Vous êtes sur la page 1sur 8

Identityandreputationsystemsonethereum

Abstract

Inthecontextofethereum,areputationisthatsomethingbywhichasmartcontractora
DAOrecognizesauser.Anidentityhereisdefinedasacollectionofreputations,who
areassociatedwitheachotherjustbyvirtueofbeingearnedbythatidentityor(we
assume)user.Weexplorethelogisticsofbuildingidentityandreputationsystemson
ethereum,withparticularemphasisplacedonavoidingandmitigatingSybilattackswith
minimalimpactonthedynamicsofreputationsystems.

Introduction

Reputationsystemsareusefulforanysmartcontract,DO,orDAOwhoneedstokeep
trackoftheirusersonanindividualbasis,likelysothattheycanrewardthemfortheir
contributions.Everysystemthatcannotusecryptographicproofofcontributionmustuse
socialproof.Socialproofinherentlyrequiresasecureidentityandreputationsystemin
ordertofunctionwithoutbeingexploited.Tobeclear,ifnoreputationisrequired,an
attackercanintroducemanyidentities(perhapsatsomeexpensetoherself)inorderto
getwhateverbenefitismadeavailabletocontributorswithouthavingactuallymadea
contribution.Asisconvention,wewillrefertothiskindofanattackasaSybilattack,and
wewillrefertotheattackerasSybil,andheridentitiesasherfacesorpersonalities(after
thebookmovieinspiredbyawomanwithmultiplepersonalitydisorder).Moreover,we
willrefertoauserwhoisnotSybilasCyril.Wewillbeginbydevelopingaschemeforthe
securityofidentityandreputationasidentifiers,andthenwillturnourattentiontothe
securityofreputationsystems.

Identityandreputation

Werealizethatinasfarasreputationisvaluableanditispossible,reputationwillbe
boughtandsold.However,theintuitionaboutreputationsisthattheyshouldnotbe
transferable,foraftertheyhavebeentransferredtheywillnotbeasreliableanindicator
oftheirownersintegrityorability.Wecouldattachreputationtopublickeysinorderto
maketheirtrademoredifficult,butthenacompromiseofthecorrespondingprivatekey
wouldcompromisetheusersreputation.Ifwemakereputationatransferrabletoken,
thisproblemwouldbesomewhatmitigatedinthecasewheretheprivatekeymighthave
beencompromised,butitalsofacilitatesthetransactionandtheftofreputation.Wecan
dobetterbyinsteadmakingidentityatransferrabletoken,andhavingreputationsbe
permanentlyattachedtoparticularidentities.Theproblemofreputationsaleismitigated
withthisapproach,iftherearemorethanonereputationsattachedtoidentities.

However,theissueofthecompromiseofasigningkeyleadingtoidentitytheftandloss
ofcontrolisstillaconcern.Wethereforewillseparatethesigningkeyfortheactivity
associatedwithanidentitysreputations(theusagekey),andthekeysrequiredto
revokeandreplacetheusagekey(thesecuritykeys),incaseofapossiblecompromise.
HerewewilluseanNofMmultisigschemeoftheuserschoosing,sothatuserscan
choosetheirpreferedtradeoffbetweenidentitysecurityandconvenienceofusagekey
revocationandreplacement.Wegototheselengthsbecausewepredictthatwithout
them,identitytheftmaybecomeasbigofanissueasbitcointheft.Notealsothatthis
schememakesthesaleofreputationandidentityevenmoredifficult,althoughitisstill
possible.Wetentativelyregardthisasasatisfactorystateofaffairs,sowewillshiftour
focusfromidentityandreputationasidentifierstothespecificationofreputation
systems.

Asocialreputationsystems

Whilewewillmainlyfocusonreputationsystemswhosepurposeistoproducesocial
proofs,itisworthnotingthatasocialreputationsystemsarepossible,forexamplewhen
cryptographicproofcanbeusedtomeasurecontribution.Thesesystems,too,canbe
subjecttoSybilattacksundersomecircumstances,aswewillsoonsee.Thisanalysis
willbetrivial,butitprovidesusefulintuitionsthatwillcarrytotheanalysisofsocial
reputationsystems.Aswehavementionedbefore,solongasreputationisvaluable,it
willbeboughtandsold.Therefore,wewillusereputationprice(orrepprice,forshort)to
refertothepriceatwhichtheparticularreputationinquestionwouldbetraded,perhaps
inferredbythebenefitsreservedfortheidentitywhichholdsthatreputation.Additionally
weintroducethenotionofreputationcost(orrepcost),whichisthecosttoaparticular
userofobtainingaparticularreputation.Notethatrepcostmaybeuserdependent,
whilereppriceisnot.Finally,wemightconsiderthatingeneralanyreputationsystem
mightrequireamembershipfeeofusersbeforegrantingthemareputation,andthe
correspondingaccessthatthatentails.

