ebook img

Wiretap channel capacity: Secrecy criteria, strong converse, and phase change PDF

0.13 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Wiretap channel capacity: Secrecy criteria, strong converse, and phase change

Wiretap channel capacity: Secrecy criteria, strong converse, and phase change Eric Graves and Tan F. Wong Abstract—This paper employs equal-image-size source parti- requirement is to view the problem as a binary hypothesis tioningtechniquestoderivethecapacitiesofthegeneraldiscrete testing of the alternate hypothesis of M and Zn being in- memoryless wiretap channel (DM-WTC) under four different dependent against the null hypothesis of M and Zn being secrecy criteria. These criteria respectively specify requirements 7 on the expected values and tail probabilities of the differences, correlated.This is an interesting case in which we would like 1 in absolute value and in exponent, between the joint probability the false positive probability given by the likelihood ratio test 0 of the secret message and the eavesdropper’s observation and n 2 tthheesceocrrrietseproiandreindgucperobbaacbkiltiotythifetshteayndwaerrdelienadkeapgeendaenndt.vSaorimateioonf PM,Zn(cid:18)(cid:26)(m,zn)∈M×Zn : PPMM(,mZn)P(mZn,z(znn)) ≥τ(cid:27)(cid:19) a distance constraints that have been previously considered in the 1 as n J literature.Thecapacitiesunderthesesecrecycriteriaarefoundto → →∞ 5 bedifferentwhennon-vanishingerrorandsecrecytolerancesare where the decision threshold τ [0,1) serves as a measure allowed.Basedonthesenewresults,weareabletoconcludethat ∈ 2 of secrecy with τ 1 begin the most secret situation. Note ] tohnelysutrnodnegrctohnevtewrosesepcrroepceyrctyritgeerniaerbaallsyedhoonldcsofnosrtrtahieniDngMt-hWetTaCil that the log-likelih→ood log2 PPMM(,mZn)P(Zmn,z(znn)) may also be used T probabilities. Under the secrecy criteria based on the expected in the hypothesis testing problem above. I values, an interesting phase change phenomenon is observed as For every (m,zn) n, define . the tolerance values vary. ∈M×Z s [c I. INTRODUCTION v(m,zn), 1− PPMM(,mZ)nP(Zmn,z(znn)) + if PM,Zn(m,zn)>0 1 The discrete memoryless wiretap channel (DM-WTC) (cid:12)(cid:12)0 (cid:12)(cid:12) if PM,Zn(m,zn)=0 (cid:12) (cid:12) 7v r(eXce,iPvYe,rZY|X,,aYnd×anZea)vecsodnrsoipstpseroZf.aAsmenedsseargeXM, aislteogbiteimseantet where |c|+ equals c if c>0 and 0 otherwise, and 734 rbeyliaZb.lyOfvreormnXusteos Yofatnhde DdiMsc-rWeetTlyC,agleatinfstne:avMesdr→oppXinng i(m,zn),(−0 log2 PPMM(,mZn)P(Zmn,z(znn)) iiff PPMM,,ZZnn((mm,,zznn))>=00. 