Contents lists available at ScienceDirect

Signal Processing: Image Communication

Signal Processing: Image Communication 39 (2015) 17–25

http://d0923-59

n CorrE-m

guoyq@kongxw

journal homepage: www.elsevier.com/locate/image

Secure spread-spectrum data embeddingwith PN-sequence masking

Ming Li, Yanqing Guo, Bo Wang n, Xiangwei KongSchool of Information and Communication Engineering, Dalian University of Technology, Dalian, Liaoning 116024, PR China

a r t i c l e i n f o

Article history:Received 21 January 2015Received in revised form30 July 2015Accepted 30 July 2015Available online 10 August 2015

Keywords:Data hidingInformation hidingPseudo-noise maskingSignal-to-interference-plus-noise ratio(SINR)Spread-spectrum embeddingSteganography

x.doi.org/10.1016/j.image.2015.07.01465/& 2015 Elsevier B.V. All rights reserved.

esponding author.ail addresses: mli@dlut.edu.cn (M. Li),dlut.edu.cn (Y. Guo), bowang@dlut.edu.cn (B@dlut.edu.cn (X. Kong).

a b s t r a c t

Conventional additive spread-spectrum (SS) data embedding has a dangerous securityflaw that unauthorized receivers can blindly extract hidden information without theknowledge of carrier(s). In this paper, pseudo-noise (PN) masking technique is adopted asan efficient security measure against illegitimate data extraction. The proposed PN-sequence masked SS embedding can offer efficient security against current SS embeddinganalysis without inducing any additional distortion to host nor notable recoveryperformance loss. To further improve recovery performance, optimal carrier design forPN-masked SS embedding is also developed. With any given host distortion budget, weaim at designing a carrier to maximize the output signal-to-interference-plus-noise ratio(SINR) of the corresponding maximum-SINR linear filter. Then, we present jointly optimalcarrier and linear processor designs for PN-masked SS embedding in linearly modifiedtransform domain host data. Finally, PN-masked multi-carrier/multi-message SS embed-ding is studied as well. The extensive experimental studies confirm our analyticalperformance predictions and illustrate the benefits of the designed PN masked optimalSS embedding.

& 2015 Elsevier B.V. All rights reserved.

1. Introduction

Information hiding, which is also called as digital dataembedding, is the process of hiding data under a cover med-ium (also referred to as host), such as image, video, or audio[1,2]. Applications may vary from annotation, copyright-marking, and watermarking to single-stream media merging(text, audio, image) and covert communication [3–8]. As ageneral encompassing comment, different applications ofinformation hiding, such as the ones identified above, requiredifferent satisfactory tradeoffs between the following fourbasic attributes of data hiding [9]: (i) payload – information

. Wang),

delivery rate; (ii) robustness – hidden data resistance to noise/disturbance; (iii) transparency – low host distortion forconcealment purposes; and (iv) security – inability byunauthorized users to detect/access the covert communica-tion channel.

The data hiding performance in terms of above fourattributes depends directly on how the data is inserted inthe host. Therefore, it is a crucial step to determine theembedding process in the design of a data hiding system.Data embedding can be performed either directly in the time(audio) or spatial (image) domain [10–14] or in a transformdomain [15–26]. While direct embedding in the original hostsignal domain may be desirable for system complexitypurposes, embedding in a transform domain may takeadvantage of the particular transform domain properties[27] and enables the powerful notion of spread-spectrum

1 Additive white Gaussian noise is frequently viewed as a suitablemodel for quantization errors, channel transmission disturbances, and/orimage processing attacks.

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–2518

(SS) data embedding when the secret signal is spread over awide range of host frequencies [28–32].

In this paper, we focus our attention on additive SSembedding in transform domain host. In direct analogy toSS digital communication systems [33], conventional addi-tive SS embedding methods use an equal-amplitude modu-lated carrier/signature to deposit one information symbolacross a group of host data coefficients or a linearly trans-formed version of the host data coefficients. Recently, adangerous security flaw of SS embedding has been alertedand investigated. Embedding a number of informationsymbols with a same carrier will create a strong basis/subspace of the hidden signal which can be tracked andanalyzed. Therefore, even without the knowledge of carrier(s), unauthorized receivers can still blindly extract theembedded data by blind signal separation (BBS) methods[34–37] or novel iterative generalized least square (IGLS)approaches [38,39]. The illegitimate blind hidden dataextraction has also been referred to as “Watermarked con-tent Only Attack” (WOA) in the watermarking securitycontext [34–37]. Thus, it raises the concerns of making SSembedding more difficult to be extracted by the illegitimateusers. Two interesting SS embedding schemes were pro-posed in [40] which attempt to withstand SS embeddinganalysis by using random-like amplitudes. However, these SSembedding schemes sacrifice recovery performance toenhance the security and consequently are sensitive to noisewhich would lead to high recovery error rates by intendedrecipients. More importantly, information leakage cannot befully prevented because information symbols are stillembedded by the same carrier.

Pseudo-noise (PN) masking technique has been proven tobe an effective technique against unauthorized data collec-tion (eavesdropping) in the context of secure wireless com-munications. Typical examples of PN masking technique aremilitary-grade communications and global-positioning sys-tems (GPS). In this work, we first develop a PN-maskedsecure SS embedding approach in which the embedded SSsingle is scrambled by random-like PN masks such that nosubspace of embedded signal can be found and tracked inthe data-embedded host. With our proposed PN masked SSembedding scheme, the performance in terms of recoverybit-error-rate (BER) at the intended receiver is maintained atalmost the same level as the conventional SS embedding (i.e.almost no performance loss), while the unauthorized userswill have BER close to 0.5 (i.e. almost perfect security). Sincethe proposed PN masked SS embedding schemes can effi-ciently minimize the likelihood that embedded data are“stolen” by the unauthorized users, they are suitable forthe applications with high security requirement, such assteganography and covert communications.

