordidtheydifferentiatebetween
AuthorKeywords
preparationandactualunlockingtime,whichprecludedade-
Smartphone;security;Android;lockscreen;fieldstudy;
tailedusabilityanalysis.
usability
Weperformedafieldstudyoveronemonthwith134partic-
ipants of mixed demographics. We collected detailed data
ACMClassificationKeywords
on unlocking, authentication speed, error counts, and types
H.5.2. Information Interfaces and Presentation: User Inter-
of errors. We provide detailed data on the influence of er-
facesInputdevicesandstrategies,evaluation
rorsonoverallauthenticationtimes,showingthat,forexam-
ple,PINandpatternusersspendsimilaramountsoftimeun-
lockingtheirdevices,consideringthehighernumberoferrors
INTRODUCTION
Due to increased processing power and storage capabilities, committedbypatternusers. Wefurthermoreshowtheinflu-
modern smartphones store a plethora of sensitive data that enceofwearingawatchonoverallsessioncount, providing
users want to prevent others from accessing [10]. This, to- proofthatmanydeviceactivationsfocusonsimpleactivities,
gether with security and usability problems of current au- suchascheckingthetime. Also, thosepatternusersthatac-
thentication systems [1, 2, 18, 22], has motivated research tivated the “stealth mode” (visited cells are not highlighted)
performedaswellastheirless-securecounterparts.
We argue that our work is complementary to the previously
mentionedstudiesandfillsanimportantgapleftbythesere-
Permissiontomakedigitalorhardcopiesofpartorallofthisworkforpersonalor searchefforts. Themaincontributionofthispaperisprovid-
classroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributed
ingabenchmarkofcurrentsmartphoneauthenticationmech-
forprofitorcommercialadvantageandthatcopiesbearthisnoticeandthefullcitation
onthefirstpage.Copyrightsforthird-partycomponentsofthisworkmustbehonored. anisms for future research to be compared against, as users
Forallotheruses,contacttheOwner/Author. areunlikelytoinvestmoreeffortintounlockingtheirsmart-
Copyrightisheldbytheowner/author(s).
CHI’16,May07-12,2016,SanJose,CA,USA phonesthantheycurrentlydo.
ACM978-1-4503-3362-7/16/05.
http://dx.doi.org/10.1145/2858036.2858267
RELATEDWORK one data point per day, which made real world applicability
In this section, we discuss the need for research on smart- ofthedatatenuousatbest. Inordertofillthisgap,Harbach
phoneauthenticationandpresentresearchonreplacingorex- etal.[12]providedthefirstrealworlddataontheamountof
tendingcurrentsystemslikePINandtheAndroidunlockpat- timeusersspendunlockingtheirsmartphonesrelativetothe
tern. We also outline work that has looked into users’ real total amount of time spent using them. They also employed
worldperformanceandperceptionsofthesemechanisms. experiencesamplingtoprovidequantitativedataonthelike-
lihood of shoulder surfing, showing that there are very few
SmartphoneAuthenticationSystems caseswhenusersperceivethisthreat. Thishighlightstheap-
Currently, the most common smartphone authentication propriatenessofcontext-dependentauthenticationsystemsas
mechanisms are PINs, the Android unlock pattern, and bio- proposedbyHayashietal.[13].
metrics (e.g., Apple’s Touch ID or Android’s Face Unlock).
Egelmanetal.[10]performedaseriesofinterviewsandon-
Astheyarewidelyused,securityandusabilityproblemsbe-
linesurveystoprovidedataonwhy(orwhynot)userschoose
come more important. These systems need to prevent unau-
tosecuretheirsmartphoneswithlockingmechanisms. They
thorized parties from gaining access to devices (security),
also quantified and correlated the breadth of sensitive data
while also minimizing the legitimate user’s burden (usabil-
storedonusers’deviceswithusers’perceptionssurrounding
ity),intermsofbothcognitiveload(e.g.,rememberingaPIN)
thatdata. Forinstance,theyshowedthataccesstoemailisa
andthetimeittakestosuccessfullyauthenticate.
heavilyunderestimatedrisk.
PINsandpatternsarebothpronetoguessingattacks,asusers
Whilethispreviousworkhaspaintedabetterpictureofhow
tend to choose easy to remember and consequently easy to
usersinteractwithsmartphonelockingmechanismsandsug-
guesssecrets[1,6,20]. Inaddition,authenticationcodesare
gestedwaysinwhichsmartphonelockingmechanismscanbe
highlysusceptibletoshouldersurfing(i.e.,apersonspyingon
improved, weareunawareofanypriorstudiesthathaveex-
the user’s input in order to steal the secret) [22]. Aviv et al.
aminedlockscreeninteractionsindetail. Thedatapresented
alsofoundthatsmudgesonthescreenscaneasilybeusedto
in this paper contributes to this line of work by filling a gap
inferanAndroidunlockpattern[2]. Finally,sidechannelat-
leftbyHarbachetal.[12]andEgelmanetal.[10]inthatwe
tacksusingbuilt-insensors(e.g.,accelerometer,microphone,
providedetailedquantitativedataonusers’interactionswith
etc.) have been proven to be efficient ways to infer a user’s
thesemechanismsinsitu.
PINorpattern[3,18].
Biometrics have recently become a popular authentication
METHOD
optionforsmartphones[8]. Whiletheysolvemanyofthese-
Our goal was to answer the following research questions in
curity and usability problems of previous mechanisms, they
ordertoprovideadeeperunderstandingofrealworldunlock-
also have some shortcomings. Besides privacy issues, re-
ingperformance, aswellastocollectbenchmarksandideas
searchersshowedthatsomebiometricsystemscanbetricked
forthedevelopmentoffuturenovellockingmechanisms:
withrelativelysimplemethods(e.g.,[11]). Inaddition,users
may perceive these systems as being insecure [17] or awk- 1. Howmuchtimedousersspendinteractingwithlockscreen
ward to use in specific situations [4, 8]. Thus, getting bio- UIelementseachday?
metricsrightishardandimportant:ifamechanismisseenas 2. Doeswearingawatchinfluencehowfrequentlythephone
low-effortandsecure,usersmaybemotivatedtoprotecttheir isactivatedorunlocked?
deviceswhentheyotherwisewouldnot[8,10]. 3. How much time do users spend on the lock screen before
beginningtounlock?
