06301705_(1)

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 6, DECEMBER 2012 1865

Steganography Integration Into a Low-Bit RateSpeech Codec

Yongfeng Huang, Chenghao Liu, Shanyu Tang, Senior Member, IEEE, and Sen Bai

AbstractLow bit-rate speech codecs have been widely used inaudio communications like VoIP and mobile communications, sothat steganography in low bit-rate audio streamswould have broadapplications in practice. In this paper, the authors propose a newalgorithm for steganography in low bit-rate VoIP audio streamsby integrating information hiding into the process of speech en-coding. The proposed algorithm performs data embedding whilepitch period prediction is conducted during low bit-rate speechencoding, thus maintaining synchronization between informationhiding and speech encoding. The steganography algorithm canachieve high quality of speech and prevent detection of steganal-ysis, but also has great compatibility with a standard low bit-ratespeech codec without causing further delay by data embeddingand extraction. Testing shows, with the proposed algorithm, thedata embedding rate of the secret message can attain 4 bits/frame(133.3 bits/second).

Index TermsG.723.1, information hiding, low bit-rate speechcodec, pitch period prediction, VoIP.

I. INTRODUCTION

N OWADAYS people are becoming more and moreconcerned about the security of private informationtransmitted over the Internet. Protecting the private informationfrom being attacked is regarded as one of the major problems inthe field of information security. Apart from encryption, digitalsteganography has been one of the solutions to protecting datatransmission over the network [1].Steganography is the science of covert communications that

conceal the existence of secret information embedded in covermedia over an insecure network. A great effort has been made toexplore the methods for embedding information in cover media,such as plaintext [2], audio files in WAV or MP3 [3], and im-ages with BMP or JPEG format [4]. In recent years, computernetwork protocols and streaming media like Voice over InternetProtocol (VoIP) audio streams were used as cover media to

Manuscript received June 23, 2012; revised August 30, 2012; accepted Au-gust 30, 2012. Date of publication September 12, 2012; date of current versionNovember 15, 2012. This work was supported in part by the National NaturalScience Foundation of China under Grant 61271392 and Grant 61272469. Theassociate editor coordinating the review of this manuscript and approving it forpublication was Prof. Jiwu Huang. (Corresponding author: S. Tang.)Y. Huang is with the Department of Electronic Engineering, Tsinghua Uni-

versity, Qing Hua Yuan, Hai Dian District, Beijing, 100084, China (e-mail:[email protected]).C. Liu and S. Bai are with the Department of Information Engineering,

Chongqing Communication Institute, Chongqing, 400035, China (e-mail:[email protected]; [email protected]).S. Tang is with the School of Computer Science, China University

of Geosciences, Wuhan City, Hubei Province, 430074, China (e-mail:[email protected]).Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TIFS.2012.2218599

embed secret messages [5], [6]. Dittmann et al. [5], for example,suggested the design and evaluation of steganography in VoIP,indicating possible threats as a result of embedding secret mes-sages in such a widely used communication protocol.The methods of speech steganography can be classified into

three categories. The first is the least significant bit (LSB) re-placement/matching method towards the pulse code modulation(PCM) format voice data [3]. The second hides a secret messagein transform domain, firstly transforming the covers data to thetransform domain, and then modifying some parameters in thedomain to embed the secret message, with often used transformincluding the Cepstrum transform [7], discrete cosine transform[8], and so on. The third is the Quantization Index Modula-tion (QIM)-based method firstly proposed by Xiao et al. [9].The QIM hides the secret message by modifying the quantiza-tion vector, which is applicable to various digital media, suchas speech, image and video. It is very suitable to informationhiding in the media compression encoding process.Although some methods have been suggested for speech

steganography, most of which dealt with high bit-rate speechformat like PCM. However, most codecs used in VoIP are thosewith low bit-rate, such as Internet low bit-rate codec (iLBC),G.723.1 and G.729A; this means existing steganographicmethods do not necessarily meet all the requirements of infor-mation hiding in VoIP. Up to now, only little attention has beenpaid to steganography in low bit-rate VoIP audio streams. Forexample, in our preliminary work, we proposed a codebookpartition algorithm called the Complementary Neighbor Vertex(CNV) algorithm for optimally dividing the vector codebookinto two subcodebooks, which are needed by QIM embedding.In general, it is more challenging to embed information in

low bit-rate VoIP streams. The first reason is the requisite forreal-time VoIP communications. Most previous steganographicalgorithms have been designed for embedding data in image oraudio files. These algorithms usually take relatively long time toprocess data embedding. So they are not suitable for steganog-raphy in VoIP streams. Secondly, only a few results have so farproved conventional steganographic algorithms could survivelow bit-rate compression. Finally, data embedding is to replacethe redundancy in the cover media with the secret message; theless the redundancy is, the more difficult information hidingbecomes. Unfortunately, all low bit-rate codecs are based onanalysis by synthesis (AbS) that uses effective methods such aslinear predictive coding (LPC) to eliminate redundancy. So con-ventional steganographic algorithms, i.e., replacing LSBs withthe secret message, are not necessarily suitable for steganog-raphy in low bit-rate VoIP audio streams.To take on these challenges, we propose a new method for

steganography in low bit-rate VoIP audio streams and design

1556-6013/$31.00 2012 IEEE

1866 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 6, DECEMBER 2012

an enhanced speech codec to integrate the information hidingfunction.The rest of the paper is organized as follows. In Section II,

related work is briefly introduced. Section III describes thepitch period prediction method in the hybrid speech codec.Section IV presents a new pitch period prediction-based al-gorithm for steganography in low bit-rate VoIP streams, andan enhanced speech codec combined with information hiding.Experimental results are discussed in Section V. Finally,Section VI concludes with a summary and directions for futurework.