It should also be understood that the host, which acts as asource of interference to the secret message of interest, isknown to the message embedder. Such knowledge can beexploited appropriately to facilitate the task of the blindreceiver at the other end and minimize the recovery errorrate for a given host distortion level, minimize host distortionfor a given target recovery error rate, maximize the Shannoncapacity of the covert channel, etc. By exploiting the knowl-edge of the second order statistics (SOS) of host, the recentlypresented Gkizeli–Pados–Medley eigen-design optimal

carrier [29,30] can maximize the signal-to-interference-noise-ratio (SINR) at the output of the correspondingmaximum-SINR linear filter. Benefiting from the legacy of[29,30], the optimal carrier design for PN-masked SS embed-ding is also studied.

The rest of the paper is organized as follows. Section 2briefly reviews the prior art on additive SS embedding. PN-masked SS embedding is present in Section 3. These resultsare generalized to multi-carrier embedding in Section 4.Section 5 is devoted to experimental studies and comparisons.A few concluding remarks are drawn in Section 6.

The following notation is used throughout the paper.Boldface lower-case letters indicate column vectors andboldface upper-case letters indicate matrices; R denotesthe set of all real numbers; ðÞT denotes matrix transpose; ILis the L� L identity matrix; sgnf�g denotes zero-thresholdquantization; Ef�g represents statistical expectation; J � J isvector norm.

2. Prior art of additive SS embedding

Consider a host image HAMN1�N2 where M is thefinite image alphabet and N1 � N2 is the image size inpixels. Without loss of generality, the image H is parti-tioned into M local non-overlapping blocks of sizeN1N2=M. Each block, H1;H2;…., HM , is to carry one hiddeninformation bit biAf71g, i¼ 1;2;…;M, respectively.Embedding is performed in a 2-D transform domain T(such as the discrete cosine transform and a wavelettransform). After transform calculation and vectorization(for example by conventional zig-zag scanning), we obtainT ðHiÞARN1N2=M ; i¼ 1;2;…;M. From the transform domainvectors T ðHiÞ we choose a fixed subset of LrN1N2=Mcoefficients (bins) to form the final host vectors xiARL,i¼ 1;2;…;M. It is common and appropriate to avoid the dccoefficient (if applicable) due to high perceptual sensitivityin changes of the dc value.

To draw a parallelism with SS communication systems,conventional SS embedding treats embedded message as theSS signal of interest transmitted through a noisy “channel”(the host). The disturbance to the SS signal of interest is thehost data themselves plus potential external noise due tophysical transmission of the watermarked data and/or proces-sing/attacking. In particular, conventional additive SS embed-ding is carried out in the transform domain by

yi ¼ Abisþxiþni; i¼ 1;…;M; ð1Þwhere information bit biAf71g is embedded in the trans-form domain host vector xiARL via additive SS embeddingby means of a (normalized) spreading sequence (carrier/signature) sARL; JsJ ¼ 1, with a corresponding embeddingamplitude A40. For the sake of generality, ni representspotential external white Gaussian noise1 of mean 0 andautocorrelation matrix σ2

nIL, σ2n40.

In an effort to reduce the interference effect of the hostsignal, the host vectors xi, i¼ 1;…;M, can be steered awayfrom the embedding carrier using an operator of the form

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–25 19

ðIL�cssT Þ with parameter cAR, i¼ 1;…;M, and the carriersARL. In parallel to (1), the composite signal of additive SSembedding on linearly transformed host data is [28,30]

yi ¼ AbisþðIL�cssT Þxiþni; i¼ 1;…;M; ð2Þ

where information symbol bit biAf71g is embedded,using amplitude A40 and (normalized) carriersARL; JsJ ¼ 1; in the ith linearly transformed host datavector ðIL�cssT Þxi. The optimal carrier s and transformparameter c to maximize the output SINR are presented inProposition 3 of [30].

The SS embedding schemes (1) and (2) have been shownto have a dangerous security flaw. Using the same carrier s toembed all information bits can generate a strong basis/sub-space of embedded signal in data-embedded host yi,i¼ 1;…;M. By analyzing observation signal yi with BBS-based algorithms [34–37] or a novel IGLS approach [39],embedded information bits can be blindly extracted byunauthorized users without the knowledge of carrier s. Usingrandom-like amplitudes [40] can weaken SS embeddinganalysis to a certain degree, but still has the problem ofinformation leakage and suffers from loss of recovery perfor-mance. To practically provide a secure SS embedding, in thenext section we adopt PN-sequence masking technique on SSembedding to protect the secret data without any notableperformance loss.

3. PN-sequence masked SS embedding

Let m¼ ½m1;m2;…;mN�T Af71gN be a PN-sequence(such as m-sequence or Gold code) of large lengthNZLM. We select M non-overlap segments from m asmask vectors m1;m2;…;mM of length L each

mi9 ½mði�1ÞLþ1;…;miL�T ; i¼ 1;…;M: ð3Þ

The PN-mask vector mi is used to scramble the SS signal ofinterest Abis by component-wise multiplication of carrier sand mi. PN-masked carrier for the ith bit bi is defined as

ci9s � mi9 ½sð1Þmið1Þ; sð2Þmið2Þ;…; sðLÞmiðLÞ�T ð4Þ

where � denotes component-wise vector multiplication.Since the mask vector mi just pseudo-randomly flips eachelement of s, then with a normalized regular carrierJsJ ¼ 1, the PN-masked carriers are also normalizedJci J ¼ 1, i¼ 1;…;M.

