Vector Quantization based Steganography for Secure Speech

Communication System

Bekkar Laskar

and Merouane Bouzid

Tamanghasset University, Tamanghasset, Algeria

Dept. Telecommunication, Electronics Faculty, USTHB University, Algiers, Algeria

Keywords: Data Hiding, Steganography, Vector Quantization, Binning Scheme, ISF Parameters, Secure Speech,

Wideband Speech Coder, MELP, AMR-WB.

Abstract: Data hiding (steganography or watermarking) involves embedding secret data into various forms of digital

media such as text, audio, image and video. In this paper we propose two variants of vector quantization

(VQ) based steganography method to hide secret speech signal in host public speech coded by the AMR-

WB (Rec. G.722.2). The secret bit stream is hidden by using the basic principle of binning scheme which is

carried out in the split-multistage vector quantization of G.722.2 immittance spectral frequencies (ISF)

parameters.

1 INTRODUCTION

Steganography is the art of hiding secret information

in a cover media without attracting attention. Indeed,

modern steganography techniques exploit the

characteristics of digital media by using them as

carriers (covers) to hold hidden information. Covers

can be of different types including text,

speech/audio, image and video. Thus, the sender

embeds secret data in a digital cover file using a key

to produce a stego-file, in such a way that an

observer cannot detect the existence of the hidden

message (Cox et al., 2008). In this work, we focalize

particularly on speech steganography techniques

which consist in hiding a secret speech signal into a

cover (host) signal.

A variety of speech/audio steganography

methods have been proposed in the past, where most

of them are based on the temporal domain, the

transform domain and the compression domain. An

extended review of the current state-of-art literature

in digital audio/speech steganography techniques in

each domain is given in (Djebbar et al., 2012). In

compression domain, speech steganography

techniques based on vector quantization (VQ) have

been getting more and more popular, since they

enhance the traditional VQ compression by adding

the ability of data hiding.

In (Geiser and Vary, 2008), Geiser and Vary

proposed a method to embed digital data in the

bitstream of an ACELP speech codec. In

(Yargicoglu and Ilk, 2010), Yargicoglu and Ilk

proposed a data hiding methods that embed secret

data during MELP coding of the speech signal. The

secret bits are hidden by using quantization index

modulation (QIM) with the multistage vector

quantization (MSVQ) of line spectral frequencies

(LSF) parameters.

In this paper, two variant of the binning scheme

approach are developed for secure speech

communication. It is about a two steganographic

binning scheme (SBS) methods by VQ codebook

division called balanced and unbalanced codebook

partitioning. Our steganographic speech system

consists in embedding secret speech signal into host

public speech coded by the Adaptive Multi-rate

Wideband (AMR-WB, ITU-T G.722.2) (Bessette et

al., 2002) speech coder. For the compression of the

secret speech, we used the 2.4 kbits/s Mixed-

Excitation Linear Predictive (MELP) (McCree,

1996) speech coder. The embedding process of the

MELP secret bit stream is carried out into the split-

multistage vector quantization (S-MSVQ) indices of

G.722.2 immittance spectral frequencies (ISF).

Laskar, B. and Bouzid, M.

Vector Quantization based Steganography for Secure Speech Communication System.

DOI: 10.5220/0006398304070412

In Proceedings of the 14th International Joint Conference on e-Business and Telecommunications (ICETE 2017) - Volume 4: SECRYPT, pages 407-412

ISBN: 978-989-758-259-2

407

2 VQ-BASED DATA HIDING:

BINNING SCHEME

APPROACH

Several data hiding methods, based on vector

quantization, have been proposed in literature (Cox

et al., 2008), (Moulin and Koetter, 2005). One of the

most popular quantization-based data hiding method

is probably the quantization index modulation

(QIM) (Moulin and Koetter, 2005). Before

presenting the basic idea of our approach, based on

the QIM binning scheme, let us first review briefly

the basics of the conventional VQ system.