II. RELATED WORKOver the past few years, a number of attempts have beenmade

to study steganography in low bit-rate audio streams. Some re-lated works are introduced below.Several MP3stego, AAC-based audio steganographic sys-

tems have been suggested in recent years [10][12]. Wang et al.[1] proposed a scheme to convey secret messages by embeddingthem in VoIP streams. The scheme divides the steganographyprocess into two steps, compressing the secret message andembedding its binary bits into the LSBs of the cover speechencoded by G.711 codec. Dittmann et al. [5] presented a moregeneral scheme for steganography in VoIP, which can be usedfor transmitting an arbitrary secret message. More recently,Huang and coworkers [7] suggested an M-Sequence basedLSB steganographic algorithm for embedding information inVoIP streams encoded by G.729A codec. With their algorithm,embedding data in a speech frame takes less than 20 us onaverage, which is negligible in comparison with the allowablecoding time of 15 ms for each frame in VoIP. In addition,Huang et al. [6] suggested an algorithm for embedding datain some parameters of the inactive speech frames encodedby G.723.1 codec. However, this algorithm is also based onthe LSB substitution of encoded audio streams. Therefore,the algorithms above would lead to obvious distortion, whichaffects the quality of steganographic speech.Xiao suggested a QIM-based steganography in low bit-rate

speech while encoding [9]. The QIM method randomly dividesthe whole codebook into two parts, each colored with white orblack. When a secret bit of 0 is embedded, the white code-word is used; the black codeword is used when a secret bit of1 is embedded. On the receiving side, the hidden bit is ex-tracted by checking which part of the codebook the codewordbelongs to. It is the first attempt to perform steganography andcompression operation in the same codec. However, this infor-mation hiding algorithm has a small hiding capacity, which isno use in practice.Our work described in this paper is the first ever effort to

explore a novel method for steganography in low bit-rate speechbased on pitch period prediction while the speech is encoded.The steganographic algorithm can not only achievemuch higherdata hiding capacity than the QIM algorithm [9], but also assurea good quality of speech.

III. PITCH PERIOD PREDICTION IN HYBRID SPEECH CODECAs pitch period prediction is required in almost all speech

analysis-synthesis (vocoder) systems, the pitch period predictoris an essential component in all speech codecs of low bit-rate.

Because of the importance of pitch period prediction, a varietyof algorithms for pitch period prediction have been proposed inthe speech processing literature [13][15]. However, accuratepredictions about the pitch period of a speech signal from theacoustic pressure waveform alone is often exceedingly difficultdue to the reasons below.1) The glottal excitation waveform is not a perfect train ofperiodic pulses. Although finding the period of a perfectlyperiodic waveform is straightforward, predicting the pe-riod of the speech waveform can be quite difficult, as thespeech waveform varies both in period and in the detailedstructure of the waveform within a period.

2) The interaction between the vocal tract and the glottal exci-tation also makes pitch period prediction difficult. In someinstances, the formants of the vocal tract can significantlyalter the structure of the glottal waveform, so that the ac-tual pitch period is unlikely to predict. Such an interac-tion is most deleterious to pitch period prediction duringfast movements of articulators while the formants are alsochanged rapidly.

3) The problem of accurately predicting the pitch period isthe inherent difficulty in defining the exact beginning andend of each pitch period during voiced speech segments.Choosing the beginning and ending locations of the pitchperiod is often quite arbitrary. The pitch period discrep-ancies are arisen from the quasi-periodicity of the speechwaveform, but also the fact that peak measurements aresensitive to the formant structure during the pitch period,whereas zero crossings of the waveform are sensitive to theformants, noise, and any DC level in the waveform.

4) Another difficulty of pitch period prediction is how to dis-tinguish between unvoiced speech and low-level voicedspeech. In many cases, transitions between unvoicedspeech segments and low-level voiced speech segmentsare very subtle, and so they are extremely hard to pinpoint.

Apart from the difficulties in measuring the pitch period dis-cussed above, pitch period prediction is also impeded by otherfactors. Although it is difficult to predict the pitch period, anumber of sophisticated algorithms have been developed forpitch period prediction. Basically, algorithms for pitch periodprediction can be classified into three categories. The first cate-gory mainly utilizes the time-domain properties of speech sig-nals, the second category employs the frequency-domain prop-erties of speech signals, and the third category uses both thetime- and frequency-domain properties of speech signals. Mostlow bit-rate speech encoders, such as ITU G.723.1 and G.729A,adopt the first type of algorithms. As an example, the pitch pe-riod prediction algorithm of ITU G.723.1 is introduced below.ITU-T G.723.1 encoder operates on frames of 240 samples

each, a speech frame is denoted by ,equal to 30 ms at an 8 kHz sampling rate. Each frame is dividedinto four subframes of 60 samples each. After accomplishinga series of processes, the input signal of a frame is con-verted to the weighted speech signal .For every two subframes (120 samples), the open loop pitch pe-riod, , is computed using the weighted speech signal .The pitch estimation is performed on blocks of 120 samples. Thepitch period is searched in the range from 18 to 142 samples.Two pitch estimations are computed for every frame, one for

HUANG et al.: STEGANOGRAPHY INTEGRATION INTO A LOW-BIT RATE SPEECH CODEC 1867

the first two subframes and the other for the last two. The openloop pitch period estimation, , is computed using the per-ceptually weighted speech . A cross correlation criterion,namely , calculated by using the maximization method[13], is used to determine the pitch period, as shown in (1).