Instead of using the same carrier s, information bitbiAf71g is embedded in host xi by means of a PN-maskedcarrier ciARL; Jci J ¼ 1

yi ¼ Abiciþxiþni; i¼ 1;…;M; ð5Þ

with an embedding amplitude A40. With random-likePN-masked carriers, the SS signal of interest Abici behaveslike white noise and no subspace of embedded signal canbe tracked from the observation data yi. Therefore, PN-masked SS embedding in (5) can efficiently preventillegitimate data extraction by unauthorized users whohave no knowledge of PN masks.

Squared Euclidean metric is rudimentary but commonchoice to measure the distortion to host. The mean-squared

(MS) distortion to the host due to the embedded data only is

D¼ EJ ðAbiciþxiÞ�xi J2 ¼ A2: ð6ÞThe MS distortion of PN-masked SS embedding

depends only on the embedding amplitudes and PN-sequence masking operation would not induce any moredistortion to host.

Recovery of the embedded information bits at theintended receiver requires use of a replica PN generatorto “strip-off” the mask by following operation:

~y i ¼ yi � mi

¼ Abis � mi � miþxi � miþni � mi

¼ Abisþ ~x iþ ~n i ð7Þ

where ~x i9xi � mi is PN-masked host vector,~n i9ni � mi. After mask removal, the PN-masked SSembedding in (7) has a similar form to the conventionalSS embedding in (1) with PN-masked host ~x i instead oforiginal host vector xi. Since PN masking operation at theintended receiver just randomly flips the host coefficientsand the external noise, the total disturbance ð ~x iþ ~n iÞ to thesignal of interest Abis would not be amplified.

The embedded bits can be recovered by looking at thesign of the output of a filter wARL

b̂i ¼ sgn wT ~y i

� �: ð8Þ

In current data hiding applications, simple matched filter(MF)

wMF ¼ s ð9Þhas been widely used by the intended receiver to recoverembedded bits. With signal of interest Abis and totaldisturbance ð ~x iþ ~n iÞ in (7) after mask removal operationat the intended receiver, the output SINR of filter w is

SINR¼ EfJAbiðwTsÞJ2gEfJwT ð ~x iþ ~n iÞJ2g

¼ A2wTssTwwT ðR ~x þσ2

nILÞwð10Þ

where

R ~x 9Ef ~x i ~xTi g ¼

1M

XMi ¼ 1

~x i ~xTi ð11Þ

is the autocorrelation matrix of PN-masked host vectors.The linear filter that offers maximum SINR at its output is[41]

wmaxSINR ¼ ðR ~x þσ2nILÞ�1s: ð12Þ

The exact maximum output SINR value attained is

SINRmax ¼ A2sT ðR ~x þσ2nILÞ�1s: ð13Þ

We can view SINRmax as a function of the embeddingcarrier s and identify the signature that maximizes theSINR at the output of the maximum SINR filter. Ourfindings are presented in the form of a proposition belowthat parallels the developments in [30] for the conven-tional SS embedding case. The proof is straightforward andomitted.

Proposition 1. Consider PN-masked SS embedding by (5).The optimal carrier soptARL that maximizes the output SINR

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–2520

of the maximum-SINR filter wmaxSINR is

sopt ¼ qL ð14Þwhere qL is the eigenvector of autocorrelation matrix R ~x in(11) with the smallest corresponding eigenvalue λL. Whensopt ¼ qL, the maximum-SINR filter with this optimal carrieris also a matched filter

wmaxSINR �wMF ¼ qL:□ ð15Þ

Now we turn our attention on PN-masked SS embed-ding on linearly transformed SS embedding which ismodeled in a form of

yi ¼ AbiciþðIL�ccicTi Þxiþni; i¼ 1;…;M; ð16Þwhere ci9s � mi is the PN-masked carrier. The hostvector xi is linearly transformed by ðIL�ccicTi Þ which isformed by the corresponding PN-masked carrier ci.

The intended receiver first removes the masks from yiby

~y i ¼ yi � mi

¼ ðAbiciþxi�ccicTi xiþniÞ � mi

¼ Abici � miþxi � mi�cðcTi xiÞðci � miÞþni � mi:

ð17ÞWith cTi xi ¼ ðs � miÞTxi ¼ sT ðmi � xiÞ and ci � mi ¼ s,

mask-removed signal in (17) can be rewritten as

~y i ¼ Abisþxi � mi�cðsT ðmi � xiÞÞsþni � mi

¼ Abisþ ~x i�cðsT ~x iÞsþ ~n i

¼ AbisþðIL�cssT Þ ~x iþ ~n i ð18Þwhere ~x i9xi � mi, ~n i9ni � mi. Now PN-masked SSembedding on linearly transformed host is similar to thenon-PN-masked one in (2).

The mean-squared distortion due to the embeddingoperation only is

D¼ E J AbiciþðIL�ccicTi Þxi� ��xi J2

n o¼ E J Abi�ccTi xi

� �ci J2

n o¼ E JAbi�csT ~x i J2n o

¼ A2þc2sTR ~xs: ð19ÞIt should be noticed that, in contrast to (6), the distortionlevel is controlled not only by A but also by c. Compared toconventional SS embedding in (5), SS embedding onlinearly transformed host uses part of available distortionto pre-suppress the interference at the embedding sideand then utilizes the remaining distortion to embedinformation bits. The joint optimal carrier s, amplitude A,and transformation parameter c design to maximize out-put SINR are summarized in Proposition 2 below whoseproof is similar to Proposition 3 in [30] and omitted.