Therehasalsobeensignificantresearchefforttosolveexist-
4. Howlongdoesasuccessfulunlockattempttake?
ing security problems of smartphone authentication. These
5. How do code length and line visibility impact successful
includeadditionalbiometricsecuritylayersforPINs[24]and
unlocktime?
Androidpatterns[7],externalhardware[5],orimprovingse-
curitybyvisualmethodslikeindirectinput[14,19,21]. This 6. How many errors do users commit when unlocking their
isjustasmallexcerptofthehugebodyofliteratureonsmart- phones?
phone authentication. However, for any of these alternative 7. Howlongdoadditionalattemptstakeandhowmuchtime
methodstobesuccessfullyadopted,adetailedunderstanding iswastedthrougherrorsintotalperday?
ofhowusersinteractwithexistingsmartphoneauthentication 8. Whichtypesoferrorsoccurmostfrequently?
mechanismsinsituisneeded. 9. How frequently does the lock screen actually prevent
someonefromaccessingthephone?
RealWorldInsights 10. Whatareusers’subjectiveviewsofthetimeittakestoun-
To provide a baseline understanding of users’ interactions lockandtheerrorscommitted?
with existing smartphone authentication systems, recent re-
In order to obtain the fine-grained field data about users’
searchhasbeguntoexaminethewaysinwhichusersinteract
unlocking behaviors needed to answer these questions, we
withthesemechanismsoutsideoflaboratoryenvironments.
hadtoinstrumentusers’primarysmartphones. Thisalsore-
In2013,vonZezschwitzetal.[23]providedfirstinsightson quired a modification of the Android operating system and
thedifferencesbetweenPINsandtheAndroidpatterninareal thusflashinganewoperatingsystemontousers’devices. To
world setting. However, in their study, users only provided facilitate such studies, the University of Buffalo offers aca-
notreached Lock Homescreenandappinteraction Screen
Timeout dismissed Code Abort off
Code too
correct s-t-u R etry short Abort
Screen Preparation Key Code
entry
on wrong
begin
Retry
Slideaborted
Device sleeping
No unlock
Figure1. Theeventsandtransitionsloggedduringdatacollection. Sessionsareframedbyscreenonandoffevents. Afterkeyentryhasbegunwith
thefirsttouchontheunlockingUIelements,theenteredcodecaneitherbecorrect,wrong,ortooshort.Ifthecodewascorrect,codeentrywasnotyet
againnecessary,ornocodeisneeded(slide-to-unlock),thelockscreenisdismissed.
demicresearchersaccesstothePhoneLabpanel1[15]ofmore After key entry begins, the entered code can either be cor-
than 200 participants who have received customized smart- rect,incorrect,ortooshort. Ifthecodeisenteredincorrectly
phones. These LG Nexus 5 phones run a modified Android five times, the user will be prevented from trying again for
Open Source Project (AOSP) build and are instrumented to 30 seconds. If a code was too short (three digits/strokes or
send log entries to a central server. The phones are period- less),Androiddoesnotcounttheattempttowardsthenumber
icallyupdatedover-the-airsothatdatacollectionandexper- of failed attempts. After the key was entered correctly, the
imentation modifications can be deployed in a flexible man- Keyguard(i.e.,thelockscreen)isdismissed. Wealsolog
ner. Participantsprimarilyincludestudentsandstafffromthe thiseventtodetectdismissalswithoutcodeentry,e.g.,when
UniversityofBuffalo. thelockscreentimeout(thetimebeforeare-authenticationis
necessary)wasnotyetreachedorwhentheuserisnotusing
We extended the instrumentation available on PhoneLab to
a code-based lock. The latter case means that the slide-to-
collect the necessary data, which was transparently dis-
unlockmethod(s-t-u)isused,whichdoesnotinvolveacode,
tributedtothepanelasanover-the-airupdateoftheirphones.
butsimplyslidingthefingeracrossthescreentodismissthe
Attheendofthestudyperiod,wealsodeployedasurveyin-
lock. Android also offers the option to entirely disable the
strumenttoparticipantstocollectqualitativeinformationon
lockscreen,whichcausesthephonetoimmediatelyshowthe
their unlocking behaviors. In the next sections, we describe
homescreenwhenpressingthepowerbutton.
ourinstrumentationandtheexitsurveyinmoredetail.
After a failed or too-short unlock attempt, a user can either
retry(ifanyattemptsareleft),orabortunlockingaltogether,
Logging which will be logged as a screen off event. Finally, if the
To gather detailed data on unlocking events, we created log phonewasunlockedsuccessfullyandsubsequentlyinteracted
entries that were modeled based on a finite state machine with,thescreenwilleventuallybeturnedoffagain(byman-
(Figure 1): a session usually starts with the screen being uallypressingthepowerbuttonorreachinganidletimeout).
turned on. Following that, a user may interact with notifi-
To capture the necessary data, we relied on log output
cationsandotherwidgetsonthelockscreen,shemayfurther
captured by the PhoneLab instrumentation. The custom
unlock the phone, or turn the screen back off. Turning the
AOSP build already included log output for certain sys-
screen off involves no interaction with lock screen elements
tem events. We reused output generated for screen-related
(i.e., the user can either press the power button or allow the
events. SCREEN ON and SCREEN OFF actions are broad-
screen to timeout), whereas further unlocking the device in-
casted when the user presses the hardware button to turn on
volvesinteractionwithlockscreenUIelements. Thisinitial
thescreenandwhenatimeoutoranotherbuttonpresscauses
interactioncanamounttoenteringthefirstdigitorcharacter,
the screen to turn off respectively. As in the previous study
inputtingthefirstpatternstroke,orsimplybeginningtoslide
by Harbach et al. [12], these events frame the user’s inter-
across the screen to unlock the phone. We denote this point
action with the device. We then modified several classes
intimeasKeyEntryBegin. Previousworkreferredtothe
in com.android.keyguard to capture the lock-related
timebeforebeginningkeyentryaspreparationtime[21].
events. Witheachevent,wecapturedadditionalinformation,
suchasthetypeoflockscreenbeingused(i.e.,PIN,pattern,
1http://phone-lab.org–lastaccess08/31/15
orslide-to-unlock),thelengthoftheenteredcode,whetheror theirphonestoWiFiontimetouploadtheloggingdata. We
notadrawnpatternwasvisible(theso-called“stealthmode”), trimmed the remaining data to only include full days from
thereasonforturningthescreenoff(bytheuseroratimeout), midnighttomidnight. Wealsocutdownthedatatothefirst
andhowmanyattemptshadalreadybeenmade.Theaccuracy 30 full days, so that we can base the analysis on an equal
of the collected data, especially the timestamps, will be dis- amountoftimeperparticipant. Finally,weremovedafurther
cussedintheLimitationssection. 8participantswhousedmultipletypesoflockscreensduring
thestudyperiod,2whousedapassword,and1whodisabled
Finally, in post processing, we converted the sequence of
the lock screen altogether, as their small numbers preclude
eventsinthelogintostatisticsabouttherespectivetransitions
comparativeanalysis.
ofthestatemachine.
