On the (Im-)Possibility of Extending Coin Toss
⋆
DennisHofheinz
1
,JörnMüller-Quade
2
,andDominiqueUnruh
2
1
CWI,CryptologyandInformationSeurityGroup,Prof.Dr.R.Cramer, Dennis.Hofheinzwi.nl
2
IAKS,ArbeitsgruppeSystemsiherheit,Prof.Dr.Th.Beth,UniversitätKarlsruhe,
{muellerq,unruh}ira.uka.de
Abstrat. We onsider theryptographi two-party protooltask ofextendinga given
oin toss. Thegoal isto generate
n
ommonrandomoinsfrom asingleuse ofanidealfuntionality whihgives
m < n
ommonrandom oinstotheparties. IntheframeworkofUniversalComposabilityweshowtheimpossibilityofseurelyextendingaointossfor
statistialandperfetseurity.Ontheotherhand,foromputationalseuritytheexistene
ofaprotoolforointossextensiondependsonthenumber
m
ofrandomoinswhihanbeobtainedforfree.
For the ase of stand-alone seurity, i.e., a simulation based seurity denition without
an environment,we present anovelprotoolfor unonditionally seure oin toss
exten-sion. Thenew protool worksfor superlogarithmi
m
,whih is optimalas we show theimpossibilityofstatistially seureointossextensionforsmaller
m
.Combiningourresultswithalreadyknownresults,weobtaina(nearly)omplete
hara-terization underwhihirumstanesointossextensionispossible.
Keywords: ointoss,universalomposability,reativesimulatability,ryptographi
pro-tools.
Table of Contents
1 Introdution... 2
2 Seuritydenitions... 4
3 TheComputationalCase ... 6
3.1 UniversalComposability ... 6
4 StatistialandPerfetCases ... 9
4.1 Stand-alonesimulatability... 9
4.2 UniversalComposability ... 12
A DetailedProofs... 17
A.1 Anauxiliarylemma... 17
A.2 ProofofLemma3... 17
A.3 ProofofLemma4... 18
A.4 ProofofTheorem6... 19
A.5 ProofofTheorem7... 21
A.6 ProofofLemma12... 23
⋆
ManuelBlum showedin[Blu81℄howtoipaoinoverthetelephoneline.His protool
guaran-teed that even ifone partydoesnotfollowthe protool,the otherpartystill gets auniformly
distributed ointoss result. This general onept of generating ommon randomness in a way
suh that nodishonest partyanditatetheresult provedveryusefulin ryptography,e.g.,in
theonstrutionofprotoolsforgeneralseuremulti-partyomputation.
Here weare interestedin thetaskofextending agivenointoss.That is,supposethat two
partiesalreadyhavethepossibilityofmakingasingle
m
-bitoin-toss.Is itpossibleforthemtoget
n > m
bitsofommonrandomness?Theanswerweomeupwithisbasially:itdepends.Therstthingtheextensibilityofagivenointossdependsonistherequiredseuritytype.
One type of seurity requirement (whih we allstand-alone simulatability here) ansimply
be that the protool imitates an idealoin toss funtionality in the sense of [Gol04℄, where a
simulatorhastoinventarealistiprotoolrunafterlearningtheoutomeoftheidealoin-toss.
A stronger type of requirement is to demand universal omposability, whih basially means
that the protool imitates an ideal oin toss funtionality even in arbitrary protool
environ-ments.Seurityin the lattersense anonvenientlybeaptured in asimulatabilityframework
like the Universal Composability framework [Can01a, Can05℄ or the Reative Simulatability
model[PW01,BPW04℄.
Orthogonal to this, onean vary the level of fullment of eah of these requirements. For
example,oneandemandstand-alonesimulatabilityoftheprotoolwithrespetto
polynomial-timeadversariesinthesensethatrealprotoolandidealfuntionalityareonlyomputationally
indistinguishable.ThisspeirequirementisalreadyfullledbytheprotoolofBlum.
Alterna-tively,oneandemand,e.g.,universalomposabilityoftheprotoolwithrespettounbounded
adversaries.Thiswouldthenyieldstatistialorevenperfetseurity.Weshowthatwhethersuh
aprotoolexistsdependsontheasymptotibehaviourof
m
.Ourresultsaresummarizedin thetable below.A yes orno indiates whetheraprotool
forointossextension exists inthat setting. Depends meansthat theanswerdependsonthe
size of the seed (the
m
-bit oin toss available by assumption), and boldfae indiates novelresults.
Seuritytype
↓
/level→
Computational Statistial Perfetstand-alonesimulatability yes depends 3
no
universalomposability depends 4
no no
Knownresultsintheperfetandstatistial ase. A folkloretheoremstates,that(perfetly
non-trivial) statistially seure oin-toss is impossible from srath (even in very lenient seurity
models). ByKitaev,this resultwasextended evento protoolsusing quantum ommuniation
(f.[ABDR04℄).[BGR96℄rstinvestigatedtheproblemofextendingaoin-toss.Theypresented
astatistially seure protool for extending agivenoin-toss (pre-shared using a VSS), ifless
than
1
6
ofthepartiesareorrupted.Notethat theirmainattentionwasontheeieny ofthe protool,sine in that senario arbitrarymulti-partyomputations and therefore in partiularoin-tossfromsrathareknowntobepossible.Theresultdoesnotapplytothetwo-partyase.
3
Cointossextensionispossibleifandonlyiftheseedhassuperlogarithmilength.
4
Coin tossextensionis impossibleif theseed doesnothavesuperlogarithmi length. Thepossibility
explained.Fortheperfetase,weshowimpossibilityofany ointossextension,nomatterhow
(in-)eient. Weshow this for stand-alone simulatability(Coro. 8) and for universal
ompos-ability(Coro.15).Nowforthestatistialase.Whendemandingonlystand-alonesimulatability,
thesituationdependsonthenumberofthealreadyavailableommonoins.Namely,wegivean
eientprotooltoextend
m
ommonoinstoanypolynomialnumber(intheseurityparam-eter),if
m
issuperlogarithmi(Th.11).Otherwise,weshowthat thereanevenbenoprotoolthatderives
m
+ 1
ommonrandomoins(Coro.8).Intheuniversalomposabilitysetting,thesituation ismorelear: weshowthat theresimply isnoprotoolthat derivesfrom
m
ommonoins
m
+ 1
oins,no matter howlargem
is (Th.14). (However,herewerestritto protoolsthatrun inapolynomialnumberofrounds.)
Known results in the omputational ase. In[Blu81℄ Blum gaveaomputationally seure
pro-tool.In [Gol04, Proposition 7.4.8℄, this protool is shown to be stand-alone simulatable, and
togetherwiththesequentialompositiontheorem[Gol04,Proposition7.4.3℄forstand-alone
sim-ulatable protools, this gives a omputationally stand-alone simulatable protool for tossing
polynomiallymanyoins.
Thismakesoin-tossextensiontrivialinthatsetting,onejustignoresthe
m
-bitoin-tossandtosses
n
-bitfromsrath.Intheomputationaluniversalomposabilitysetting,ithasbeenshownin[CF01℄that
oin-tossannotbeahievedfromsrath. However,theyshowedthatgivenasuientlylarge
om-monreferenestring(CRS),bitommitmentispossible.Fromthisitiseasyto seethat suha
CRS(andthereforealsoasuientlylargeoin-toss)anbeextendedtoanypolynomiallength.