Proposition 2. Consider PN-masked SS embedding in line-arly transformed host data by (16). Let q1;q2;…;qL beeigenvectors of R ~x in (11) with corresponding eigenvaluesλ1Zλ2Z⋯ZλL. For any hidden message-induced distortionbudget D, the optimal carrier s, amplitude A, and

transformation parameter c to maximize SINR are

sopt ¼ qL; ð20Þ

copt ¼λLþσ2

nþD�ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðλLþσ2

nþDÞ2�4λLDq

2λL; ð21Þ

A¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiD�c2λL

p: ð22Þ

When sopt ¼ qL and c¼ copt, then maximum SINR filtersimplifies to

wmaxSINR ¼wMF ¼ qL:□ ð23Þ

Adaptive optimal carrier design can significantlyimprove recovery performance. However, when hostimage is changed, the embedder needs to re-design theoptimal carrier and re-transmit it to the intended receivervia a secure channel. This operation may be not applicablefor some data hiding applications. In the most commondata hiding cases, the carrier is pre-defined and known forboth embedder and receiver. For any arbitrary carrier s, theoptimal separation of distortion budget to maximize theoutput SINR is presented in the form of a propositionbelow.

Proposition 3. Consider PN-masked SS embedding in line-arly transformed host data by (16) and matched filterwMF ¼ s is used for embedded bits recovery. For any hiddenmessage-induced distortion budget D and carrier s, theoptimal amplitude A and transformation parameter c tomaximize SINR are

copt ¼αþσ2

nþD�ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðαþσ2

nþDÞ2�4αDq

2α; ð24Þ

A¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiD�c2α

p; ð25Þ

where α9sTR ~xs.□

Proof. With SS embedding signal (17) after mask removal,the output SINR of MF is

SINR¼ EfJAbi J2gEfJsT ððIL�cssT ÞxiþnÞJ2g

ð26Þ

¼ A2

sT ððIL�cssT ÞR ~x ðIL�cssT Þþσ2nIÞs

¼ A2

sTR ~xs�2csTR ~xsþc2sTR ~xsþσ2n

¼ A2

α�2cαþc2αþσ2n

ð27Þ

where we define α9sTR ~xs. By ApplyingD¼ A2þc2sTR ~xs¼ A2þc2α into (27), we obtain

SINR¼ D�c2αα�2cαþc2αþσ2 ð28Þ

By direct differentiation of (28) with respect to c and root

selection, we obtain c¼ αþσ2n þD�

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðαþσ2

n þDÞ2 �4αDp

2α in (24).□

14 16 18 20 22 24 26 28 30 3210−7

10−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

Fig. 1. BER versus per-block distortion due to embedding (512�512Baboon, L¼63, external noise variance is fixed at σ2n ¼ 3 dB).

2 With block MS distortion D, the peak signal-to-noise ratio (PSNR)of the image due to embedding can be calculated byPSNR¼ 20log10ð255Þ�10log10ðD=64Þ. The embedding distortion to attackdistortion ratio (WNR) measure can also be easily obtained byWNR¼ 10log10ðD=64=σ2

nÞ.3 With σ2

n ¼ 3 dB, the initial peak signal-to-noise ratio (PSNR) ofimage (i.e. no hidden data, only external noise) is 45:1 dB.

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–25 21

4. Multi-carrier PN-masked SS embedding

In this section, we extend our studies to the emergingconcept of multi-carrier/multi-message SS embedding. Weconsider K distinct message bit sequences, fbk;1; bk;2;…;

bk;Mg, k¼ 1;2;…;K , bk;iAf71g, i¼ 1;…;M, each of lengthM bits. The K message sequences may be to be delivered toK distinct corresponding recipients or they are just Kportions of one large message sequence to be transmittedto one recipient. In particular, the ith bit from each of Ksequences, b1;i;…; bK ;i, is simultaneously embedded in theith transform-domain host vector xi with correspondingamplitude Ak;iZ0 and PN-masked carrier ck;i ¼ sk � mi,miAf71gL; skARL, Jsk J ¼ 1, k¼ 1;…;K , i¼ 1;…;M:

yi ¼XKk ¼ 1

Ak;ibk;ick;iþxiþn

¼XKk ¼ 1

Ak;ibk;isk

!� miþxiþn; i¼ 1;2;…;M: ð29Þ

The contribution of each individual embedded messagebit bk;i to the composite signal is Ak;ibk;ick;i and the MSdistortion to the original host data due to the embeddedmessage-k alone is

Dk ¼ EfJAk;ibk;ick;i J2g ¼1M

XMi ¼ 1

A2k;i; k¼ 1;…;K:

Under a statistical independence assumption acrossmessage bits, the mean-squared distortion of the originalimage due to the total multi-message insertion is

Dt ¼XKk ¼ 1

Dk ¼1M

XKk ¼ 1

XMi ¼ 1

A2k;i: ð30Þ

After PN-mask removal, the intended receiver of the kthmessage can use a filter wkARL to recover embedded bits

b̂k;i ¼ sgn wTk yi � mi� �� �

¼ sgn wTk

XKk ¼ 1

Ak;ibk;iskþxi � miþn � mi

!( )

¼ sgn Ak;ibk;iðwTkskÞþ

XKj ¼ 1;jak

Ak;ibk;iwTksjþwT

k ~x iþwTk~n i

8<:

9=;

ð31Þwhere ~x i9xi � mi, ~n i9ni � mi. Similar to the finding inProposition 6 of [30], the conditionally optimal carriers aresummarized in the form of Proposition 4 below.