Inthefiguresthatfollow,weplotpossiblefunctionalrelationshipsbetweenrep
priceandrepcost.Theredcurve,commontoallplots,reflectsthecostofmembership
incurredbyauseruponenteringthereputationsystem.Theslopeofthegreencurve,
alsocommontoallplots,reflectsthebaselinereppricetorepcostratio.Finally,theblue
curvereflectstherelationshipbetweenreppriceandrepcost,asimplementedinthe
particularreputationsystem.Thedifferenceinslopesandtheintersectionpointbetween
thegreenandbluecurvesareofparticularimportanceintheanalysis.

Ifforanygivenusertherelationshipbetweenreppriceandrepcostisconcave
down(asinFigure1),thenthereisarepcostwhichmaximizesthereppricetorepcost
ratio.Thisincentivizesuserstospreadamultipleoftheoptimalcostevenlyacross
multipleaccounts,therebymakingthereputationsystemvulnerabletoSybilattack.If,on

theotherhand,therelationshipbetweenreppriceandrepcostisconcaveup(asin
Figure2),thenusershaveanincentivetohaveonlyonereputation,andSybilattacks
areprevented.Thedrawbackinthiscaseisthatuserswhoareabletoinvestmorerep
costaregoingtodevalueorfarovershadowthereppriceofuserswhoarenotableto
investasmuchcost,sincetheratioofreppricetorepcostisunbounded.However,
thereisathirdpossibility(Figure3):lettherelationshipbe(atleastasymptotically)
linear.Thisisappealingbecauseitsimultaneouslyboundsthereppricetorepcost
rationandpreventsSybilattacksbyincentivizinguserstohaveonlyonereputation.

WeobservethatBitcoinisanasocialreputationoftheconcaveupvariety,withhashing
powerasreputation.Herethereppricecanberegardedasapiethatissplitbetween
theusersinproportiontotheirreputation,becausereppriceisdeterminedbytheshare
ofnewlymintedbitcointhatitrepresents.Theeffectisthatuserswhoareabletoinvest
moreintomininggetmuchhigherrepprice,andtheytherebydilutethereppriceof
otherusers.IfBitcoinsproofofworkfunctionwasrestrictedtoCPUs,therelationship
betweenreppriceandrepcostwouldbemuchclosertolinear.

Cyriltokens

Ideally,mitigatingSybilattacksonsocialreputationsystemsdiscouragesasmuchas
possibleSybilsmalicioususeofthesystemwithoutdiscouragingcontributors.Thefirst
classofapproachesthatwewilldiscussaimstomakeitmoreexpensiveforSybiltouse
thereputationsystem.Therearetwogoodwaysweknowofaccomplishingthis.One
simplymakestheuserpayamembershipfeeinordertomakebeingSybilmore
expensive,eitherthroughproofofburnorapaymentdirectlytothenetwork.Theotheris
todesignchallengesortasksthatareresistanttoSybilattacksbecausetheyarerelatively
easyforhumanstodoasmallnumberoftimes,butarehardtoautomate.Thefirstsetof
solutionsisstraightforwardandisguaranteedtobeeffectivetoapointbuttheobvious
drawbackisthatallusershavetopay.Solutionsofthesecondtype,ontheotherhand,are
challengingtoimplementsustainablyparticularlybecauseofadvancesinmachinelearning.
Iftheycanbedoneinaconvenientway,however,itwillsaveuserscoinandhelpkeep
Sybiloffofthenetwork.Wewillfocusonsolutionsofthesecondtype,bycreatingasystem
oftokens,calledCyriltokens,thatareissueduponcompletionofsuchtasks.

AtaskbasedCyriltokenissaidtobebrokenifitiseasyforSybiltoearnthesetokens.A
brokenCyriltokenisthereforecheap,andthiswillbereflectedinitspriceifitistradedona
market.Ontheotherhand,agoodCyriltokenisexpensive,providingevidencethatSybil
hasnotyetsucceededinautomatingitsproduction.Thereisastrongincentiveforpeople
toautomatethecreationofthesetokensinasfarastheyareexpensiveorrequiredfor
accesstoreputationsystems.Iftheysucceedinautomatingthecompletionofthese
challenges,thepriceoftheCyriltokenswillfall,revealingthatthechallengehasbeen
broken.

IfareputationsystemrequiresCyriltokensforuse,insteadofdestroyingthem,itmaysell
themontothemarket,therebyinterferingwiththepricethatismeanttosignalwhetheror
nottheCyriltokenisbroken.Therefore,Cyriltokensshouldbeawareofwhethertheyare
beingtransferredonto/fromtheexchangeornotiftheyaretransferredoutsideofthe
exchange,theyshouldautomaticallybecomevoid.Whileenforcingsucharulehasits
difficulties,ithastheadditionalbenefitofmakingthepricesignalstronger,bynearly
eliminatingthepossibilityofadarkmarketforCyriltoken.