For these remaining 134 participants, we preprocessed log
Survey datatodiscardeventsthatwereoutoforder. Thiswaslikely
At the end of our data collection period, PhoneLab staff caused by threading issues in AOSP, causing a log entry for
emailed participants a link to our online exit survey. As an a later event to be written before an earlier event’s entry.
incentivetocompletethesurvey,wecreatedadrawingfora ThisanddataaccuracyarediscussedintheLimitationssec-
$100 Amazon gift card. The survey asked for demographic tion. Overall, cleaning amounted to 1.2% of events being
information and subjective views of the unlocking process. discardedonaverage§.
These views were collected by proposing statements (e.g.
”unlocking my phone is easy”) to which participants were Nsurvey 71
asked to indicate their agreement on a 7-point Likert scale. Age 19–70years,Mdn=35years
We also asked them open-ended questions to tell us about Gender 36 female
situations in which they were particularly happy to have a 34 male
lockscreenorparticularlyannoyedtohavealockscreen,and 1 n/a
Ethnicity 38 white
whetherthereareanysituationsinwhichtheyusuallystrug-
5 hispanic/latino
gletounlocktheirphones.
4 black/africanamerican
1 nativeamerican
StatisticalAnalysis 14 asian/pacificislander
6 other
Therearetwowaysinwhichwecanlookatthelogdatawe
3 n/a
collected: (a)summarizingtheloggeddatabyuserandcom-
paring users’ performances; and (b) using the raw informa- Highestdegree 11 highschooldiplomaorless
40 bachelor/masterdegree
tion. The lattercase allows for a detailedview on theentire
18 doctorate/professionaldegree
distribution of all collected events, while the former results 2 n/a
inalossofinformation. However,moststatisticaltestsoper- Annualhouseholdincome 34 lessthan$50k
ateundertheassumptionthatindividualdataitemsareinde- 19 lessthan$100k
13 morethan$100k
pendentofeachotherandthusrequiresummarizingperuser
5 n/a
first. Whenlookingatdifferencesbetweenlockscreentypes,
we will thus be relying on data that is first summarized per Nlogging 134
user. However,toalsogiveasenseofthetotaldistributionof Lockscreentype 32 PIN
thetimingdatawegathered,wewillalsobediscussingsome 35 pattern
67 slide-to-unlock
oftherawvalues. Toclarifywhichdatawassummarizedper
Avg.PINlength 4.25 digits(sd=.84,range:4-8)
user first and which was not, per user data will be marked Avg.patternlength 5.9 cells(sd=1.81,range:4-9)
withasectionsign(§)andrawdatawithadoubledagger(‡). Patternstrokesinvisible 8 participants
Whenconductingsignificancetesting,weusenon-parametric Table1.Participantdemographics.
tests, as many of the distributions we encountered were
Of the 134 participants we received log data from, 71 com-
skewed. Unlessotherwisenoted,weuseKruskal-Wallisrank
pleted our survey. We suspect that the remaining 63 par-
sum tests for between-groups analyses with more than two
ticipants were not motivated to respond, as prior PhoneLab
groups and Wilcoxon rank sum tests for data between two
data collection has been effortless for them. Table 1 gives
groups as well as post-hoc pairwise testing with Holm cor-
anoverviewoftheavailabledemographics. Thetypeoflock
rections. As there is no immediate effect size measure for
screenuseddidnotdifferbasedondemographicproperties.
Kruskal-Wallis tests, we report the effects of the significant
√
pairwisetestsusingther=Z/ N metric. In total, we observed an average of 246.4 events per user
per day (σ = 181.7, Mdn = 197.7, ranging from 33.4 to
RESULTS 1,170.9)§. In the following, we will present data to answer
WecollecteddatafromJuly15toAugust31,2015usingthe theresearchquestionsintroducedabove.
PhoneLabpanel,yieldingdatafrom202individuals. Before
analysis, we cleaned the data to exclude participants from TotalTimeSpentUsingandUnlocking
whom we received less than 30 full days of data (57 users). Our first goal was to examine the actual time participants
These participants are likely to either not have installed the spendinteractingwiththesecuritymechanismperdaytoget
AOSP update during the period of study or did not connect an idea of how much effort they are investing. Previously,
Method #sessions/day #unlocks/day Timespentunlocking/day[s] LockedSessionLength[s] UnlockedSessionLength[s]
Slide 75.0(56.2,58.4) 42.6(29.6,35.8)∗ 8.0(5.2,6.8)∗+ 87.1(178.4,44.6) 287.3(225.8,224.6)∗
Pattern 76.8(54.7,62.5)∗ 44.1(32.0,36.5)+ 64.1(87.8,48.7)∗ 58.1(65.4,32.0) 316.4(177.4,264.2)
PIN 53.2(44.7,46.1)∗ 29.5(22.6,22.3)∗+ 58.3(69.2,36.8)+ 59.6(59.5,38.7) 537.0(762.3,331.3)∗
Overall 70.3(53.8,57.1) 39.9(29.2,31.8) 34.7(61.8,18.3) 73.0(133.9,39.1) 354.6(423.6,259.5)
χ2=7.0,p=.03 χ2=7.5,p=.02 χ2=89.6,p<.0001 χ2=2.43,p=.30 χ2=6.75,p=.034
r∗2=.32 r+2=.47,r∗=.41 r+2=.79,r∗=.75 2 r∗2=.25
Table2.Sessionstatisticsbytypeoflockscreenused(mean,σ,Mdn)§.Pairwisesignificantlydifferentcellsineachcolumnaremarkedwithanasterisk∗
orplussign+(holm-correctedWilcoxonranksumtestsp<.05).