Proposition 4. Consider additive SS embedding by (29). Letq1;q2;…;qL be eigenvectors of R ~x with corresponding eigen-values λ1Zλ2Z⋯ZλL. Orthogonal carriers that condition-ally maximize the output SINR of the maximum SINR filterwmaxSINR;k are

soptk ¼ qL�kþ1; k¼ 1;…;K : ð32ÞWhen soptk ¼ qL�kþ1, maximum SINR filter simplifies tomatched filter

wmaxSINR;k �wMF;k ¼ soptk :□ ð33Þ

5. Experimental studies

To carry out an experimental study of the developmentspresented in the previous sections, we consider the familiargray-scale 512�512 “Baboon” image as a host example. Weperform 8�8 block DCT single-carrier embedding over all63 bins except the dc coefficient. Hence, our carrier length isL¼63 and we embed 5122=82 ¼ 4096 bits. For the sake ofgenerality, we also incorporate white Gaussian external noiseof variance σ2

n ¼ 3 dB. We evaluate the performance of eightdifferent embedding schemes: (i) conventional SS embed-ding in (1) with an arbitrary carrier sarb, (ii) PN-maskedconventional SS embedding in (5) with an arbitrary carriersarb, (iii) conventional SS embedding in (1) with an optimalcarrier sopt, (iv) PN-masked conventional SS embedding in(5) with an optimal carrier sopt, (v) linearly transformed SS(LTSS) embedding in (2) with an arbitrary carrier sarb and theoptimal transformation parameter copt, (vi) PN-masked LTSSembedding in (16) with an arbitrary carrier sarb and theoptimal transformation parameter copt, (vii) LTSS embeddingin (2) with an optimal carrier sopt and the optimal transfor-mation parameter copt, (viii) PN-masked LTSS embedding in(16) with an optimal carrier sopt and the optimal transformparameter copt. In all examined SS embedding schemes, MFis utilized to recover embedded bits.

Fig. 1 shows the recovery BER created by the embeddedmessage for above six embedding schemes as a function ofthe MS distortion D per-block.2

We recall that the per-block distortion is due to theembedding only (see (6)) and the variance of external noiseis fixed at σ2

n ¼ 3 dB in the experiment.3 It is demonstratedfrom Fig. 1 that the proposed PN-masked SS embedding

14 16 18 20 22 24 26 28 30 3210−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

Fig. 2. BER versus per-block distortion due to embedding (512�512Bridge, L¼63, external noise variance is fixed at σ2n ¼ 3 dB).

14 16 18 20 22 24 26 28 30 3210−8

10−7

10−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

Fig. 3. BER versus per-block distortion due to embedding (512�512Boat, L¼63, external noise variance is fixed at σ2n ¼ 3 dB).

16 18 20 22 24 26 28 30 32 34 3610−7

10−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

Fig. 4. BER versus per-block distortion due to embedding (averagefindings over a data set of more than 1500 images [42,43], L¼63, externalnoise variance is fixed at σ2n ¼ 3 dB).

0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 110−7

10−6

10−5

10−4

10−3

10−2

10−1

SSIM

BE

R

Conventional, s , w/o PN mask

Conventional, s , w. PN mask

Conventional, s , w/o PN mask

Conventional, s , w. PN mask

LTSS, s , c , w/o PN mask

LTSS, s , c , w. PN mask

LTSS, s , c , w/o PN mask

LTSS, s , c , w. PN mask

Fig. 5. BER versus SSIM (average findings over a data set of more than1500 images [42,43], L¼63, external noise variance is fixed at σ2n ¼ 3 dB).

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–2522

schemes have almost the same BER performance as theircorresponding conventional SS embedding counterparts (seethe same color curves). This observation means that the use ofPN-sequence mask would not notably affect the BER perfor-mance of SS embedding, i.e., no obvious performance loss. InFigs. 2 and 3, we repeat the same experiment for gray-scale512�512 “Bridge” and “Boat” images and the same conclu-sions can be drawn. Now we examine the average perfor-mance of the proposed PN-masked SS embedding algorithmsover a large image database. The experimental image data setconsists of more than 1500 8-bit gray-scale photographicimages ([42,43] combined) which have great varieties (e.g.outdoor/indoor, daylight/night, natural/man-made) and differ-ent sizes. Recovery performance plots are given in Fig. 4.Similar conclusion can be drawn as in previous individualimage host experimentations.

To assess the perceptual quality of the data-embeddedimages, in Fig. 5 we repeat the experiment in Fig. 4, but usestructural similarity (SSIM) index [44] as perceptual criterioninstead of the per-block distortion. The proposed similar

conclusion can be drawn. To further illustrate the perceptualdistortion due to the SS embedding, in Fig. 6 we displaysome meaningful images. Fig. 6(a) is the original clean512�512 Baboon image, Fig. 6(b) is the stego image withdata embedded by optimal LTSS, and Fig. 6(c) is the stegoimage with data embedded by proposed optimal PN-maskedLTSS. From Fig. 6, we do not observe obvious perceptualdistortion. All these results illustrate that the proposed PNmask method would not introduce any additional distortionthan the conventional SS embedding. To show the suitabilityof the proposed PN masked SS embedding scheme fordifferent transform domains, in Fig. 7 we repeat the sameexperiment of Fig. 1 but data are embedding on the discretewavelet transform (DWT) domain. Comparing Fig. 7 withFig. 1, it can be found that the performance is almost thesame which means that the proposed PN masked SSembedding scheme is also suitable for other transformdomains. To evaluate robustness of the proposed PN-

14 16 18 20 22 24 26 28 30 3210−7

10−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

Conventional, s , w/o PN mask

Conventional, s , w. PN mask

Conventional, s , w/o PN mask

Conventional, s , w. PN mask

LTSS, s , c , w/o PN mask

LTSS, s , c , w. PN mask

LTSS, s , c , w/o PN mask

LTSS, s , c , w. PN mask

Fig. 7. BER versus per-block distortion due to embedding (512�512Baboon, embedding on DWT domain, carrier of length L¼63, externalnoise variance is fixed at σ2n ¼ 3 dB).