0 and ϕn : n be the encoding and decoding functions . respectiveYly em→plMoyed at X and Y, where = [1 : 2nR] is We see that all the secrecy requirements discussed above can 1 the message set and M is uniformlydistribuMted over . The be compactly specified in terms of the tail probabilities and 0 7 transmission reliability requirement is specified by M expected values of v(M,Zn) and i(M,Zn): 1 v: Pr{ϕn(Yn)6=M}≤ǫn (1) S1(δn):PM,Zn({v(M,Zn)>δn})→0 Xi wrehqeurieremǫnen∈ta(s0se,s1s)esdehnoowtemsuthche oenrreomr atoyleleraanrncea.bTouhteMsecfrreocmy SS23((δlnn))::PEMM,,ZZnn([vi((MM,,ZZnn))]=>klnPM),Zn0−PMPZnk1 ≤δn r { } → a Zn. This requirement is often quantified by measuring the S4(ln):EM,Zn[i(M,Zn)]=I(M;Zn) ln ≤ level of “independence” between M and Zn based on either the variation distance wexhpeercetaδtinon∈w(.0r.,t.1]P,Mln,Z∈n.(N0,o∞te)t,haatndS2E(δMn,)Zna[n·]ddSen4(oltne)s tahree 1 P P P thevariationdistanceanddivergence(leakage)constraints,re- M,Zn M Zn 1 2k − k spectively,whileS (δ )andS (l )correspondtothesecrecy 1 n 3 n 1 , PM,Zn(m,zn) PM(m)PZn(zn) requirementsspecifiedbythehypothesistestingproblemusing 2 | − | the likelihood and log-likelihood ratios, respectively. (m,znX)∈M×Zn Clearly these secrecy requirements are related to each or the divergence D(PM,Zn PMPZn) = I(M;Zn) between other. For example, S (δ ) implies S (δ ), and S (l ) im- k 1 n 2 n 3 n PM,Zn and PMPZn. Another way of quantifying the secrecy plies S (l ). Then, by Pinsker’s inequality, S (l ) implies 4 n 4 n erTEiarcincsFgG.rWraavo@enusgfiilss.wweiidtthhuDAerpmayrtmReensteaorfchEleLcatrbi,caAldaenldphCi,omMpDute2r0E78n3g,ineUe.rSin.Ag,. Sin2eq(cid:18)uqaliltny2,lnS2(cid:19)2(δnif) linmp∈lies(0S,1l(n2√2δ).n)Siimf iδlanrly by0.MAalrskoovb’ys University ofFlorida,Gainesville, FL32611,[email protected] elementarylogarithm inequalities, S (δ ) imp→lies S 2δn if T.F.WongwassupportedbytheNationalScienceFoundationunderGrant δ 0. Thus for vanishing toleranc1es,nS –S are e3sselnnt2ially CCF-1320086. Eric Graves was supported by a National Research Council n → 1 4 (cid:0) (cid:1) Research Associateship AwardatArmyResearch Lab. equivalent. Special cases of these secrecy requirementshave been con- that lim (R ,ǫ ,η ) = (R,ǫ,η), for i 3,4 . We n→∞ n n n ∈ { } sideredintheliterature.Forexample,requiringǫ 0in(1), have instead chosen to present the “weak” versions of these n S (nr ) is the equivocation constraint originally→considered criteria, simply because their proofs trivially recover their 4 l in [1]. Six secrecy requirements S –S are more recently “strong” counterparts. That is the ǫ-secrecy capacity under 1 6 considered in [2]. Setting ǫ 0, S is S (l ) for some S (l) orS (l), mustbe less than the ǫ-secrecycapacityunder n 1 4 n 3 4 lln →00,,SS2 iissSS2((δln))ffoorrssoommeeδ→lnn→00,,Sa3nidsSS3(ilsn)Sfo(lrs)omfoer S3 nlogn2n and S4 nlogn2n -respectively. Yet, the secrecy snom→e ln 4 0. 