(1)

The index which maximizes the cross correlation, ,is selected as the open loop pitch estimation for the appropriatetwo subframes.While searching for the best index, preference isgiven to smaller pitch periods to avoid choosing pitch multiples.Maximums of are searched for beginning with .For every maximum found, its value is compared to thebest previous maximum found, . The following pseudocode shows how it works:

if

then (if

then (select , )

)

else (if

then (select , )

)

Using the pitch period estimation, , a closed loop pitchpredictor is computed. The pitch predictor in G.723.1 is afifth order pitch predictor. The pitch prediction contributionis treated as a conventional adaptive codebook contribution.For subframes 0 and 2, the closed loop pitch lag is selectedfrom around the appropriate open loop pitch lag in the range of. For subframes 1 and 3, the closed loop pitch lag is coded

differentially using 2 bits and may differ from the previoussubframe lag only by , 0, or [10].

IV. PITCH PERIOD PREDICTION-BASED STEGANOGRAPHYALGORITHM

A. Embedding Algorithm

In the process of G.723.1 encoding, the open-loop pitchestimation is conducted first, followed by closed-loop pitchprediction. The open-loop pitch estimation computes theopen-loop pitch period of a frame of speech signal

. For each frame, two pitch periodsare computed by using the first two subframes and the lasttwo subframes, respectively. The method for computing theopen-loop pitch period is described below.First, a cross correlation criterion is computed by using

(1), and then it searches for the open-loop pitch following theprocedures below [13]:1) Suppose , , ;

2) Using (1), compute . If

(2)

and

(3)

then , and .3) Set , if , return to 2), otherwise stop.Having obtained the pitch period of a frame of speech

signal , search for the closed-loop pitchperiod and embed information.The closed-loop pitch period of a subframe is defined by ,

, 1, 2, 3, and its open-loop pitch period is ,, 1, representing the open-loop pitch periods of the first twosubframes and the last two subframes, respectively. Adjusting

yields

(4)

The closed-loop pitch period is assigned a value closeto the open-loop pitch period . The values for oddsubframes and for even subframes are obtained from differentranges as shown in (5).

(5)

The minimum value of is 17, and its maximum is 143. Thenumber of is equal to the number of elements in , denotingby . represents the th element in ,

.The pitch prediction contribution is treated as a conventional

adaptive codebook contribution. For subframes 0 and 2, theclosed loop pitch lag is selected around the appropriate openloop pitch lag in the range and coded using 7 bits. For sub-frames 1 and 3, the closed loop pitch lag is coded differentiallyusing 2 bits and may differ from the previous subframe lag onlyby , 0, or [13]. The quantized and decoded pitch lagvalues are referred to as from this point on. The pitch pre-dictor gains are vector quantized using two codebooks with 85or 170 entries for the high bit rate and 170 entries for the lowbit rate. The 170 entry codebook is the same for both rates. Forthe high rate, if is less than 58 for subframes 0 and 1 or ifis less than 58 for subframes 2 and 3, then the 85 entry code-

book is used for the pitch gain quantization. Otherwise, the pitchgain is quantized using the 170 entry codebook. We studied thepitch distribution probabilities of closed-loop pitch period of un-touched G.723.1 VoIP speeches, and Fig. 1 shows the pitch dis-tribution probability results for four types of untouched G.723.1VoIP speeches, each with 250 samples.

1868 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 6, DECEMBER 2012

Fig. 1. Pitch distribution probabilities of four types of untouched G.723.1 VoIPspeech samples.

TABLE IDATA EMBEDDING AT DIFFERENT EMBEDDING BIT-RATES

In search for the closed-loop pitch period, data embedding isaccomplished by adjusting the searching range of the pitchprediction of a subframe according to the secret bit informa-tion to be embedded. For instance, if the secret information to beembedded is 0, the subframe search is performed on the evenelements in ; if the secret information is 1, the odd elementsin are searched. In G.723.1, each frame has four sub-frames, , all subframesrequire searching for the closed-loop pitch, so that data embed-ding can be performed on part of or all subframes. Therefore, wepropose a series of solutions for steganography at four differentembedding bit-rates, as shown in Table I, while the 15 strate-gies are randomly selected, the average data embedding rate isaround 2.1 bits/frame, not 4 bits/frame.On the basis of the steganography solutions listed in Table I,

a new data embedding algorithm is proposed below.

Step 1: Step 0: generate a random , ,then choose a steganography solution accordingto and Table I.

