Steganography in Inactive Frames of VoIP Streams Encoded by Source Codec

Transcription

1 Seganography n Inacve Frames o VoIP Sreams Encoded by Source Codec Y. F. Huang, Shanyu Tang, and Jan Yuan, Senor Member, IEEE Absrac Ths paper descrbes a novel hgh capacy seganography algorhm or embeddng daa n he nacve rames o low b rae audo sreams encoded by G.723. source codec, whch s used exensvely n Voce over Inerne Proocol (VoIP). Ths sudy reveals ha, conrary o exsng houghs, he nacve rames o VoIP sreams are more suable or daa embeddng han he acve rames o he sreams, ha s, seganography n he nacve audo rames aans a larger daa embeddng capacy han ha n he acve audo rames under he same mpercepbly. By analysng he concealmen o seganography n he nacve rames o low b rae audo sreams encoded by G.723. codec wh 6.3kbps, he auhors propose a new algorhm or seganography n deren speech parameers o he nacve rames. Perormance evaluaon shows embeddng daa n varous speech parameers led o deren levels o concealmen. An mproved voce acvy deecon algorhm s suggesed or deecng nacve audo rames akng no packe loss accoun. Expermenal resuls show our proposed seganography algorhm no only acheved perec mpercepbly bu also ganed a hgh daa embeddng rae up o 0 bs/rame, ndcang ha he daa embeddng capacy o he proposed algorhm s very much larger han hose o prevously suggesed algorhms. Index Terms seganography, nacve rames, audo sreams, VoIP S. Tang s wh he Faculy o Compung, London Meropolan Unversy, London N7 8DB, UK (correspondng auhor o provde phone: ; ax: ; e-mal: s.ang@ londonme.ac.uk).

2 I. INTRODUCTION Sreamng meda, such as Voce over Inerne Proocol (VoIP) sreams, are broadcased lve over he Inerne and delvered o end-users. Secury remans one o he man challenges wh hs new echnology. Wh he upsurge o VoIP applcaons avalable or use n recen years, VoIP sreams become one o he mos neresng cover obecs or modern seganography. Dgal seganography n low b rae audo sreams s commonly regarded as a challengng opc n he eld o daa hdng. There have been several seganography mehods o embeddng daa n audo sreams. For example, Wu e al. [] suggesed a G.7-based adapve speech normaon hdng approach. Aok [2] proposed a echnque o lossless seganography n G.7 encoded speeches. Ma e al. [3] ramed a seganography mehod o embeddng daa n G.72 encoded speeches. All hese mehods adop hgh b rae audo sreams encoded by he waveorm codec as cover obecs, n whch pleny o leas sgncan bs exs. However, VoIP are usually ransmed over low b rae audo sreams encoded by he source codec lke ITU G.723. codec o save on nework bandwdh. Low b rae audo sreams are less lkely o be used as cover obecs or seganography snce hey have ewer leas sgncan bs han hgh b rae audo sreams. Lle eor has been made o develop algorhms or embeddng daa n low b rae audo sreams. Chang e al. [4] embedded normaon n G.729 and MELP audo sreams. Huang e al. [5] proposed a seganography algorhm or embeddng normaon n low b rae audo sreams. Bu hese seganography algorhms have consrans on he daa embeddng capacy, ha s, her daa embeddng raes are oo low o have praccal applcaons. Thus he man ocus o hs sudy was o work ou how o ncrease he daa embeddng capacy o seganography n low b rae audo sreams.

3 The res o hs paper s organsed as ollows. Secon 2 summarses some relaed work, dscussng he possbly o embeddng daa n he nacve rames o low b rae audo sreams. In Secon 3, he mpercepbly o he seganography algorhm or embeddng daa n he nacve audo rames s analysed. Our proposed seganography algorhm s presened n Secon 4. Secon 5 deals he expermenal se-up and perormance evaluaon resuls. Fnally, he paper ends wh conclusons and drecons or uure work n Secon 6. II. RELATED WORK Analyss by synhess (ABS)-based speech normaon hdng approach was adoped o embed speech daa n an orgnal speech carrer, wh good ecency n seganography and good qualy o oupu speech [6]. Recenly, lnear predcve coecens were subsued wh secre speech daa by usng a ABS speech codng scheme [7], bu he expermenal resuls avalable are very lmed. Kräzer, Dmann, and Vogel [8] argued ha he nacve voce o a speech was no suable or beng used as a cover obec or seganography owng o an obvous dsoron o he orgnal speech. By conras, Huang e al. [9] suggesed an algorhm or embeddng normaon n some parameers o he speech rame encoded by ITU G.723. codec, whou leadng o dsncon beween nacve voces and acve voces. I seems ha Kräzer, Dmann, and Vogel s opnons [8] and Huang and co-workers resuls [9] conradc each oher. Such a conradcon can be arbued o he deren speech codecs ha were used o compress and encode audo sgnals. In [8], audo sreams were encoded by a pulse-code modulaon (PCM) codec; bu an ITU G.723. source codec was used o encode audo sreams n [9]. The PCM codec s based on he waveorm model ha samples, quanzes and encodes audo sgnals drecly; he sample value represens he orgnal volume o he sgnal. In hs case,