cantlymoretimeafterhavingunlockedthedevice,theyalso
unlocklessfrequently.Thissuggestsadifferentusagepattern
150 fortheseparticipants.
ay
D Finally, for each session, we also collected the termination
per reason, that is, how the screen was turned off. On average,
Sessions 100 MethSolidde pbuartttoicni)painnts61te.9rm%inoaftesdessseisosniosn(sσth=em2s6e.l8v%es,(uMsdinng=the71p.o8w%er,
mber of PPaINttern rtiaonngidnigdfnroomt d1if.f8e%r stiogn9ifi5.c7an%tl)y§.beAtwseeexnpelcotcekd,stchriesenprotyppoers-
Nu (Kruskal-Wallisχ2 =2.4, p=.3)§.
ge 50 2
a
Aver WearingaWatch
Whendesigningthisstudy,werealizedthattherecouldbea
mitigating factor for the difference between activations and
0
unlocks observed in previous studies: users regularly wear-
0 1 2
Time Spent Unlocking per Day log(x+1) [log(s)] ing watches may rely less on their phones to tell the time.
Also, having a smartwatch would further reduce this need
Figure2. Relationshipbetweentimespentunlockingperdayandaver-
agenumberofsessionsperdaygivendifferentlockscreentypes§. by showing notifications on the watch. So, as a side note
to the main aim of this paper, we asked in the exit survey
whether participants regularly wear a watch or smartwatch.
Out of the 71 participants who responded to the survey, 31
Harbach et al. [12] reported that users spend an average of
indicated they wear a watch, and two a smartwatch, on a
2.6minutesperdayunlockingtheirdevices. However, their
regular basis. While this precludes looking at the specific
analysisconsideredthetotaltimefromturningthescreenon
effect of smartwatches on activations, we found that the 33
to dismissing the lock screen. It stands to reason that only
participantswithanytypeofwatch(61.3activationsperday,
afractionofthattimeisactuallydevotedtoenteringacode.
σ = 56.3, Mdn = 46.6) activated their phones significantly
Table2givesanoverviewofsessionsandunlocksperpartic-
less frequently (W = 456, p < .05, r = .23) than partici-
ipantper day. Bothsessionsand unlocksdiffersignificantly
pants without watches (68.1 activations per day, σ = 31.0,
by type of lock screen used. The number of activations and
Mdn = 64.6). Thenumberofunlocksperday,however,did
unlocksperdayalignwiththefindingsofHarbachetal.[12].
notdiffersignificantlybetweenthesetwogroups.
However,ourmoredetaileddatashowthatmuchlesstimeis
actuallyspentunlockingthedeviceperdayonaverage.
TimeNeededBeforeanUnlockAttempt
Furthermore,patternandPINuserspredictablyspentsignif- As already shown above, our data provides a more detailed
icantly longer unlocking their devices than slide-to-unlock lookattheunlockprocess.Sonext,wewantedtoseewhether
users (Table 2). Interestingly, the overall effort in terms of thetimebeforekeyentrybeginsdiffersforlockscreentypes.
timespentforthecode-basedlockingmechanismsissimilar. Thisfirststepafterturningthescreenonhasbeenreferredto
Itisimportanttonote,thatwhilemostparticipantsspentless as“preparation.” Dependingonauser’sintention, heorshe
than100secondsperday,afewparticipantsspentalmost400 may look at notifications first and then unlock the phone, or
secondsormore(Figure2).Thefigurealsoshowsthatourset unlock the phone straight away; the preparation phase mea-
ofparticipantsincludedbothinactiveandveryactiveusers. sures how long she spent on the locked screen prior to initi-
atinganunlockattempt(i.e.,startingtodrawapattern,enter
Toputthetimespentunlockingintoperspective,Table2also
aPIN,orslidingherfinger). InFigure1,thisisthetransition
providesanoverviewofparticipants’averagesessionlengths
between“screenon”and“keyentrybegin.”
bylockscreentype. Thelengthofsessionswithunlocksdif-
feredsignificantlybetweenlockscreentypes:PINusers’ses- We observed that preparation times vary widely. Averaging
sionstookalmosttwiceaslongassessionsbyslide-to-unlock perparticipant,themeanpreparationtimeis71.5secondsper
users. Patternusers’sessionlengthsfellinbetween. Session attempt(σ=131.9, Mdn=38.7)§ whentheparticipantdoes
lengthsofnon-unlocksessionsdidnotdiffersignificantlybe- not unlock the device. If an unlock follows the preparation,
tween lock screen types. So, while PIN users spend signifi- however,meanpreparationtimeperparticipantandattemptis
3000
4000
2000 Method Method
nt Slide nt Slide
u u
o Pattern o Pattern
C C
PIN PIN
2000
1000
0 0
−2 −1 0 1 2 0 1000 2000 3000
Preparation Time [log(s)] Successful Unlock Time [ms]
Figure3. Distributionofpreparationtimebetweenlockscreentypes‡. Figure4.Distributionofsuccessfulunlockattempttimebetweendiffer-
Thelog-scalex-axisshowssub-secondintervalstotheleftofzeroand entlockscreentypes‡.
longerintervalstotheright.
Thetimeasuccessfulunlocktakesisverydifferentbetween
22.7seconds(σ = 84.1, Mdn = 3.4)§. Asthemedianvalue lockscreentypes(K-Wχ2 =107.6, p<.0001,pairwisetests
2
suggests,themajorityofpreparationintervalsisbrieferthan all p < .0001,pairwiseeffectsizes.74 < r < .82)§: slide-to-
four seconds and likely represents the cases where users do unlock users take 230ms on average (σ = 43,Mdn = 224),
notchecknotificationsfirst. pattern users 910ms (σ = 625, Mdn = 739), and PIN users
1,963ms (σ = 1,665,Mdn = 1,535). The differences also
Figure 3 shows a log-scale histogram of preparation times
become apparent in the histogram of all successful unlock
separated by lock screen types used when an unlock attempts(Figure4)‡. Giventhelongtailsofthedistributions,
followed‡.Preparationtimessignificantlydifferbetweenlock Table3liststhequantilesforeachunlockmethod‡.
screentype(K-Wχ2 = 22.2,p < .0001)§,withpairwisetests
2
yieldingsignificantresults(p < .05,.24 < r < .42)between
Method 25% 50% 75% 90%
all lock screen types. Median preparation time for slide-to-
unlockis2.2s,4.3sforpattern,and9.8sforPIN. Slide 120 150 210 420
Pattern 510 660 880 1,225
One explanation for the differences we observed in prepa- PIN 920 1,200 1,729 2,380
ration times are recall processes. For instance, preparation
Overall 158 480 880 1,397
times for slide-to-unlock might be so brief because there is
almostnorecallinvolved;onecanslidethefingeracrossthe Table3.Quantilesofsuccessfulunlocktimesinmilliseconds‡.
device in the right spot without even looking at the screen.