2 3 4 5 6 7 8 9 1010−8

10−7

10−6

10−5

10−4

10−3

10−2

10−1

100

External noise variance in dB

BE

R

Conventional, s , w/o. PN maskConventional, s , w. PN maskConventional, s , w/o. PN maskConventional, s , w. PN maskLTSS, s , c , w/o. PN maskLTSS, s , c , w. PN maskLTSS, s , c , w/o. PN maskLTSS, s , c , w. PN mask

Fig. 8. BER versus external noise level (512�512 Baboon, embedding onDCT domain, carrier of length L¼63, distortion is fixed at 20 dB).

Fig. 6. (a) Original 512�512 Baboon image, (b) image with 4096 bits data embedded by optimal LTSS method (per-block distortion is D¼ 20 dB), (c) imagewith 4096 bits data embedded by PN-masked optimal LTSS method (per-block distortion is D¼ 20 dB).

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–25 23

masked embedding schemes and investigate the effect of theexternal noise on the performance, in Fig. 8 we keep thedistortion at D¼ 20 dB and show the BER of each embeddingscheme as a function of external noise variance σn

2varying

from 0 dB to 10 dB. First of all, at different noise levels, theproposed PN mask schemes would not notably degrade theperformance comparing their counterparts. Moreover, as thenoise level increases, the performance of ðsopt; coptÞ embed-ding schemes (i.e. red curves) becomes worse. This meansthat, for a larger noise level, we need to increase the distortionto maintain the same BER performance. For other six embed-ding schemes (black, green, and blue curves), since theinterference from the host to the embedded signal is verylarge and dominant, performances of these six schemes donot notable change with varying noise level.

To demonstrate the security offered by PN-sequencemasking, we adopt IGLS-based algorithm [39] which hasbeen shown to have better performance than BSS-basedalgorithms. We keep the Baboon image as the host and dataare embedded with optimal carrier via (i) Circular Water-marking (CW) scheme proposed in [40] to enhance SSembedding security, (ii) non-PN-masked SS embedding,and (iii) PN-masked SS embedding. The intended receiverknows carrier and uses non-blind matched filter to recoverembedded bits. The unauthorized/adversary receiver has noknowledge of carrier and uses IGLS-based algorithm toblindly extract embedded bits. The BER performances ofintended receiver and unauthorized receiver are shown inFig. 9. If the intended receiver and the unauthorizedreceiver have similar BER, then the embedding schemehas a low security level; If the intended receiver has lowBER and the unauthorized receiver has much higher BER,then the embedding scheme has a high security level; If theunauthorized receiver has BER as high as 0.5 which meansthat what he received are garbage, then the embeddingscheme has perfect security. While the unauthorized recei-ver can successfully recover data hidden by CW scheme andnon-PN-masked SS embedding (almost the same BER as theintended receiver), PN-masked SS can fail the unauthorizedreceiver and provides perfect security for SS embedding. InFig. 10, we repeat the same experimentation with arbitrarycarrier and the same results can be found. In Fig. 11, instead

14 15 16 17 18 19 2010−8

10−7

10−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

CW, intended receiverCW, unauthorized receiverW/o. PN mask, intended receiverW/o. PN mask, unauthorized receiverW. PN mask, intended receiverW. PN mask, unauthorized receiver

Fig. 9. BER versus per-block distortion due to embedding (512�512Baboon, optimal carrier of length L¼63, external noise variance is fixed atσ2n ¼ 3 dB).

24 25 26 27 28 29 3010−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

CW, intended receiverCW, unauthorized receiverW/o. PN mask, intended receiverW/o. PN mask, unauthorized receiverW. PN mask, intended receiverW. PN mask, unauthorized receiver

Fig. 10. BER versus per-block distortion due to embedding (512�512Baboon, arbitrary carrier of length L¼63, external noise variance is fixedat σ2n ¼ 3 dB).

14 15 16 17 18 19 20 21 2210−7

10−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

CW, intended receiverCW, unauthorized receiverW/o. PN mask, intended receiverW/o. PN mask, unauthorized receiverW. PN mask, intended receiverW. PN mask, unauthorized receiver

Fig. 11. BER versus per-block distortion due to embedding (averagefindings over a data set of more than 1500 images [42,43], optimalcarrier of length L¼63, external noise variance is fixed at σ2n ¼ 3 dB).

14 16 18 20 22 24 26 28 30 32 34 3610−7

10−6

10−5

10−4

10−3

10−2

10−1

100

Distortion (per−block) in dB

BE

R

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

Conventional, s , w/o. PN mask

Conventional, s , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

LTSS, s , c , w/o. PN mask

LTSS, s , c , w. PN mask

Fig. 12. BER versus allowable per-message per-block distortion (averagefindings over a data set of more than 1500 images [42,43], L¼63, K¼16,external noise variance is fixed at σ2n ¼ 3 dB).

