measuringhttp

TheEmperor'sNewPasswordCreationPolicies:AnEvaluationofLeadingWebServicesandtheEectofRoleinResistingAgainstOnlineGuessingDingWangandPingWangSchoolofEECS,PekingUniversity,Beijing100871,Chinawangdingg@mail.
nankai.
edu.
cn;pwang@pku.
edu.
cnAbstract.
WhilemuchhaschangedinInternetsecurityoverthepastdecades,textualpasswordsremainasthedominantmethodtosecureuserwebaccountsandtheyareproliferatinginnearlyeverynewwebservices.
Nearlyeverywebservices,nomatterneworaged,nowenforcesomeformofpasswordcreationpolicy.
Inthiswork,weconductanextensiveempiricalstudyof50passwordcreationpoliciesthatarecur-rentlyimposedonhigh-prolewebservices,including20policiesmainlyfromUSand30onesfrommainlandChina.
Weobservethatnotwositesenforcethesamepasswordcreationpolicy,thereislittlerationaleundertheirchoicesofpolicieswhenchangingpolicies,andChinesesitesgenerallyenforcemorelenientpoliciesthantheirEnglishcounterparts.
Weproceedtoinvestigatetheeectivenessofthese50policiesinresistingagainsttheprimarythreattopasswordaccounts(i.
e.
onlineguessing)bytestingeachpolicyagainsttwotypesofweakpasswordswhichrepresenttwotypesofonlineguessing.
Ourresultsshowthatamongthetotal800testinstances,541onesareaccepted:218onescomefromtrawlingonlineguessingattemptsand323onescomefromtarget-edonlineguessingattempts.
Thisimpliesthat,currently,thepoliciesenforcedinleadingsiteslargelyfailtoservetheirpurposes,especiallyvulnerabletotargetedonlineguessingattacks.
Keywords:Userauthentication,Passwordcreationpolicy,Passwordcracking,Onlinetrawlingguessing,Onlinetargetedguessing.
1IntroductionTextualpasswordsareperhapsthemostprevalentmechanismforaccesscontrolinabroadspectrumoftoday'swebservices,rangingfromlowvaluenewsportalsandftptransfers,moderatevaluesocialcommunities,gamingforumsandemailstoextremelysensitivenancialtransactionsandgenomicdataprotection[27].
Thoughitsweaknesses(e.
g.
,vulnerabletoonlineandoineguessing[42])havebeenarticulatedasearlyasaboutfortyyearsagoandvariousalterna-tiveauthenticationschemes(e.
g.
,multi-factorauthenticationprotocols[26,52]andgraphicalpasswords[56])havebeensuccessivelysuggested,password-basedauthenticationrmlystaysasthedominantformofuserauthenticationovertheThisisthefullversionofourpaperthatistobeappearedinProc.
of20thEuropeanSymposiumonResearchinComputerSecurity(ESORICS2015),Vienna,Austria,Sept.
21-25,2015.
2DingWangandPingWangInternet.
Duetobotheconomicalandtechnicalreasons[25],itwillprobablystilltaketheleadonwebauthenticationintheforeseeablefuture.
Ithaslongbeenrecognisedthatsystem-assignedpasswordsarehardlyusable[1,5],yetwhenusersareallowedtoselectpasswordsbythemselves,theytendtopreferpasswordsthatareeasilymemorable,shortstringsbutnotarbitrarilylong,randomcharactersequences,renderingtheaccountsprotectedbyuser-generatedpasswordsathighriskofcompromise[6,17,54].
Itisararebitofgoodnewsfromrecentpasswordstudies[16,47,50]that,ifproperlydesigned,passwordcreationpoliciesdohelpuserselectmemorableyetsecurepasswords,alleviatingthisusability-securitytension.
Unsurprisingly,nearlyeverywebser-vice,nomatterneworaged,followsthefashionandnowenforcessomeformofpasswordcreationpolicy.
Generally,apasswordcreationpolicy1iscomposedofsomepasswordcompositionrulesandapasswordstrengthmeter(seeFig.
1).
Theformerrequiresuser-generatedpasswordstobesatisedwithsomecomplexity(e.
g.
,acombinationofbothlettersandnumbers)andnudgesuserstowardsselectingstrongpasswords[10,39],whilethelatterprovidesuserswithavisual(orverbal)feedback[16,50]aboutthepasswordstrengthduringregistration.
Fig.
1.
AtypicalexampleofpasswordcreationpolicyHowever,towhatextentcanthewidely-deployedpasswordcreationpoliciesontheInternetberelieduponhaslongbeenanopenissue.
In2007,Furnell[19]initiatedaninvestigationintothepasswordpracticeson10popularwebsitesandfoundthat,passwordrulesandmetersarevastlyvariableamongtheexaminedsitesandnoneofthemcanperformideallyacrossalloftheevaluatedcriteria.
In2010,BonneauandPreibush[8]conductedtherstlarge-scaleempiricalstudyofpasswordpolicyimplementationissuesinpractice.
Byexamining150dierentwebsites,theyobservedthatbadpasswordpracticeswerecommonplaceandparticularly,highlyinconsistentpolicieswereadoptedbyindividualsites,whichsuggeststhatthereisalackofwidelyacceptedindustrystandardsforpasswordimplementations.
Atthemeantime,FlorencioandHerley[18]investi-gatedtherationaleunderlyingthechoicesofpasswordpoliciesamong75high-prolewebsitesandfoundthat,greatersecuritydemands(e.
g.
,thesitescale,thevalueprotectedandthelevelofseverityofsecuritythreats)generallydonotconstitutethedominantfactorforselectingmorestringentpasswordrules.
Instead,theseInternet-scale,highvaluewebservices(e.
g.
,e-commercesiteslikePaypalandonlinebankingsiteslikeCitibank)acceptrelativelyweakpasswords1Weuse"passwordpolicy"and"passwordcreationpolicy"interchangeably,anddon'tconsiderotherpasswordpolicieslikestorage[4],expiration[12]andrecovery[45].
TheEmperor'sNewPasswordCreationPolicies3andthesesitesbearingnoconsequencesfrompoorusability(e.
g.
,governmentanduniversitysites)usuallyimplementrestrictivepasswordrules.
Togureoutwhetherleadingwebsitesareimprovingtheirpasswordman-agementpoliciesastimegoeson,in2011Furnell[20]madeaninvestigationinto10worldwidetop-rankingsitesandcomparedtheresultswiththoseofthestudy[19]heperformedin2007.