4 n n → 6 3 n cap(cid:16)acities u(cid:17)nder S3 (cid:16)nlogn2n (cid:17)and S4 nlogn2n can also be Thenm→ajority of known secrecy capacity results under the achievedunderS3(ln(cid:16))andS(cid:17)4(ln)fors(cid:16)omeln (cid:17) 0,using[8, → abovesecrecy requirementsare for cases with vanishingerror Theorem 17.11]. tolerance, ǫ 0, and secrecy tolerance, l 0, ln 0, Write n → n → n → or δn 0. These results are nicely summarizedin [2], which C ,maxI(X;Y) shows→thatthesecrecycapacitiesunderS1–S6 (seefootnote1) 1 PX oftheDM-WTCareallgivenbymax I(U;Y) I(U;Z), C ,maxI(U;Y) I(U;Z) PU,X − 2 PU,X − ❝ ❝ where U X Y,Z. Here we are mainly interested in where the alphabet of U in the definition of C can be cases where both the error tolerance ǫ and secrecy tolerance U 2 n assumed to have . Also restrict ǫ [0,1) and δ δ , l or ln are non-vanishing, on which only a few partial |U| ≤ |X| ∈ ∈ n n n [0,1], and r [0, ). Then the following theorems give our l resultsexist.TheoldestsuchresultdatesbacktoWyner’sorig- ∈ ∞ main results regarding the secrecy capacities: inalpaper[1],in whichthesecrecycapacityunderS (nr )of 4 l the degraded DM-WTC (PY,Z|X = PZ|YPY|X) is calculated Theorem 1. The ǫ-secrecy capacity under S1(δ) of the DM- for the case of ǫ 0. The ǫ-secrecy capacity under S (l ) WTC is given by n 4 n → of the degraded DM-WTC is obtained in [3] for the case of C if δ <1 ln 0. Thiscase hasalso beenextendedto the generalDM- C (δ), 2 WnT→C in [4] and [5]. The ǫ-secrecy capacity under S (δ) of 1 (C1 otherwise. 2 the degraded DM-WTC is found in [6]. for all ǫ. In this paper, we determine the secrecy capacities for the Theorem 2. The ǫ-secrecy capacity under S (δ) of the DM- generalDM-WTCundertheabovefoursecurityrequirements, 2 S –S ,withnon-vanishingtolerances.Alloftheseresultscan WTC is given by 1 4 be straightforwardly obtained using our recently developed C if ǫ+δ <1 equal-image-size source partitioning techniques [4], [7]. Fur- C2(ǫ,δ), 2 ther,theǫ-secrecycapacityforeachofthesefourrequirements (C1 otherwise. isunique.UnderS andS thestrongconversepropertyholds, Theorem3. Theǫ-secrecycapacityunderS (nr )oftheDM- 1 3 3 l while it does not under S and S in general. In addition, WTC is given by 2 4 under S and S , the capacity can be broken into distinct 2 4 C (r ),C +min(r ,C C ) phases depending on the error tolerance. For instance, under 3 l 2 l 1− 2 S2 the capacity of the channel is either equal to the capacity for all ǫ. of the channel with vanishing error, or the capacity of the Theorem4. Theǫ-secrecycapacityunderS (nr )oftheDM- channel with no secrecy requirement. We call this interesting 4 l WTC is given by phenomenon a phase change. r C (ǫ,r ),C +min l ,C C . II. MAIN RESULTS 4 l 2 1 ǫ 1− 2 (cid:18) − (cid:19) For i 1,2,3,4 , we call (fn,ϕn) a (n,R ,ǫ ,S (η ))- Theorems 1 and 4 state that the ǫ-secrecy capacities of the n n i n ∈{ } code if the domain of fn (i.e., ) is