Step 2: Step 1: according to , decide the embeddingbit-rate and where to embed the secret bit stream

, i.e., which i is the subframe inthe frame, .

Step 3: Step 2: suppose the bit in the bit stream isembedded in the subframe of the frame, data embedding is conducted by using the

following algorithm.Step 4: Step 3: if , then data are embedded in the

subframe of the frame, i.e., the pitchperiod of the subframe is searched upon.

ifif

ifif (6)

ifif

ifif (7)

If , then data are embedded in the subframeof the frame, i.e., the pitch period of the subframe issearched upon .

ifif

ifif

ifif

ifif (8)

Step 4: repeat Step 3 until the completion of data embeddingof the secret message .For steganography using the data embedding algorithm

above, errors in predicting speech pitch periods can be esti-mated in theory. As G.723.1 samples at 8 KHz, analysis ofthe closed-loop pitch period prediction shows data embeddingwould lead to one sampling-point error. So the absolute error

in predicting pitch period caused by data embeddingcan be computed by

(9)

If the pitch period is , the maximum of is26.144 Hz, and the relative error is 5.882%;If the pitch period is , the maximum of is

0.394 Hz, and the relative error is 0.699%.Therefore, the error in pitch frequency as a result of adjusting

pitch prediction is proportional to the pitch frequency of speech

HUANG et al.: STEGANOGRAPHY INTEGRATION INTO A LOW-BIT RATE SPEECH CODEC 1869

signal, but the error has a little impact on speech synthesis, par-ticularly for those speech signals with lower pitch frequency. Inthe literature [15], the average error of the most advanced algo-rithms for predicting pitch periods is found to be samples,indicating that the pitch period prediction error arising from thedata embedding algorithm is within the normal range.

B. Extracting Algorithm

The sender embeds the secret message in the low bit-ratespeech streams encoded by G.723.1, and the bit streams con-taining the message are then sent to the receiver who extractsthe secret message following the algorithm below.

Step 1: Step 1: using a negotiating mechanism, thereceiver acquires the data embedding algorithm(steganography solution) for the current speechframe .

Step 2: Step 2: compute the pitch periodsof four subframes , , , ofthe speech frame f[m] decoded by G.723.1.

Step 3: Step 3: according to the data embedding algorithm, decide which of the four subframes ,, , contains the secret message,

and determine the bits of the message using thefollowing formula

(10)Step 4: Step 4: repeat Step 3 until completion of decoding all

speech frames, following by the bit streams of thesecret message to be convertedto the secret message .

C. Design of the Coder With SteganographyA joint information embedding and lossy compression

method is suggested in the literature [16], but no attemptshave been made to study data embedding integrating into lowbit-rate speech encoding. By using a data embedding algorithmbased on pitch period prediction, we here develop the G.723.1low bit-rate speech codec with data embedding functionality,i.e., the embedding and extracting of the secret message areintegrated into G.723.1 speech codec.To achieve data embedding while encoding in G.723.1, our

specially designed secret information preprocessing module,steganography solution selecting module, updating module,and secret information bit stream framer module are insertedinto a normal G.723.1 speech coder, as shown in Fig. 2. Thepitch period prediction module in the codec is also modifiedso as to enable search for the closed-loop pitch upon the pitchperiod updating set, thus realizing data embedding. Similarly,in order to achieve secret data extraction, the novel pitch periododd-even deciding module, steganography solution selectingmodule, secret data extraction module, and secret informationpostprocessing module are built into the G.723.1 decoder, asshown in Fig. 3. Fig. 2 illustrates information embedding inte-grating into G.723.1 coder, whereas Fig. 3 shows informationextraction along with G.723.1 decoding.In the process of information embedding and speech en-

coding, the secret message are compressed

Fig. 2. G.723.1 coder with information embedding.

Fig. 3. G.723.1 decoder with information extraction.

to form the secret data bit stream , whichis divided into segments according to the data embeddingalgorithm. The secret segments are then embedded into speechstreams by adjusting pitch period prediction.In the process of speech decoding and information extraction,

G.723.1 decoder computes the pitch period of a subframe ,, 1, 2, 3, in the current frame , decides the odd-even

nature of the pitch period of the subframe by using the pitchperiod odd-even deciding module, determines the hidden databit according to the odd-even nature of and the steganog-raphy solution . The hidden data bit is then used to extractthe secret information, , by using the secret informationpostprocessing module.

1870 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 6, DECEMBER 2012

Fig. 4. Comparisons of time-domain amplitude lots of a 3-second CSM sampleat different embedding bit-rates.

V. RESULTS AND DISCUSSION

A. Test Samples and Conditions

To evaluate the performance of the proposed steganographicalgorithm, we employed different speech sample files with PCMformat as cover media for steganography to conduct experi-ments. The speech samples are classified into four groups, Chi-nese SpeechMan (CSM), Chinese SpeechWoman (CSW), Eng-lish Speech Man (ESM), and English Speech Woman (ESW).Each group contains 100 pieces of speech samples with lengthof 3 seconds, and 100 pieces of 10-second speech samples, andthe four groups total 800 speech samples. Each speech samplewas sampled at 8000 Hz and quantized to 16 bits, and saved inPCM format. Those speech samples with length of 3 seconds aredefined as the Sample-3 sample set; the Sample-10 samplecontains 10-second speech samples.In our experiments, ITU G.723.1 codec operated at 6.3 kbps,

without silence compression. Fifteen solutions for data embed-ding proposed in Table I were used to conduct steganographyat four different embedding bit-rates (1 bit/frame, 2 bits/frame,3 bits/frame, and 4 bits/frame). Secret data were embedded intoeach audio frame by randomly choosing different embeddingbit-rates and steganography solutions at equal probability.