Disappointingly,hereportedthat,duringthefour-yearinterveningperiodtherehasbeenhardlyanyimprovementinpasswordpracticeswhilethenumberofwebservicesandsecuritybreacheshasincreasedgreatly.
In2014,CarnavaletandMannan[11]investigatedtheproblemoftowhatextentthecurrentlydeployedpasswordstrengthmetersarelackofsounddesignchoicesandconsistentstrengthoutcomes.
Theysystematicallyevaluated13metersfrom11high-prolewebservicesbytestingabout4millionpasswordsthatareleakedfrompopularonlineservicesaswellasspecicallycomposedpasswords.
Itisfoundthatmostmetersintheirstudyare"quitesimplisticinnatureandapparentlydesignedinanad-hocmanner,andbearnoindicationofanyseriouseortsfromtheseserviceproviders"[11].
Fortunately,mostmeterscancorrectlyassignsensiblescorestohighlyweakpopularpasswords,e.
g.
,atleast98.
4%ofthetop500passwords[9],suchaspassword,123456,iloveyouandqwerty,areconsidered"weak"or"veryweak"byeverymeter.
Motivations.
However,mostoftheexistingworks[8,18–20]wereconductedveyearsago,whiletheonlineworldhasevolvedrapidlyduringtheinterveningperiod.
Inearly2010,Twitterhad26millionmonthlyactiveUsers,nowthisgurehasincreasedtenfold;2InNov.
2010,Gmailhad193millionactiveusers,nowthisgurereaches500million;3InApril2010,Xiaomi,aprivatelyownedsmartphonecompanyheadquarteredinBeijing,China,juststartedup,nowithasbecometheworld's3rdlargestsmartphonemaker(rankedafterAppleandSamsung)andthereare100millionXiaomiusersworldwidewhorelyonitscloudservice.
4Allthesethreesiteshaverecentlybeenthevictimsofhackingandleakedlargeamountsofusercredentials[37,40,43].
Aswewilldemonstrate,theyall(aswelleightothersitesexaminedinthiswork)havechangedtheirpoliciesatleastonceduringthepastveyears.
Moreover,atthattimehowtoaccuratelymeasurepasswordstrengthwasanopenproblemandtherewerefewreal-lifepassworddatasetspubliclyavailable,andthusthemethodologiesusedintheseearlierworksarefarfromsystematic(mature)andsatisfactory.
ThesolerecentworkbyCarnavaletandMannan[11]mainlyfocusesonex-aminingpasswordmetersfrom13sites,payinglittleattentiontotheotherpartofpasswordpolicies(i.
e.
,passwordcompositionrules).
Duetothefactthatapassword(e.
g.
,Wanglei123)measured"strong"bythepasswordmeterofasite(e.
g.
,AOL)mayviolatethepasswordruleofthissite,nallyitisstillrejectedbythesite.
Inaddition,manysites(e.
g.
,Edas,AOLandSohu)enforcemandatorypasswordrulesbutsuggestivemeters,apasswordmetered"weak"mightpassthepasswordruleofthesesites,andnallythis"weak"passwordisstillaccepted.
2http://www.
statista.
com/statistics/282087/3http://thefusejoplin.
com/2015/01/choose-google-gmail-yahoo-mail/4https://www.
techinasia.
com/xiaomi-miui-100-million-users/4DingWangandPingWangConsequently,thequestionofhowwellthesesitesactuallyrejectweakpasswordsandwithstandonlineguessingremainsunanswered.
Anotherlimitationofexistingworksisthatlittleattentionhasbeengiventonon-Englishwebservices.
Astypicalhieroglyphics,Chinesehasbeenthemainlanguageusedinatotalofover3.
64millionwebservicesuntil2014andabout0.
95millionnewwebservicesthatstartedupin2014(whichmeans0.
95Mnewpasswordpoliciescomeoutandimpactoncommonusers.
)[24].
What'smore,Chinesewebusers,whohavereached649millionbytheendof2014[13],havebeenthelargestInternetpopulationintheworldandaccountforaquarteroftheworld'stotalnetizens.
Therefore,itisimportant(andinteresting)toinvestigatewhat'sthestrengthsandweaknessesofthecurrentpasswordpoliciesinChinesewebservicesascomparedtotheirEnglishcounterparts.
Ourcontributions.
Themaincontributionsofthisworkareasfollows:(1)First,weproposeasystematic,evidence-groundedmethodologyformea-suringpasswordcreationpoliciesandinvestigatethestatusquoofpoliciesenforcedby50leadingwebservices(withspecialemphasisonChinesewebservices)withatotaloftenapplicationdomains.
Wendthat,generally,gamingsites,emailsites,e-commercesitesandnon-protorganizationsmanagewiththeleastrestrictivepasswordrules,whilethesitesofITmanufacturersimposethemoststringentones;Webportals,emailsites,e-commercesitesandtechnicalforumstendtoprovideexplicitfeedbacksofthepasswordstrengthtousers,whilesitesofsecuritycompanies,ITmanufacturersandacademicservices,ironically,oftendonotbothertoprovideuserswithanypieceofinformationaboutpasswordstrength.
(2)Second,weexplorethedierencesinpasswordpolicychoicesbetweenEnglishsitesandChinesesites.
ComparedtotheirEnglishcounterparts,Chinesesites,ingeneral,aremoreundaunted(audacious)intheirpass-wordrulechoices,whilethereisnosignicantdierencebetweenthesetwogroupsofsiteswithregardtothepasswordmeterchoices.
(3)Third,weemploystate-of-the-artpasswordcrackingtechniques(includ-ingtheprobabilistic-context-free-grammar(PCFG)basedandMarkov-Chain-based)tomeasurethestrengthofthe16testingpasswordsthatareusedtorepresenttwoprimarytypesofonlinepasswordguessingattempts.
Thisprovidesareliablebenchmark(ordering)oftheactualstrengthofthesetestingpasswordsbeyondintuitive(heuristic)estimatesasopposedtopreviousworkslike[11,20].
Weobservethatmostofthemetersoverestimatethestrengthofatleastsomeofthese16passwords,renderingthecorrespondingwebservicesvulnerabletoonlineguessing.
Thestructureofthispaperisasfollows:OurmethodologyiselaboratedinSec.
2;OurresultsarepresentedinSec.
3.
TheconclusionisdrawninSec.
4.
