You are on page 1of 15

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO.

2, FEBRUARY 2007 391

Optimal Signature Design for


Spread-Spectrum Steganography
Maria Gkizeli, Member, IEEE, Dimitris A. Pados, Member, IEEE, and Michael J. Medley, Senior Member, IEEE

AbstractFor any given host image or group of host images and [1], [2], the broad common objective of most steganographic
any (block) transform domain of interest, we find the signature applications is a satisfactory tradeoff between hidden message
vector that when used for spread-spectrum (SS) message embed- resistance to noise/disturbance, information delivery rate, and
ding maximizes the signal-to-interference-plus-noise ratio (SINR)
at the output of the corresponding maximum-SINR linear filter. host distortion for concealment purposes. Message embedding
We establish that, under a (colored) Gaussian assumption on the can be performed either directly in the time (audio) or spa-
transform domain host data, the same derived signature minimizes tial (image) domain [3][7] or in a transform domain (for ex-
host distortion for any target message recovery error rate and ample, for images we may consider full-frame discrete Fourier
maximizes the Shannon capacity of the covert steganographic link. transform (DFT) [8][12], full-frame discrete cosine transform
Then, we derive jointly optimal signature and linear processor
designs for SS embedding in linearly modified transform domain (DCT) [13], block DFT or DCT [14][18], or wavelet trans-
host data and demonstrate orders of magnitude improvement forms [19][21]). Direct embedding in the original host signal
over current SS steganographic practices. Optimized multisigna- domain may be desirable for system complexity purposes, while
ture/multimessage embedding in the same host data is studied as embedding in a transform domain may take advantage of the
well. particular transform domain properties [23].
Index TermsAuthentication, covert communications, data In this paper, we focus our attention on transform domain
hiding, distortion, linear filters, signal-to-interference-plus-noise spread-spectrum (SS) embedding methods for image steganog-
ratio (SINR), spread spectrum, steganography, watermarking. raphy. In a broad sense, any steganographic system for which
the secret signal is spread over a wide range of host image fre-
quencies can be referred to as an SS embedding system. Once
I. INTRODUCTION
the transform embedding domain is selected, the hidden mes-
TEGANOGRAPHY is the process of embedding a secret sage can be applied to the host data through an additive [8], [12],
S digital signal (hidden message) in another digital signal
(image or audio) called cover or host. Unlike general
[14][17], [22] or multiplicative [9][11] rule. In the literature,
additive SS embedding methods either directly apply (add) the
digital watermarking applications, steganography attempts to message/watermark to several host coefficients [17] or in direct
establish covert communication between trusting parties and analogy to SS digital communications systems use an amplitude
imposes the requirement of concealing the existence of the modulated signature to deposit one information symbol across
embedded message. a group of host data coefficients [8], [12], [14], [16] or a linearly
As in any watermarking operation, the first step in the de- transformed version of the host data coefficients [22].
sign of a steganographic system is to determine the embed- SS embedding algorithms for blind image steganography
ding process. This is a crucial task since host-carrier proper- (that is, hidden message recovery without knowledge of the
ties, message detector design and performance depend directly original image) have been based on the understanding that the
on the way the message is inserted in the host data. While each host signal acts as a source of interference to the secret message
watermarking application has its own individual requirements of interest. Yet, it should also be understood that this inter-
ference is known to the message embedder. Such knowledge
can be exploited appropriately to facilitate the task of the blind
receiver at the other end and minimize the recovery error rate
Manuscript received December 21, 2004; revised July 2, 2006. This work for a given host distortion level, minimize host distortion for a
was supported by the U.S. Air Force Research Laboratory under Agreements
F30602-03-2-0023 and FA8750-04-1-0091. This paper was presented in part
given target recovery error rate, maximize the Shannon capacity
at the IEEE International Conference on Image Processing (ICIP), Singapore, of the covert steganographic channel, etc. Indeed, if we were to
October 2004. The associate editor coordinating the review of this manuscript place the steganography application in an information theoretic
and approving it for publication was Dr. Benoit Macq.
M. Gkizeli was with the Department of Electrical Engineering, State Uni-
context, it could be viewed as a communications-with-side-in-
versity of New York at Buffalo, Buffalo, NY 14260 USA. She is now with the formation problem [24][26]. Optimized embedding methods
Department of Electronics, Technological Education Institute of Crete, Chania, can facilitate host interference suppression at the receiver side
73133 Greece (e-mail: mgkizeli@chania.teicrete.gr).
D. A. Pados is with the Department of Electrical Engineering, State Uni-
when knowledge of the host signal is adequately exploited in
versity of New York at Buffalo, Buffalo, NY 14260 USA (e-mail: pados@ system design.
eng.buffalo.edu). In this paper, for any given image (or set of images), (block)
M. J. Medley is with the Department of Electrical Engineering, State Uni-
versity of New York Institute of Technology, Utica, NY 13504 USA (e-mail:
transform domain, and host bins, we derive the additive em-
Michael.Medley@sunyit.edu). bedding signature that maximizes the output signal-to-interfer-
Digital Object Identifier 10.1109/TIP.2006.888345 ence-plus-noise ratio (SINR) of the linear maximum-SINR re-
1057-7149/$25.00 2007 IEEE
392 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 2, FEBRUARY 2007