4 he nacve voce canno be used o embed normaon snce wll lead o obvous dsoron. However, he source codec s a hybrd codec, whch s based on he source model. Ths codec compresses he speech a a very low b rae and perorms on a rame-by-rame bass; each rame s encoded no varous parameers raher han he sample volumes. Thus he volume o he speech does no change mpercepbly even hough her nacve audo rames conan hdden normaon. The heorecal analyss above suggess ha seganography n he nacve rames o low b rae audo sreams would aan a larger daa embeddng capacy an approprae seganography algorhm were used. The res o hs paper deals our successul eor on such a new seganography algorhm or embeddng daa n he nacve rames o low b rae audo sreams encoded by ITU G.723. source codec. III. PRINCIPLE OF STEGANOGRAPHY IN INACTIVE AUDIO FRAMES A. Hangover Algorhm or Deecng Acve Voces To reduce nework bandwdh n VoIP applcaons, some source codecs nroduce slence compresson durng he nacve perod o audo sreams. The slence compresson echnque has wo componens: voce acvy deecon (VAD) and comor nose generaor [0]. The VAD s used o decde wheher he curren audo rame s an acve voce by comparng he energy o he rame (Enr) wh a hreshold (Thr), as shown n (): VAD 0 Enr Thr Enr Thr () VAD = 0 means he rame s an nacve voce; oherwse, he rame s an acve voce. The energy o he curren rame, Enr, s compued by

5 239 '2 Enr e ( n) (2) 80 n60 where e (n) s he oupu sgnal o he ne mpulse response (FIR) ler whose npu sgnal s he curren rame, {s[n]} n= The FIR ler compues e (n) usng (3) 0 ' e ( n) s[ n] ano[ ] s[ n ] n (3) where A no (Z) = {a no [] = 0,,0} s he auocorrelaon coecen vecor o he ler. The hreshold n (), Thr, s gven by Thr Nlev log2 28 I Nlev 28 I 28 Nlev 6384 I Nlev 6384 (4) where Nlev s he nose sze o he curren rame, and s updaed by s prevous value, he energy o he prevous rame Enr -, and he sel-adapve lag Aen. Nlev s dened as ollows: Nlev 0.25 Nlev Nlev 0.75 Enr Nlev else Enr.0325 Nlev Aen 0 Nlev (5) Nlev else where Aen = [0, 6], and Nlev s lmed o a value beween 28 and 307. Acve voce Vcn Hcn Slence compresson Enr Thresh Enr < Thresh Normal encodng Slence encodng Fg.. Illusraon o Hangover algorhm In general, Hangover algorhm s used or deecng nacve voces o avod nose peaks beng exended [0]. I an audo rame s deermned o be an nacve voce, he rame s encoded no a slence nser descrpon (SID) rame by

6 usng he slence compresson algorhm. Havng receved he SID rame, he decoder generaes a comorable nose a he recevng end. The Hangover algorhm s llusraed n Fg.. The rs row n Fg. shows he classcaon o voce duraon beore slence compresson. Hcn s he Hangover-rame number o nacve voces when an acve voce begns o change o an nacve voce n he speech. The second row s an esmae o he energy and he hrd row ncludes he correspondng codec algorhms. An audo sream s acually dvded no rames beore beng encoded. For nsance, wh G.723 codec he audo sream s dvded no rames 30 ms n lengh. Suppose he audo sream F conans N rames, F = { = 0,, N}. I he energy (Enr) o he rame s less han he hreshold value (Thr), Enr < Thr, he rame s he rs rame o an nacve voce. Ths rame s dened as a Hcn rame n Hangover algorhm and s hen encoded by usng he normal codec algorhm raher han he slence compresson algorhm. I subsequen rames are sll nacve voces, Hangover algorhm wll no perorm slence compresson unl he sxh rame. In oher words, Hangover algorhm sars o encode he sxh rame o he nacve voce no a SID rame unl he nex acve voce emerges. The rs ve rames ( s o 5 h ) o he nacve voce are sll encoded no Hangover rames, denoed by *Hcn. The acve voce o he audo sream s encoded no acve rames *A by usng he normal codec algorhm. Accordng o Hangover algorhm, audo rames are classed no hree ypes, acve voce rame A, Hangover rame Hcn, and slence compresson rame S. The audo speech F can be expressed as F = { A, s = 0,, N, = 0,, N 2, N = N + N 2 } (6) The speech F s hen encoded no F* by usng Hangover algorhm, whch can be wren as F (F) (7)

7 F * * A * Hcn * SID,, l,..., n,,..., n 2, l,..., n 3, N n n 2 n 3 B. Denons o Inacve and Acve Frames The slence compresson echnque s an oponal uncon or he source codec. In ac, mos source codecs do no use slence compresson n VoIP applcaons. All audo rames are encoded unormly by usng he normal encodng algorhm regardless o wheher hey are acve voces or nacve voces. Thus wo ypes o rames are oupued when he speech sream F s encoded by he source codec. For example, ITU G.723. codec encodes he speech no wo ypes o rames, acve rames and nacve rames, whou usng he slence compresson algorhm. Denon : The acve rame *A s encoded by he source codec rom he acve voce o he speech. I s expressed as * A A 0,..., N (8) Denon 2: The nacve rame *S s encoded by he source codec rom he nacve voce o he speech. I s expressed as * S S 0,..., N 2 (9) As he speech s dvded no nacve voces and acve voces by VAD, all he voces are encoded unormly by he source codec o orm audo rames, n whch nacve rames can be dsngushed rom acve rames. Combnng (6) - (9) yelds F* = { *A, *S = 0,, N, = 0,, N 2, N = N + N 2 } (0) C. B Dsrbuon Paerns o Inacve Frames