B. Results and Analysis

Fig. 4 shows comparisons of the time-domain amplitudespectrum of an original 3-second CSM sample with those of thestego 3-second CSM samples at four different data embeddingbit-rates. Almost no distortion occurred in the time domain as aresult of data embedding in the speech sample; no differencesbetween the original speech sample and the stego speech sam-ples in the time-domain spectrum were perceived, indicatingthat our proposed steganography algorithm had no or very littleimpact on the quality of the original speech.We used the perceptual evaluation speech quality (PESQ)

value to assess the subjective quality of the stego speech sam-ples. Figs. 5 and 6 shows the PESQ values for the originalspeech samples after G.723.1 codec without any data embed-ding and the stego speech files processed by G.723.1 with dataembedding by means of the proposed steganography algorithm(detailed in Section IV), when the 3-second and the 10-secondspeech samples were used as cover media, respectively. The

Fig. 5. PESQ values for 3-second samples using the proposed steganographyalgorithm.

Fig. 6. PESQ values for 10-second samples using the proposed steganographyalgorithm.

black curves are the PESQ values for the original speech sam-ples without data hiding. Steganography was carried out at fourdifferent data embedding bit-rates (red curve: 1 bit/frame, greencurve: 2 bits/frame, blue curve: 3 bits/frame, navy curve: 4 bits/frame,). As Figs. 5 and 6 show, for the two types of speech covermedia, the variations in PESQ between the original speech filesand the stego speech files were so small, which means the pro-posed steganography algorithm has little effect on PESQ.Figs. 7 and 8 show comparisons of PESQ values between

using the proposed steganography algorithm and using theCNV algorithm (yellow curve) presented in the literature [9]for 3-second samples and 10-second samples, respectively.There were no obvious discrepancies in the PESQ value

HUANG et al.: STEGANOGRAPHY INTEGRATION INTO A LOW-BIT RATE SPEECH CODEC 1871

Fig. 7. Comparisons of PESQ values for 3-second samples between using theproposed steganography algorithm and using the CNV algorithm [9].

Fig. 8. Comparisons of PESQ values for 10-second samples between using theproposed steganography algorithm and using the CNV algorithm [9].

without (black curve: no hiding) and with data embedding attwo different embedding bit-rates (blue curve: 3 bits/frame,navy curve: 4 bits/frame). As Figs. 7 and 8 show, the variationsin PESQ between the original speech files and the stego speechfiles were so small, indicating that the proposed informationhiding along with speech compression encoding had no or verylittle impact on the quality of the synthesized speech.Tables IIV list the PESQ values for the original speech sam-

ples and the stego speech files obtained by using the proposedsteganography algorithm, when the 3-second and the 10-second

TABLE IIPESQ STATISTICS AT 1 BIT/FRAME DATA EMBEDDING BIT-RATE

TABLE IIIPESQ STATISTICS AT 2 BITS/FRAME DATA EMBEDDING BIT-RATE

TABLE IVPESQ STATISTICS AT 3 BITS/FRAME DATA EMBEDDING BIT-RATE

speech samples were used as cover media, respectively. Thestatistical results were obtained for steganography experimentsconducted at four different data embedding bit-rates. The PESQvalues ranged from 2.9 to 4.1. On average, data hiding hadless effect on the PESQ values of the male speech samplesthan the female speech samples. This is probably due to thefact that the pitch frequency of female speech has a greaterrange, and changes more quickly than male speech. Analysisof Tables IIV shows, as the data embedding bit-rate increases,the average worsening change in PESQ increasesfor 3s sam-ples, ; for 10s samples,

. The maximum of theaverage worsening change in PESQ is 0.50%, and the average

1872 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 6, DECEMBER 2012

TABLE VPESQ STATISTICS AT 4 BITS/FRAME DATA EMBEDDING BIT-RATE

TABLE VIPESQ STATISTICS USING THE STEGANOGRAPHY ALGORITHM PRESENTED IN [9]

TABLE VIICOMPARISONS OF CHANGES IN PESQ BETWEEN THE PROPOSEDSTEGANOGRAPHY ALGORITHM AND THE ONE PRESENTED IN [9]

change in PESQ is within the standard error in PESQ for thespeech samples without data hiding. This also means data hidinghas a negligible effect on PESQ.Table VI lists PESQ statistical results for the stego speech

files obtained by using the steganography algorithm presentedin [9], with cover media having the lengths of 3 and 10 seconds.Similarly, data embedding with the proposed algorithm led toa small change in PESQ, and the average change in PESQ isalso within the standard error in PESQ for the speech sampleswithout data hiding. However, the previous steganography algo-rithm [9] resulted in a larger change in PESQ than our proposedalgorithm, and so it had a slightly high impact on PESQ.Table VII lists comparisons of changes in PESQ between the

proposed steganography algorithm and the CNV algorithm pre-sented in [9]. At the same embedding bit-rate with 3-second