of cardinality 2nRn, DM-WTC under S (δ) and S (nr ) are invariantto the value and the pair satisfy both (1) anMd Si(ηn). Further we say of ǫ [0,1) for al1l valid valu3es olf δ and rl, respectively. In athcehiervaatebleerirforthesreecreexciysts(RaEsSe)q-tureipnlcee (oaf,b(n,c,)Rn∈,ǫnR,3Sii(sηnS))i-- aonthderS∈3w(onrrdls),.tAhelthsotruognhgicnovnavriearnsteopfrothpeeretryrohrotlodlseruanndceer,Sth1e(δǫ)- caonddelsimsuch thaRt lim,ǫn→,η∞n(R=n,(ǫan,,bη,nc))=if(ia,b,c3),4if .iT∈h{e1n,2th}e, tsheecrelecaykacgapearcaittey run.dIenr sSp3e(cnifirlc), itshenoǫn-s-etrcirveicayllycadpeapceintydeunntdoenr ǫR-sseuccrehcnyth→caa∞tptah(cid:0)ceitRnyEuSnn-dternirpt(cid:1)lhee(aRp,pǫr,oηp)riaisteSS-ia∈(·c)h{iisevtha}belme.aximum iSt3s(antrulr)atienscraetasCes,llitnheear(lnyoans-saecfruent)ctcioapnaociftyrloffrotmheCd2iscurnettiel i 1 Note that for S and S , the above definition corresponds memoryless channel (DMC) ( ,P , ). 3 4 Y|X to what is called “weak” secrecy in the literature [2]. If When the secrecy requiremeXntsare wYeakened to S (δ) and 2 “strong” secrecy is desired, the definition could be modi- S (nr ) from S (δ) and S (nr ) respectively, Theorems 2 4 l 1 3 l fied to that the RES-triple (R,ǫ,η) is S -achievable when and 4 show that the strong conserve property no longer holds i there exists a sequence of (n,R ,ǫ ,S (η ))-codes such fortheDM-WTCastheǫ-secrecycapacitiesgenerallydepend n n i n onthevalueofǫ.UnderS (δ),theǫ-secrecycapacityremains and 2 at C as long as ǫ [0,1 δ). However, for ǫ [1 δ,1), 2 ∈ − ∈ − Ω (r), (m,zn) n : the ǫ-secrecy capacity value experiences an abrupt phase n ∈M×Z change,increasingtoC1 asifthereisnosecrecyrequirement. n PM,Zn(m,zn) 2n(r+λn)PM(m)PZn(zn) , Restricting to within either of the two value ranges, the ǫ- ≤ secrecy capacity under S2(δ) is invariant to ǫ. then o r U=n0defrorSa4l(lnǫrl),[0th,e1)ǫ.-sNeocrteectyhactatphaicsitaylsroeminacilnusdeast Cth2ewcahseens PM,Zn(Ωn(r))≤PQn QSn(r) +µn. l of strong secrec∈y (S (l ) with l 0) and bounded leakage For proving achievability in Th(cid:0)eorems(cid:1)2 and 4, we will 4 n n (S (l ) with l =l). Thus the stro→ng conversepropertyholds make use of the following lemma to simplify discussions: 4 n n when rl = 0 as proven in [4] and [5]. For any fixed rl Lemma 6. For i 2,4 , if the RES-triple (R,0,η) is S - ∈ i (0,C1 C2), the ǫ-secrecy capacity increases from C2 +rl achievable, then th∈e{RES-}triple (R,γ,(1 γ)η) is also S - − i to C and then levels off as ǫ increases in the range [0,1). − 1 achievable for any γ [0,1). TheDM-WTCexhibitsa phasechangefromwherethestrong ∈ conversepropertyholdstowhereitdoesnot.Whenr C A. Proof of Theorem 1 l 1 ≥ − C , the ǫ-secrecy capacity value remains at C for all ǫ 2 1 (Direct) For any δ [0,1) and ǫ [0,1), the RES- ∈ [0,1), and the DM-WTC exhibits