8 Ths secon dscusses wheher he /0 dsrbuon paern o an nacve rame s smlar o ha o an acve rame he nacve voce o a speech s encoded no nacve rames. Frsly, we analysed he sascal probably o /0 presenaon n nacve rames. Assumng an audo sream s dvded no N rames, among hem here are N nacve rames and N 2 acve rames,.e. N = N + N 2. The audo sream s denoed by F * = { *A, *S = 0,, N, = 0,, N 2 }. Suppose each rame consss o M bs, namely * ={b 0,, b,, b M b = 0, }. The average probably o presenaon n all he nacve rames s compued by usng () N M b b, MN 0 0 () where b, denoes he h n he h nacve rame o he sream. So he average probably o 0 presenaon n all he nacve rames s gven by (2) b0 b Table lss he expermenal resuls o he sascal probables o presenaon n he nacve rames and acve rames encoded by G.723. codec, respecvely. Ten speech sample les were used or he expermens, wh each le beng esed sx mes n order o work ou he average probables o presenaon n he nacve and acve rames. Table Average probables o presenaon n nacve and acve rames Speech le Acve rame Inacve rame No Mean Varance Mean Varance

9 The average probables shown n Table ndcae ha here was no obvous derence n he 0/ presenaon probably beween he acve rames and nacve rames o low b rae audo sreams. In oher words, we could no dsngush nacve rames rom acve rames by he /0 presenaon probably. Secondly, we examned he probably o /0 umpng n nacve rames. The /0 umpng probably expresses he chance ha he b changes o 0 or 0 o nversely wll occur n an nacve rame. Smlarly, he audo sream s denoed by F * = { *A, *S = 0,, N, = 0,, N 2 }, and * s = {b 0,, b,, b M b = 0, }. Then he average probably o /0 umpng n all he nacve rames o he speech le s calculaed by N2 M c b b 00% M N (3) where b denoes he h b n he nacve rame and M denoes he b number o he nacve rame, such as M = 92 or G.723. codec wh 6.3 kbps. Table 2 shows he average probables o /0 umpng n he nacve rames and acve rames o low b rae audo sreams, respecvely. Each mean probably s based on sx repeaed expermens on a speech le. The resuls sugges boh nacve rames and acve rames were ndsngushable n erms o he umpng probably. Table 2 Average probables o /0 umpng n nacve and acve rames Speech le Acve rame Inacve rame no Mean Varance Mean Varance

10 Fnally, we suded he run-lengh sascal characer o 0/ n nacve rames. The run-lengh sascal mehod was used o calculae he run-lenghs o connuous 0 or presenaon n nacve rames. Assumng he audo sream s denoed by F * = { *A, *S = 0,, N, = 0,, N 2 } and * s = {b 0,, b,, b M b = 0, }, sases he ollowng equaon: b b b k k b kr k b b k kr (4) where = 0,, R, and R denoes he run-lengh o b k n an nacve rame. Then he run-lengh o b k (b k = 0, ) n he nacve rame s equal o he number o bs rom b k o b k+r-. The dsrbuon paern o he run-lengh n nacve rames s dened as he probably o varous run-lenghs presenng n all nacve rames o he speech le, gven by N2 0 0 ( ) M ( ) 00% 0,... R (5) M N 2 0 N2 ( ) M ( ) 00% 0,... R (6) M N 2 0 where (), = 0, denoes he percen o he run-lengh o he b 0 or beng equal o n all nacve rames, and M 0 () and M () denoe he numbers o he run-lengh o he b 0 or beng equal o n all nacve rames, respecvely. The Mann-Whney-Wlcoxon (M-W-W) es, one o he bes-known non-paramerc sgncance ess, was used o evaluae wheher he derence n run-lengh probably dsrbuons beween he nacve rames and acve rames o a speech le s ndsngushable. To have 95 percen condence,.e. wh a condence coecen (-) o 0.95, where s called he level o sgncance, we hereore requre z( - /2) = z(0.975) =.960, where z s he percenle o he sandard normal dsrbuon. Hence, he sandardzed es sasc z *.960, wo dsrbuons do no der.

11 Table 3 descrbes he dsrbuon paerns o he run-lenghs o 0 and n all nacve rames and acve rames, and he M-W-W es resuls or comparng he probably dsrbuons beween he nacve rames and acve rames or our speech samples, respecvely. Snce z *.960 or all he cases, we conclude ha he probably dsrbuons or boh nacve rames and acve rames do no der, ndcang ha he nacve rames and acve rames had a smlar run-lengh paern or each speech le. To summarse, he above hree expermens on b dsrbuon paerns ndcae ha he b dsrbuon o nacve rames s smlar o ha o acve rames or he same speech les. Oherwse saed, s hghly unlkely o use he /0 dsrbuon paern o dsngush he nacve rames rom acve rames o low b rae audo sreams. Table 3 Run-lengh paerns n nacve and acve rames Probably Frame ype Speech le no () Inacve Acve () Inacve Acve (2) Inacve Acve (2) Inacve Acve (3) Inacve Acve (3) Inacve Acve (4) Inacve Acve (4) Inacve Acve (> 4) Inacve Acve (> 4) Inacve Acve Tes sasc ( z * )