PatternmightbequickerthanPINbecausetheuserneedsto InfluenceofCodeLengthandLineVisibility
recallthePINfirst,whilewithapattern,theuseronlyneeds Another aspect worth looking at is the correlation between
tofindthestartingposition. However,thereisanalternative codelengthandsuccessfulunlocktime. Unfortunately,only
hypothesis:aswecannotprovecausation(anadditionalstudy 2 PIN users did not have a PIN of length four, which pre-
isneeded),slowerpeoplemaysimplybemorepredisposedto cludesstatisticalanalysisforthislockscreentype. However,
usingPINsratherthanpatterns. pattern users showed more variety: eight participants chose
a pattern of length 4, thirteen of length 5, two of length 6,
fouroflength7,twooflengtheightandsevenoflength9,al-
DurationofSuccessfulUnlocks
lowingustotakeacloserlookattimingdifferencesbetween
Next, we look at successful unlock attempts to establish a
thosegroups. Tothisend,wefitalinearmodelusingpattern
baselinethatfuturemechanismscanusetoevaluatetheirper-
length to explain average time for a successful unlock (Fig-
formance. The data we collected includes 160,746 unlock ure5, F(1,33) = 36.53, p < .0001,adjustedR2 = .51)§. Per
sessions, of which 75,298 (46.9%) involved entering an un-
additionalcellusedinthepattern,thesuccessfulunlocktime
lockcode. Theremaining53.1%pertainedtouserswithout
increasesby147msonaverage.
code-based lock screens or unlock attempts where entering
thecodewasunnecessary(timeoutnotyetreached). Ofthese Finally, therewasnosignificantdifferencebetweenthesuc-
code-based unlock sessions, 68,739 (91.2%) were immedi- cessful unlock times for participants with hidden pattern
ately successful (i.e., unlocking the device on the first try) strokes(i.e.,“stealthmode”;8participants)andthosewithout
andtook1.18secondsonaverage(σ = 33.98, Mdn = .82)‡. (27participants, W = 135, p = 0.30)§. Whilethisindicates
In the remaining code-based unlock sessions, participants that our stealth mode participants did not observably take
aborted their unlock attempts in 768 (1.0%) cases, and an- more time to correctly enter their patterns, this data is con-
other5,790(7.7%)unlocksessionsinvolvedatleastoneer- foundedbythefactthatthesetwogroupswereself-selected:
ror.Thesewillbediscussedinmoredetailinthenextsection. thereisnoevidencethatforcingallparticipantstousestealth
Besides using their phones less frequently, the lower error
ock [ms]3000 yyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy===================================−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−5555555555555555555555555555555555533333333333333333333333333333333333 +++++++++++++++++++++++++++++++++++ 111111111111111111111111111111111114444444444444444444444444444444444477777777777777777777777777777777777***********************************xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx rPaItNe upsroevrsidsepsenadnoutnhleorckeixnpglainsastiiomnilaorfawndhyevtehneaovliettrlaelllotiwmeer
nl
uccessful U2000 tulhonanlnogcetkrh.ewTpihathitsteearanlscouhspesreross’vs,iiodenveseananntdhooapudpgdohitriPtouInnNaitlyuasfteotrersmitmapktpseroalolvsneogmeterankttoe:
S
Time for 1000 rseadvuecuinpgtothaefiefrtrhorofrattheeiorfupnalottcekrninugsteirmsewpoeurlddaaylloownatvheermagteo.
ge Thismayalsomakethepatternlockmoreattractiveforusers
Avera 0 currentlynotusingacode-basedlockingmechanism.
0 1 2 3 4 5 6 7 8 9
Pattern Length
TypesofErrorsCommitted
Figure5.Regressionmodelbuilttoexplainsuccessfulunlocktimebased Looking at the types of errors committed, PIN and pattern
onpatternlength§.
users show similar behaviors. Table 5 lists the most com-
monerrors forbothmechanisms. Singleerrors arethemost
frequent,accountingformorethan70%oferrorsinbothun-
mode would not increase their unlock times. That is, those
lockingmethods‡. Interestingly,forbothmethods,singleer-
who chose to enable stealth mode were likely already com-
rors without successful unlocks are in the top five. Those
fortableenteringtheirpatterns.
mostlikelyrepresentaccidentalinputswithoutthetrueinten-
tiontounlock(e.g.,“pocketdialing”). Trulycriticalerrors–
NumberofUnlockErrorsCommitted
thoseattemptswherethelockscreenwasnotdismissedafter
Inthissection, weexaminethenumberoferrorsthe67par-
fivefailedattempts–arediscussedinthenextsection.
ticipantsofcode-basedlockscreenscommitted. In6,558in-
stances (11.5%), unlock attempts did not result in success
Method Errortype Count Proportion
immediately or at all. Pattern users committed errors much
morefrequently(5,667of46,921attempts,12.1%)thanPIN Pattern fs 2736 48.3%
users (891 of 28,376 attempts, 3.1%)‡. On average, pattern ts 1353 23.9%
users needed 1.13 attempts (sd = .06, Mdn = 1.11) to suc- t 385 6.8%
ffs 269 4.8%
cessfully unlock their devices, while PIN users only needed tts 192 3.4%
1.04attempts(sd = .03, Mdn = 1.03,W = 1064, p <.0001, fts 135 2.4%
r = .88)§. Thus, given the average number of unlocks per f 120 2.1%
day presented above, a PIN user would commit 1.3 errors PIN fs 477 53.5%
(σ = 1.2, Mdn = 1.0, rangingfrom.07to4.1), whileapat- ts 291 32.7%
tern user would commit 8.0 (σ = 6.6, Mdn = 4.9, ranging f 30 3.4%
from0.4to26.4)§. tts 27 3.0%
t 21 2.4%
Table5. Typesoferrorsbyunlockingmethod, explainingmorethan
TimetoRecoverfromErrors 90%oftotalerrors. f=failed(codewrong),t=too-short(codetooshort
Unlock errors make the unlocking process take longer, as tobecountedasproperattempt),s=success‡.