However,itwasunlearwhat theminimumsize requiredfromtheCRSortheoin-tossis.
NotethatthereisasubtledierenebetweenthenotionofaCRSandaoin-toss.ACRSis
randomnessthat isavailabletoallpartiesatthebeginningoftheprotool,whilewithoin-toss
the randomness is only generated when all parties agree to run the oin-toss. This makesthe
oin-toss atually the stronger primitive, sine in somesituations it is neessary to guarantee
that not even orrupted parties learn the outomes of the oin-toss prior to a given protool
step.
In [Cle86℄, the task of oin-toss is onsidered in a senario slightly dierent from ours:
in[Cle86℄,protoolpartiipantsmaynotabortprotoolexeutionwithoutgeneratingoutput.In
that setting, [Cle86℄ showthat oin-tossis generallynotpossibleevenagainstomputationally
limitedadversaries.However,tothebestofourknowledge,anextensionofagivenointosshas
notbeenonsideredsofarin theomputationalsetting.
Our results in the omputational ase. We answer the question onerning the minimal size
neessary for a oin-toss to be extensible: If an
m
-bit oin-toss funtionality is given, andm
is not superlogarithmi,then it is already impossiblefor the parties to derive
m
+ 1
ommonrandomoins(inauniversallyomposableway)fromit(Th.6).However,wealsoshowthatunder
strengthenedomputationalassumptions,there areprotoolsthat extend
m
toanypolynomialnumber(intheseurityparameter)of ommonrandomoins,if
m
issuperlogarithmi(Th. 5).Inthatsense,wegivetheremainingparts foraompleteharaterizationoftheomputational
ase.
Notation
Afuntion
f
isnegligible,ifforanyc >
0
,f
(
k
)
≤
k
−
c
forsuientlylarge
k
(i.e.,f
∈
k
−
ω
(1)
).
f
isoverwhelming,if1
−
f
isnegligible(i.e.,f
∈
1
−
k
−
ω
(1)
).
f
isnotieable,ifforsomec >
0
,f
(
k
)
≥
k
−
c
forsuientlylarge
k
(i.e.,f
∈
k
−
O
(1)
).Note
that funtionsexists,whihareneithernegligiblenornotieable.
f
ispolynomiallybounded,ifforsomec >
0
,f
(
k
)
≤
k
c
forsuientlylarge
k
(i.e.,f
∈
k
O
(1)
f
ispolynomially-large,ifthereisac >
0
s.t.f
(
k
)
c
≥
k
forsuientlylargek
(i.e.,f
∈
k
Ω
(1)
).
f
issuperpolynomial,ifforanyc >
0
,f
(
k
)
> k
c
forsuientlylarge
k
(i.e.,f
∈
k
ω
(1)
).
f
issuperlogarithmi,iff /
log
k
→ ∞
(i.e.,f
∈
ω
(log
k
)
).Itiseasytoseethatf
issuperlog-arithmiifandonlyif
2
−
f
isnegligible.
f
is superpolylogarithmi, if for anyc >
0
,f
(
k
)
>
(log
k
)
c
for suiently large
k
(i.e.,f
∈
(log
k
)
ω
(1)
).
f
isexponentially-small,ifthere exists ac >
1
,s.t.f
(
k
)
≤
c
−
k
forsuientlylarge
k
(i.e.,f
∈
Ω
(1)
−
k
= 2
−
Ω
(
k
)
).
f
issubexponential,ifforanyc >
1
,f
(
k
)
< c
k
forsuientlylarge
k
(i.e.,f
∈
o
(1)
k
= 2
o
(
k
)
).
2 Seurity denitions
In this setion we roughly sketh the seurity denitions used throughout this paper. We
dis-tinguish between twonotions: stand-alone simulatability asdened in [Gol04℄, 5
and Universal
Composability(UC) asdened in[Can01a℄.
Stand-alonesimulatability.In[Gol04℄adenitionfortheseurityoftwo-partyseurefuntion
evaluationsisgiven(alledseurityinthemaliiousmodel).Wewillgiveasketh,formoredetails
wereferto [Gol04℄.
Aprotoolonsistsoftwopartiesthatalternatinglysendmessagestoeahother.Theparties
mayalso invokean idealfuntionality,whih is givenasan orale(in ourases,they invokea
smalleroin-tosstorealisealargerone).
We saytheprotool
π
stand-alonesimulatablyrealisesaprobabilistifuntionf
,iffor anyeient adversary
A
that mayreplae none orasingleparty, there is an eient simulatorS
s.t.forallinputsthefollowingrandomvariablesareomputationallyindistinguishable:
Therealprotoolexeution.Thisonsistsoftheviewoftheorruptedpartiesuponinputs
x
1
andx
2
forthepartiesandtheauxiliaryinputz
fortheadversary,togetherwiththeoutputsI
oftheparties.Theidealprotool exeution.Herethesimulatorrstlearntheauxiliaryinput
z
andpossiblytheinputfortheorruptedparty(the simulatormustorrupt thesamepartyasthe
adver-sary).Then heanhoosetheinputof theorrupted partyfortheprobabilistifuntion
f
,theotherinputsarehosenhonestly(i.e.,therstinputis
x
1
iftherstpartyisunorrupted, andtheseondinputx
2
iftheseondpartyis).Thenthesimulatorlearnstheoutput
I
off
(weassumetheoutputtobeequalforallparties).It may now generate afake view
v
of the orrupted parties. The ideal protool exeutionthenonsistsof
v
andI
.Of ourse, in our ase the probabilisti funtion
f
(the oin-toss) has no input, so the abovedenitiongetssimpler.
What wehaveskethedaboveis whatweallomputational stand-alonesimulatability.We
further dene statistial stand-alone simulatability and perfet stand-alone simulatability. In
these ases we donot onsider eient adversaries and simulators, but unlimited ones. Inthe
aseofstatistial stand-alonesimulatabilitywerequiretherealand idealprotool exeutionto
be statistially indistinguishable (and not only omputationally ), and in the perfet ase we
evenrequirethesedistributions tobeidential.
UniversalComposability. Inontrasttostand-alonesimulatability,UniversalComposability
[Can01a℄isamuhstriterseuritynotion.Themaindiereneistheexisteneofanenvironment,
thatmayinteratwithprotool andadversary(orwithidealfuntionalityandsimulator)
5
Infat,[Gol04℄doesnotusethenamestand-alonesimulatabilitybutsimplyspeaksaboutseurityin
themaliousmodel.Weadoptthenamestand-alonesimulatabilityforthispapertobeabletobetter
advantageof aversatileomposition theorem(theUniversalCompositionTheorem[Can01a℄).
Weonlyskeththemodelhereandreferto [Can01a℄fordetails.
Aprotoolonsistsofseveralmahinesthatmay(a)getinputfromtheenvironment,(b)give
outputtotheenvironment(bothalsoduringtheexeutionoftheprotool),and()sendmessages
toeahother.