12 D. Seganography n Inacve Frames The source codec lke ITU G.723. s operaed on a rame-by-rame bass. Each rame encoded by G.723. codec has 240 audo samples ha are encoded accordng o PCM. Frs o all, each rame s lered by a hgh-pass ler o remove he DC componen and s hen dvded no our subrames o 60 samples each. A enh order lnear predcve codng (LPC) ler s compued usng he unprocessed npu sgnal or every subrame, and he las subrame s quanzed usng a predcve spl vecor quanzer. For every wo subrames (20 samples), he weghed speech sgnal s used o compue he open-loop pch perod. A harmonc nose shapng ler s hen consruced usng he open-loop pch perod compued prevously, and a closed-loop pch predcor s consruced accordng o he mpulse response creaed by he nose shapng ler. Fnally, boh he pch perod and he derenal value are ransmed o he decoder and he non-perodc componen o he excaon s approxmaed. Aer compleon o hese operaons, all speech parameers such as LPC, Pulse sgn (Pamp) and Pulse poson (Ppos) and so on, are obaned. Accordng o (6) - (0), he speech, F = { A, S = 0,, N, = 0,..., N 2, N = N + N 2 }, s npued no G.723. codec, hen he b sream F* = { *A, *S = 0,..., N, = 0,, N 2, N = N + N 2 } s oupued wh wo ypes o rames, nacve rames and acve rames. Moreover, he b allocaon o he nacve rame s smlar o ha o he acve rame. The b allocaon o G.723. codec wh 6.3kb/s s lsed n Table 4. Table 4 B allocaon o G.723. codec wh 6.3kb/s Parameers Subrame 0 Subrame Subrame 2 Subrame 3 Suboal (bs) Adapve codebook lags (Olp/Aclg) LPC ndces (Ls) Grd ndex (Grd) 4 All he gans combned (Mamp) Pulse posons (Ppos) Pulse sgns (Pamp) Toal

13 The nex sep s o deermne whch speech parameers o nacve rames are suable or daa embeddng. All he speech parameers are sored no hree mpercepbly levels o seganography n erms o he dsance o sgnal o nose rao (DSNR), whch s dened as he derence n sgnal o nose rao (SNR) beween he orgnal speech and sego speech. Close analyss o he daa n Table 5 shows he mpercepbly levels o seganography or deren parameers o he nacve rames are wdely deren. So s possble o choose deren parameers and varous parameer bs o embed daa on demand o praccal applcaons. In shor, he parameers marked wh level -2 are suable cover obecs or seganography. Table 5 Impercepbly levels o seganography n varous parameers o nacve rames Number o bs Olp (s) Ls (s2) Aclg (s3) Grd (s4) Mamp (s5) Ppos (s6) IV. OUR PROPOSED ALGORITHM FOR STEGANOGRAPHY IN INACTIVE FRAMES Our seganography model s llusraed n Fg. 2, where VAD, daa embeddng and audo rame encodng are carred ou sequenally n he speech codng process. The sender samples an audo sgnal and encodes no a PCM ormaed audo sream, F = { = 0,, N}. The VAD algorhm s hen used o deec he nacve voce n he sream. I he curren rame s an nacve voce, he rame s marked wh S; oherwse, s marked wh A. As a resul, he audo sream s dvded no a sequence o rames, F = { A, S = 0,, N, = 0,, N 2, N = N + N 2 }. All he rames are hen encoded unormly by G.723. codec no a low b rae sream, whch s called he orgnal speech, F* = { *A, *S = 0,, N, = 0,, N 2, N = N + N 2 }.

14 The low b rae sream conans wo ypes o rames, nacve rames and acve rames. Accordng o he rame ype, wo deren seganography algorhms are hen used, respecvely, o embed he secre normaon, S = (s, s 2, s, s n ), s (0, ), n he sream. Tha s, he algorhm suggesed below s used o embed normaon n nacve rames; he algorhm 2 presened n [5] s used or seganography n acve rames. The low b rae sream wh hdden ~ ~ ~ normaon s called he sego speech, denoed by F ~,,..., recever receves he sego speech, rom whch he secre normaon s nally exraced. 2, whch s ransmed usng VoIP. Aerwards, he Fg. 2. Flowchar o seganography n nacve and acve rames To sum up, he seganography process has hree sub-processes, voce acvy deecon, daa embeddng and exracng. The correspondng algorhms are dealed below.

15 A. Improved VAD Algorhm Hangover algorhm s normally used or voce acvy deecon n he speech codng process. To synchronse he embeddng and exracon n seganography, s very mporan o keep he VAD resul conssen beween he sender and recever because an nconssen VAD resul wll resul n errors n he exracng process. Some acors, such as packe loss, seganography and so on, may have an mpac on he VAD resul. So an mproved VAD algorhm called he resdual energy mehod s suggesed below. The resdual energy mehod adops he auocorrelaon coecen, whch s no aeced by he sae o he codec, o deec he nacve voce n he speech. The coecen vecor o he FIR ler on (3), A no (Z), s compued by Levson-Durbn algorhm as ollows: Rp[0] Rp[] Rp[0] Rp[] Rp[0] R p[0] Rp[9] a R [9] [0] a p Rp 0 E 0 0 p (7) As (7) reveals, he auocorrelaon sum o he rame, R p [ ], mus be compued n advance by usng (8) n order o oban A no (Z). k3 k R [ ] R [ ] 0,...,0 (8) p where R [] denoes he auocorrelaon o he subrame. As a rame consss o our subrames, each o whch has auocorrelaon coecens, all he auocorrelaon coecens or he rame can be descrbed as R [], = 0,, 0, = 0,, 3. To compue he coecens or he rs