2.1 Vector Quantization

A k-dimensional VQ of rate R bits/sample is a

mapping of k-dimensional Euclidean space 

into a

finite codebook Y = {y

, …, y



} composed of L =

codevectors (Gersho and Gray, 1992). The

design principle of a VQ consists of partitioning the

k-dimensional space of source vectors x into L non

overlapping cells {R

,..., R



} (partition) and

associating with each cell R

a unique codevector y

such that the total average distortion D is minimized

(Gersho and Gray, 1992).

Various algorithms for the optimal design of VQ

have been developed in the past. The most popular

one is certainly the LBG algorithm (Gersho and

Gray, 1992). This algorithm is an iterative

application of the two optimality (nearest neighbor

and centroid) conditions such as the partition and the

codebook are iteratively updated.

2.2 Binning Scheme based on VQ

Codebook Partition

The considered steganographic binning scheme

(SBS) in this work is the one that modify the VQ

indices by codebook partitioning to hide secret bits

sequence (Geiser and Vary, 2008).

To embed a message of n bits per input cover

vector, the basic idea of the SBS is to partition first

the main VQ codebook Y into 2

disjoint sub-

codebooks Y

(i = 0,…, 2

1), by referring to a user-

key K called also "stego-key". Then, for each input

cover vector, a sub-codebook is selected according

to the steganographic bit to be embedded. The

traditional VQ search procedure is then done using

the sub-codebook for the input vector. Notice that in

the one bit embedding case (n = 1 bit), the VQ

codebook Y is partitioned in two sub-codebooks Y

and Y

. Then, the input vector is coded using the

nearest codevector from Y

and Y

according to

whether the secret bit is 0 or 1, respectively.

Figure 1 presents an example of SBS codebook

partitioning for embedding one bit per VQ index

according to stego-key K = {1, 0, 0, 1, 1, 0, 1, 0}.

Many codebook partitioning techniques for VQ-

based data hiding have been proposed in the past. In

(Kim, 2002), Jo and Kim proposed a method to

improve imperceptibility by partitioning each pair of

the codevectors with the shortest Euclidean distance

into three sub-codebooks, according to a given

threshold. They modified the VQ indices to embed

secret bits, while only two sub-codebooks are used

in the data hiding procedure. In (Wang, 2007), Wang

et al. developed an efficient method which also

partitions the VQ codebook into sub-codebooks to

modify the VQ indices in order to carry secret bits.

In the next section, we present our two SBS

codebook partitioning methods inspired from the

basic idea of Wang, Jain and Pan (Wang, 2007) VQ-

based data hiding approach applied for image

watermarking.

3 PROPOSED VQ-BASED DATA

HIDING METHOD

Our SBS system can be divided into three phases:

pretreatment (VQ codebook partitioning), message

embedding and message extraction. For simplicity,

the description below is limited to embedding only

one bit into each input cover vector.

Figure 1: SBS codebook (L = 8) partitioning in the case of

one bit embedded per VQ index.

3.1 Pretreatment Phase

The pretreatment phase consists in partitioning the k-

SECRYPT 2017 - 14th International Conference on Security and Cryptography

408

dimensional VQ codebook Y into two disjoint sub-

codebooks Y

and Y

according to a secret key K =

, …, k

}, k

 {0, 1} (i = 1,…, L). This division is

carried out implicitly where each codevector y

of Y

is assigned to sub-codebook

To ensure that the partition has a minimum effect

on the imperceptibility, an object function must be

minimized. It is about the total distortion caused by

the embedding process, formulated in our work as:

),(),(

1min

0min

YyDYyDD













where D

min

, Y

) is the minimal squared Euclidean

distance between the i

codevector of Y and all the

vectors of a sub-codebook Y

(m = 0 or 1).

3.1.1 Balanced Codebook Partitionning

According to a stego-key K = {k

,…, k

}, the

codebook Y of L codevectors is split into two sub-

codebooks Y

and Y

containing the same number of

codevectors. In this method, named Balanced

Codebook Partitioning (BCP), the key K must

contain the same number of "0" and "1" which is

L/2. Each codevector y

(1 ≤ i ≤ L) of Y will then be

assigned to sub-codebook Y

݇݅

. The basic steps of our

BCP of SBS system are given below.