the user needs to recover. This includes realizing that an
error was made, potentially remembering the correct code, Too-short errors are responsible for 43.8% of pattern users’
andthenretrying. Table4givesanoverviewofthetimead- failedunlockattempts. Whilenotcountedtowardsthelock-
ditional unlocks take for PINs and patterns. Pattern users outthreshold(i.e.,thenumberoffailedattemptstheusercan
spend twice as much time on successive attempts, whereas make before the device prevents additional attempts), these
PINusersspendonlyaboutathirdmore. Giventheaverage 2,487 (5.3%)‡ of all pattern entry attempts are delaying the
errorrate,unlockingerrorscontribute13.4seconds(20.9%) unlock process, likely due to slipping up while entering the
to each pattern user’s daily unlocking time of 64.1 seconds pattern by lifting the finger too early. Last, enabling stealth
onaverage§. ForPINusers,thistimeamountsto3.5seconds mode(i.e.,whenthepatternstrokesareinvisible)didnotob-
(6.0%)oftheoverallaverageof58.3seconds§. Patternusers servably influence the number of additional attempts neces-
thus need to invest more than three times as much effort in sarytounlockthedevice(W =73, p=.17)§.
termsoftimetocompensateforerrors.
AccessPrevented
The last aspect of our data analysis concerned whether we
Method Firstattempt
would actually be able to observe instances where the lock
Pattern 791 σ=371 Mdn=703 screenpreventedaccesstothedevice. Suchunlockattempts
PIN 1,952 σ=1,765 Mdn=1,500
have been called critical errors in previous work (e.g., [9])
Method Furtherattempts and are usually assumed to happen after three attempts. On
Pattern 1,699 σ=2,805 Mdn=1,069 Android, however, critical errors occur after having entered
PIN 2,668 σ=1,141 Mdn=2,413 aPINorpatternincorrectly5times, excludingattemptsthat
Table4.Meanunlockingtimeperuserbasedonattemptandlockscreen are too short. The user then has to wait 30 seconds before
typeinmilliseconds§. beingabletoattemptkeyentryagain.Criticalerrorsinafield
Sentiments of users with code−based lockscreen
Like being
protected 5% 8% 87%
Is fast 15% 23% 62%
Is easy 26% 10% 64%
Is annoying
sometimes 44% 13% 44%
Could be easier 38% 18% 44%
Could be quicker 36% 28% 36%
Get code wrong
frequently 62% 15% 23%
Have problems
unlocking sometimes 85% 3% 13%
Is difficult 74% 13% 13%
Would use more
secure but longer 44% 28% 28%
100 50 0 50 100
Percentage
Sentiments of users without code−based lockscreen
Would use lock if quicker 22% 16% 62%
Would use lock if more secure 59% 25% 16%
100 50 0 50 100
Percentage
Response do not agree at all 2 3 4 5 6 fully agree
Figure6.Sentimentsofsurveyparticipants.Theywereaskedtoratetheiragreementtotherespectivestatementsona7-pointnumericalscale.
study scenario are either cases where a legitimate user has as their lock screen. Figure 6 gives an overview of senti-
troublerememberingorenteringthecode(falsepositives),or ments toward code-based lock screens that the 71 respon-
anunauthorizeduserwastryingtogetaccess(truepositives). dents provided. Participants appreciate the protection that
their lock screens afford them and they mostly think their
WeobservednocriticalerrorsforPINusersandonlysixcrit-
lockscreensarefastandeasy. However, theyalsoindicated
ical errors committed by five different pattern users. These
that lock screens can be annoying sometimes and that they
includedalargenumberofadditionaltooshortattempts,sug-
wouldliketohaveeasierandquickeralternatives. Mostpar-
gestingthattheymayresultfromunintentionalinputorsome-
ticipants did not feel that they get codes wrong frequently
onelikeachildplayingwiththedevice.
orhaveproblemsunlockingtheirdevices. Code-basedlocks
On the other hand, we can look at all unlock attempts for were also not perceived as being difficult. The responses of
code-based lock screens that were aborted with at least one participantswithacode-basedlockscreendidnotdiffersig-
failedentryattemptasanupperboundofaccessattemptsthat nificantlybasedontheirrespectivetypeoflockscreen.
wereprevented. Weobservedsuchcasesfor28patternusers
Theremainingitemaskedparticipantswithcode-basedlock
and for 12 PIN users. Individual PIN users had as many as
screens if they were ready to invest more time into using
8sucheventswhileonepatternuserhadabortedattemptsin
a more secure alternative. The majority of respondents did
30instances. Thesenumberswouldsuggestthatlockscreens
notagreewiththisstatement. Similarly,respondentswithout
are effective at keeping others from using the phone. How-
code-based locks would not adopt a secure lock screen if it
ever, it is very likely that they include instances where our
wasjustmoresecure.However,mostofthemwouldbeready
participants aborted themselves. Of course, this does also
toadoptoneifittooklessoftheirtime.
not account for cases where simply seeing an active, code-
basedlockscreendetersanunwantedpersonfromusingthe We also asked participants to describe situations in which
device. A dedicated investigation is necessary to determine theywereparticularlyhappytohaveacode-basedlockscreen
lockscreenefficacyintermsofitssecuritygoals. on their device. The 18 respondents who gave accounts of
suchinstancesmostlydescribedsituationswhereaphonewas
Altogether,ourdatasetseemstosuggestthattrulycriticaler-
misplacedorlostandthecodepreventedattackersfromeither
rors are very rare events for most users: they do not happen
abusing or keeping it in the participant’s view. One partici-
withinagivenmonth. Yet,lockscreensmaystillbeeffective
pant said: “Dropped my phone while walking on [a beach].
indeterringunwanteduse.