ceiver filter.1 We show that under a (colored) Gaussian assump-


tion on the host bins, this same signature minimizes the receiver
bit-error rate (BER) at any mean-square (MS) image distortion
level, minimizesconverselythe MS image distortion at any
target BER, and maximizes the Shannon capacity of the covert
link. We then generalize our findings and present a novel joint
signature and linear host data projection optimization scheme.
In this present work, we consider only scalar parametrized host
data projection as in [22]. Finally, we extend signature-only as
well as joint signature and host-projection optimization to mul-
tiuser (multiple-signature) embedding. Our emphasis is directed
primarily toward low complexity, sequential, conditional mul-
tiuser optimization.
The rest of the paper is organized as follows. Section II
presents the core signature and embedding optimization results.
These results are generalized to multiple signature embedding
in Section III. Section IV is devoted to experimental studies
and comparisons. A few concluding remarks are drawn in
Section V.

II. SIGNATURE OPTIMIZATION FOR SS EMBEDDING


In this section, we develop an optimized SS steganographic
system. To draw a parallelism with conventional SS communi-
cations systems, in SS watermark (message) embedding the wa-
termark can be regarded as the SS signal of interest transmitted
through a noisy channel (the host). The disturbance to the SS
signal of interest is the host data themselves plus potential ex-
ternal noise due to physical transmission of the watermarked Fig. 1. (a) Baboon image example H2f 0; 1; . . . ; 255g . (b) Host
data and/or processing/attacking. The purpose of the watermark 2
data autocorrelation matrix (8 8 DCT, 63-bin host).
detector is to withstand the influence of the total end-to-end dis-
turbance and recover the original hidden message.

A. Signal Model and Notation


Consider a host image that is to be wa- (1)
termarked where is the image alphabet and
is the image size in pixels. Fig. 1(a) shows a grayscale ba- where denotes statistical expectation (here, with respect
boon image example in . to over the given image ) and is the transpose operator.
Without loss of generality, the image is partitioned into It is easy to verify that in general , where
local blocks of size pixels. Each block is the size-L identity matrix; that is, is not constant-value
is to carry one hidden information bit diagonal or white in field language. For example, 8 8 DCT
, respectively. Embedding is with 63-bin host data formation (exclude only the dc coefficient)
performed in a real 2-D transform domain . After transform for the baboon image in Fig. 1(a) gives the host autocorrelation
calculation and conventional zig-zag scanning vectorization, matrix in Fig. 1(b).
we obtain . From the
transform domain vectors we choose a fixed subset B. Signature Optimization
of coefficients (bins) to form the final host
vectors (for example, it is common Consider direct additive SS embedding of the form
and appropriate to exclude the dc coefficient, , from
(2)
the host ).
The autocorrelation matrix of the host data is an important
where is the message bit embedded in the host
statistical quantity for our developments and is defined as fol-
is the bit amplitude, , is the (normal-
lows:
ized) embedding signature to be designed, and
1The problemand solutionparallels the eigen-signature design for represents potential external white Gaussian noise2 of variance
code-division multiple-access (CDMA) wireless communications [28]. In SS
steganography, however, the disturbance (transform-domain host) is readily 2Additive white Gaussian noise is frequently viewed as a suitable model for
available at the embedder for manipulation and statistical exploitation and in quantization errors, channel transmission disturbances, and/or image processing
contrast to the CDMA problem may be plausibly characterized as Gaussian. attacks.
GKIZELI et al.: OPTIMAL SIGNATURE DESIGN 393