2OurmethodologyAsthereislittleresearchonstudyingpasswordpracticesandtheapproachesusedinthefewpioneeringworks[8,11,18,20]arefarfromsystematicandmayTheEmperor'sNewPasswordCreationPolicies5bedemodedoverthepastveyears,inthefollowingwetakeadvantageofstate-of-the-arttechniquesandelaborateonasystematicmethodologyformeasuringpasswordpolicies.
Asfarasweknow,forthersttimeseveralnewapproaches(e.
g.
,theuseoflarge-scalereal-lifepasswordsascorroborativeevidence,theuseoftargetedonlineguessingtomeasurepasswordstrength,andtheclassicationandselectionoftestingpasswords)areintroducedintothisdomain.
2.
1SelectingrepresentativesitesToinvestigatethestatusquoofpasswordcreationpoliciesdeployedintoday'sInternet(withspecialemphasisonChinesewebservices),rstofallweselectedtenthemesofwebservicesthatwearemostinterestedinandthatarealsohighlyrelevanttoourdailyonlinelives:webportal,ITcorporation,email,secu-ritycorporation,e-commerce,gaming,technicalforum,socialforum,academicserviceandnon-protorganization.
Then,foreachthemewechooseitstop5sitesaccordingtotheAlexaGlobalTop500siteslistbasedontheirtracranking(http://www.
alexa.
com/topsites).
Somecompanies(e.
g.
,MicrosoftandGoogle)mayoervariousservices(e.
g.
,email,search,news,productsup-port)andhaveafewaliatedsites,fortunatelytheygenerallyrelyonthesameauthenticationsystem(e.
g.
,WindowsLiveandGoogleAccount)tomanageallconsumercredentialsandwecanconsiderallthealiatedsitesasone.
Similarly,foreachthemewealsochooseitstop10sitesthatareamongtheAlexaTop500Chinesesitesranklist.
Inthisway,thereare15leadingsitesselectedforeachtheme:5fromEnglishsitesand10fromChinesesites.
Further,werandomlyselected5sitesoutofthese15sitesforeachtheme,resultingin50sitesusedinthiswork(seeTable5):20fromEnglishsitesand30fromChinesesites.
Wenotethatthoughourselectedwebsiteshaveawidecoverage,yetmanyoth-erthemesarestillleftunexplored,suchase-banking,e-healthande-government.
Theprimaryreasonwhywedoesnotincludethemisthat,theyrelyheavilyonmulti-factorauthenticationtechniquesinwhichpasswordsplayamuchlesscriticalrole.
Inaddition,thenumberofsitesallocatedforeachthemeisalsolimited.
Nonetheless,oursamplecharacterizesthecurrentmostrecognisedandleadingportionoftheonlinewebservices,whichattractthemajorityofthevisittrac[28,31].
Therefore,thepasswordpracticesusedbythesesiteswillimpactonthemajorfractionofend-usersandmayalsobecameamodelforotherlessleadingsites(whichgenerallyarewithlesstechnical,capitalandhumanresources).
Furtherconsideringtheamountofworkincurredforonesite,aninspectionof50sitesisreallynotaneasytask,letaloneaninitialstudylikeours(asthereisnosophisticatedproceduretofollow,wehavetocarryoutaniterativeprocessofdatacollection).
Inthefuturework,weareconsideringtoincreasethenumberofsitesforeachthemeto10orpossibly20,andtheinvestigationresultsaswellasasetofevidence-supported,practicablepolicyrecommendationswillbemadeavailableatthecompanionsitehttp://wangdingg.
weebly.
com/password-policy.
html.
2.
2MeasuringpasswordpolicystrengthThetaskofmeasuringstrengthofapolicyisgenerallyaccomplishedbyevaluat-ingstrengthofthepassworddatasetgeneratedunderthispolicy,andanumber6DingWangandPingWangofmethodsfortacklingthelatterissuehavebeenproposed,includingstatistical-basedones(e.
g.
,guessingentropyandα-guesswork[6])andcracking-basedones(e.
g.
,[34,53]).
However,thesemethodsallrequireaccesstoarealpassworddatasetwithsucientsize.
Fortunately,wenotethatFlorencioandHerley[18]'ssimplemetric—Nmin·log2Cmin—isnotsubjecttothisrestrictionandsucientforourpurpose,whereNministheminimumlengthallowedandCministhecardinalityoftheminimumcharsetimposed.
5Forinstance,thestrengthofapolicythatrequiresauser'spasswordtobenoshortthan6andmustcontainaletterandanumberis31.
02(=6·log236)bits.
Thismetricwellcharacterizestheminimumstrengthofpasswordsallowedbythepolicy,providingalowerboundofthepolicystrength.
Weadoptthismetricinourwork.
Table1.
BasicinformationaboutthesevenpassworddatasetsusedinthisworkDatasetServicesLocationLanguageWhenleakedHowleakedTotalpasswordsRockyouSocialUSAEnglsihDec.
14,2009SQLinjection32,603,387TianyaSocialChinaChineseDec.
4,2011Hackerbreached30,233,6337k7kGamingChinaChineseDec.
2,2011Hackerbreached19,138,452DodonewEcommerceChinaChineseDec.
3,2011Hackerbreached16,231,271CSDNProgrammingChinaChineseDec.
2,2011Hackerbreached6,428,287DuowanGamingChinaChineseDec.
1,2011Insiderdisclosed4,982,740YahooPortalUSAEnglishJuly12,2012SQLinjection453,4912.
3Exploitingreal-lifepassworddatasetsOurworkreliesonsevenpassworddatasets,atotalof124.
9millionreal-lifepasswords(seeTable1),totrainthecrackingalgorithmsandlearnsomebasicstatisticsaboutuserpasswordbehaviorsinpractice.
FivedatasetsofChinesewebpasswords,namelyTianya(31.
7million),7k7k(19.
1million),Dodonew(16.
3million),Duowan(8.
3million)andCSDN(6.
4million),wereallleakedduringDec.
2011inaseriesofsecuritybreaches[36].
TianyaisthelargestsocialforuminChina,7k7k,DodonewandDuowanareallpopulargamingforumsinChina,andCSDNisawell-knowntechnicalforumforChineseprogrammers.
TwodatasetsofEnglishwebpasswords,namelyRockyou(32.
6million)andYahoo(0.
5million),wereamongthemostfamousdatasetsinpasswordresearch[35,53].
Rockyouisoneoftheworld'slargestin-gamevideoandplatformforpremiumbrandslocatedinUS,anditspasswordsweredisclosedbyahackerusingaSQLinjectioninDec.