Input data:

- Database of cover vectors X = {x

,…, x

- Sequence of secret bits M = {m

, m

,…}.

Step 1:

- Generate randomly S keys {K

,…, K

} such as the

numbers of "0" and "1" are identical in each key.

Step 2:

- Embed the given secret bits M = {m

, m

,…} into

VQ indices of the nearest codevectors of cover

vectors X by using respectively the S stego-keys.

Step 3:

- Evaluate the performance of each key K

(i = 1,…,

S) according to the embedding total distortion D

Step 4:

- Select and save the best stego-key which generated

the smallest total distortion.

Notice that the embedding procedure used in Step 2

is given below in sub-section 3.2.

The BCP method was developed to ensure a

minimum overall distortion, however it does not

ensure minimal degradation for each codevector of

the codebook Y. To add this characteristic, we

proposed a codebook partitioning method which

ensures a minimum overall distortion while ensuring

minimal degradation for each codevector of Y. This

method was named Unbalanced Codebook

Partitioning (UCP).

3.1.2 Unbalanced Codebook Partitionning

The UCP method splits the codebook Y into two

sub-codebooks which does not have the same

number of codevectors. The basic idea is to find a

minimal degradation for each codevector of the

codebook Y. Thus, for each index ݅ (1 ≤ ݅ ≤ ܮ), we

must find the index ݆ (1 ≤ ݆ ≤ ܮ, ݅ ≠ ݆) such as the

distance between the pair of codevectors y

and y

(d(y

, y

)) is minimal. The two codevectors of indices

݅ and ݆ will then be assigned to two different sub-

codebooks. Each codevector must belong to only

one sub-codebook. These two sub-codebooks will

permit to generate the stego-key K which will be

used in the embedding and extraction procedures.

The steps of our UCP approach are as follow.

Input data:

- VQ codebook of L codevectors.

Step 1:

- Find all pairs of indices i, j (1 ≤ ݅, j ≤ ܮ, ݅ ≠ ݆) such

as the distances d(y

, y

) are minimal.

Step 2:

- For each index pair (i, j), put the index i in a group

("0") and the index j in the other group ("1").

- Label each index by its group number:

Label(i) = 0 and Label(j) = 1.

Step 3:

- Generate the stego-key K = {k

,…, k

} by using the

labels of the L indices : k

= Label(i), for i = 1,…, L.

3.2 Embedding Phase

Similar to (Kim, 2002) and (Wang, 2007), we use

the codevectors indices in Y

or Y

to hide bit "0" or

bit "1", respectively. The embedding procedure steps

are given below.

Input data:

- VQ codebook of L codevectors.

- Secret stego-key K = {k

, k

, . . ., k

- Database of cover vectors X = {x

,…, x

- Sequence of secret bits M = {m

, m

,…}.

Vector Quantization based Steganography for Secure Speech Communication System

409

Step 1:

- For the vector x

and the secret bit m

, find the

nearest codevector y for x

from Y with the condition

= m

(i.e., use Y

if m

= 0 or Y

if m

= 1).

- Send the stego-index of the selected codevector y

to the reception side.

Step 2:

- Repeat Step 1 until all the secret bits have been

treated.

3.3 Extraction Phase

To extract the bit hidden in a received stego-index i,

one have only to know the value of k

in the key K.

The extraction procedure steps are given below.

Input data:

- Secret stego-key K = {k

, k

, . . ., k

- Quantization stego-indices : I = {i

, i

,….}.

Step 1:

- For the j

index i

of the obtained carrier

codevector

determine the hidden bit: m

Step 2:

- Repeat Step 1 until all the hidden bits M = {m

,…} have been extracted.

4 SPEECH STEGANOGRAPHIC

SYSTEMS: APPLICATION OF

THE BCP AND UCP METHODS

In this section, we evaluate the performance of our

speech steganographic systems, designed based on

SBS by codebook partitioning and called "SBS-CP".

These systems were developed separately by the

BCP and UCP methods presented above.