Ithinkthephonewas leftonarailingbecauseitdidhavea
lockfeature.” Theothertypeofsituationseveralparticipants
mentionedwaswhenvisitors,suchasdates,familymembers
QualitativeData
or children are present and the code then affords them pro-
After finishing log data collection, we sent out survey invi-
tection: “When I went to visit my 5 year old cousin and he
tations to all PhoneLab participants. As mentioned before,
always wanted to play with my phone.” Another participant
71 completed the survey. Of those, 39 used PIN or pattern
put it differently: “Always [happy to have it]. I just wish it Thus,lookingatourdata,itappearsthatthethreemostcom-
worked properly.” These statements also implicitly commu- monly used mechanisms to lock smartphones allow users to
nicatethetrade-offbetweensecurityandusability: whilethe makeatradeoffbetweensecurityandunlockingspeedorcon-
eventsduringwhichthecodeprotectedtheirdeviceanddata veniencegiventheirunlockfrequency. Thesubjectiveviews
are rare, the level of security it affords them is appreciated oftheunlockingprocessappeartosupportthis: participants
while the effort is accepted. It could also be that encounter- with a code-based lock screen were generally satisfied with
ingsucheventsisanimportantprecursorormotivatingfactor their unlocking experience and few considered the process
foradoptingorkeepingacode-basedlockscreen. problematic. Participantswhodonotcurrentlyhaveacode-
based lock screen indicated that they would be most moti-
Finally,weaskedourexitsurveyrespondentstodescribesit-
vatedtoadoptamoresecurealternativeifittooklesstime.
uations in which they particularly struggle to unlock their
devices. The 32 open-ended responses all describe contexts We also provided a code-length based model for pattern un-
thatcanmakeinteractingwithmobiledevicesmoredifficult lockingthatwillallowacomparisonofthetimeittakestoun-
in general. These situations involve either being distracted lock given the pattern length. Our data additionally showed
(“while driving”), being in a rush (“when I want to take a that whether or not the lines are visible during pattern en-
video/picture fast”), having limited dexterity (“when I have trydidnotobservablyinfluenceunlockingtimeorerrorrates
tousemylefthandinsteadofmyrighthand”),orwhenenvi- (thoughthismaybeconfoundedbyself-selectionbias). Ap-
ronmental conditions make recognizing touch input difficult proximationsforthiswerepreviouslyonlyavailablefromlab
(“rain, sweat”, “heat, humidity”, “when I have gloves, in studies. In addition, a dedicated, more detailed deconstruc-
winter”). Tworespondentsalsopointedoutlimitationsofthe tion of factors influencing the usability of the pattern lock
availableuserinterface. Theymentionedthatitisdifficultto screencouldbeofinterest,astheworkofvonZezschwitzet
hitthecorrectkeysindifficultlightingconditions(“whenit’s al.[22]suggeststhatseveralotherpropertiesofpatterns(for
brightout”)andingeneral(“Ijustaccidentallyhit[numbers] exampleknight-movesorcorners)impactsecurity.
nexttotheonesIwantfrequently”). Anothersaid, “thelast
AnotherinterestingaspecttonoteisthatmostPINuserslim-
digitofmypasswordandtheenterbuttonareincloseprox-
itedthemselvestofourdigits. TheprevalenceofPINsofthat
imitytoeachotherandIoftenpressenterinsteadofthedigit,
length in other applications likely influence this. As the rel-
whichisannoying.”
ativelysmallsearchspaceforfour-digitPINsandthelackof
Therefore, based on respondents’ subjective points of view, strongratelimitingonAndroidlockscreensaresecuritycon-
thepotentialforimprovedsolutionsliesinprovidingquicker cerns, futureworkshouldlookatavenuesforimprovingthe
alternatives to users who currently do not have lock screens security of these users while maintaining acceptable usabil-
and improving the performance and UI of existing lock ity. One possibility could be an improved PIN lock screen
screensindifficultsituationsandcontexts. thatusesspacedrepetitiontechniquestoeffortlessly“teach”
users one or two additional digits over time, similar to the
approachpresentedbySchechterandBonneau[16].
Intermsofunlockingerrors,theusersinourdatasetseemto
DISCUSSION
be well trained on their chosen locking mechanisms, as rel-
Inthispaper,weprovideddeeperinsightsintotheindividual
atively few errors were committed, and virtually none could
parts of the smartphone unlocking process. Going beyond
be considered critical. We only observed six attempts that
theinsightspreviousstudieshaveprovided, wewereableto
causedtheusertowaitbeforeattemptingfurtherkeyentries
providedetailedbenchmarksforthecurrentlyavailablelock
acrosstheonemonthdatacollectionperiod.Thisimpliesthat
screens on the Android operating system. Overall, the data
novellockscreenproposalsneedtobeabletoachievesimi-
presentedshowsthatsmartphoneuseingeneralandunlock-
larerrorratesinfieldenvironments,giventheamountoftime
ing behavior in particular varies widely between and within
usersarespendingunlockingtheirphonesonanaverageday.
users. Thelargerangesofmeasureddataandthecorrespond-
Additionally,ouranalysisdemonstratedapotentialtoreduce
ingstandarddeviations, aswellasthelongtailofthedistri-
thetimewastedbypatternusers: reducingthenumberofer-
butionsareatestamenttothat.
rors they commit every day could save them up to a fifth of
However, clear patterns also emerged that show differences theirtotaldailyauthenticationtime.
based on lock screen type: PIN users interact with their
Thequalitativeinsightswegatheredfromourparticipantsfur-
phones for longer, but less frequently. PIN users also need
thermore provide starting points for improving their experi-
longer to prepare unlocking and successful unlocks take
ence. Oneexamplecouldbetodetectwhenthephoneisheld
longer, butarealsolesserror-prone. Ontheoppositeendof
intheless-usedhandandthendisplayingthepatternframein
thespectrum,slide-to-unlockusersinteractwiththeirphones
a smaller space to make it more readily reachable. Another
morefrequentlybutforlesstime,whilealsoneedingverylit-
examplewouldbetodisablelockingafterthephonewasun-
tle time to unlock due to the lack of a code. Pattern users
locked once while being in a car or riding a bike. Lastly,
appear to be using their phones as frequently as users with-
improved touch screen hardware could help to improve on
outacode-basedprotectionandforsimilaramountsoftime.
environmentalinfluences,suchaslimitedcontrastduetoam-
However, theyspendaboutthesameoverallamountoftime
bientlightingconditionsormisdetectionoftouchinputfrom
unlockingtheirdevicesasPINusersbecausetheycommital-
moisthands.
mostsixtimesasmanyerrors.