2009[3].
Thisdatasetistherstsourceoflarge-scalereal-lifepasswordsthatarepubliclyavailable.
YahooisoneofthemostpopularsitesintheworldknownforitsWebportal,searchengineandrelatedserviceslikeYahooMail,YahooNewsandYahooFinance.
Itattracts"morethanhalfabillionconsumerseverymonthinmorethan30languages".
ItspasswordswerehackedbythehackergroupnamedD33DsinJuly2012[55].
Wewillpayspecialattentiontothissitebecauseithaschangeditspasswordpolicy,asfaraswecanconrm,atleastthreetimesduringthepastveyears.
2.
4MeasuringpasswordstrengthEssentially,thestrengthofapasswordisitsguessingresistanceagainsttheassumedattacker.
Thisequalstheuncertaintythisattackerhastogetridof,andnaturallytheideaofshannonentropywassuggestedtomeasurepassword5Thisimplicitlyassumesthatusersareleast-eortones.
TheEmperor'sNewPasswordCreationPolicies7strength,calledNISTentropy[10].
Later,NISTentropywasfoundtocorrelatespoorlywithguessresistanceandcanatbestserveasa"roughruleofthumb"[34,53].
Incontrast,theguess-numbermetric,whichisbasedonpasswordcrackingalgorithms(e.
g.
,PCFG-basedandMarkov-based[35]),wasshowntobemuchmoreeective,andithasbeenusedinanumberoffollowingworkslike[38,47].
However,wenotethatthetraditionaluseofguess-numbermetricgenerallyimplicitlyassumesthattheattackerisarandom,trawlingattackerAtra(i.
e.
,nottargetingaselecteduser).
Inmanycasesthisisapparentlynotrealistic.
ForatargetedattackerAtar,withtheknowledgeofthenameofthetargetuser,shecandrasticallyreducetheguessnumberrequiredtondtherightpassword.
Inthiswork,weconsiderthesetwokindsofattackerandsupposethatthetargetedattackerknowoftheuser'sname.
Thisassumptionisreasonablebecause,forAtartolaunchatargetedattack,hemustknowsomespecicinformationaboutthevictimuserUv,andUv'snameisno-doubtthemostpubliclyavailabledata.
Totakeadvantageofnameinformationincracking,weslightlymodifythePCFG-basedandMarkov-basedalgorithmsbyspeciallyincreasingtheprobabil-ityofthename-relatedlettersegments.
ThiscanbeeasilyachievedinPCFG-basedattacks[35].
Forinstance,assumingthevictim'snameis"wanglei",afterthePCFG-basedtrainingphase,onecanincreasetheprobabilityoftheitem"L4→wang"inthePCFGgrammarstothatofthemostpopularL4segmentandsimilarly,theitem"L7→wanglei"tothatofthemostpopularL7segment.
Algorithm1:OurMarkov-Chain-basedgenerationoftargetedguessesInput:AtrainingsetTS;AnamelistnameList;Thevictimuser'snamevictimName;Thesizekoftheguesslisttobegenerated(e.
g.
,k=107)Output:AguesslistLwiththekhighestrankeditems1Pre-Training:2forname∈nameListdo3trieTree.
insert(name)4forpassword∈TSdo5forletterSegment∈splitToLetterSegments(password)do6ifInTrieTree(letterSegment)then7ifisFullName(letterSegment)then8password.
replace(letterSegment,victimName.
fullName)9ifisSurName(letterSegment)then10password.
replace(letterSegment,victimName.
surName)11ifisFirstName(letterSegment)then12password.
replace(letterSegment,victimName.
firstName)13OrdinaryMarkov-Chain-basedtrainingonthepre-trainedsetTSusingGood-TuringsmoothingandEnd-Symbolnormalization(see[51]);14ProducealistLwithtop-kguessesindecreasingorderofprobability.
However,forMarkov-basedattackssincethereisnoconcreteinstantiationof"lettersegments"duringtraining,wesubstituteallthenamesegments(in-cludingfull,sur-andrstnames)intrainingpasswords(weuse2MDuowanpasswordsand2MCSDNpasswordstogetherastrainingsets)withthevictim'scorrespondingnamesegmentsbeforetraining.
Forinstance,"zhangwei0327"is8DingWangandPingWangTable2.
Twotypesofpasswordsmodelingtwokindsofguessingattacks('Guessrank'istheorderinwhichthecorrespondingattackerwilltrythatguess;'–'=notexist)UserPasswordGuessrankinGuessrankinGuessrankinGuessrankintrawlingPCFGtrawlingMarkovtargetedPCFGtargetedMarkov12345611321234567893213520131468910TypeAwoaini121930423(Hotspot)iloveyou4334724359password8416434194woaini1314737116150132736password1231700236834657236679wanglei281595641wanglei12313929358523247wanglei14262786999345016TypeBwanglei121695462359711120558(Name-based)Wanglei12330208096222672323392wang.
lei30154778562392287915379205wanglei@1235291970–19271855109Wanglei@123––1927186206144Table3.
PopularityofTypeApasswordsinreal-lifepassworddatasetsHotspotTianyaDodonew7k7kDuowanRockyouYahoo(31.
7M,2011)(16.
3M,2011)(19.
1M,2011)(8.
3M,2011)(32.
6M,2009)(0.
5M,2012)PasswordRankFreq.
RankFreq.
RankFreq.
RankFreq.
RankFreq.
RankFreq.
12345613.
98%11.
45%13.
79%13.
43%10.
89%10.
38%12345678940.
59%30.
32%40.
63%30.
62%30.
24%60.
05%520131470.
19%50.
19%60.
34%60.
28%4150.
01%50900.
00%woaini170.
09%260.
04%150.
09%180.
07%36260.
00%–0.
00%iloveyou490.
04%1060.
01%530.
03%450.
03%50.
15%160.
03%password860.
02%230.
04%980.
02%870.
02%40.
18%20.
18%woaini13142950.
01%180.
05%720.
02%570.
03%873480.
00%–0.
00%password123200450.
00%80040.
00%224620.
00%143820.
00%13840.
00%1530.
01%replacedwith"wanglei0327","zhao@123"isreplacedwith"wang@123",and"pingpku@123"isreplacedwith"leipku@123",where"wang"and"lei"isUv'ssurnameandrstnameinChinesePinyin,respectively.