Thereal protool exeutiononsistsof aprotool
π
,anadversaryA
and anenvironmentZ
.Heretheenvironmentmayfreelyommuniatewiththeadversary,andthelatterhasfullontrol
overthe network,i.e., it may deliver,delayordropmessagessentbetweenparties. Weassume
theauthentiated modelin this paper,sotheadversarylearnstheontentof themessagesbut
maynotmodifyit.When
Z
terminates,itgivesasinglebitofoutput.Theadversarymayhoosetoorruptparties atanypointintime. 6
The ideal protool exeution is dened analogously,but insteadof aprotool
π
there is anidealfuntionality
F
andinsteadoftheadversarythereisasimulatorS
.Thesimulatoranonlylearnandinueneprotooldata,if(a)thefuntionalityexpliitlyallowsthis,or(b)itorrupts
aparty(note that the simulator may onlyorrupt the sameparties as the adversary).In the
latterase,thesimulatoranhooseinputsintothefuntionalityinthenameofthatpartyand
getstheoutputsappartainingtothatparty.Intheaseofunorruptedparties,theenvironment
isinontrol oftheorrespondingin- andoutputoftheidealfuntionality.
We say aprotool
π
universally omposably (UC)-implements an idealfuntionalityF
(orshort
π
isuniversallyomposableifF
islearfromtheontext),ifforanyeientadversaryA
,there is aneientsimulator
S
, s.t. for alleientenvironmentsZ
and all auxiliaryinputsz
for
Z
,the distributions ofthe output-bitofZ
in thereal and theidealprotool exeutionareindistinguishable.
What hasbeenskethed aboveweallomputational UC. Wefurther denestatistial and
perfet UC. In these notions,we allow adversary, simulator and environment to be unlimited
mahines.Further,in theaseofperfetUC,werequirethedistributionsoftheoutput-bitof
Z
tobeidential inrealandidealprotool exeution.
The Ideal Funtionality for Coin Toss. Todesribethetaskofimplementingauniversally
omposableoin-toss,wehaveto denetheidealfuntionalityof
n
-bitoin-toss.Inthefollowing,let
n
denote apositiveinteger-valuedfuntion.Below is an informal desription of our ideal funtionality for a
n
-bit oin toss. First, thefuntionalitywaitsforinitializationinputsfrombothparties
P
1
andP
2
.Assoonasbothparties have this way signalled their willingness to start, the funtionality seletsn
oins in form ofan
n
-bit stringκ
uniformly andsends thisκ
to the adversary.(Note that aointoss doesnotguaranteeserey ofanykind.)
Ifthefuntionalitynowsent
κ
diretlyandwithoutdelaytotheparties,thisbehaviourwouldnot be implementable by any protool (this would basially mean that the protool output is
immediatelyavailable,evenwithoutinteration).Sothefuntionalityletstheadversarydeide
whentodeliver
κ
toeahparty.Notehowever,thattheadversarymaynotinanywayinuenethe
κ
that isdelivered.A moredetaileddesriptionfollows:
6
Itisthenalledanadaptiveadversary.Iftheadversaryanonlyorruptpartiesbeforethestartofthe
protool,wespeakofstatiorruption.Allresultsinthispaperholdforbothvariantsoftheseurity
Idealfuntionality
CT
n
(n
-bit Coin Toss)1. Wait until there have been
init
inputs fromP
1
andP
2
. Ignore messages from the adversary,but immediatelyinformtheadversaryabouttheinit
.2. Selet
κ
∈ {
0
,
1
}
n
uniformlyandsend
κ
totheadversary.Fromnowon:on therst (and only the rst)
deliver to 1
message from the adversary, sendκ
to
P
1
,on therst (and only the rst)
deliver to 2
message from the adversary, sendκ
to
P
2
.Using
CT
n
,weanalsoformallyexpresswhatwemeanbyextendingaointoss.Namely: Denition1. Letn
=
n
(
k
)
andm
=
m
(
k
)
be positive, polynomially bounded and omputablefuntionssuhthat
m
(
k
)
< n
(
k
)
forallk
.Thenaprotoolisauniversallyomposable(
m
→
n
)
-ointossextensionprotoolifitseurelyandnon-triviallyimplements
CT
n
byhavingaessonly toCT
m
.Thisseurity anbe omputational, statistial orperfet.Byanon-trivial implementationwemeanaprotoolthat, withoverwhelmingprobability,
guaranteesoutputsifnopartyisorruptedandallmessagesaredelivered.(Alternatively,onemay
alsoonsiderprotoolsthatprovideoutputwithoverwhelmingprobability.)Thisrequirementis
usefulsinewithoutit,atrivialprotoolthatdoesnotgenerateanyoutputformallyimplements
everyfuntionality.(Cf.[CLOS02℄and[BHMQU05,Setion5.1℄formoredisussionandformal
denitionsofnon-triviality.)
Onunlimitedsimulators.Following[BPW04℄,wehavemodelledstatistialandperfet
stand-aloneandUCseurityusingunlimitedsimulators.Anotherapproahistorequirethesimulators
to bepolynomialin the running-timeof theadversary. All ourresults applyalso to that ase:
Fortheimpossibilityresults,thisisstraightforward,sinetheseuritynotiongetsstriterwhen
thesimulatorsbeomemorerestrited.Theonlypossibilityresultforstatistial/perfetseurity
isgivenin Theorem11.There, thesimulator weonstrutis in fatpolynomialin theruntime
oftheadversary.
In thefollowingsetions,weinvestigate theexisteneof suh ointoss extensionprotools,
dependingonthedesiredseurity level(i.e.,omputational/statistial /perfet seurity)and
theparameters
n
andm
.3 The Computational Case
3.1 Universal Composability
Inthe following,we need theassumption of enhanedtrapdoorpermutations with dense
pub-li desriptions (alled ETD heneforth). Roughly, these are trapdoor permutations with the
additional properties that (i) onean hoose the publi keyin an oblivious fashion, i.e., even
giventhe ointosses weused itis infeasibleto invertthefuntion,and (ii)thepubli keysare
omputationallyindistinguishablefrom randomstrings.
Denition2 (Enhaned trapdoor permutations with dense publi desriptions). A
system of enhaned trapdoor permutations with dense publi desriptions (ETD) onsists of
the following eient algorithms: a key generation algorithm
I
that (given seurity parameterk
) generates publi keyspk
and orresponding trapdoorstd
(we treatpk
andtd
as eientlyomputable funtionsto failitate notation), anda domain sampling algorithm
S
that givenpk
outputsanelementin the domainof
pk
,satisfyingthe following:Forany non-uniformprobabilistipolynomial-timealgorithm
A
thereisanegligible funtionPermutations.
Pr
(pk
,
td)
←
I
(1
k
) :
pk
isavalid publikeyand
td
=
pk
−1
≥
1
−
µ
(
k
)
,
and any valid publikeyisapermutation.Almost uniform sampling. For any valid publi key
pk
that an be output byI
(
1
k
)
, the
statistial distanebetweentheoutputof
S
(pk
)
andrandomly hosenelementsinthedomain(=range)of
pk
isboundedbyµ
(
k
)
.Enhanedhardness. For all
k
∈
NPr
(pk
,
td)
←
I
(1
k
)
, y
←
S
(pk
)
, x
′
←
A
(
1
k
,
pk
, y, r
) :
pk
(
x
′
) =
y
≤
µ
(
k
)
Here
r
denotesthe randomnessusedbyS
.Densepublidesriptions.Thereisapolynomiallybounded,eientlyomputablefuntion
s
(not dependingon
A
)s.t.Pr
(pk
,
td)
←
I
(1
k
) :
A
(
1
k
,
pk
) = 1
−
Pr
h
pk
← {
0
,
1
}
s
(
k
)
uniformly:
A
(
1
k
,
pk
) = 1
i
≤
µ
(
k
)
.