TABLE VIIIDIFFERENCES IN PESQ BETWEEN NORMAL EN- AND DECODING AND DATA

HIDING USING DIFFERENT ALGORITHMS

speech samples, the overall average standard error for the stegospeech files using the proposed steganography algorithm was1.60%, 4.04% less than the CNV algorithm, with both algo-rithms leading to 0.96% change in PESQ; for 10-second speechsamples, the average worsening changes in PESQ of CSM andCSW with the proposed algorithm were smaller, those of ESMand ESW were bigger, the overall worsening change in PESQwas 0.05% larger, and the standard error (0.84%) was 0.02%larger in comparison with CNV. With the embedding bit-ratereaching 4 bits/frame, the average worsening change in PESQof 3-second speech samples with the proposed algorithm was0.26% larger, and the overall standard error (1.61%) was 4.03%smaller compared with CNV; for 10-second speech samples, theaverage worsening change in PESQ was 0.33% larger, and theoverall standard error (0.90%) was 0.08% bigger than CNV.Table VIII lists differences in PESQ between normal en-

and decoding and data hiding using different algorithms. Whenusing the proposed steganography algorithm, the average wors-ening change in PESQ and the standard error of both 3s and10s speech samples were within the range of the standard errorof normal en- and decoding. For the algorithm presented in[9], this was the case for the 10s speech samples only. In com-parison with the previous algorithm, the proposed algorithmhad less impact on PESQ at lower data embedding bit-rates;when the data embedding bit-rate increased to 4 bits/frame, theaverage worsening change in PESQ was 0.295% larger, andthe overall average standard error was 1.975% less than theprevious algorithm.To evaluate the security of the proposed steganography algo-

rithm, we employed the latest steganalysis method [17][20],which uses Derivative Mel-Frequency Cepstral Coefficients(DMFCC)-based Support Vector Machine (SVM) to detectaudio steganography. SVM set RBF core function as its defaultparameter.The test samples used were 501 CSM samples (300 as

training samples, and 201 as test samples), 533 CSW samples(300 as training samples, and 233 as test samples), 819 ESMsamples (600 as training samples, and 219 as test samples), 825ESM samples (600 as training samples, and 225 as test sam-ples), and Hybrid samples containing CSM, CSM, ESM andESW samples. These five sorts of speech samples were usedas the cover media in which data embedding at 4 bits/frametook place by using the proposed steganography algorithm and

HUANG et al.: STEGANOGRAPHY INTEGRATION INTO A LOW-BIT RATE SPEECH CODEC 1873

TABLE IXSTEGANALYSIS RESULTS OF THE LITERATURE [6] ALGORITHM USINGDMFCC AT DIFFERENT DETECTION WINDOWS (DATA EMBEDDING

RATE OF 3 BITS/FRAME)

TABLE XSTEGANALYSIS RESULTS OF THE PROPOSED ALGORITHM USING DMFCC ATDIFFERENT DETECTION WINDOWS (DATA EMBEDDING RATE OF 3 BITS/FRAME)

the one presented in [6]. The steganalysis results are listed inTables IX and X.In the experiments, we used LIBSVM Version 3.0 [21]. In

the SVM-scale of LIBSVM, the lower is , the upper is 1, andthe other parameters used are default values. In the SVM-trainof LIBSVM, the svm_type is C-SVC, the kernel_type is RBF(radial basis function), the cost is 1000, the epsilon is 0.00001,and the other parameters used are default values.As Table IX shows, when the detection window length was

150 frames, the accuracy of DMFCC in detecting steganographyusing the algorithm suggested in [6] reached 80% for all thefive types of speech samples, and increased further to over 90%at detection window length of 300 frames. This indicates thatDMFCC is very effective in detecting the old steganographyalgorithm [6].Table X shows the accuracy of DMFCC in detecting

steganography with the proposed algorithm barely achieved53% for five types of speech samples, with the maximumaccuracy up to 56%, indicating that the proposed steganog-raphy algorithm is unlikely to be detected by DMFCC audiosteganalysis.We also adopted the latest DMFCC audio steganalysis,

Second-order derivative-based Markov approach for audiosteganalysis [22], [23], to detect VoIP steganography with theproposed steganographic algorithm, and the results are pre-sented in Table XI. As Table XI shows, the average accuracyof Markov-DMFCC steganalysis in detecting steganographywith the proposed algorithm just reached 51% for five differenttypes of speech samples, with the maximum accuracy up to

TABLE XISTEGANALYSIS RESULTS OF THE PROPOSED ALGORITHM USING THEMARKOV-DMFCC APPROACH [22], [23] AT DIFFERENT DETECTION

WINDOWS (DATA EMBEDDING RATE OF 3 BITS/FRAME)

Fig. 9. Comparisons of steganalysis results of two algorithms using DMFCCat different detection window lengths.