16 subrame, needs o oban he daa o hree connuous subrames. The connuous subrames, (-)h, h, (+)h subrames, can be combned o orm a sequence, ThreeSubFrm, whch conans 80 samples. When =0, he (-)h subrame belongs o he prevous rame. I he predecessor o he curren rame s los, an error wll occur n calculang he auocorrelaon coecens o he curren rame. Ths s because Hangover algorhm has memory and error propagaon would resul rom los or delayed packes. In an aemp o solve hs problem, we sugges an mproved saeless algorhm or compung he auocorrelaon R []. The algorhm s descrbed n deal below. Frs, ThreeSubFrmH (n) s compued hrough wndowng/applyng a Hammng wndow n he sequence o rames, gven by ThreeSubFr mh ( n) ThreeSubFr m ( n) HammWndo w( n), n 0,...,79 (9) Second, he auocorrelaon coecens o he subrame are compued by 79 n R [ n] ThreeSubFrmH ( ) ThreeSubFrmH( n ) (20) where n = 0,,0, whch s he number o auocorrelaon. Thrd, a whe nose s used o adus he rs coecen, R [0], as shown n (2): 025 R [ 0] R [0] (2) 024 And a bnomal wndow s used o adus he oher coecens by means o he ollowng equaon R [ n] R[ n] Bnormal[ n], n 0 (22) Equaon (22) ndcaes ha 80 samples needed or compung R [n] are locaed n wo connuous rames, he prevous rame and he curren rame. As he speech has he shor erm saonary propery, he samples n he prevous rame can

17 be replaced wh one o he curren rame n (9). Thereore, even he prevous rame s los, wll no aec he compuaonal resuls o auocorrelaon coecens or he curren rame. Fnally, once all he auocorrelaon coecens o our subrames are obaned or a rame, he resdual energy o he rame can be compued by Yule-Walker equaon: 0 0 [0] [9] [0] [9] [0] [] [0] [] [0] 0 E a a R R R R R R R R R (23) where R [] ( = 0,, 0) denoes he auocorrelaon coecens, and a ( =,,0) are he LPC coecens. The algorhm or solvng Yule-Walker equaon s descrbed below. Sep, nalsaon: = 0, E = R [], a = 0 and =,, 0 Sep 2, compue k usng (24) E R a R k ] [ ] [ 0 (24) I < 0, hen 0 ] [ 0 R a Sep 3, compue a by a k a k a (25) Sep 4, compue E as ollows: E k E ) ( 2 (26) And = +, < 0, go o Sep 2. Oherwse, he nal resdual energy E yelds; E = E when = 0.

18 A new mehod o deecng acve voces s hen suggesed here, ha s, comparng he hreshold wh he resdual energy o he rame raher han he energy o he rame, as shown n (27) Inacve rame Acve rame E Thr E Thr (27) where E denoes he resdual energy o he curren rame, and Thr denoes he hreshold, whch s an emprcal value obaned rom expermens. The above VAD mehod ha s based on he resdual energy nsead o he rame energy s only relaed o he R [] ( = 0, 0) o he curren rame when he resdual energy s compued. So he mproved VAD mehod s no aeced by packe loss, hereby guaraneeng he VAD resul o be conssen beween he sender and recever. B. Embeddng Algorhm The embeddng process s dvded no our seps as shown n Fg. 2. Sep, voce acvy deecon. The speech wh PCM orma s dvded no rames, F = {,, }. Each rame,, s npued no he VAD deecor ha adops he resdual energy algorhm above. The rame s marked wh A s deermned o be an acve voce; oherwse, he rame s marked wh S. The rames are dened as A S else s an acve voce (28) The sequence o he rames wh marks s hen obaned, gven by F = { A, S = 0,, N, = 0,..., N 2 } (29)

19 Sep 2, encodng all rames by G.723. codec. Regardless o he rame ype, all he rames, A and S, are encoded by usng he sandard G.723. algorhm wh 6.3kbps. The resulng low b rae audo sream conanng acve and nacve rames s hen oupued rom he codec. The low b rae audo sream s expressed as } 0,...,,,..., ) ( ), ( { ) ( } 0,...,,,...,, { 2 * 2 * * * N N F F N N F S A S A (30) Sep 3, embeddng normaon n rames. Accordng o he rame ype, wo deren seganography algorhms are used o embed normaon n he rames. They are expressed as A A S S S S S S * * * * 2 * * * * ), ( ~ ), ( ~ (3) The expresson ( *, S) means he algorhm s used o embed he secre normaon S n he nacve rame. The expresson 2 ( *, S) denoes he algorhm 2 s used o embed he normaon n he acve rame. So he sego speech s gven by F ~...,, ~, ~ ~ 2 (32) Sep 4, encapsulaon and sendng. The nacve rames and acve rames wh hdden normaon are encapsulaed n VoIP packes },..., ), ~ ( { n p p P, whch are ransmed over he Inerne. C. Exracng Algorhm The exracon o secre normaon rom he sego speech s he nverse process o he embeddng algorhm, and s dvded no he ollowng hree seps.