Exponentially-hardETDaredenedanalogously, exeptthat werequiretheonditionsabove
tohold for all subexponential-time
A
andan exponentially-smallµ
.Lemma3. There isaonstant
d
∈
Ns.t. the following holds:AssumethatETDexist,s.t.thesizeofthe iruitsdesribingthe ETDisboundedby
s
(
k
)
forseurity parameter
k
. 7Then thereis aprotool
π
usinga uniformommon referene string(CRS) of lengths
(
k
)
d
,
s.t.
π
seurelyUC-realisesabit ommitmentthat anbeusedpolynomially many times.A protool forrealisingbit ommitmentusingaCRShasbeengivenin [CLOS02℄. Toshow
thislemma, weonlyneedtoreviewtheironstrutionto see,that aCRS oflength
s
d
is indeed
suient.Fordetails,seeAppendix A.2.
Lemma4. Let
s
(
k
)
beapolynomially boundedfuntion, thatisomputableintimepolynomialin
k
.Assumeoneof the following holds:
ETDexistand
s
isapolynomially-large funtion.Exponentially-hard ETDexistand
s
isasuperlogarithmi funtion.Then there also exist aonstant
e
∈
N independent ofs
and ETD, s.t. the sizeof the iruitsdesribingthe ETDisboundedby
s
(
k
)
e
for seurity parameter
k
.This is shownby saling theseurity parameterof the originalETD. The proof is givenin
AppendixA.3.
Theorem5. Let
n
=
n
(
k
)
andm
=
m
(
k
)
be polynomially bounded andeiently omputable funtions.Assumeoneof the following onditions holds:
m
ispolynomially-large andETDexist,or
m
issuperpolylogarithmi andexponentially-hard ETDexist.Thenthereisapolynomial-timeomputationallyuniversallyomposableprotool
π
for(
m
→
n
)
-ointossextension.
7
Bythesize oftheiruitswemeansthetotalsizeoftheiruitsdesribingboththekeygeneration
andthe domainsamplingalgorithm.Notethat thentriviallyalso thesize of theresulting keysand
Proof. Let
d
beas in Lemma 3.Letfurthere
be asin Lemma4. Ifm
ispolynomially-largeorsuperpolylogarithmi,then
s
:=
m
1
/
(
de
)
ispolynomially-largeorsuperlogarithmi,resp. So,by
Lemma4thereareETD,s.t.thesizeoftheiruitsdesribingtheETDisboundedby
s
e
=
m
1
/e
.
Then,byLemma3thereisaUC-seureprotoolforimplementing
n
bitommitmentsusingan(
m
1
/d
)
d
=
m
-bitCRS.
Given
n
bitommitments,thefollowingprotoolπ
UC-realisesann
-bit oin-toss(basedontheprotool of[Blu81℄):Uponinput
(init
)
, partyP
1
ommitston
randombitsr
1
.Uponinput(init
)
,andafterP
1
hasommitteditself,partyP
2
sendsn
randombitsr
2
toP
1
.ThenP
1
unveilsthebits
r
1
.Theoutputofthepartiesisthen
-bitstringr
=
r
1
⊕
r
2
,where⊕
denotesthebit-wise exlusiveor.Itiseasytosee,thatthisprotoolUC-realisesan
n
-bitoin-toss.Weonlyroughlyskeththesimulator
S
:Assoonasallunorruptedpartiesgotinput(init
)
,S
learnswhatvaluer
theidealn
-bitoin-tosshas.WhenP
1
isorgetsorrupted,S
learnsthevaluer
1
assoonasP
1
ommits,sothesimulated
r
2
anbehosenasr
1
⊕
r
.WhenP
2
isorgetsorrupted,butP
1
isunorrupted atleastduringtheommitmenttor
1
,thesimulatorS
unveilvaluer
1
tor
2
⊕
r
.Intheasethat bothpartiesgetorrupted,theenvironmentdoesnotlearnthevaluefromtheidealoin-toss,sothesimulatoransimplyhose itto be
r
1
⊕
r
2
.Furthermore, an
m
-bit CRS an be trivially implemented using anm
-bit oin-toss. UsingtheCompositionTheorem weanput theaboveonstrutionstogetherand getaprotool that
UC-realisesan
n
-bit oin-tossusinganm
-bitoin-toss.⊓
⊔
Note that given stronger, but possibly unrealisti assumptions, the lower bound for
m
inTheorem 5 an be dereased. If we assume that for any superlogarithmi
m
, there are ETDs.t.thesizeoftheiriruitsisboundedby
m
1
/d
(where
d
istheonstantfromLemma3),wegetoin-tossextensionevenforsuperlogarithmi
m
(usingthesameproofasforTheorem5,exeptthatinsteadofLemma4weusethestrongerassumption).
However,weannotexpetanevenbetterlowerboundfor
m
,asthefollowingtheoremshows:Theorem6. Let
n
=
n
(
k
)
andm
=
m
(
k
)
be funtions withn
(
k
)
> m
(
k
)
≥
0
for allk
,andassume that
m
isnot superlogarithmi (i.e.,2
−
m
is non-negligible). Then there isno
non-trivial polynomial-time omputationally universally omposable protool for
(
m
→
n
)
-oin tossextension.
Proof (sketh). Assume for ontradition that protool
π
, with partiesP
1
andP
2
usingCT
m
, implementsCT
n
(withm, n
as in the theorem statement). LetA
1
bean adversary onπ
that, takingtheroleofaorruptedpartyP
1
,simplyreroutesallommuniationofP
1
(witheitherP
2
orCT
m
)totheprotoolenvironmentZ
1
andthusletsZ
1
takepartasP
1
in therealprotool.Imagineaprotoolenvironment
Z
1
,runningwithπ
andA
1
asabove,thatkeepsandinternal simulationP
1
ofP
1
and lets this simulation take part in the protool (throughA
1
). After a protoolrun,Z
1
inspetstheoutputκ
1
ofP
1
andomparesittotheoutputκ
2
oftheunorruptedP
2
.Inarealprotoolrunwith
π
,A
1
,andZ
1
,wewillhaveκ
1
=
κ
2
withoverwhelmingprobability sineπ
non-triviallyimplementsCT
n
,andCT
n
guaranteesommonoutputs.SoasimulatorS
1
, runningin theidealmodel withCT
n
andZ
1
, mustbeabletoahievethat theidealoutputκ
2
(thatisideallyhosenbyCT
n
andannotbeinuenedbyS
1
)isidentialtowhatthesimulationP
1
ofP
1
insideZ
1
outputs.Inthatsense,S
1
mustbeabletoonvineP
1
toalsooutputκ
2
.Tothisend,
S
1
mayandmustfakeaompleterealprotoolommuniationasA
1
woulddeliver ittoZ
1
(andthus, toP
1
).end, anadversary
A
2
onπ
withorruptedP
2
is employed that forwardsall ommuniationofP
2
witheitherP
1
orCT
n
toZ
2
.Internally,Z
2
nowsimulatesS
1
(andnotP
2
!)fromaboveand aninstaneCT
n
ofthetrustedhostCT
n
.ReallthatS
1
,givenatargetstringκ
byCT
n
,mimis anunorruptedP
2
alongwithaninstaneofCT
m
.Inthatsituation,S
1
anonvineanhonestP
1
withoverwhelmingprobabilitytoeventuallyoutputκ
.Chanes are
2
−
m
that the
CT
m
-instane made upbyS
1
outputsthe sameseedasthe realCT
m
inarunofZ
2
withπ
andA
2
.Sowithprobabilityatleast2
−
m
−
µ
fornegligibleµ
,insuh arun,Z
2
observesaP
1
-outputκ
thatisidentialtotheoutputoftheinternallysimulatedCT
n
. Butthen,byassumptionabouttheseurityofπ
,thereisalsoasimulatorS
2
forA
2
andZ
2
that providesZ
2
withanindistinguishableview.Inpartiular,in anidealrun withS
2
andCT
n
,Z
2
observesequaloutputs fromCT
n
andCT
n
with probabilityatleast2
−
m
−
µ
′
for negligible
µ
′
.