another phase change after triple (C ,ǫ,δ) being S∈-achievable follow∈s directly from [8, 2 1 which the strong converse property holds again. Theorem 17.11], which in particular shows the RES-triple (C ,0,0) is S -achievable.On the other hand, the RES-triple III. PROOFS OF THEOREMS 2 1 (C ,ǫ,1)isS -achievablesinceC isthechannelcapacityfor 1 1 1 We prove the converses in Theorems 1–4 by employing the DMC ( ,P , ), and δ =1 correspondsto no secrecy Y|X the following strong Fano’s inequality developed in [4] and X Y constraint. information stabilization result developed in [7]: (Converse) To prove that C (δ) is an upper bound on the 1 Strong Fano’s inequality. For any (fn,ϕn) of rate R that ǫ-secrecycapacityunderS1(δ), firstapplyLemma5toobtain gives Pr ϕn(Yn) = M ǫ over the DM-WTC, there exist values τn, µn, and λn which converge to 0 as n increases, scauabrradsenintdaolmi{tyiinsdaext mQo6nstthpao}tlyr≤naonmgieasloinvenr,aζnni→nde0x, asentdQanwinhdoesex ΩPsuMnc(h,0Z)ntha(anΩtdnP(Q0M)Sn,)(Z0n)(aΩ1snd(e0ρfi)n)nefd≤orinsPoLQmenme(cid:0)mQρnaSn(50.)W(cid:1)0e,+dalµuseno,thofaovSre1(tshδeat)st. Thus S (δ) and≥Lemm−a 5 together imply→that 1 1 R , q :R I(M;Yn Q =q )+ζ Qn (cid:26) n ∈Qn ≤ n | n n n(cid:27) PQn QSn(0) ≥PM,Zn(Ωn(0))−µn ≥1−ρn−µn. (2) satisfying P ( R) 1 ǫ ζ . Butthe(cid:0)nthestr(cid:1)ongFano’sinequalityand(2)togethergivethe Qn Qn ≥ − − n existence of a q such that Information stabilization. For the (fn,ϕn) pair, random n ∈Qn iinnddeexx Qsunb,seatndiZndex setsQatnisfaybinogveP,ther(e eZx)ist ξ1n →ξ0:a1ndan R≤ n1I(M;Yn|Qn =qn)+ζn (3) 1) PZZnn:|QPnZQ(nZnˆ|Qn⊆n(q(znQn)n||qqnn))≥=.ξ1n−2−ξHnQ,(ZnwnhQ|eQrnne=Zq≥ˆnn)(q,−n),n (cid:8)zn ∈ n1I(M;Zn|Qn =qn)≤τn (4) 2) tPhere (ex˜is(tqs )qa) M˜1(qn)ξ , an⊆d P M(cid:9)(msqati)sfy=.ing sfoinrcelarPgeQnenQouRngh∩nQSna(n0d)ǫ ≥ [10,−1)ǫ. −Coζmnb−ininρgn −(3)µannd>(40) M|Qn M n | n ≥ − n M|Qn | n ξn (cid:0) (cid:1) ∈ 3) P2−H(M|Qn(=˜qnn)(mfo,rqea)cmh,mq ∈) M˜(qn),1and ξ where walilthǫ[8, L[0e,m1)m.aO1n7.t1h2e],oitthfeorllhoawnsd,thwathRen≤δC=2+1,ζnth+e ηstnrofnogr Zn|M,Qn Z n | n ≥ − n . ∈ 2Z˜−nH(m(Z,nq|Mn),Q,n=qnz)n,∈ Zn : PZn|M,Qn(zn|m,qn) =ξn Fasanion’sthienesqtuaanlditayrd(is.etr.,on(g3))cognivveersseRar≤gumCe1n+t foζnr tfhoer DalMl ǫC, (cid:8) ( ,P , ). for each qn ∈QZn. (cid:9) X Y|X Y B. Proof of Theorem 2 Obtained through the information stabilization result in the appendix, the following lemma will also be needed: (Direct) The RES-triple (C ,ǫ,δ) is S -achievable, once 2 2 again, by [8, Theorem 17.11], for ǫ+δ < 1. For ǫ+δ 1, Lemma 5. For any r ≥ 0, there exist τn → 0, µn → 0, and the RES-triple (C ,ǫ,1 ǫ) is S -achievable by Lemm≥a 6, λn →0 satisfying nλn →∞ such that by defining since the RES-trip1le (C −,0,1) is S2 -achievable. 