20 Sep, recevng and decapsulaon. The VoIP packes, P = {p, p 2,, p n }, are receved, buered and hen decapsulaed by he recever. The decapsulaon algorhm s descrbed as ~ ~ ~ F { ( p ),,..., n} (33) ~ ~ Sep 2, decodng and acve rame deecon. The buered rames F {,..., n} are coped o he decodng buer and decoded no he PCM ormaed audo sream F ' { ',..., n}. The mproved VAD mehod s hen ' ' A ' S used o dsngush beween nacve rames and acve rames, F,,..., N,,..., N }. { 2 Sep 3, exracng secre normaon. The nacve and acve rames o he low b rae audo ~ ~ ' ' A ' S sream F {,..., n} are dened by reerrng o F,... N,,..., N }. The secre { 2 ~ ~ normaon s hen exraced rom F {,..., n} by usng Algorhm and Algorhm 2. The Algorhms and 2 are used o exrac he secre normaon rom he nacve rames and he acve rames, respecvely. V. PERFORMANCE ANALYSIS OF STEGANOGRAPHY IN INACTIVE FRAMES In our expermens, voce acvy deecon, daa embeddng and audo encodng operaons were conduced n sequence or each speech sample by means o he correspondng algorhms dealed n Secon IV. Two parameers, mpercepbly and daa embeddng capacy, were used o evaluae he perormance o he proposed seganography algorhm. Tweny speech samples les wh PCM orma were employed as cover obecs or seganography, and hey are classed no our groups, Group, Group 2, Group 3 and Group 4 (Table 6). Secre normaon was embedded n he nacve rames o he speech les, he mpercepbly o he resulng sego les

21 was hen evaluaed, and he daa embeddng capacy was esmaed accordngly or each speech le. The expermenal resuls are dscussed n deal below. Table 6 Numbers o nacve rames o 20 PCM speech les Group Speech le name Fle lengh (s) Number o nacve rames (30 ms) Average no o nacve rames MC 0 83 MC2 0 0 Group MC MC MC WC 0 39 WC Group 2 WC WC WC ME 0 69 ME Group 3 ME ME ME WE 0 5 WE Group 4 WE WE WE To very he mpercepbly o seganography n varous parameers o nacve rames, he same secre normaon was embedded n each parameer o he 20 speech les encoded by G.723. codec, and he DSNR values o he resulng sego speech les were hen compued. The DSNR s dened as he derence n sgnal o nose rao (SNR) beween he orgnal speech and he sego speech, gven by DSNR SNR b SNR (34) a where SNR b and SNR a are he sgnal nose raos o he orgnal speech and he sego speech, respecvely. Fg. 3 shows he resuls o expermens on he 20 speech les lsed n Table 6, wh he horzonal axs represenng he number o bs o he parameer ha are replaced by secre normaon. Expermens ndcae ha n mos nsances he

22 DSNR value beween he orgnal speech and he sego speech was so small ha he dsoron o he sego speech was unlkely o be perceved as long as approprae parameer bs o nacve rames were used o embed he secre normaon. The overall rend n DSNR was upward wh ncreasng b numbers o embeddng. The parameers wh DNSR values o less han 0.5dB were chosen o embed normaon. DSNR (db) Ls Olp Mamp Ppos Pamp Grd B numbers o embeddng Fg. 3. DSNR or seganography n varous parameers o nacve rames As shown n Fg. 3, when he b number o hdden normaon n he Ls parameer was no more han 3 bs, he DSNR value was under 0.5 db; however, DSNR rose sgncanly when more han 4 bs o normaon were embedded. Ths means no more han 3 bs o normaon should be embedded n he Ls parameer. For he Olp and Mamp parameers, even replacng b n each subrame wh secre normaon (amounng o 4 bs hdden normaon per rame) resuled n a larger DSNR value, ndcang ha boh parameers are no suable or daa embeddng. By lookng a he DSNR curves n Fg. 3 and he mpercepbly levels o seganography n Table 5, we realsed ha all bs o he Ppos, Pamp and Grd parameers could be used o embed normaon n nacve rames. We hereore seleced ve parameers o nacve rames (Table 7) o carry ou urher seganography expermens.

23 Table 7 Parameers o he nacve rame perecly suable or daa embeddng Parameer name Ls Grd H_Ppos L_Ppos Pamp Toal bs Number o bs As a rame o G.723. wh 6.3 kbps has 92 bs, and he oal number o replaceable parameer bs n an nacve rame s 0 bs, he daa embeddng capacy rao C r or he nacve rame s deermned by C r = Embeddng bs / Toal bs = 0/92 = 52.6% (35) A. Impercepbly Accordng o he mproved VAD algorhm, we couned he number o nacve rames or each speech le (Table 6), and encoded hese les no low b rae sreams usng G.723. codec wh 6.3kbps. Fve parameers o he nacve rame (Table 7) were seleced o embed normaon. We hen evaluaed he mpercepbly o he sego speech les n erms o subecve qualy and obecve qualy. Subecve Qualy The A/B/X es mehod, ITU P.860 recommendaon [2], was ulsed o assess he subecve qualy o he sego speech les. Ths mehod s descrbed n deal as ollows. Suppose here are hree ypes o speech les, denoed by A, B, and X, respecvely. A represens he sego speech le conanng hdden normaon, B denoes he orgnal speech le whou any hdden normaon, and X s eher A or B. Fve evaluaors were employed o lsen he speech les, and hen asked o decde wheher X s A or B. Speech samples were chosen randomly rom he our groups o speech les lsed n Table 6. Each eser made 20 udgmens n oal, some o whch were successul and he oher udgmens were alures. These alure udgmens nclude negave alures and posve alures. Table 8 shows he percenage o alures o deny he sego speech le.