Ourbasicideaisthatthepopularityofname-basedpasswordsinthetrainingsetslargelyreectstheprobabilityofthetargetedusertouseaname-basedpassword,andthecleverattackerAtarwillbaseonthisprobabilitytoexploitUv'sname.
OurMarkov-basedalgorithmfortargetedonlineguessingisshownasAlgorithm1.
Onecaneasilyseethat,basedonouridea,besidesChinesePinyinnames,thisalgorithmcanbereadilyextendedtoincorporatenamesinanyotherlanguage(e.
g.
,"JamesSmith"inEnglish),andtoincorporateotheruser-specicdata(suchasaccountnameandbirthdate)tomodelamoreknowledgeabletargetedattacker.
Toavoidambiguity,weonlyconsidernamesegmentsnoshorterthan4.
Todeterminewhetherapasswordpickedfromthetrainingsetincludesanameornot,werstbuildaname-basedTrie-treebyusingthe20millionhotelreservationdataleakedinDec.
,2013[22].
Thisnamedatasetconsistsof2.
73millionuniqueChinesefullnamesandthusisadequateforourpurpose.
Wealsoadd504ChinesesurnameswhichareociallyrecognizedinChinaintotheTrie-tree.
ThesesurnamesareadequateforustoidentifytherstnamesofChineseusersintheTrie-treetobeusedinPCFG-basedtargetedguessgeneration.
TheEmperor'sNewPasswordCreationPolicies9Table4.
PopularityofTypeBpasswordsinreal-lifedatasetsNamedictionaryTianyaDodonew7k7kDuowanAverageRockyouYahooAverageChineseEnglishPinyinsurname(len≥4)6.
34%10.
04%7.
14%8.
44%7.
99%1.
38%1.
29%1.
34%Pinyinfullname(len≥4)9.
87%15.
90%11.
42%13.
42%12.
65%5.
37%3.
61%4.
49%Pinyinnametotal(len≥4)10.
91%18.
06%14.
81%14.
92%14.
68%5.
36%4.
21%4.
78%2.
5SelectingtestingpasswordsAswehavementionedinSection2.
3,wemeasurehowthe50passwordpoliciesweareinterestedinareresistanttotwotypesofguessingattacker,i.
e.
,atrawlingattackerAtraandatargetedattackerAtar(withthevictim'sname).
TheaimofAtraistobreakasmanyaccountsaspossiblewithafewpasswordtrials[6],whileAtarintendstobreakthesingleaccountofthegivenvictimuserUv.
Tobeeective,Atrawouldtrythemostpopularpasswordsindecreasingorderofprobabilitywithregardtothetargetingpopulation,whileAtarwouldtrythemostpopularpasswordsindecreasingorderofprobabilitywithregardtothespecicuser.
AsshowninTable2,weuseTypeApasswords(wecallhotspotpasswords)torepresenttheattemptsAtrawilltryandTypeBpasswords(wecallChinese-Pinyin-name-basedpasswords)torepresenttheattemptsAtarwilltry,respectively.
Asrevealedin[51],Chinesewebuserscreateanewtypeofpasswords,named"Chinese-stylepasswords",suchaswoaini,5201314andwanglei123basedontheirlanguage.
Notethat,"wanglei"isnotarandomstringoflength7butahighlypopularChinesename,amongthetop-20listofChinesefullnames[49];"520"soundsas"woaini"inChinese,equivalentto"iloveyou"inEnglish;"1314"soundsas"foreverandever"inChinese.
Thus,both"woaini1314"and"5201314"mean"Iloveyouforeverandever".
SuchpasswordsareextremelypopularamongChineseusers(seeTable3)andthusareasdangerousasinternationallybadpasswordslikeiloveyouandpassword123.
Inthefollowingweshowwhythesetwotypesofpasswordsareweakandcanreallyserveasrepresentativesofpasswordattemptsthattheaforementionedtwotypesofattackerwouldtry.
Table3revealsthat,alltheeightTypeApasswordsareamongthetop-200ranklistinatleastonewebservices.
Morespecically,alltheTypeApasswords(exceptwoaini1314andpassword123)areamongthetop-100ranklistinthefourChinesewebservices,whilewoaini1314isonlyslightlylesspopular(i.
e.
,witharank295)inTianyaandEnglishservices,andpassword123iscomparativelymuchmorepopularinEnglishservices,i.
e.
,witharank153inYahooandarank1384inRockyou,respectively.
Besidespopularity,theseeightTypeApasswordsarealsodierentinlength,culture(language)andcompositionofcharsets.
Therefore,theywellrepresentthecharacteristicsofpotentialpasswordsthatatrawlingattackerAtrawouldtry.
AsstatedinSection2.
4,tomodelatargetedguessingattackerAtar,wemainlyfocusonthecasethatAtarknowsofthevictim'sname.
Withoutlossofmuchgenerality,weassumethevictimisaChinesewebuser,named"wanglei".
FromTable4(andseemoredatain[51])wecanseethatChineseusersreallylovetoincludetheir(Pinyin)namesintopasswords:anaverageof14.
68%ofChineseusershavethishabit.
Thatis,givenatargeteduser,itiscondenttopredictthat10DingWangandPingWangthereisachanceof14.
68%thatsheincludeshernameintoherpassword,andAtarwouldgaingreatadvantagebymakinguseofthisfact.
Weconservativelydealwiththeambiguitiesduringthenamematching.
Forinstance,therearesomeEnglishsurnames(e.
g.
,Lina)maycoincidewithaChinesefullname,andwetakenoaccountofsuchnameswhenprocessingEnglishdatasets.
Well,howdoesauseruseshername,whichcanbeseenasaword,tobuildapasswordThereareadozenofmanglingrulestoaccomplishthisaim,andthemostpopularones[14,30]includeappendingdigitsand/orsymbols,capitalizingtherstletter,leetetc.
ThisresultsinoureightTypeBpasswords.
OnecanseethattheguessrankinMarkov-basedtargetedattack(seethelastcolumninTable2)quiteaccordswiththerankofgeneraluserbehaviorsassurveyedin[14].
ThisimpliestheeectivenessofourMarkov-basedtargetedattackingalgorithm.
2.
6CollectingdatafromsitesToobtainrst-handdataonpasswordpolicypractices,wecreaterealaccountsoneachsite,readthehtml/PHP/Javascriptsourcecodeoftheregistrationpage,andtestsamplepasswordstoseethereactionofthemeterwhenavailable.
Wenotethattherearemanyunexpectedbehaviorsofsites.