This isaontradition,asbothoutputs areuniformly and independentlyhosen
n
-bit strings,and
n
≥
m
+ 1
.⊓
⊔
4 Statistial and Perfet Cases
4.1 Stand-alonesimulatability
Westarto withanegativeresult:
Theorem7. Let
m < n
befuntionsinthe seurityparameterk
.Ifm
isnotsuperlogarithmi,thereis notwo-party
n
-bitoin-toss protoolπ
(not evenan ineientone) that uses anm
-bitoin-tossandhas the following properties:
Non-triviality.If no partyisorrupted,the probability that the parties givedierent,invalid
or nooutput isnegligible (byinvalid output wemean outputnot in
{
0
,
1
}
n
).
Seurity. For any (possibly unbounded) adversary orrupting one of the parties there is a
negligiblefuntion
µ
,s.t.foreveryseurityparameterk
andeveryc
∈ {
0
,
1
}
n
,theprobability
for protool output
c
isatmost2
−
n
+
µ
(
k
)
.
If we require perfet non-triviality (the probability for dierent or no outputs is
0
) andperfet seurity (the probability for agiven outputc
is at most2
−
n
), suh aprotool
π
does not exist,evenif
m
issuperlogarithmi.Proof (sketh). Itissuienttoonsiderthease
n
=
m
+ 1
.Without loss of generality, we an assume that the available
m
-bit oin toss is only usedat theend of theprotool. Similarly, weanassume that in thehonest ase, theparties never
outputdistintvalues.Adetailedproofforthesestatementsanbefoundin thefullproof.
Toshowthetheorem,werstonsiderompletetransripts oftheprotool.Byaomplete
transript wemean all messagessent during the run of a protool, exluding thevalue of the
m
-bitoin-toss.Wedistinguishthreesetsofompletetransripts:thesetA
oftransriptshavingnon-zeroprobabilityfortheprotooloutput
0
n
,theset
B
oftransriptshavingzeroprobabilityofoutput
0
n
andzeroprobabilitythattheprotoolgivesnooutput,andtheset
C
oftransriptshavingnon-zeroprobabilityofgivingnooutput. Notethat, sineforaompletetransript,the
protool outputonlydependsonthe
m
-bit oin-toss,anyof theabovenon-zeroprobabilitiesisatleast
2
−
m
.
Foranypartialtransript
p
(i.e.,asituationduring therunoftheprotool),wedenethreevalues
α
,β
,γ
. The valueα
denotes theprobability with whih aorrupted Aliean enforeatransript in
A
startingfromp
, the valueβ
denotesthe probability with whih a orruptedBob an enforea transript in
B
, andthe valueγ
denotes the probability that the ompleteprotool transriptwillliein
C
ifno-oneisorrupted. Weshowindutivelythatforanypartialsimpliity,weassumethat
2
−
m
is notonlynon-negligible,but notieable(inthefull proof,the
generalaseisonsidered).Sineatransriptin
C
givesnooutputwithprobabilityatleast2
−
m
,
theprobabilitythattheprotoolgeneratesnooutput(intheunorruptedase)isatleast
2
−
m
γ
.
By the non-triviality ondition, this probability is negligible, so
γ
must be negligible, too. So(1
−
α
)(1
−
β
)
is negligible,too. Thereforemax
{
1
−
α,
1
−
β
}
mustbenegligible. Fornow, weassumethat
1
−
α
isnegligibleor1
−
β
isnegligible(forthegeneralase,seethefullproof).If
1
−
α
isnegligible,theprobabilityforoutput0
n
isatleast
2
−
m
α
.Sine
α
isoverwhelmingand
2
−
m
notieable,thisisgreaterthan
2
−
n
=
1
2
2
−
m
byanotieableamountwhihontradits
theseurityproperty.
If
1
−
β
is negligible, we onsider the maximum probability a orrupted Bob an ahievethattheprotooloutputisnot
0
n
.Bytheseurityproperty,thisprobabilityshould beatmost
(2
n
−
1)2
−
n
plusanegligibleamount,whihisnotoverwhelming.However,sineeverytransript
in
B
givessuhanoutputwithprobability1
,theprobabilityofsuhisβ
,whihisoverwhelming,inontraditionoftheseurityproperty.
Theperfetaseisprovensimilarly.
⊓
⊔
Thefullproofisgivenin AppendixA.5.
Corollary8. By a non-trivial oin-toss protool we mean a protool s.t. (in the unorrupted
ase) the probability that the parties give no or dierent output is negligible. By a perfetly
non-trivialoin-tossprotool wherethisprobability iszero.
Let
m
be notsuperlogarithmi andn > m
. Then there is no non-trivial protool realisingn
-bitoin-tossusing anm
-bitoin-tossin thesense of statistial stand-alonesimulatability.Let
m
be any funtion (possibly superlogarithmi) andn > m
. Then there is no perfetlynon-trivial protool realising
n
-bit oin-toss using anm
-bit oin-toss in the sense of perfetstand-alone simulatability.
Proof. AstatistiallyseureprotoolwouldhavetheseuritypropertyfromTheorem7andthus,
ifnon-trivial,ontraditTheorem7.Analogouslyforperfet seurity.
⊓
⊔
However,notallislost:
Now we will prove that there exists a protool for oin toss extension from
m
ton
bitwhih is statistially stand-alone simulatably seure. The basi ideais to have the parties
P
1
andP
2
ontributerandomstringstogenerateonestringwithsuientlylargemin-entropy(the min-entropyofarandomvariableX
is denedasmin
x
−
log
Pr
[
X
=
x
]
).Therandomnessfrom this stringisthen extrated usingarandomness extrator.Interestinglythe amountof perfetrandomness(i.e.,thesizeofthe
m
-bitoin-toss)oneneedstoinvestissmallerthantheamountextrated.Thismakesointossextensionpossible.
Toobtaintheointossextensionweneedaresultaboutrandomnessextratorsabletoextrat
onebitof randomnesswhileleavingtheseedreusablelikeaatalyst.