1 2 S(r), q : 1I(M;Zn Q =q ) r+τ (Converse) On the other hand, to prove that C2(ǫ,δ) is an Qn n ∈Qn n | n n ≤ n upper bound on the ǫ-secrecy capacity under S (δ), observe (cid:26) (cid:27) 2 that S (δ) implies . 2 (cid:12)(cid:12)n11lFoogr2aanny−nonn1-nleogga2tibvne(cid:12)(cid:12)λ≤nλ→n.0,an>0,andbn>0,an=λn bn means δ ≥kPM,Zn −PMPZnk1 ≥ PM,Zn(m,zn)−PM(m)PZn(zn) together with (8) implies that there must be a qn ∈QRn such ≥(m,zn)∈MX×Zn\Ωn(0)PM,Zn(m,zn) 1−2−nλn that n1I(M;Zn|Qn =qn)≤ rl1+ αnǫlogζ2nn. (9) (m,zn)∈MX×Zn\Ωn(0) (cid:0) (cid:1) Again by the strong Fano’s inequality−, for−this q we also = 1−2−nλn [1−PM,Zn(Ωn(0))]. (5) have(3).Hencecombining(3)and(8)with[8,Lemmna17.12], Thus (cid:2)combining L(cid:3)emma 5 and (5) gives it follows that r + αlog n P S(0) 1 δ 2−nλn µ . R C + l n 2 +ζ . Qn Qn ≥ − − − n ≤ 2 1 ǫ ζ n n − − As a result, if ǫ(cid:0)+δ <(cid:1)1, then there must exist a q n n The strong Fano’s inequality alone also gives the bound R ∈ Q such that (3) and (4) are simultaneously satisfied since ≤ C +ζ . Thus we have 1 n PQn(QRn ∩QSn(0))≥1−ǫ−δ−ζn−2−nλn −µn >0 R min C + rl+ αnlog2n +ζ , C +ζ . ≤ 2 1 ǫ ζ n 1 n forallsufficientlylargen. Henceagainby[8, Lemma17.12], (cid:26) − − n (cid:27) we have R C2 +ζn +τn. If ǫ+δ 1, then the strong IV. CONCLUSIONS ≤ ≥ Fano’s inequality (i.e., (3)) gives R C +ζ . 1 n Employing the recently developed techniques of equal- ≤ image-size partitioning, we obtained the ǫ-secrecy capacities C. Proof of Theorem 3 under S (δ), S (δ), S (nr ), and S (nr ) of the DM-WTC (Direct)The RES-triple ((C +min(r ,C C ),ǫ,r) is 1 2 3 l 4 l S achievabledueto[8,Theore2m17.13],wl hic1h−sho2wsthatlthe fornon-vanishingǫ, δ, and rl. The secrecycriteria considered 3 include the standard leakage and variation distance secrecy RES-triple ((C +min(r ,C C ),0,r) is S achievable. (Converse) O2n the othler h1a−nd, 2to provle that3C (r ) is an constraints often employed in the literature. Our new results 3 l show that both the capacity value and the strong converse upper bound on the ǫ-secrecy capacity under S (nr ) of the 3 l property of the DM-WTC are in fact dependent on the DM-WTC, we notethat Lemma5 and S (nr ) directlyimply 3 l secrecy criterion adopted. We conjecture that the interesting PQn QSn(rl) ≥PM,Zn(Ωn(rl))−µn ≥1−ρn−µn (6) phasechangephenomenonobservedincaseswherethestrong conserve property does not hold is commonplace in many for som(cid:0)e ρ (cid:1)0. Thusas beforethe strongFano’sinequality n other multi-terminal DMCs. → and(6)togethergivetheexistenceofaq satisfying(3) n n ∈Q and APPENDIX 1 I(M;Zn Q =q ) r +τ (7) A. Proof of Lemma 5 n n k n n | ≤ We need the following lemma to prove Lemma 5: since P R S(r ) 1 ǫ ζ ρ µ > 0. Qn Qn ∩Qn l ≥ − − n − n − n Hencebyapplying[8,Lemma17.12]tothecombinationof(3) Lemma7. LetQ bearandomindexrangingover ,whose n n (cid:0) (cid:1) Q and (7), we