Forexample,insomesites(e.
g.
,Edas,EasycharandYahoo)thedescriptionsofpasswordpoliciesarenotexplicitlygiven(ortheinformationexplicitlygivenarenotcomplete),andadditionaldataaboutpoliciescanonlybeextractedfromthefeedbacksoftheserverafteronehaveactuallyclickedthe"submit"button.
Consequently,forallsitesandeverypasswordtestinginstance,wepressthe"submit"buttondownandtakenoteoftheresponsetoavoidmissinganythingimportant.
Initially,consideringthegreatamountofmanualworkloadinvolved,weat-tempttoautomatethecollectionofdatafromeachsitebyusingPHP/Pythonscriptsorwebspiders.
However,wehavetoabandonthisideamainlyduetofourreasons:1)Alargeportionofsites(38%)preventautomatedregistrationbyrequiringuserstosolveCAPTCHApuzzleswhenregistration;2)18%sitesneedtoinputthevericationcodereceivedbyuser'smobilephonetoaccomplishtheregistration;3)8%sitesinvolveavericationcodetobereceivedbytheuser'semailbeforetheusercaninputthepassword;4)Informationdisplayedoneachsiteishighlyheterogeneous,asdemonstratedinSection3,notwositessharethesamepasswordpolicy,andthusbatchprocessinghardlyworks.
Asaresult,thewholedatacollectionprocessismanuallyhandled.
Toassureaccuracy,everyprocessisconductedatleasttwice(atintervalsofmorethanoneweek)andthecollecteddataallhasbeencrossvalidatedbytheauthors.
3OurresultsInthissection,werstpresentthestatusquoofthepasswordpoliciesemployedinthe50webservicesstudied,andthenexaminetheeectivenessofthesepoliciesinresistingagainstonlineguessingattacks.
AllofthedatawerecollectedfromtheseservicesbetweenthemonthsofJan.
toFeb.
in2015.
3.
1PasswordcompositionrulesinthewildForeachpasswordcompositionrule,weinvestigatethefollowingsixcommonrequirements:lengthlimits,charsetrequirement,whetherrulesareexplicitlys-TheEmperor'sNewPasswordCreationPolicies11tated,whetherallowingsymbols,whetherusingablacklistandwhetherdeterringtheuseofpersonaldata.
TheresultsaresummarizedinTable5.
Lengthlimits.
Allsitesbutoneimposeaminimumlengthlimit.
60%sitesrequirepasswordstobenoshorterthan6,30%sitesrequirepasswordstobenoshorterthan8,withtheremaining8%sitesragingfrom5,7to9.
Itisinterestingtoseethat,allsitesfromtheITcorporationcategoryenforceaminimum-8lengthlimit.
IsthisbecausethattheseservicescarethesecurityofuseraccountmorethanotherservicesexaminedWewillexplorethisquestionlater.
Atthemeantime,72%sitesimposeamaximumlengthlimitnolargerthan64,asfarastheycanbeidentied.
Surprisingly,22%sitesdonotallowpasswordstobelongerthan16.
Asitiscognitivelyimpossibleforcommonuserstoremembercomplexnon-linguisticstringsyetattackvectorsareincreasing,passphraseshaverecentlyreceivedmuchinterestandshowntobemoreuseablethanpasswords[29,33,48],andactually,theyhavebeenusedsuccessfullyandgainpopularity(seehttp://correcthorsebatterystaple.
net/).
However,passphrasesarehighlylikelytoexceedsuchmaximumlengthlimitsandthusareprohibited.
Furtherconsideringthat,increasingthepasswordlengthisgenerallymoreeectiveinenhancingpasswordsecuritythanextendingthecharsets[23,44],itismoreadvisabletosetamaximumlengthlimitthatislargeenough(e.
g.
,64).
Charsetrequirement.
Amongthe50sites,23sites(46%)implementsomecharsetrequirements.
Onceagain,allsitesfromtheITcorp.
categoryenforceacharsetrequirement,whileothercategoriesdonotshowthisfeature.
Remarkably,3Chinesesitesrequirethatadigit-onlypasswordcannotbeshorterthansomeminimumlength(e.
g.
,9).
ThismaybeduetotheirinsightintothefactthatChineseusershighlylovetousedigit-onlypasswords—accordingtooneofourearlierworks[51],anaverageof52.
93%Chineseusersusedigit-onlypasswords.
Symbolacceptance.
Itisperhapssurprisingtonotethatfoursites(includingbothEnglishandChinesesites)preventsymbolstobeincludedintopasswords.
Theoretically,amongthe95printableASCIIcharacters,33onesaresymbols,excludingwhichwouldlargelyreduceanattacker'ssearchspace.
Ithasalsobeenestablishedempiricallythatpasswordswithsymbol(s)aregenerallymuchsecurethanpasswordswithnosymbol[38,53].
TheonlyplausiblereasonforforbiddingsymbolsthatwecanimagineistopreventSQLinjectionattacks,yetsuchattackscanbewellpreventedbyproperlyhandlingtheescapecharacters.
Itisreallybeyondcomprehensionwhythesefoursitesforbidsymbols.
Usingblacklist.
AsrecommendedinNIST-800-63[10],ablacklistofsucientsize(e.
g.
,atleast50,000)ishighlydesirableinpreventpopularpasswordswhichareparticularlyvulnerabletostatisticalattacks[46].
US-CERTalsosuggesttheuseofblacklist[39].
However,only16sites(32%)imposeablacklistandnoneoftheirblacklistsareadequate.
Forinstance,theblacklistofTwitteronlycontains370badpasswordsandironically,theblacklistofIEEEonlyconsistsofthefamous"password".
Alsonotethat,allemailsitesimposeablacklist;33%Chineseservicesimposeablacklist,andthisgureforEnglishservicesis30%.
Checkinguserinfo.
AshighlightedinbothNIST-800-63[10]andNIST-800-118[44],usestendtousetheirpersonaldata(e.
g.
,accountnameandpersonal12DingWangandPingWangTable5.
Anoverviewofthepasswordcompositionrulesintheselectedwebservices('-'meansalengthlimitoflargerthan64;''meansnocharsetrequirement;'Blacklist'meansalistofbannedpopularpasswordsorstructures(e.
g.
,repetition);'Userinfo'considerstwotypesofauser'spersonalinformation,i.
e.
nameandaccountname.