54%, which means the proposed steganographic algorithm isunlikely to be detected by Markov-DMFCC steganalysis. Thiswas probably due to the ineffectiveness of Markov-DMFCCsteganalysis through analyzing Markov transition features, indetecting the proposed steganographic algorithm, which usesthe pitch lag parameters substitution.Fig. 9 shows comparisons of steganalysis results of two al-

gorithms using DMFCC at different detection window lengthswhen Hybrid speech samples were used as cover media. As thedetection window length increased, the accuracy of DMFCCin detecting the steganography algorithm presented in [6] im-proved significantly; the detection accuracy attained 90% whenthe detection window length reached 200 frames. By contrast,DMFCC was not effective in detecting the proposed steganog-raphy algorithm at different detection window lengths.Fig. 10 shows the pitch distribution probabilities of G.723.1

VoIP samples (duration of 20 seconds) without and with dataembedding. No obvious changes in the statistical propertyof the closed-loop pitch periods in the speech samples afterG.723.1 codec without or with data embedding had been found

1874 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 6, DECEMBER 2012

Fig. 10. Pitch distribution probabilities of G.723.1 VoIP samples (duration of20 seconds) without and with data embedding.

TABLE XIISTEGANALYSIS RESULTS OF THE PROPOSED ALGORITHM USING

SVM AT DIFFERENT DETECTION WINDOWS

for four types of VoIP audio samples, indicating that the pro-posed steganographic system retains the statistical property oforiginal closed-loop pitch periods.We carried out extra steganalysis experiments. As our

proposed steganographic algorithm is based on pitch periodprediction, pitch statistical characteristic-based steganalysiswas specially designed in a way that suppose eavesdroppersknow our steganographic algorithm (Kerckhoffs-compliant),with VoIP samples of 3s, 5s, 10s, 20s and 30s in length with andwithout steganography being available, through analyzing pitchlag of VoIP samples with and without steganography eaves-droppers obtained the first-order pitch statistical characteristics,which were classified by using SVM (similar to DMFCC de-tection method in setup), and the detection results are presentedin Table XII. As the table shows, at five different detectionwindow lengths, the accuracy in detecting steganographywas below 70%, indicating that our proposed steganographicalgorithm is capable of standing against steganalysis.

VI. CONCLUSIONSIn this paper, we have proposed a new method for steganog-

raphy in low bit-rate VoIP streams based on pitch periodprediction. On the basis of ITU G.723.1, a widely used lowbit-rate speech codec, we have developed a much-improvedG.723.1 speech codec with the information hiding function-ality. Fifteen solutions for steganography have been suggestedto perform on VoIP speech samples at four data embeddingbit-rates taking into account the characteristics of G.723.1. The

experimental results have shown that the worsening changein PESQ of the stego speech files obtained by using the pro-posed steganography algorithm was within 1.2%, indicatinglittle impact on the quality of speech. In comparison with aprevious algorithm [9], the proposed steganography algorithmhas been found to have slightly larger effect on PESQ for 3sspeech samples, but have less effect for 10s speech samplesat 3 bits/frame data embedding rate; the worsening changein PESQ was 0.298% higher as the data embedding bit-ratereaching 4 bits/frame (33.3% increase than the old algorithm).Steganalysis tests using DMFCC-SVM have shown that theproposed steganography algorithm could prevent from beingdetected by steganalysis. Investigation into the applicability ofthe proposed algorithm to other low bit-rate speech codecs shallbe the subject of future work. The steganalysis performancewith different classifiers such as Fishers linear classifier andlogistic regression shall be part of future work.

REFERENCES[1] C. Wang and Q. Wu, Information hiding in real-time VoIP streams,

in Proc. 9th IEEE Int. Symp. Multimedia, Taichung, Taiwan, 2007, pp.255262.

[2] S. Zander, G. Armitage, and P. Branch, A survey of covert channelsand countermeasures in computer network protocols, IEEE Commun.Surveys Tutorials, vol. 9, no. 3, pp. 4457, 3rd Quarter, 2007.

[3] N. Aoki, A technique of lossless steganography for G.711 telephonyspeech, inProc. 4th Int. Conf. Intelligent InformationHiding andMul-timedia Signal Processing (IIHMSP2008), 2008, pp. 608611.

[4] P.-C. Su and C.-C. J. Kuo, Steganography in JPEG2000 compressedimages, IEEE Trans. Consumer Electron., vol. 49, no. 4, pp. 824832,Nov. 2003.

[5] J. Dittmann, D. Hesse, and R. Hillert, Steganography and steganalysisin voice over IP scenarios: Operational aspects and first experienceswith a new steganalysis tool set, in Proc. SPIE Security, Steganog-raphy, and Watermarking of Multimedia Contents VII, Mar. 2005, vol.5681, pp. 607618.

[6] Y. F. Huang, S. Tang, and J. Yuan, Steganography in inactive framesof VoIP streams encoded by source codec, IEEE Trans. Inf. ForensicsSecurity, vol. 6, no. 2, pp. 296306, Jun. 2011.

[7] Y. Su, Y. Huang, and X. Li, Steganography-oriented noisy resistancemodel of G.729a, in Proc. 2006 IMACS Multi-Confe. ComputationalEngineering in Systems Applications, Beijing, China, 2006, pp. 1115.

[8] L. Liu, M. Li, Q. Li, and Y. Liang, Perceptually transparent infor-mation hiding in G.729 bitstream, in Proc. 4th Int. Conf. IntelligentInformation Hiding and Multimedia Signal Processing, Harbin, China,2008, pp. 406409.