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–2524

of focusing on one image, we carry out the experiment withimage data set consisting of more than 1500 images. Theaverage performance results shown in Fig. 11 illustrate thatour proposed PN masking SS embedding can efficientlyimprove the security.

Finally, we consider the problem of multi-carrier SSembedding. We wish to hide K¼16 data messages with eachmessage having its own individual orthogonal embeddingcarrier. In Fig. 12 we examine the average performance of theproposed multi-carrier SS embedding algorithms over a largeimage database. The results in Fig. 12 reiterate the impor-tance of PN masking operation and jointly optimized carrierand host-data manipulation.

6. Conclusions

We considered the problem of embedding data in a digitalhost via SS embedding in an arbitrary transform domain. PN-sequence masked SS embedding was first proposed to

enhance the security and prevent illegitimate data extractionby unauthorized users. Then, adaptive optimal carrier designwas developed to maximize the output SINR with any giventotal distortion budget. To take these findings one step further,we also extended our effort to cover multi-carrier/multi-message embedding and developed carrier optimization algo-rithm which can, once again, improve the probability of erroras well as enhance the security. As a brief concluding remark,PN-sequence masked SS embedding is a very efficient securedata hiding approach to protect embedded data withoutinducing more distortion to host nor affecting recoveryperformance. Optimal carrier design utilizes SOS of host andcan significantly improve SINR and consequently reduce BER.

Acknowledgments

This work is supported by the Fundamental ResearchFunds for the Central Universities (Grant no. DUT14RC(3)

M. Li et al. / Signal Processing: Image Communication 39 (2015) 17–25 25

103), the Open Fund of Artificial Intelligence Key Labora-tory of Sichuan Province (Grant no. 2012RZJ01), theNational Natural Science Foundation of China (Grant no.61172109 and 61402079), and the Foundation for Innova-tive Research Groups of the NSFC (Grant no. 71421001).

References

[1] M.D. Swanson, M. Kobayashi, A.H. Tewfik, Multimedia data-embedding and watermarking technologies, Proc. IEEE 86 (June)(1998) 1064–1087.

[2] F.A.P. Petitcolas, R.J. Anderson, M.G. Kuhn, Information hiding: asurvey, Proc. IEEE (Special Issue on Identification and Protection ofMultimedia Information) 87 (July) (1999) 1062–1078.

[3] I.J. Cox, M.L. Miller, J.A. Bloom, Digital Watermarking, Morgan-Kaufmann, San Francisco, CA, 2002.

[4] G.C. Langelaar, I. Setyawan, R.L. Lagendijk, Watermarking digitalimage and video data: a state-of-the-art overview, IEEE SignalProcess. Mag. 17 (September) (2000) 20–46.

[5] N.F. Johnson, S. Katzenbeisser, A survey of steganographic techni-ques, in: S. Katzenbeisser, F. Petitcolas (Eds.), Information Hiding,Artech House, Norwood, MA, 2000, pp. 43–78.

[6] C. Cachin, An information-theoretic model for steganography, in:Proceedings of 2nd International Workshop on Information Hiding,Portland, OR, April 1998, pp. 306–318.

[7] G.J. Simmons, The prisoner's problem and the subliminal channel,in: Advances in Cryptology, Proceedings of CRYPTO'83, Plenum, NewYork, NY, 1984, pp. 51–67.

[8] J. Fridrich, Steganography in Digital, Media, Principles, Algorithmsand Applications, Combridge Univeristy Press, Combridge, UK, 2010.

[9] Y. Wang, P. Moulin, Perfectly secure steganography: capacity, errorexponents, and code constructions, IEEE Trans. Inf. Theory 54 (June)(2008) 2706–2722.

[10] L.M. Marvel Jr., C.G. Boncelet, C.T. Retter, Spread spectrum imagesteganography, IEEE Trans. Image Proc. 8 (August) (1999) 1075–1083.

[11] M. Kutter, S. Winkler, A vision-based masking model for spread-spectrum image watermarking, IEEE Trans. Image Proc. 11 (January)(2002) 16–25.

[12] J.R. Smith, B.O. Comiskey, Modulation and Information Hiding inImages, Lecture Notes on Computer Science 1174 (1996) 207–226.

[13] M. Wu, B. Liu, Data hiding in binary image for authentication andannotation, IEEE Trans. Multimed. 6 (August) (2004) 528–538.

[14] N. Nikolaidis, I. Pitas, Robust image watermarking in the spatialdomain, Signal Proc. 66 (May) (1998) 385–403.

[15] I.J. Cox, J. Kilian, F.T. Leighton, T. Shannon, Secure spread spectrumwatermarking for multimedia, IEEE Trans. Image Proc. 6 (December)(1997) 1673–1687.

[16] M. Barni, F. Bartolini, A. De Rosa, A. Piva, Optimum decoding anddetection of multiplicative watermarks, IEEE Trans. Signal Proc. 51(April) (2003) 1118–1123.

[17] M. Barni, F. Bartolini, A. De Rosa, A. Piva, A new decoder for theoptimum recovery of nonadditive watermarks, IEEE Trans. ImageProc. 10 (May) (2001) 755–766.

[18] M. Barni, F. Bartolini, A. De Rosa, A. Piva, Capacity of full frame DCTimage watermarks, IEEE Trans. Image Proc. 9 (August) (2000)1450–1455.

[19] J. Hernandez, M. Amado, F. Pérez-González, DCT-domain watermarkingtechniques for still images: detector performance analysis and a newstructure, IEEE Trans. Image Proc. 9 (January) (2000) 55–68.