)WebServicesLen.
limitsCharsetRulesAcceptUsingCheckingMinMaxRequirementExplicitSymbolblacklistuserinfoSina616YesYesNoNoChina.
com6-1+lower,1+upper,1+digitNoaYesNoNoWebTecent616Notanumberwithlenhttp://thepasswordproject.
com/leaked_password_lists_and_dictionariesTheEmperor'sNewPasswordCreationPolicies193.
Allan,C.
:32millionRockyoupasswordsstolen(Dec2009),http://www.
hardwareheaven.
com/news.
phpnewsid=5264.
Bauman,E.
,Lu,Y.
,Lin,Z.
:Halfacenturyofpractice:WhoisstillstoringplaintextpasswordsIn:Lopez,J.
,Wu,Y.
(eds.
)ISPEC2015,LNCS,vol.
9065,pp.
253–267.
Springer-Verlag(2015)5.
Bishop,M.
,VKlein,D.
:Improvingsystemsecurityviaproactivepasswordchecking.
Computer&Security14(3),233–249(1995)6.
Bonneau,J.
:Thescienceofguessing:Analyzingananonymizedcorpusof70millionpasswords.
In:IEEES&P2012.
pp.
538–552.
SanFrancisco,USA(May21-232012)7.
Bonneau,J.
,Herley,C.
,vanOorschot,P.
,Stajano,F.
:Passwordsandtheevolutionofimperfectauthentication.
Comm.
oftheACM58(7),78–87(2015)8.
Bonneau,J.
,Preibusch,S.
:Thepasswordthicket:Technicalandmarketfailuresinhumanauthenticationontheweb.
In:Proc.
WEIS2010(June7-82010)9.
Burnett,M.
:10,000toppasswords(June2011),https://xato.
net/passwords/more-top-worst-passwords/10.
Burr,W.
,Dodson,D.
,Perlner,R.
,Polk,W.
,Gupta,S.
,Nabbus,E.
:NISTSP800-63–electronicauthenticationguideline.
Tech.
rep.
,NIST,Reston,VA(April2006)11.
Carnavalet,X.
,Mannan,M.
:Fromveryweaktoverystrong:Analyzingpassword-strengthmeters.
In:Proc.
NDSS2014.
pp.
1–16.
SanDiego,CA,USA(2014)12.
Chiasson,S.
,vanOorschot,P.
C.
:Quantifyingthesecurityadvantageofpasswordexpirationpolicies.
Designs,CodesandCryptography(2015),inpress,Doi:http://dx.
doi.
org/10.
1007/s10623-015-0071-913.
CNNIC:CNNICReleasedthe35thStatisticalReportonInternetDevelopmentinChina(Feb2015),http://www.
apira.
org/news.
phpid=173214.
Das,A.
,Bonneau,J.
,Caesar,M.
,Borisov,N.
,Wang,X.
:Thetangledwebofpasswordreuse.
In:Proc.
NDSS2014.
pp.
1–15(2014)15.
DISAforDoD:Applicationsecurityanddevelopment.
Tech.
rep.
,DefenseInfor-mationSystemsAgency(DISA),Reston,VA(July,2013),doi:http://www.
stigviewer.
com/stig/application_security_and_development/16.
Egelman,S.
,Sotirakopoulos,A.
,Beznosov,K.
,Herley,C.
:Doesmypasswordgouptoeleven:theimpactofpasswordmetersonpasswordselection.
In:Proc.
CHI2013.
pp.
2379–2388.
ACM(2013)17.
Florencio,D.
,Herley,C.
:Alarge-scalestudyofwebpasswordhabits.
In:Proc.
WWW2007.
pp.
657–666.
ACM(2007)18.
Florencio,D.
,Herley,C.
:WheredosecuritypoliciescomefromIn:Proc.
ACMSOUPS2010.
pp.
1–14.
ACM,Redmond,Washington,USA(July14-162010)19.
Furnell,S.
:Anassessmentofwebsitepasswordpractices.
Computers&Security26(7),445–451(2007)20.
Furnell,S.
:Assessingpasswordguidanceandenforcementonleadingwebsites.
ComputerFraud&Security2011(12),10–18(2011)21.
Furnell,S.
,B¨ar,N.
:Essentiallessonsstillnotlearnedexaminingthepasswordpracticesofend-usersandserviceproviders.
In:Proc.
HAS2013,LNCS,vol.
8030,pp.
217–225.
Springer(2013)22.
Goldman,J.
:ChineseHackersPublish20MillionHotelReservations(Dec2013),http://www.
esecurityplanet.
com/hackers/chinese-hackers-publish-20-million-hotel-reservations.
html23.
Goodin,D.
:Anatomyofahack:Howcrackersransackpasswordslike"qead-zcwrsfxv1331"(May,2013),http://arstechnica.
com/security/2013/05/how-crackers-make-minced-meat-out-of-your-passwords/2/20DingWangandPingWang24.
Haikun,C.
:Multiplythetotalto3.
647milliononchinesewebsites(Feb,2015),http://www.
changhaikun.
com/index.
php/2015/04/03/multiply-the-total-to-3-647-million-on-chinese-web-sites/25.
Herley,C.
,VanOorschot,P.
:Aresearchagendaacknowledgingthepersistenceofpasswords.
IEEESecurity&Privacy10(1),28–36(2012)26.
Huang,X.
,Xiang,Y.
,Chonka,A.
,Zhou,J.
,Deng,R.
H.
:Agenericframeworkforthree-factorauthentication:Preservingsecurityandprivacyindistributedsystems.
IEEETrans.
ParallelDistrib.
Syst.
22(8),1390–1397(2011)27.
Huang,Z.
,Ayday,E.
,Fellay,J.
,Hubaux,J.
P.
,Juels,A.
:Genoguard:Protectinggenomicdataagainstbrute-forceattacks.
In:Proc.
IEEES&P2015.
pp.
1–1628.
Ihm,S.
,Pai,V.
S.
:Towardsunderstandingmodernwebtrac.
In:Proc.
ACMSIGCOMM2011.
pp.
295–312.
ACM(2011)29.
Jakobsson,M.
,Akavipat,R.
:Rethinkingpasswordstoadapttoconstrainedkeyboards.
Proc.
IEEEMoST2012pp.
1–11(2012)30.
Jakobsson,M.
,Dhiman,M.
:Thebenetsofunderstandingpasswords.
In:Proc.
HotSec2012.
pp.
1–6.
USENIXAssociation(2012)31.
Jiang,Q.
,Tan,C.
H.
,Phang,C.
W.