In our applications, the main purpose is to hide a

secret speech signal coded by the 2.4 kbps MELP

into a host public speech coded by the AMR-WB

Rec. G.722.2 (Bessette et al., 2002). The embedding

is done during the S-MSVQ quantization of the

G.722.2 ISF parameters (ISFs) by using a secret

stego-key K. Notice that in all simulations, we used

the G.722.2 in mode 12.65 kbits/s where the ISFs

are coded by an S-MSVQ of 46 bits/frame.

Recall, that the G.722.2 ISF parameters are

quantized by a split multistage vector quantizer (S-

MSVQ) with 1

order MA predictor. The G.722.2 S-

MSVQ uses 7 codebooks, where 2 codebooks at the

first stage (named here CB

and CB

) and 5

codebooks (named CB

, CB

) at

the second stage.

4.1 Performance Evaluation Criteria

Performance evaluation of the implemented speech

steganographic systems will be done according to

two criteria: the hiding capacity represented by the

embedding rate of the secret speech and the

transparency (imperceptibility) represented by the

perceptual quality of the speech stego-signal

synthesized by the G.722.2 with embedding

procedure.

The total embedding rate R is given by the ratio

of the number of hidden secret bits and the length of

the host speech coder frame:

sbitsnnR /50

1020









(2)

where n is the total number of ISF S-MSVQ

quantization stego-indices used in all the embedding

process. Note that the highest embedding rate, which

can be obtained by this SBS-CP method, is reached

when we hide one bit in each of the 7 S-MSVQ

codebook quantization indices, i.e., 7 bits hidden per

frame. The maximum total embedding rate would

then be: R = n  350 bits/s.

On the other hand, for imperceptibility, we use

the ITU-T Rec. P.862.2 known under the

abbreviation WB-PESQ (Wide Band extension of

Perceptual Evaluation Speech Quality) (ITU-T Rec.

P.862.2., 2005) to evaluate the coded cover/stego

speech signals quality. The hidden speech signal is

imperceptible if a listener is unable to distinguish

between the cover and the stego speech signals;

which means that the WB-PESQ difference between

the two cover/stego signals is negligible.

The performance of the steganographic S-MSVQ

quantizer will be also evaluated by the well-know

average spectral distortion (SD) measure. The

spectral distortion of each frame i is given, in

decibels, by (Paliwal and Atal, 1993), (Cheraitia and

Bouzid, 2014):

1()

10 log ,

()

jnN























(3)

where S(e



n/N

) and Ŝ(e



n/N

) are respectively the

original and quantized power spectra of the LPC

SECRYPT 2017 - 14th International Conference on Security and Cryptography

410

synthesis filter, associated with the i

frame of

speech signal.

Generally, we can get transparent quantization

quality if we maintain the three following conditions

(Paliwal and Atal, 1993): 1)- The average spectral

distortion (SD) is about 1 dB, 2)- No Outliers frames

with SD greater than 4 dB, 3)- The percentage of

Outlier frames having SD within the range of 2-4 dB

must be less than 2%.

4.2 Performances of Steganographic

SBS-CP Systems Implemented in

G.722.2 S-MSVQ Quantizer

For a given embedding rate, we performed an

optimization procedure of our steganographic

systems which consists in finding the best choice of

S-MSVQ codebooks that can be used in the hiding

process to obtain the best possible performance.

The speech database used in the following

experiments consists of 60 minutes of speech taken

from the international TIMIT database (f

= 16 kHz)

(Garofolo et al., 1988). To construct the ISF

database, we used the same LPC analysis function of

the G.722.2, where a 16-order LPC analysis, based

on the autocorrelation method, is performed every

analysis frame of 20 ms. Thus, a database of 180000

ISF vectors was constructed.

For embedding rates varying between 1 and 7

bits/frame, the SD performances of speech

steganographic SBS-CP systems implemented in

G.722.2 S-MSVQ are shown in Table 1. The SBS-

CP systems were designed respectively by our BCP

and UCP methods. For comparative evaluation, the

performance of a conventional steganographic

system designed by a random codebook partitioning

(RCP) method was also included in Table 1. For a

secret bits sequence, the wideband ISF vectors of

dimension 16 are quantized by the same G.722.2 S-

MSVQ quantizer of seven codebooks, denoted in the

Table as follows (CB

, CB

,…, CB

). For

example, the notation "2 (0-0-0-1-1-0-0)" means that

for an embedding rate of 2 bits/frame, the codebooks

and CB

of the S-MSVQ 2

stage were

selected as best choice to be used in hiding 2 bits per

each frame.