Lemma9. For every
m
there exists afuntionh
m
:
{
0
,
1
}
m
× {
0
,
1
}
m
−1
→ {
0
,
1
}
,
(
s, x
)
7→
r
suhthatforauniformlydistributed
s
andforanx
withamin-entropyofatleastt
thestatistialdistane of
s
k
h
m
(
s, x
)
andthe uniformdistributionon{
0
,
1
}
m
+1
isatmost
2
−
t/
2
/
√
2
.
Proof. Let
h
m
(
s, x
) :=
h
s
1
. . . s
m
−1
, x
i ⊕
s
m
. Hereh·
,
·i
denotes the inner produt and⊕
the addition overGF(2)
. It is easy to verify thath
m
(
s,
·
)
onstitutes a family of universal hashfuntions [CW79℄,where
s
is theindexseletingfrom that family.ThereforetheLeftoverHashLemma[ILL89,Sti02℄guaranteesthatthestatistialdistanebetween
s
k
h
m
(
s, x
)
andtheuniformdistributionon
{
0
,
1
}
m
+1
isboundedby
1
2
√
2
·
2
−
t
= 2
−
t/
2
/
√
2
Withthisfuntion
h
m
asimpleprotoolispossiblewhihextendsm
(
k
)
ointossestom
(
k
)+1
ifthefuntion
m
(
k
)
issuperlogarithmi.Theorem10. Let
m
(
k
)
beasuperlogarithmi funtion, thenthereexistsa onstantroundsta-tistially stand-alone simulatable protool that realises an
(
m
+ 1)
-bit oin-toss using anm
-bitoin-toss.
Proof. Let
h
m
beas in Lemma9.Then thefollowingprotoolrealisesaointoss extensionbyonebit. Assume
m
:=
m
(
k
)
wherek
istheseurityparameter.1.
P
1
uniformlyhoosesa
∈ {
0
,
1
}
⌊
m
−1
2
⌋
andsendsa
toP
2
2.
P
2
uniformlyhoosesb
∈ {
0
,
1
}
⌈
m
−1
2
⌉
andsendsb
toP
1
3. Ifonepartyfailstosendastringofappropriatelengthorabortsthenthisstringisassumed
bytheotherpartytobeanall-zerostringoftheappropriatelength
4.
P
1
andP
2
invokethem
-bit ointoss funtionalityand obtainauniformly distributeds
∈
{
0
,
1
}
m
. Ifoneparty
P
i
failsto invoketheointoss funtionality oraborts, then theotherpartyhooses
s
at random5. Both
P
1
andP
2
omputes
k
h
m
(
s, a
k
b
)
andoutputthisstring.Similartoonstrution7.4.7in[Gol04℄theprotoolisonstrutedinawaythattheadversary
is not able to abort the protool (not even by not terminating). Hene we ansafely assume
that the adversarywill send somemessage of the orret length and will invokethe oin toss
funtionality. Weassumetheadversaryto orrupt
P
2
,orruption ofP
1
ishandled analogously. FurtherweassumetherandomtapeofA
tobexedinthefollowing.Duetotheseassumptionsthere exists a funtion
f
A
:
{
0
,
1
}
⌊
m/
2⌋
→ {
0
,
1
}
⌈
m/
2⌉
for eah real adversary
A
suh that themessage
b
sentinstep2oftheprotoolequalsf
A(
a
)
.Thereisnolossin generalityifweassumetheviewofthepartiestoonsistsofjust
a, b, s
andtheprotooloutputtobes
k
h
m
(
s, a
k
b
)
.Nowfor a spei adversary
A
with xed random tape the output distribution of the realprotool (i.e., view and output) is ompletely desribed by the following experiment: hoose
a
∈ {
R
0
,
1
}
⌊
m/
2⌋
,let
b
←
f
A(
a
)
,hooses
R
∈ {
0
,
1
}
m
(
k
)
,let
r
←
s
k
h
m
(
s, a
k
b
)
andreturn((
a, b, s
)
, r
)
.We nowdesribethesimulator. Todistinguish thetherandom variablesin theidealmodel
from their real ounterparts, we deorate them with a
∼
, e.g.,˜
a,
˜
b,
s
˜
. The simulator in theidealmodel obtainsastring
˜
r
R
∈ {
0
,
1
}
m
+1
from theideal
n
-bit oin-tossfuntionalityand sets˜
s
=
r
1
. . . r
m
. Thenthesimulator hoosesa
˜
R
∈ {
0
,
1
}
⌊
m
−1
2
⌋
andomputes˜
b
=
f
A(˜
a
)
bygiving˜
a
toasimulatedopyoftherealadversary.If
h
m
(˜
s,
a
˜
k
˜
b
) = ˜
r
m
+1
thenthesimulatorgivess
˜
tothe simulatedrealadversaryexpetingtheointoss.Thenthesimulatoroutputstheview(˜
a,
˜
b,
˜
s
)
.If however,h
m
(˜
s,
a
˜
k
˜
b
)
6
= ˜
r
m
+1
thenthesimulatorrewindstheadversary,i.e.,thesimulatorhoosesafresh
˜
a
R
∈ {
0
,
1
}
⌊
m
−1
2
⌋
and againomputes˜
b
=
f
A
(
a
)
.Ifnowh
m
(˜
s,
˜
a
k
˜
b
) = ˜
r
m
+1
thesimulatoroutputs
(˜
a,
˜
b,
˜
s
)
. If againh
m
(˜
s,
a
˜
k
˜
b
)
6
= ˜
r
m
+1
then the simulator rewinds the adversary again. If afterk
invoations of the adversary no triple(˜
a,
˜
b,
s
˜
)
wasoutput, the simulator aborts andoutputs
fail
.Toshowthatthesimulatorisorret,wehavetoshowthatthefollowingtodistributionsare
statistiallyindistinguishable:
((
a, b, s
)
, r
)
asdened intherealmodel,and((˜
a,
˜
b,
˜
s
)
,
r
˜
)
.Byonstrutionof thesimulator, itis obviousthatthetwodistributionsareidential under
theonditionthat
r
m
= 0
,r
˜
m
= 0
and that thesimulator doesnotfail. Thesameholds givenr
m
= 1
,˜
r
m
= 1
andthatthesimulatordoesnotfail.Thereforeitissuienttoshowtwothings:(i) the statistial distane between
r
and the uniform distribution onn
bits is negligible, and(ii) the probability that that the simulator fails is negligible. Property (i) is shown using the
propertiesoftherandomnessextrator
h
m
.Sinea
ishosenatrandom,themin-entropyofa
isat least
⌊
m
−1
2
⌋ ≥
m
2
−
1
, sothemin-entropyofa
k
b
isalso at leastm
boundedby
2
−
m/
4−1
/
2
/
√
2 = (2
−
m
)
1
/
4
/
2
.Sineforsuperlogarithmi
m
itis2
−
m
negligible,this
statistialdistaneisnegligible.
Property(ii )istheneasilyshown:From(i)wesee,thataftereahinvoationoftheadversary
thedistributionof
h
m
(˜
s,
a
˜
k
˜
b
)
isnegligiblyfarfromuniform.Sotheprobabilitythath
m
(˜
s,
˜
a
k
˜
b
)
6
=
˜
r
m
isat mostnegligibly higherthan1
2
. Sine theh
m
(˜
s,
˜
a
k
˜
b
)
in thedierent invokations oftheadversaryareindependent,theprobabilitythat
h
m
(˜
s,
˜
a
k
˜
b
)
6
= ˜
r
m
aftereahativationisneglibiglyfarfrom
2
−
k
.Sothesimulatorfailsonlywithnegligibleprobability.