,Sutanto,J.
,Wei,K.
K.
:Understandingchineseonlineusersandtheirvisitstowebsites:Applicationofzipf'slaw.
InternationalJournalofInformationManagement33(5),752–763(2013)32.
Johns,R.
:Illogicalappleidpasswordrules(May,2012),https://discussions.
apple.
com/thread/378549433.
Keith,M.
,Shao,B.
,Steinbart,P.
:Abehavioralanalysisofpassphrasedesignandeectiveness.
J.
oftheAssoc.
forInf.
Syst.
10(2),2(2009)34.
Kelley,P.
G.
,Komanduri,S.
,Mazurek,M.
L.
,Shay,R.
,Vidas,T.
,Bauer,L.
,Christin,N.
,Cranor,L.
F.
,Lopez,J.
:Guessagain(andagainandagain):Measuringpasswordstrengthbysimulatingpassword-crackingalgorithms.
In:Proc.
IEEES&P2012.
pp.
523–537.
IEEE(2012)35.
Ma,J.
,Yang,W.
,Luo,M.
,Li,N.
:Astudyofprobabilisticpasswordmodels.
In:Proc.
IEEES&P2014.
pp.
538–552.
IEEE(2014)36.
Martin,R.
:AmidWidespreadDataBreachesinChina(Dec2011),http://www.
techinasia.
com/alipay-hack/37.
Mathew,J.
S.
:15,000twittercredentialsstolenandleaked,hackerpromisesmoresoon(Aug,2013),http://www.
itcmt.
com/2013/08/23/15000-twitter-credentials-stolen-and-leaked-hacker-promises-more-soon/38.
Mazurek,M.
L.
,Komanduri,S.
,Vidas,T.
,Cranor,L.
F.
,Kelley,P.
G.
,Shay,R.
,Ur,B.
:Measuringpasswordguessabilityforanentireuniversity.
In:Proc.
CCS2013.
pp.
173–186.
ACM(Nov4–82013)39.
McDowell,M.
,Hernan,S.
,Rafail,J.
:SecurityTip(ST04-002):ChoosingandProtectingPasswords(2013),https://www.
us-cert.
gov/ncas/tips/ST04-00240.
Millward,S.
:Xiaominowhas100millionusersofitsandroid-basedmobileos(Feb,2015),https://www.
techinasia.
com/xiaomi-miui-100-million-users/41.
Mirante,D.
,Cappos,J.
:Understandingpassworddatabasecompromises.
Tech.
rep.
,PolytechnicInstituteofNYU,McLean,VA(2013),doi:https://isis.
poly.
edu/~jcappos/papers/tr-cse-2013-02.
pdf42.
Morris,R.
,Thompson,K.
:Passwordsecurity:Acasehistory.
Comm.
oftheACM22(11),594–597(1979)43.
Rhodan,M.
:Nearly5milliongooglepasswordsleakedonrussiansite(Sep,2014),http://time.
com/3318853/google-user-logins-bitcoin/44.
Scarfone,K.
,Souppaya,M.
:NISTSP800-118:Guidetoenterprisepasswordmanagement.
Tech.
rep.
,NIST,Reston,VA(Aug2013)TheEmperor'sNewPasswordCreationPolicies2145.
Schechter,S.
,Brush,A.
B.
,Egelman,S.
:It'snosecret.
measuringthesecurityandreliabilityofauthenticationviasecretquestions.
In:Proc.
IEEES&P2009.
pp.
375–390.
IEEE,Oakland,California(May16-192009)46.
Schechter,S.
,Herley,C.
,Mitzenmacher,M.
:Popularityiseverything:Anewapproachtoprotectingpasswordsfromstatistical-guessingattacks.
In:Proc.
HotSec2010.
pp.
1–8(2010)47.
Shay,R.
,Bauer,L.
,Christin,N.
,Cranor,L.
F.
,Forget,A.
,Komanduri,S.
,Mazurek,M.
,Melicher,W.
,Segreti,S.
M.
,Ur,B.
:Aspoonfulofsugartheimpactofguidanceandfeedbackonpassword-creationbehavior.
In:Proc.
CHI2015.
pp.
2903–291248.
Shay,R.
,Komanduri,S.
,Durity,A.
L.
,Huh,P.
S.
,Mazurek,M.
L.
,Segreti,S.
M.
,Ur,B.
,Bauer,L.
,Christin,N.
,Cranor,L.
F.
:CanlongpasswordsbesecureandusableIn:Proc.
ACMCHI2014.
pp.
2927–2936.
ACM(2014)49.
Top500chinesepinyinnames(Jan,2015),http://www.
data.
ac.
cn/zrzy/g22.
asp50.
Ur,B.
,Kelley,P.
G.
,Komanduri,S.
,etal.
:Howdoesyourpasswordmeasureuptheeectofstrengthmetersonpasswordcreation.
In:Proc.
USENIXSecurity2012.
pp.
65–80.
Bellevue,WA,USA,(August8-102012)51.
Wang,D.
,Cheng,H.
,Wang,P.
:UnderstandingPasswordsofChineseUsers:Characteristics,SecurityandImplications(Jan2015),http://t.
cn/RzSlpDz52.
Wang,D.
,He,D.
,Wang,P.
,Chu,C.
H.
:Anonymoustwo-factorauthenticationindistributedsystems:Certaingoalsarebeyondattainment.
IEEETrans.
Depend.
Secur.
Comput.
(2014),http://dx.
doi.
org/10.
1109/TDSC.
2014.
235585053.
Weir,M.
,Aggarwal,S.
,Collins,M.
,Stern,H.
:Testingmetricsforpasswordcreationpoliciesbyattackinglargesetsofrevealedpasswords.
In:Proc.
CCS2010.
pp.
162–175.
ACM(October4-82010)54.
Yan,J.
,Blackwell,A.
F.
,Anderson,R.
J.
,Grant,A.
:Passwordmemorabilityandsecurity:Empiricalresults.
IEEESecurity&privacy2(5),25–31(2004)55.
Yap,J.
:450,000userpasswordsleakedinYahoobreach(July2012),http://www.
zdnet.
com/article/450000-user-passwords-leaked-in-yahoo-breach/56.
Zhu,B.
,Yan,J.
,Bao,G.
,Mao,M.
,Xu,N.
:Captchaasgraphicalpasswords–anewsecurityprimitivebasedonhardAIproblems.
IEEETrans.
Inform.
ForensicsSecurity9(6),891–904(2014)