have R < C +r +ζ +τ . The strong Fano’s cardinality is at most polynomialin n, and V be any discrete 2 l n n inequalityalso gives the boundR≤C1+ζn. Thuswe obtain random variable distributed over V. Then there exist λn →0 R≤min{C2+rl+ζn+τn,C1+ζn} for all ǫ. and ξn′ →0 such that nλn →∞ and . D. Proof of Theorem 4 P (v,q ) :P (v q )= P (v) V,Qn n ∈V ×Qn V|Qn | n λn V (Direct) First note the RES-triple 1 ξ′. ≥ −(cid:0)(cid:8)n (cid:9)(cid:1) (C2+min 1r−lǫ,C1−C2 ,0,1r−lǫ is S4 achievable Note that λn and ξn′ both depend only on the polynomial (cid:16)due to [8,(cid:16) Theorem 1(cid:17)7.13]. H(cid:17)ence the RES-triple cardinality bound on . n (C +min rl ,C C ,ǫ,r is S -achievable by Q 2 1−ǫ 1− 2 l 4 Proof: Let α > 0 be such that nα. First (cid:16)Lemma 6. (cid:16) (cid:17) (cid:17) write = (v,q ) :P (v|Qqn)|>≤n2αP (v) (Converse) To prove C4(ǫ,rl) upper-bounds the ǫ-secrecy and A= (v,q )n ∈V ×Q:nP V|Q(nv q|)n<n−2αPV(v) . capacity under S4(nrl) of the DM-WTC, notice that S4(nrl) ThenB (cid:8) n ∈V ×Qn V|Qn | n V (cid:9) implies (cid:8) (cid:9) . r 1I(M;Zn) 1I(M;Zn Q ) αlog n PV,Qn (v,qn)∈V ×Qn :PV|Qn(v|qn)=λn PV(v) l ≥ n ≥ n | n − n 2 ≥1−(cid:0)(cid:8)PV,Qn(A∪B) (cid:9)(cid:1)(10) 1 α ≥ nI(M;Zn|Qn =qn)PQn(qn)− n log2n where λn = 2nαlog2n. Thus the lemma is verified by (10) if qnX∈QRn we canshowthatPV,Qn(A∪B)→0.In particular,we doso 1 α by bounding P ( ) n−α and P ( ) n−2α, and min I(M;Zn Q =q )P ( R) log n (8) V,Qn A ≤ V,Qn B ≤ ≥qn∈QRn n | n n Qn Qn − n 2 setting ξn′ =n−α+n−2α. To bound P ( ), note that for all (v,q ) , wherenα isthecardinalityboundon n.Butfromthestrong V,Qn A n ∈A Fano’s inequality, we have P ( RQ) 1 ǫ ζ . This P (q ) n−2α (11) Qn Qn ≥ − − n Qn n ≤ since since (m,zn,q ) Ξ . Thus Lemma 5 results from (16) by n n ∈ settingτ =4λ +2ξ andµ =3ξ′ +4ξ ,becausewehave P (v) P (v q )P (q ) n2αP (v)P (q ). n n n n n n V ≥ V|Qn | n Qn n ≥ V Qn n from (14) Then the upper bound on P ( ) follows from (11) as below: V,Qn A PM,Zn(Ωn(r)) P ( )= P (v q )P (q ) ≤PM,Zn,Qn(Ξn∩Γn∩Ωn(r)×Qn)+3ξn′ +4ξn V,Qn A (v,Xqn)∈A V|Qn | n Qn n ≤PQn QSn(r) +3ξn′ +4ξn. P (v q )n−2α n−α. B. Proof of L(cid:0)emma(cid:1)6 ≤ V|Qn | n ≤ (v,Xqn)∈A (n,FRor ,(1i γ)ǫ∈+γ,S{2(,(41}, γ)wl e))-cocdaen (fˆnco,nϕsntr)u,ctgivena The upper bound on PV,Qn(B) follows similarly in that that thnere e−xistsna (n,Rni,ǫn,−Si(lnn))-code (fn,ϕn). Whence P ( )= P (v q )P (q ) the lemma follows by the definition of the RES-triples. V,Qn B V|Qn | n Qn n Letting Mˆ be a random variable distributed identical, but (v,Xqn)∈B independent, to M. The new encoder, fˆn, is constructed by P (v)P (q )n−2α n−2α. ≤ V Qn n ≤ setting it equal to f(M) with probability 1 γ and to f(Mˆ) (v,Xqn)∈B with probability γ. While the new decoder ϕ−ˆn =ϕn. Clearly, an error will likely occur if fˆ(M) is set equal to Apply