[20] C. Qiang, T.S. Huang, An additive approach to transform-domaininformation hiding and optimum detection structure, IEEE Trans.Multimed. 3 (September) (2001) 273–284.

[21] C.B. Adsumilli, M.C.Q. Farias, S.K. Mitra, M. Carli, A robust errorconcealment technique using data hiding for image and videotransmission over lossy channels, IEEE Trans. Circuits Syst. VideoTechnol. 15 (November) (2005) 1394–1406.

[22] J.J. Eggers, B. Girod, Quantization effects on digital watermarks,Signal Proc. 81 (February) (2001) 239–263.

[23] P. Moulin, M.K. Mihçak, A framework for evaluating the datahidingcapacity of image sources, IEEE Trans. Image Proc. 11 (September)(2002) 1029–1042.

[24] S. Pereira, S. Voloshynovskiy, T. Pun, Optimized wavelet domainwatermark embedding strategy using linear programming, In:Proceedings of SPIE Wavelet Applications Conference, vol. 4056,Orlando, FL, April 2000, pp. 490–498.

[25] P. Moulin, A. Ivanović, The zero-rate spread-spectrumwatermarkinggame, IEEE Trans. Signal Proc. 51 (April) (2003) 1098–1117.

[26] X.G. Xia, C.G. Boncelet, G.R. Arce, A multiresolution watermark fordigital images, In: Proceedings of IEEE International Conference onImage Processing (ICIP), vol. 1, Santa Barbara, CA, October 1997,pp. 548–551.

[27] C. Fei, D. Kundur, R.H. Kwong, Analysis and design of watermarkingalgorithms for improved resistance to compression, IEEE Trans.Image Proc. 13 (February) (2004) 126–144.

[28] H.S. Malvar, D.A. Florencio, Improved spread spectrum: a newmodulation technique for robust watermarking, IEEE Trans. SignalProc. 51 (April) (2003) 898–905.

[29] M. Gkizeli, D.A. Pados, M.J. Medley, SINR, bit error rate, and Shannoncapacity optimized spread-spectrum steganography, In: Proceedingsof IEEE International Conference on Image Processing (ICIP), Singa-pore, October 2004, pp. 1561–1564.

[30] M. Gkizeli, D.A. Pados, M.J. Medley, Optimal signature design forspread-spectrum steganography, IEEE Trans. Image Proc. 16 (Febru-ary) (2007) 391–405.

[31] L. Wei, D.A. Pados, S.N. Batalama, M.J. Medley, Sum-SINR/sum-capacity optimal multisignature spread-spectrum steganography,in: Proceedings of SPIE, Mobile Multimedia/Image Processing,Security, and Applications Conference, SPIE Defense & SecuritySymposium, vol. 6982, Orlando, FL, March 2008, pp. 0D1–0D10.

[32] A. Valizadeh, Z.J. Wang, Correlation-and-bit-aware spread spectrumembedding for data hiding, IEEE Trans. Inf. Forensics Secur. 6 (June)(2011) 267–282.

[33] S. Glisic, B. Vucetic, Spread Spectrum CDMA Systems for WirelessCommunications, Artech House, Norwood, MA, 1997.

[34] F. Cayre, C. Fontaine, T. Furon, Watermarking security: theory andpractice, IEEE Trans. Signal Proc. 53 (October) (2005) 3976–3987.

[35] L. Pérez-Freire, P. Comesana, J.R. Troncoso-Pastoriza, F. Pérez-Gon-zález, Watermarking security: a survey, LNCS Trans. Data HidingMultimed. Secur., 4300(October), 2006, 41–72.

[36] M. Barni, F. Bartolini, T. Furon, A general framework for robustwatermarking security, ACM J. Signal Proc. – Spec. Sect.: Secur. DataHiding Technol. 83 (October) (2003) 2069–2084.

[37] L. Pérez-Freire, F. Pérez-González, Spread-spectrum watermarkingsecurity, IEEE Trans. Inf. Forensics Secur. 4 (March) (2009) 2–24.

[38] M. Gkizeli, D.A. Pados, S.N. Batalama, M.J. Medley, Blind iterativerecovery of spread-spectrum steganographic messages, In: Proceed-ings of IEEE International Conference on Image Processing (ICIP),vol. 2, Genova, Italy, 11–14 September, 2005, pp. 1098–1101.

[39] M. Li, M. Kulhandjian, D.A. Pados, S.N. Batalama, M.J. Medley, J.D.Matyjas, On the extraction of spread-spectrum hidden data in digitalmedia, In: Proceedings of International Conference on Communica-tions (ICC), Ottawa, Canada, June 2012.

[40] P. Bas, F. Cayre, Achieving subspace or key security for WOA usingnatural or circular watermarking, In: Proceedings of ACM Multimediaand Security Workshop, Geneva, Switzerland, September 2006.

[41] D.G. Manolakis, V.K. Ingle, S.M. Kogon, Statistical and AdaptiveSignal Processing, McGraw-Hill, New York, 2000.

[42] G. Schaefer, M. Stich, UCID–an uncompressed colour image database, In:Proceedings of SPIE, Storage and Retrieval Methods and Applications forMultimedia, San Jose, CA, January 2004, pp. 472–480.

[43] USC-SIPI Image Database. Available: ⟨http://sipi.usc.edu/database/database.cgi?volume=misc⟩.

[44] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Simoncelli, Image qualityassessment: from error visibility to structural similarity, IEEE Trans.Image Process. 13 (April (4)) (2004) 600–612.