24 Table 8 Percenages o alures usng A/B/X es Group Group 2 Group 3 Group 4 Teser 55% 30% 50% 55% Teser 2 30% 55% 65% 30% Teser 3 60% 30% 60% 35% Teser 4 55% 45% 50% 60% Teser 5 45% 40% 60% 55% Average 49% 40% 57% 47% Close analyss o he daa lsed n Table 8 shows he average percenage o alure udgmens was percen. Ths means was mpossble o dsngush he sego speech rom he orgnal speech by usng he A/B/X esng mehod when secre normaon was embedded n nacve rames. The resuls also ndcae ha he subecve qualy o he proposed algorhm or seganography n nacve rames was close o ha o he orgnal speech. Table 9 Tesng resuls wh ITU P.862 Group Speech le name MOSLQO value Average MOSLQO MC MC Group MC MC MC WC WC Group 2 WC WC WC ME 4.25 ME Group 3 ME ME ME WE 4.43 WE Group 4 WE WE WE

25 We also adoped he ITU P.862 recommendaon o measure he subecve qualy o he sego speech. The recommendaon descrbes an obecve mehod or predcng he subecve qualy o narrow-band speech codecs. I uses he percepual evaluaon speech qualy (PESQ) value o assess he subecve qualy o he sego speech. As he PESQ s no well mached wh mean opnon score (MOS), PESQ-lsenng qualy obecve (LQO) s recommended o evaluae he qualy o he sego speech. The PESQ s hen mapped o he MOSLQO value. The esng resuls wh he ITU P.862 mehod are lsed n Table 9. Accordng o he ITU P.862 sandard, he MOSLQO value o he orgnal speech s equal o 4.5. As shown n Table 9, he average MOSLQO value o he sego speech was esmaed o be So he derence n MOSLQO beween he orgnal speech and he sego speech was so mnor (3.8%) ha dsoron resuled rom seganography n nacve rames was mpercepble. Obecve Qualy To evaluae urher he mpercepbly o he sego speech, we compared he specrum beween he orgnal speech and he sego speech n he requency and me doman. For nsance, he specrums o he MC speech le havng 83 nacve rames wh and whou hdden normaon are shown n Fg. 4. Careul analyss o Fg. 4 shows very lle dsoron occurred n he me doman as a resul o daa embeddng n nacve rames; however, we could no perceve any derences beween he orgnal speech and he sego speech n he requency doman. Ths suggess seganography n nacve rames a a daa embeddng rae o 0 bs/rame had no or very lle mpac on he qualy o he orgnal speech.

26 Fg. 4. Specrum comparsons n he me- and requency-doman The mean cepsrum dsoron (MCD) merc [] was used o measure he obecve qualy o he sego speech. The MCD s dened as MCD N N 0 ln0 2 p k ~ c( ) c( ) 2 (36) where N s he number o audo rames, and c() and c ~ ( ) are he cepsrum coecens o he orgnal speech and he sego speech, respecvely, and p s he order o c(). Table 0 lss he MCD resuls o he sego speech les wh normaon embedded n he nacve rames; each average MCD value n he hrd column s he arhmec mean o MCDs obaned rom sx repeaed expermens on an orgnal speech le. As Table 0 shows, all he MCD values and varances o he sego speech les were relavely small, ndcang ha he proposed seganography algorhm or embeddng normaon n he nacve rames acheved perec mpercepbly.

27 Table 0 MCD values o sego speech les Group Speech le name Average MCD Group Group 2 Group 3 Group 4 MC.769 MC MC3.57 MC4.222 MC5.639 WC.39 WC2.209 WC3.64 WC4.450 WC5.24 ME ME ME ME ME WE WE WE WE WE MCD Mean Varance B. Daa Embeddng Capacy Usng (34), we compued he daa embeddng capacy or each nacve rame. The lengh o he nacve rame encoded by G.723. codec a 6.3kbps was 92 bs; among hem, 0 bs were used o embed normaon. The ollowng paragraphs descrbe how o deermne he average daa embeddng capacy or seganography n he acve rames and nacve rames o low b rae audo sreams, respecvely. Suppose he rame number o he orgnal speech s L, and he number o nacve rames s D, hen he number o acve rames s L-D. And, N bs o he nacve rame, such as N = 0, are used o embed normaon. Meanwhle, M bs o he acve rame are used o embed normaon. The average daa embeddng capacy o he speech le can hen be dened as he daa embeddng rae v n bs per second (bps), gven by v ( D N ( L D) M ) / ( L 92) (37)

28 Several oher algorhms, such as CNV [5], MELP [8], [3], and parameer-lsb [9], [4]-[6], were prevously suggesed or embeddng normaon n low b rae audo sreams encoded by ITU-T G However, hese algorhms are suable or seganography n acve rames only, achevng deren levels o daa embeddng. For comparson purposes, hese prevously suggesed algorhms and our proposed seganography algorhm were adoped o embed daa n he speech sample les lsed n Table 6, respecvely. Fg. 5 shows he comparsons n daa embeddng capacy beween our proposed algorhm (denoed as HF ) and he oher algorhms. 800 Average daa embeddng rae (bps) HF MELP CNV Para.-LSB Algorhms Fg. 5. Comparsons n daa embeddng raes beween he proposed algorhm HF and oher algorhms As Fg. 5 shows, he daa embeddng rae o our proposed algorhm HF was much hgher han hose o he oher algorhms. Ths s because he proposed seganography algorhm made good use o he redundancy n he nacve rames o low b rae audo sreams.