Itfollowsthattherealandtheidealprotoolexeutionareindistinguishable,andtheprotool
stand-alonesimulatablyimplementsan
(
m
+1)
-bitoin-toss.The idea of the one bit extension protool an be extended by using an extrator whih
extratsalargeramountofrandomness(whilenotneessarilytreatingtheseedlikeaatalyst).
This yields onstant round oin toss extension protools. However, the simulator needed for
suh a protool doesnot seem to be eient, even if the real adversary is. Toget aprotool
thatalsofulls boththepropertyofomputationalstand-alonesimulatabilityand ofstatistial
stand-alonesimulatability,weneedasimulatorthat iseientiftheadversaryis.
Below we give suh aoin toss extension protool for superlogarithmi
m
(
k
)
whih issta-tistially seure and omputationaly seure, i.e., the simulator for polynomial adversaries is
polynomiallybounded, too.Thebasiideahereisto extratonebit atatime in polynomially
manyrounds.
Theorem11. Let
m
(
k
)
besuperlogarithmi, andp
(
k
)
beapositive polynomially-boundedfun-tion, then there exists a statistially and omputationally stand-alone simulatable protool that
realisesan
(
m
+
p
)
-bitoin-tossusing anm
-bitoin-toss.Proof. Let
h
m
beas in Lemma9.Then thefollowingprotoolrealisesaointoss extensionbyp
(
k
)
bits.1.
for
i
= 1
to
p
(
k
)
do
(a)
P
1
uniformlyhoosesa
i
∈ {
0
,
1
}
⌊
m
−1
2
⌋
andsendsa
i
toP
2
(b)
P
2
uniformlyhoosesb
i
∈ {
0
,
1
}
⌈
m
−1
2
⌉
andsendsb
i
toP
1
() If one party fails to send a string of appropriate length or aborts then this string is
assumedbytheotherpartytobeanall-zerostringoftheappropriatelength
2.
P
1
andP
2
invoke them
-bit oin toss funtionality and obtain a uniformly distributeds
∈ {
0
,
1
}
m
. If one party
P
i
fails to invoke the oin toss funtionality or aborts, then theotherpartyhooses
s
at random3.
P
1
andP
2
omputes
k
h
m
(
s, a
1
k
b
1)
k
. . .
k
h
m
(
s, a
p
(
k
)
k
b
p
(
k
))
and outputthisstring.WeonlyroughlyskeththedierenestotheproofofTheorem10.Foreahprotoolroundthe
simulatorfollowsthestrategydesribedintheproofofTheorem10(i.e.,thesimulatorrewindsthe
adversarybyoneround,iftheoin-tossproduedisnottheorretone.)Thenusingstandard
hy-bridtehniquesitanbeshownthatthissimulatorindeedgivesanindistinguishableideal
proto-olrun.Hereitisonlynoteworthythatweusethefatthat
s
k
h
m
(
s, a
1
k
b
1)
k
. . .
k
h
m
(
s, a
p
(
k
)
k
b
p
(
k
))
isstatistiallyindistinguishablefromtheuniformdistributiononm
+
p
bits.However,thisfollowsdiretlyfrom Lemma 9andthefat thateah
a
i
k
b
i
hasmin-entropyat least⌊
m
−1
2
⌋
evengiventhevaluesofall
a
µ
k
b
µ
forµ < i
.⊓
⊔
4.2 Universal Composability(statistial/perfetase)
In the aseof statistial seurity, adversaryand protool environmentare allowed to be
partiesto haltafter apolynomialnumberofativations. However,notethat wedo notimpose
any restritions on the amount of omputational work these parties perform in one of those
ativations.
Theproofofthisstatementisdonebyontradition.Furthermore,theproofissplitupinto
anauxiliarylemmaandtheatualproof.Intheauxiliarylemma, weshowthatwithoutlossof
generality, aprotool for statistially universally omposable oin toss extension has aertain
outerform.Thenweshowthat anysuhprotool(ofthispartiularouterform)isinseure.
For the following statements, we always assume that
m
=
m
(
k
)
, n
=
n
(
k
)
are arbitrary funtions,onlysatisfying0
≤
m
(
k
)
< n
(
k
)
forallk
.Wealsorestrittoprotoolsthatproeedinapolynomialnumberofrounds.That is,byaprotool wemeanin thefollowingoneinwhih
eah partyhaltsafter atmost
p
(
k
)
ativations,wherep
(
k
)
isapolynomialwhihdependsonlyontheprotool. (Asstatedabove,thepartiesarestill unboundedin eah ativation.)Westart
withahelping lemmawhoseproofisavailablein AppendixA.6.
Lemma12. If there isa non-trivial statistially universally omposable protool for
(
m
→
n
)
-ointossextension, thenthereisalso oneinwhih eahparty
has only oneonnetion tothe other party andoneonnetion to
CT
m
,in eah ativation sends either an
init
message toCT
m
or some message to the other party,sendsineahprotoolrunatmostonemessageto
CT
m
,andthisisalwaysaninit
message, the internal state of eah of the two parties onsists only of the view that this party hasexperienedsofar, and
after
P
i
sendsinit
toCT
m
,it does not furtherommuniate withP
3−
i
(fori
= 1
,
2
and in aseofno orruptions).Weproeedwith
Lemma13. There isno non-trivial statistially universally omposable protool for
(
m
→
n
)
-ointossextensionwhih meetsthe requirementsfromLemma12 .