Socialreputationsystems

Themainchallengewhenbuildingsocialreputationsystems,aswehavealludedto,is
avoidingtheeffectiveexploitationandsubsequentcooptingofthesystembySybiland
hermanypersonalities.Wehavealreadymentionedthatareputationsystemmay
requiretheexpenditureofCryiltokensforitsuse,howeverthisisntsufficientmitigation
againstSybilattacksforareputationsystemwhosedesignisinherentlyvulnerable.For
example,anaiveimplementationmightletanyuserincreaseanyotherusersreputation
inproportiontotheirownreputation.WhilerequiringCyriltokensincreasesthecostto
Sybilofusingthesystem,thiscostincreaseslinearlyinthenumberofnodes,whileSybil
mayincreaseherreputationmuchfaster,inthenumberofreputablenodesintroduced.
Inthiscase,wehaveaninstanceofFigure2,whereanincreaseincosttoSybilleadsto
anacceleratingincreaseinherrepprice.Wethereforemustfindmethodsbywhich
socialreputationscanensurethatSybilsreppriceincreaseslinearlyinherrepcost.
Preferably,thismitigationwillnotcrippletheintendeddynamicsofthereputationsystem.

Wehavediscoveredonesuchsolution,whichhappenstobebothcompellingand
elegant.Itrequiresthatthereputationsystemgiveusersareputationallowance,an
exhaustiblebudgetforgivingotherusersreputation.Reputationallowancemustbe
earned,anditmustcostSybilmorethantheincreaseinreppricethatshemightreceive
fromincreasingherownreputation.Specifically,areputationsystemshouldrequirethe
expenditureofCyriltokensduringtheactivitythatearnsrepallowanceinaneffortto
makeitpossibleforthereputationsystemtobeprofitableforCyril,butunprofitablefor
Sybil.Howpreciselyasocialreputationsystemaccomplishesthiscoordinationwillhave
tobeimplementationspecific,andthereforewillnotreceivetreatmenthere.


Allothersolutionsthatwevecomeupwithhaveeitherbeencomputationallydifficultor
havefailedtomaketheincreaseinreppriceforSybil,gainedbyaddingreputable
nodes,linearintherepcostinvested.Wewilllistsomeofthesesolutionshere,for
reference.Keepingtrackofpairsofuserstoinsurethattheydonotincreasetheirrep
toomuchisanexpensivemitigation,anditonlyreducesthebenefittoSybiltobe
quadraticinthenumberofnodes,inthebestcase.Havingethereumcreateinstances
whereonlychosenuserscanincreaseeachothersreputationincreasesthecostor
waitingtimetoSybil,butdoesnotchangethefundamentaldynamicsofthereputation
systemitmerelymakesthesameinstancesofexploitationoccurprobabilistically.Using
machinelearningtoidentifyabusivebehavioursisexpensivetothepointwhereitis
unfeasibletobuildintotheethereumcontractsofthereputationsystems,althoughit
providesfuzzypromise.Finally,itisalwayspossibletoemployaweboftrustmodel,in
thiscaserequiringacentralizedinitializationprocesstobootstrapthenetwork.Whilea
weboftrustmaybepromisingforsmallercommunities,itiscumbersomeandinhibitive
onalargerscale.

Discussion

MembershipfeesofferasimplemitigationofSybilattacks.Additionally,tofurther
preventSybilattacksandcentralizationofrepprice,itisimportantforreputationsystems
tohaveanasymptoticallylinearrelationshipbetweenreppriceandrepcost.Thisistrue
ofbothasocialandsocialreputationsystems.Insocialreputationsystems,requiring
Cyriltokensfornetworkactivitythatmightearnauserreputationallowancecan
successfullymitigateSybilattacks,iftheamountofCyriltokensrequiresiscalibrated
correctly.However,thiscalibrationwillbecomemoredifficultovertimeasitbecomes
harderforanygivenusertoprovethattheyarenotautomated.Therefore,thesooner
weestablishaworkingidentityandreputationsystem,themorelikelywearetofind
successwithoutexcessivedifficulty.Oncewecancheaplyautomatethecompletionofall
Turingtests,however,securereputationsystemswillbecomeunprofitableforCyril.

Conclusion

Wehavedevelopedaschemethatmakesitpossibletocreateworkingidentityand
reputationsystems,whichcannotbeexploitedbyautomatedattackers,solongas
workingtestsforautomationcanbedesigned.Thisdoesnotthenmeanthattheycannot
beattackedbydedicatedhumanattackers.Theonlymitigationagainstthesemore
sophisticatedattacksistorequirethatareputationsystemsmembershavesomany
reputationsattachedtotheiridentitiesthattheycouldnotpossiblyhavemanyidentities
inthesystem.Thiswillwork,butisonlypossibleinamatureidentityandreputation
system.Alternatively,usersofareputationsystemcouldberequiredtoconstantly

broadcasteverythingtheyaredoing,sothattheycanconvincethenetworkthattheyare
nottakingonmultipleidentities,butthisisnotanattractivesolutioninasfaraswecan
solvethesameproblembymoreefficientmeans.

Vous aimerez peut-être aussi