29 I s worh menonng ha he daa embeddng capacy o seganography n nacve rames s lmed by he number o nacve rames o he orgnal speech le. Research ound percen o a VoIP sesson were nacve rames, so seganography n he nacve rames could aan a hgher daa embeddng rae han oher algorhms, whch s n agreemen wh our expermen resuls. VI. CONCLUSIONS In hs paper, we have suggesed a hgh capacy seganography algorhm or embeddng daa n he nacve rames o low b rae audo sreams encoded by G.723. source codec. The expermenal resuls have shown ha our proposed seganography algorhm can acheve a larger daa embeddng capacy wh mpercepble dsoron o he orgnal speech, compared wh oher hree algorhms. We have also demonsraed ha he proposed seganography algorhm s more suable or embeddng daa n nacve audo rames han n acve audo rames. However, beore he proposed algorhm comes no praccal use n cover VoIP communcaons, s necessary o explore how o assure he negry o hdden messages n he case o packe loss, whch shall be he subec o uure work. ACKNOWLEDGMENT Ths work was suppored n par by grans rom he Naonal Naural Scence Foundaon o Chna (nos and ) and he Naonal Basc Research Program o Chna (no. 2007CB30806), and a gran rom Brsh Governmen (no. kp6367). REFERENCES [] Z. Wu, and W. Yang, G.7-based adapve speech normaon hdng approach, Lecure Noes n Compuer Scence, vol. 43, pp , 2006.

30 [2] N. Aok, A echnque o lossless seganography or G.7 elephony speech, n Fourh Inernaonal Conerence on Inellgen Inormaon Hdng and Mulmeda Sgnal Processng (IIHMSP2008), pp , [3] L. Ma, Z. Wu, and W. Yang, Approach o hde secre speech normaon n G.72 scheme, Lecure Noes n Compuer Scence, vol. 468, pp , [4] P. Chang, and H. Yu, Dher-lke daa hdng n mulsage vecor quanzaon o MELP and G.729 speech codng, n Conerence Record o 36h Aslomar Conerence on Sgnals, Sysems and Compuers, vol. 2, pp , Nov [5] B. Xao, Y.F. Huang, and S. Tang, An Approach o Inormaon Hdng n Low B rae Speech Sream, n IEEE GLOBECOM 2008, IEEE Press, pp , Dec [6] Z. Wu, W. Yang, and Y. Yang, ABS-based speech normaon hdng approach, Elecroncs Leers, vol. 39, no. 22, pp , [7] Z. Wu, W. Gao, and W. Yang, LPC parameers subsuon or speech normaon hdng, The Journal o Chna Unversy o Poss and Telecommuncaons, vol. 6, no. 6, [8] C. Kräzer, J. Dmann, T. Vogel, and R. Hller, Desgn and evaluaon o seganography or voce-over-ip, n Proceedngs o IEEE Inernaonal Symposum on Crcus and Sysems, pp , May [9] C. Bao, Y.F. Huang, C. Zhu, Seganalyss o Compressed Speech, n IMACS Mulconerence on Compuaonal Engneerng n Sysems Applcaons (CESA), pp. 5-0, Oc [0] ITU-T Recommendaon G.723. Annex A. Avalable: hp:// [] N. Kawak, H. Nagabuch, and K. Ioh, Obecve qualy evaluaon or low-b-rae speech codng sysems, IEEE Journal on Seleced Areas n Communcaons, vol. 6, no. 2, pp , 988. [2] Z.M. Lu, B. Yan, and S.H. Sun, Waermarkng combned wh CELP speech codng or auhencaon, IEICE Transacons on Inormaon and Sysems, vol. E88-D, no. 2, pp , [3] J. Dmann, D. Hesse, and R. Hller, Seganography and seganalyss n Voce over IP scenaros: Operaonal aspecs and rs experences wh a new seganalyss ool se, In Secury, Seganography, and Waermarkng o Mulmeda Conens VII, Elecronc Imagng Scence and Technology, pp , [4] M.U. Celk, G Sharma, A.M. Tekalp, and E. Saber, Lossless generalzed-lsb daa embeddng, IEEE Transacons on Image Processng, vol. 4, ssue 2, pp , 2005.

31 [5] H. Tan, K. Zhou, Y.F. Huang, D. Feng, and J. Lu, A cover communcaon model based on leas sgncan bs seganography n Voce over IP, n Proceedngs o he 9h Inernaonal Conerence or Young Compuer Scenss, pp , Nov [6] L.Y. Ba, Y.F. Huang, G. Hou, B. Xao, Cover channels based on Jer eld o he RTCP header, n Proc. o IEEE Inernaonal Conerence on Inellgen Inormaon Hdng and Mulmeda Sgnal Processng, pp , 2008.