[9] B. Xiao, Y. Huang, and S. Tang, An approach to information hidingin low bit-rate speech stream, in Proc. 2008 IEEE Global Telecommu-nications Conf., New Orleans, LA, 2008, pp. 15.

[10] D. Yan, R. Wang, and L. Zhang, Quantization step parity-basedsteganography for MP3 audio, J. Fund. Inform., vol. 97, no. 12, pp.114, 2009.

[11] Fabien Petitcolas Mar. 28, 2012 [Online]. Available: http://www.petit-colas.net/fabien/steganography/mp3stego/

[12] M. Sheikhan, K. Asadollahi, and R. Shahnazi, Improvement of em-bedding capacity and quality of DWT-based audio steganography sys-tems, World Appl. Sci. J., vol. 10, no. 12, pp. 15011507, 2010.

[13] ITU.ITU-T Recommendation G.723.1. Dual Rate Speech Coderfor Multimedia Communication Transmitting at 5.3 and 6.3 kbit/s1996 [Online]. Available: http://www.itu.int/rec/T-REC-G.723.1-200605-I/en

[14] ITU.ITU-T Recommendations G.729. Coding of Speech at 8 kbit/sUsing Conjugate-Structure Algebraic-Code-Excited Linear-Prediction(CS-ACELP) 2007 [Online]. Available: http://www.itu.int/rec/T-REC-G.729/e

[15] R. P. Ramachandran and P. Kabal, Pitch prediction filters in speechcoding, IEEE Trans. Acoustics Speech Signal Process., vol. 37, no. 4,pp. 467478, Apr. 1989.

[16] A. Maor and N. Merhav, On joint information embedding and lossycompression, IEEE Trans. Inf. Theory, vol. 51, no. 8, pp. 29983008,Aug. 2005.

HUANG et al.: STEGANOGRAPHY INTEGRATION INTO A LOW-BIT RATE SPEECH CODEC 1875

[17] Y. Huang, S. Tang, C. Bao, and Y. J. Yip, Steganalysis of compressedspeech to detect covert voice over Internet protocol channels, IET Inf.Security, vol. 5, no. 1, pp. 2632, 2011.

[18] Q. Liu, A. H. Sung, andM. Qiao, Temporal derivative-based spectrumand mel-cepstrum audio steganalysis, IEEE Trans. Inf. Forensics Se-curity, vol. 4, no. 3, pp. 359368, Sep. 2009.

[19] Y. Huang, S. Tang, and Y. Zhang, Detection of covert voice over In-ternet protocol communications using sliding window-based steganal-ysis, IET Commun., vol. 5, no. 7, pp. 929936, 2011.

[20] H. Yong-feng, Y. Jian, and C. Mingchao, Key distribution in thecovert communication based on VoIP, Chinese J. Electron., vol. 20,no. 2, pp. 357361, 2011.

[21] C.-C. Chang and C.-J. Lin, LIBSVM: A Library for Support VectorMachines [DB/OL] Oct. 22, 2011 [Online]. Available: http://www.csie.ntu.edu.tw/~cjlin/libsvm/

[22] Q. Liu, A. H. Sung, and M. Qiao, Novel stream mining for audio ste-ganalysis, in ACMMultimedia. NewYork: ACM, 2009, pp. 95104.

[23] Q. Liu, A. H. Sung, and M. Qiao, Derivative-based audio steganal-ysis, ACM Trans. Multimedia Comput., Commun., Appl., vol. 7, no.3, pp. 18:118:9, 2011.

Yongfeng Huang received the Ph.D. degree in com-puter science and engineering from Huazhong Uni-versity of Science and Technology in 2000.He is an Associate Professor in the Department

of Electronic Engineering, Tsinghua University, Bei-jing. His research interests include VoIP, P2P IP TV,multimedia network security, and next-generation In-ternet. He has published five books and over 70 re-search papers on computer network and multimediacommunications. As one of the principal researchers,he has designed and constructed the China Education

and Research Network (CERNET), which is the second largest computer net-work in China.Dr. Huang is the principal/joint grant holder of ten externally funded research

projects.

Chenghao Liu received the B.E. and M.S. degreesin the Department of Information Engineering,Chongqing Communication Institute, Chongqing,China.His research interests mainly focus on information

hiding and image processing.

Shanyu Tang (A08M08SM10) received thePh.D. degree from Imperial College London in 1995.He is a Distinguished Professor in the School of

Computer Science, China University of Geosciences.He is dedicated to adventurous research in fractalcomputing methods for covert communications,network security, and bio-informatics.Dr. Tang is the principal grant holder of four ex-

ternally funded research projects. He has contributedto 70 scientific publications36 refereed journalpapers including IEEE TRANSACTIONS and IEE/IET

journal papers.

Sen Bai received the B.E. degree in mathematicsfrom Sichuan University, China, in 1985, and theM.S. and Ph.D. degrees in applied mathematics,control theory and control engineering from theChongqing University, China, in 1998 and 2002,respectively. He is a Professor in the Department ofInformation Engineering, Chongqing Communica-tion Institute.Dr. Bais research interests mainly focus on

information hiding, image processing, and patternrecognition.