Proof. Assume for ontradition that
π
, usingCT
m
, is a statistially universally omposable implementationofCT
n
, andalsosatisestherequirementsfrom Lemma12.Assumeaxedenvironment
Z
0
that givesbothpartiesinit
inputandthenwaitsforboth partiestooutputaointossresult.ConsideranadversaryA
0
thatdeliversallmessagesbetween thepartiesimmediately.TheresultingsettingD
0
isdepitedinFigure 1.Denote the protool ommuniation in a run of
D
0
, i.e., the ordered list of messages sent betweenP
1
andP
2
,bycom
.Denotebyκ
1
andκ
2
thenaloutputsoftheparties.ForM
⊆ {
0
,
1
}
n
andapossibleprotoolommuniationprex
c
,letE
(
M, c
)
betheprobabilitythattheprotool outputsareidentialandinM
,providedthat theprotoolommuniationstartswithc
,i.e.,E
(
M, c
) :=
Pr
[
κ
1
=
κ
2
∈
M
|
c
≤
com]
,
where
x
≤
y
meansthatx
isaprexofy
.Note thatthepartieshave,apartfrom theirommuniation
com
,onlytheseedω
∈ {
0
,
1
}
m
provided by
CT
m
for omputingtheir naloutputκ
. Sowemay assumethat there is a deter-ministifuntionf
forwhihκ
1
=
κ
2
=
f
(com
, ω
)
withoverwhelmingprobability.Foraxedprotoolommuniation
com
=
c
,onsiderthesetP
1
ω
ω
P
2
com
CT
m
A
0
κ
2
Z
0
κ
1
P
1
ω
ω
P
2
Z
1
A
1
CT
m
Z
0
κ
2
κ
1
Fig.1. Left: The initial setting
D
0
for the statistial ase. (Some onnetions whih are not importantforourproofhavebeenomitted.)Right: SettingD
1
withaorruptedP
1
.SettingD
2
(withP
2
orruptedinsteadofP
1
)isdenedanalogously.of improbable outputs after ommuniation
c
. Then obviously|
M
c
| ≥
2
n
−
2
m
≥
2
n
−1
. By
denition of theidealoutput(i.e., theoutput of
CT
n
in theidealmodel),this impliesthat for suientlylargeseurityparametersk
,theprobabilitythatκ
1
=
κ
2
∈
M
c
isatleast2
/
5
.(Here, any number stritly between0
and1
/
2
would have doneas well.) Otherwise, an environmentould distinguishreal and idealmodel bytesting for
κ
1
=
κ
2
∈
M
c
. SineE
(
M
c
, ε
)
is exatly that probability, wehaveE
(
M
c
, ε
)
≥
2
/
5
forsuientlylargek
.Also,E
(
M
c
, c
)
isnegligibleby denition,soM
c
satisesE
(
M
c
, ε
)
−
E
(
M
c
, c
)
≥
1
3
(1)forsuientlylarge
k
.Sine the protool onsists by assumption onlyof polynomially many rounds,
c
is a list ofsize atmost
p
(
k
)
foraxedpolynomialp
.This meansthat thereisaprexc
ofc
and asinglemessage
m
(eithersentfromP
1
toP
2
orvie versa)suhthatcm
≤
c
andE
(
M
c
, c
)
−
E
(
M
c
, cm
)
≥
1
3
p
(
k
)
(2)forsuientlylarge
k
. Intuitively,thismeansthat ataertainpointduring theprotoolrun,asinglemessage
m
hadasigniantimpatontheprobabilitythat theprotooloutputisinM
c
.Note that suh an
m
must be either sent byP
1
orP
2
. So there is aj
∈ {
1
,
2
}
, suh that for innitely manyk
, partyP
j
sends suh anm
with probability at least1
/
2
. We desribeamodiation
D
j
of settingD
0
. InsettingD
j
, partyP
j
is orrupted and simulated (honestly) insideZ
j
.Furthermore,adversaryA
j
simplyrelaysallommuniationbetweenthissimulationinside
Z
j
andtheunorruptedpartyP
3−
j
. ForsupplyinginputstothesimulationofP
j
and to the unorruptedP
3−
j
, asimulation ofZ
0
is employed insideZ
j
. The situation (forj
= 1
) is depitedinFigure1.Sine
D
j
is basially only a re-grouping ofD
0
, the random variablescom
,ω
, andκ
i
are distributedexatlyasinD
0
,sowesimplyidentifythem.Inpartiular,inD
j
,forinnitelymanyk
, there is with probability at least1
/
2
a prexc
and a messagem
sent byP
j
ofcom
thatNowweslightlyhangetheenvironment
Z
j
intoanenvironmentZ
′
j
.EahtimethesimulatedP
j
sendsamessagem
toP
3−
j
,Z
′
j
heksforall subsetsM
of{
0
,
1
}
n
whether
∃
M
⊆ {
0
,
1
}
n
:
E
(
M, c
)
−
E
(
M, cm
)
≥
1
3
p
(
k
)
,
(3)where
c
denotestheommuniationbetweenP
j
andP
3−
j
so far. If(3)holdsatsomepointforthersttime,thenZ
′
j
tossesaoinb
uniformlyatrandom,andproeedsasfollows:if
b
= 0
,thenZ
′
j
keepsgoingjust asZ
j
would have.Inpartiular,Z
′
j
thenlets
P
j
sendm
toP
3−
j
. However,ifb
= 1
, thenZ
′
j
rewinds the simulationofP
j
to the pointbefore that ativation,and ativates
P
j
againwith freshrandomness,therebylettingP
j
sendapossiblydierentmessage
m
′
.Inthefurtherproof,
c
,m
, andM
refertothese valuesforwhih(3)holds.
In any ase, after having tossed the oin
b
one,Z
′
j
remembers the setM
from (3), anddoesnothek(3)again. Aftertheprotool nishes,
Z
j
outputseither(
⊥
,
⊥
)
(if (3)wasneverfullled),or
(
b, β
)
fortheevaluationβ
oftheprediate[
κ
1
=
κ
2
∈
M
]
(i.e.,β
= 1
itheprotool givesoutput, theprotooloutputsmathandlieinM
).Nowbyourhoieof
j
,Pr
[
b
=
6
⊥
]
≥
1
/
2
forinnitelymanyk
.Also, Lemma 12 guarantees that the internal state of the parties at the time of tossing
b
onsistsonlyof
c
.So,whenZ
′
j
hashosenb
= 1
,andrewoundthesimulatedP
j
,theprobabilitythat at the end of the protool
κ
1
=
κ
2
∈
M
is the same asthe probability of that eventin thesettingD
j
undertheonditionthattheommuniationcom
beginswithc
¯
.Thisprobabilityagainisexatly
E
(
M,
¯
c
)
bydenition. Similarly, whenZ
′
j
hashosenb
= 0
, the probability that at the end of the protoolκ
1
=
κ
2
∈
M
isthesameastheprobabilityof thateventin thesettingD
j
under theonditionthattheommuniation
com
beginswith¯
cm
,i.e.E
(
M,
¯
cm
)
. ThereforejustbeforeZ
′
j
hoosesb
(i.e.,when¯
c
andM
arealreadydetermined),theprobabilitythat at the end we will have
β
= 1
∧
b
= 1
is1
2
E
(
M,
c
¯
)
and the probability ofβ
= 1
∧
b
=
0
is1
2
E
(
M,
¯
cm
)
. Therefore the dierene between these probabilities is at least1
2
E
(
M,
c
¯
)
−
E
(
M,
¯
cm
)
≥
1
3
p
(
k
)
.Sinethisboundonthediereneoftheprobabilitiesalwaysholdswhen
b
6
=
⊥
,byaveragingweget
Pr
[
β
= 1
∧
b
= 1
|
b
6
=
⊥
]
−
Pr
[
β
= 1
∧
b
= 0
|
b
6
=
⊥
]
≥
1
3
p
(
k
)
andusingthefatthat
Pr
[
b
6
=
⊥
]
≥
1
2
forinnitelymanyk
wethenhavethatPr
[
β
= 1
∧
b
= 1]
−
Pr
[
β
= 1
∧
b
= 0]
≥
1
6
p
(
k
)
(4)forinnitelymany
k
whenZ
′
j
runs withthereal protoolasdesribedabove.We show that no simulator
S
j
an ahieveproperty(4) in theideal model, whereZ
′
j
runswith
CT
n
andS
j
. Todistinguish randomvariables during arun ofZ
′
j
in theideal model fromthosein thereal model, weaddatildeto arandomvariablein arun of
Z
′
j
intheidealmodel,e.g.,
˜
b
,β
˜
.Sinetheprotool
π
isnon-trivial,foranyS
j
ahievingindistinguishabilityofrealandidealmodel, wean assumewithoutlossofgeneralitythat
S
j
alwaysdeliversoutputs.Byonstrutionof
˜
b
andκ
,thevariable˜
b
andthetuple( ˜
M , κ
)
areindependentgiven˜
b
6
=
⊥
.Hene,sine