These comparative results show that the

performances of steganographic SBS-CP systems

those designed by the BCP. On the other hand, the

systems designed by UCP and BCP methods

outperform the systems designed by the traditional

RCP. Indeed, they can both achieve the transparent

quantization quality until an embedding rate of 3

bits/frame.

4.3 Performance Evaluation of G.722.2

with SBS-CP Systems Implemented

in S-MSVQ Quantizer

The cover public speech database used in the

following evaluations is composed of 10 speech

sequences of 32s extracted from the same TIMIT

database. The secret bit stream was generated by the

2.4 kbps MELP from a speech sequence of f

= 8

kHz extracted from a phonetically balanced Arabic

speech database (Boudraa et al., 1992).

Table 2 presents WB-PESQ performance

comparative evaluation of the global G.722.2 where

its ISF parameters were quantized by the 46

bits/frame steganographic S-MSVQ with SBS-CP

systems designed respectively by BCP and UCP.

Notice that an embedding rate of 0 bits/frame means

the original standard G.722.2 without steganography

(i.e assessment of the cover speech signal).

Table 1: Performance of steganographic SBS-CP systems designed respectively by UCP, BCP and RCP methods.

Embedding

rate

(Bits/frame)

SBS-CP systems by UCP SBS-CP systems by BCP SBS-CP systems by RCP

Av. SD

(dB)

Outliers (in %)

Av. SD

(dB)

Outliers (in %)

Av. SD

(dB)

Outliers (in %)

2 - 4 dB > 4 dB

2 – 4 dB > 4 dB 2 – 4 dB > 4 dB

1 (0-0-0-1-0-0-0) 0.95 1.13 0.0005 0.95 1.11 0.001 0.96 1.32 0.001

2 (0-0-0-1-1-0-0) 0.99 1.45 0.002 1.00 1.42 0.001 1.03 1.93 0.005

3 (0-0-0-1-1-1-0) 1.03 1.80 0.005 1.05 1.83 0.007 1.09 2.66 0.010

4 (0-0-1-1-1-1-0) 1.09 2.48 0.008 1.11 2.55 0.007 1.17 4.43 0.005

5 (0-0-1-1-1-1-1) 1.15 3.19 0.009 1.17 3.35 0.007 1.24 5.84 0.015

6 (0-1-1-1-1-1-1) 1.23 5.57 0.017 1.25 6.14 0.027 1.37 11.03 0.120

7 (1-1-1-1-1-1-1) 1.30 7.61 0.035 1.33 8.85 0.041 1.51 17.46 0.465

Vector Quantization based Steganography for Secure Speech Communication System

411

Table 2: Performance of the global G.722.2 with

steganographic SBS-CP implementation.

Embedding rate

(Bits/frame)

G.722.2 with

SBS-CP by BCP

G.722.2 with

SBS-CP by UCP

WB-PESQ WB-PESQ

0 3.790 3.790

1 3.798 3.705

2 3.823 3.744

3 3.687 3.814

4 3.719 3.680

5 3.756 3.747

6 3.766 3.720

7 3.676 3.720

For all embedding rates, these simulation results

show that the overall quality of stego-speech is

almost identical to quality of cover public speech;

which means that our proposed steganographic

techniques are practically imperceptibles. Most WB-

PESQ scores of the stego-signals are between 3.67

and 3.82. Hence, a good speech quality was obtained

and no degradation was caused by the embedding

process. On the other hand, steganographic SBS-CP

systems designed by the UCP yields slight

improvement to the G.722.2 WB-PESQ performance

compared to SBS-CP with balanced partitioning.