Overall, one could argue that the available mechanisms suit Another limitation concerns accuracy. The accuracy of the
the needs of smartphone users. Room for improvement ap- timestampsinourlogfilesislimitedbythetemporalresolu-
pearstoexistfortheerrorratesofpatternusers,whichcould tion of the AOSP runtime and by its threading and process
save a considerable portion of their daily unlocking time. scheduling. While the logging infrastructure offers times-
Also,therelativelyfewparticipantsinthelongtailofthetime tamps at millisecond resolution, context switches between
spentunlockingdistributioncouldbeaworthwhiletargetfor threadsandprocessesmaycausethecreationofalogentryto
eitherimprovedsecurity(betteruseofthetimespent)orus- bedelayedfromtheactualuseraction. Bycollectingtimes-
ability (reducing the time). Finally, given the overall reluc- tamps twice in several parts of the logging instrumentation,
tancytoadoptcode-basedlockscreensthathasbeenobserved we found that context switches occur during the logging of
inthisandseveralotherstudiestothisday,itisunlikelythat 9.8% of events and caused an average delay in logging of
anovelmechanismneedingmoretimetounlockthanthecur- 494milliseconds(sd =238,Mdn=496)‡,insimilarpropor-
rentlyavailableoptionswouldseemuchadoption. tionsforalleventswecollected. Thisresultsineventsbeing
outoforderiftheyareshorterthanthedelay, inwhichcase
Thus, an important research challenge to tackle is to create
weexcludedthemfromtheanalysisduringpreprocessing. If
novel designs that take less or at most the same amount of
eventsarelongerthanthedelay,theseinstancesappeartobe
time as current methods, while providing protection against
shorter than they actually were in our data. As the extent
some of the threats current methods offer no protection for.
ofthisbiasislimitedandoccurringwithsimilarfrequencies
Atthesametime,researchonlockscreensneedstoshiftthe
for all events and users, we argue that the conclusions we
focusmoretowardsrealworldusability(authenticationtime
draw are still valid, although reported averages are likely to
anderrorratesinsitu)insteadof,forexample,mainlyaiming
beabout50msbrieferthantheactualdurationoftheevent.
to reduce the risk of shoulder surfing. Therefore, using lon-
gitudinalfieldstudiestovalidateproposalsseemsparamount,
CONCLUSION
as we observed great variance in our baseline data. A pos-
This paper provided the first detailed look at in situ perfor-
siblewayforwardwouldbetofocusmoreonunderstanding
mance ofAndroid lockscreenimplementations. Webelieve
whatmakescurrentlockscreensfast(e.g. otherthanmotor
thatwewereabletoprovideabenchmarkagainstwhichnovel
memory), what real-world security problems exist, and with
mechanisms can be evaluated. Given the preferences our
whatfrequencytheyoccur.
users expressed, it is unlikely that solutions requiring users
tospendmoretimeperdayunlockingtheirdeviceswillfind
significantadoption. Ourdataalsoshowedpotentialforim-
LIMITATIONS provementsthatcaninspirenewsolutions.
The data presented in this paper is limited in a few ways.
We were able to show that unlocking itself takes much less
First, due to the non-trivial nature of collecting this data in
time than previously approximated. Our data also suggests
thefield,allourparticipantscamefromthePhoneLabpanel.
that users interacting with their devices more frequently did
The panel is primarily recruited from students and staff of
soforshorteramountsoftimeandchoseaquickerlockscreen
theUniversityofBuffaloandisthusnotrepresentativeofall
type. However,overall,bothPINandpatternusersspentthe
smartphone users. Lock screen interactions and subjective
sameamountoftimeunlockingtheirdevicesonaverage.Yet,
viewsmaybedifferentforusersnotrepresentedbyoursam-
PIN users need more than twice as long before beginning
ple,althoughthereisnoindicationinpreviousworkthatthis
the unlock process, possibly to recall their PIN. We found
isaconcernforsmartphonesecurity.
that users wearing a watch on a regular basis activate their
Furthermore,allparticipantswereusingthesametypeofde- phones less frequently and that using stealth mode does not
vice, an LG Nexus 5. It is likely that performance of espe- influencetheunlockingtimeofparticipantsusingthepattern
ciallythepatternlockcandifferbasedondevicesize. Also, lock. Weprovideamodelfortheinfluenceofcode-lengthon
ourdatacollectionwaslimitedtoa30dayinterval. Thus,we successful unlock time for the pattern lock screen and show
cannotbesureoftheinfluenceofinfrequentevents–suchas how errors contribute a large amount of overall unlock time
travel, mental state, or illness – on the collected data. Last, forthismethod. Finally, wefoundlittledirectevidencethat
ourinsightspertainprimarilytothelockscreensavailableon lockscreenspreventsomeoneelsefromaccessingthephone
Android4.4andtheNexus5inJulyandAugust2015. While and provided insights into participants’ subjective views of
thepreviousworkbyHarbachetal.hasshownthattheseare theunlockingprocess.
themostusedlockscreensforAndroidusers,otherplatforms
providealternatives. Forexample,Apple’sTouchIDprovides ACKNOWLEDGEMENTS
averyfastyetalsosecurelockingmechanism. Ifitwaspos- We would like to thank the PhoneLab team at University of
sible to overcome the restrictions imposed by iOS, repeat- Buffalofortheirsupportingatheringthedata.Thisworkwas
ing this study on Apple devices would certainly be worth- partlysupportedbyafellowshipwithintheFITweltweitpro-
while. The performance of slide-to-unlock and PIN users gram of the German Academic Exchange Service (DAAD),
likelytranslatestousersofdevicesofsimilarsizeoniOSand as well as by the U.S. National Science Foundation under
Windows,astheuserinterfacesforunlockingareverysimi- awardCNS-1318680.
lar. However, itisalsopossiblethatthegroupsofusersthat
decide to use different operating systems also exhibit differ-
entperformancecharacteristics.