Lemma 7 three times with V = M, V = Zn, and f(Mˆ). On the other hand, the probability of error will revert V =(M,Zn), respectively. Writing to that of (fn,ϕn) if fˆ(M) is set equal to f(M). Thus the probability of error for (fˆn,ϕˆn) is at most (1 γ)ǫ +γ. Γn , (m,zn,qn)∈M×Zn×. Qn : Letting PZn,M be the joint distribution o−f Znn,M for n PPMM,|ZQnn|(Qmn|(qmn),z=.nλ|qnnP)M=λ(mn P),Ma,nZdn(m,zn), fPiˆnnduacasen(dd1bP−yfγn).P,BwZunet,Mctahne+nw,γrfPiotZerntthhPeeMjvo,aiwnritahdtiilioestnrtihdbeuistmtiaoannrcgoeif,nZalns,rMemafoinr . M Zn where λ is obPtaZinne|Qdni(nznL|eqmn)m=aλ7n, PwZenh(azvne)o k(1−γ)PZn,M +γPZnPM −PZnPMk1 n =(1 γ) PZn,M PZnPM 1. − k − k PM,Zn,Qn(Γn)≥1−3ξn′. (12) And for divergence Next define D((1 γ)PZn,M +γPZnPM PZnPM) Ξn , (m,zn,qn)∈M×Zn×Qn :qn ∈QZn, ≤(−1−γ)D(PZn,M||PZnP|M| )+γD(PZnPM||PZnPM) n m ˜(q ), and zn ˆn(q ) ˜n(m,q ) =(1 γ)D(PZn,M PZnPM). n n n − || ∈M ∈Z ∩Z o with the corresponding Z, ˜(q ), ˆn(q ), and ˜n(m,q ) Qn M n Z n Z n as given in the information stabilization result summarized in REFERENCES Section III. Similar to before, [1] A.Wyner,“Thewire-tapchannel,”BellSyst.Tech.J.,vol.54,pp.1355– 1387,Oct.1975. PM,Zn,Qn(Ξn)≥1−4ξn. (13) [2] M. Bloch and J. Laneman, “Strong secrecy from channel resolvability,” IEEETrans.Info.Theory,vol.59,pp.8077–8098, Dec.2013. Combining (12) and (13) gives [3] V.TanandM.Bloch,“Informationspectrumapproachtostrongconverse theoremsfordegradedwiretapchannels,”inProc.52ndAnnualAllerton PM,Zn,Qn(Ξn∩Γn)≥1−3ξn′ −4ξn. (14) Conf.Comm.,Con.,andComp.,pp.747–754,Sep.2014. [4] E. Graves and T. F. Wong, “Equal-image-size source partition- From here note that for any (m,zn,q ) Ξ Γ , n n n ing: Creating strong fano’s inequalities for multi-terminal discrete ∈ ∩ memoryless channels,” ArXiv e-prints, Dec. 2015. Available at PM,Zn(m,zn) 2n(r+λn)PM(m)PZn(zn) https://arxiv.org/abs/1512.00824 . ≤ [5] Y.-P. Wei and S. Ulukus, “Partial Strong Converse for the Non- implies Degraded Wiretap Channel,” ArXiv e-prints, Oct. 2016. Available at https://arxiv.org/abs/1610.04215 . 1 log PZn,M|Qn(zn,m|qn) r+4λ , (15) [6] M.Hayashi,H.Tyagi,andS.Watanabe,“Strongconverseforadegraded n 2 P (zn q )P (mq ) ≤ n wiretap channel via active hypothesis testing,” in Proc. 52nd Annual Zn|Qn | n M|Qn | n AllertonConf.Comm.,Con.,andComp.,pp.148–151,Sep.2014. because (m,zn,q ) Γ . And then in turn, for all [7] E. Graves and T. F. Wong, “Information stabilization of images over n n (m,zn,q ) Γ Ξ ,∈ discrete memoryless channels,” in Proc. IEEE Int. Symp. Info. Theory, n ∈ n∩ n pp.2619–2623, Jun.2016. 1 [8] I. Csisza´r and J. Ko¨rner, Information Theory: Coding Theorems for r+4λ I(M;Zn Q =q ) 2ξ (16) DiscreteMemorylessSystems.CambridgeUniversityPress,2nded.,2011. n n n n ≥ n | −

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.