5 CONCLUSION

In this paper, we proposed two variants of VQ-based

speech steganography binning schemes for G.722.2

secure speech communication system. The

simulation results showed that the two

steganographic SBS-CP methods by UCP and BCP

can generate stego-speech signals with similar

quality to cover speech signals; which means that

the resulting stego-speech is indistinguishable from

the original cover speech. Hence, the two proposed

variants of SBS-CP method can ensure a high

transparency with a maximal embedding rate of 7

bits/frame (350 bits/s).

Robustness against intentional and non-

intentional attacks has not been investigated in this

work; it will be studied in future research.

REFERENCES

Bessette, B., Salami, R., Lefebvre, R., Jelínek, M., Rotola-

Pukkila, J., Vainio, J., Mikkola, H., Järvinen, K., 2002.

The adaptive multirate wideband speech codec (AMR-

WB), IEEE Transactions on Speech and Audio

Processing, vol. 10, no. 8, pp. 620-636.

Boudraa, M., Boudraa, B., Guerin, B., 1992. Mise en place

de phrases arabes phonetiquement equilibrées. In

JEP'92, XIXèmes Journées d'Etude sur la Parole.

Bruxelles.

Cheraitia, S., Bouzid, M., 2014. Robust coding of

wideband speech immittance spectral frequencies.

Speech Communication, Elsevier, vol. 65, pp. 94-108.

Cox, I. J., Miller, M. L., Bloom, J. A., Fridrich, J., Kalker,

T., 2008. Digital Watermarking and Steganography,

Second Edition, Morgan Kaufmann Publishers, USA.

Djebbar, F., Ayad, B., Meraim, K. A., Hamam, H., 2012.

Comparative study of digital audio steganography

techniques, EURASIP Journal on Audio, Speech, and

Music Processing, Springer, vol. 25, pp. 1-16.

Garofolo J. S., et al., DARPA TIMIT Acoustic-phonetic

Continuous Speech Database. National Institute of

Standards and Technology (NIST), Gaithersburg,

October 1988.

Geiser, B., Vary, P., 2008. High rate data hiding in

ACELP speech codecs. In ICASSP’2008, IEEE

International Conference on Acoustics, Speech and

Signal Processing. pp. 4005-4008. USA.

Gersho, A., Gray, R. M., 1992. Vector quantization and

Signal compression, Kluwer Acad. Publishers, USA.

ITU-T Recommendation G.722.2. Wideband coding of

speech at around 16 kb/s using Adaptive Multi-rate

Wideband (AMR-WB), 2003.

ITU-T Recommendation P.862.2. Wideband Extension to

Recommendation P.862 for the Assessment of

Wideband Telephone Networks and Speech Codecs,

Geneva, 2005.

Kim, Jo. M., 2002. A digital image watermarking scheme

based on vector quantisation. IEICE Trans. on Inf. and

Systems, vol. E85-D, pp. 1054-1056.

McCree, A., Truong, K., George, E. B., Barnwell, T. P.,

Viswanathan, V., 1996. A 2.4 kbits/s MELP Coder

Candidate for the New U.S. Federal Standard. In

ICASSP'96, IEEE International Conference on

Acoustics, Speech and Signal Processing. pp. 200-203.

Moulin, P., Koetter, R., 2005. Data-Hiding Codes.

Proceedings of The IEEE, Vol. 93, pp. 2083-2126.

Paliwal, K. K., Atal, B. S., 1993. Efficient vector

quantization of LPC parameters at 24 bits/frame. IEEE

Transactions on Speech and Audio Processing, vol. 1,

no. 1, pp. 3-14.

Wang, F. H., Jain, L. C., Pan, J. S., 2007. A novel VQ-

based watermarking scheme with genetic codebook

partition. Journal of Network and Computation

Applications (JNCA), Elsevier, vol. 30, no. 1, pp. 4-23.

Yargıcoglu, A. U., Ilk, H. G., 2010. Hidden data

transmission in mixed excitation linear prediction

coded speech using quantisation index modulation,

IET Information Security, vol. 4, Issue 3, pp. 158–166.

SECRYPT 2017 - 14th International Conference on Security and Cryptography

412