. The mean-squared distortion of the original image due to is the optimum (minimum probability of error) bit detector [30]
the embedded data only is with probability of error

(3)

With signal of interest and total disturbance in (2), (10)


the linear filter that operates on and offers maximum SINR at
its output [27] is where . We see that is a
monotonically decreasing function of . If we now view
as a function of the embedding signature , then from
Proposition 1 and (10), when the probability of error of
(4) the optimum detector is minimized to

The exact maximum output SINR value attained is (11)

and optimum detection reduces to

(12)
(5) Conversely, for any preset probability of error level
minimizes the host distortion due to the hidden message to
We propose to view as a function of the embedding
signature , and identify the signature that maxi- (13)
mizes the SINR at the output of the maximum SINR filter. Our
findings are presented in the form of a proposition below that We conclude that under a Gaussian transform-domain host
parallels the developments for code-division-multiple-access data assumption, the minimum eigenvector of the host data
(CDMA) codeword optimization in [28] and [29]. The proof is autocorrelation matrix, when used as the embedding signa-
straightforward and omitted. ture, allows message recovery with the minimum possible
Proposition 1: Consider additive SS embedding according to BER and it does so by trivial signature
(2). Let be eigenvectors of in (1) with cor- (eigenvector) matched filter detection. Conversely, the mes-
responding eigenvalues . For any hidden sage-induced image distortion is minimized for any given
message-induced distortion level , a signature that maximizes target BER.
the output SINR of the maximum SINR filter is If necessary, further BER improvements below
for any fixed distortion can be attained
(6) via error correcting coding of the information bits at the
expense of reduced information bit payload. The maximum
When , the output SINR is maximized to possible host payload in bits that still allowstheoretically
for asymptotically large number of image blocks message
recovery with arbitrarily small probability of error is where
is the Shannon capacity of the covert link
(7) in bits per embedding attempt. We recall that identifies
the information conveyed about the input variable by the
and maximum SINR data filtering simplifies to received vector and denotes the input variable probability
distribution function. For Gaussian host data
(8) and average image distortion constraint , we can calculate
[31]
(14)
In summary, Proposition 1 says that the minimum eigen-
vector of the host data autocorrelation matrix, when used as where is the determinant operator. We can show that the
the embedding signature, sends the output SINR to its max- signature choice is also optimal in terms of capacity,
imum possible value . At the same time, maximum i.e., maximizes the capacity of the covert link, and, there-
SINR filtering becomes plain signature (eigenvector) matched fore, the maximum allowable payload for the host vectors
filtering. . The result is presented in the form
If, in addition, we are allowed to assume that is Gaussian, of Proposition 2 below whose proof is given in the Appendix.
, then Proposition 2: Consider additive SS embedding by (2) with
. Let be eigenvectors of with
(9) corresponding eigenvalues . For any
394 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 2, FEBRUARY 2007

hidden message-induced distortion level , the signature (signature , parameter ) that maximizes the output SINR of
maximizes the covert channel capacity to the maximum SINR filter is

(20)
bits per symbol embedding (15)
and

C. Signature Optimization for Linearly Transformed Host (21)


Data
When and , the output SINR is maximized to
In an effort to further reduce host distortion for a given target
probability of error during message recovery, we modify the
previous developments and attempt to steer the host data vectors (22)
away from the embedding signature using an operator of
the form [22] where both the parameter and
and maximum SINR data filtering simplifies to
the signature , are to be designed.3 In parallel
to (2), the composite signal now becomes

(16)

and the mean-squared distortion due to the embedding operation (23)


only is
The target distortion is achieved when the bit amplitude is set
at .
(17) Proposition 3 shows that the optimum signature assign-
ment is still the minimum eigenvector of and maximum
We observe that, in contrast to (3), the distortion level is con- SINR filtering still reduces to plain signature (eigenvector)
trolled not only by but by and , as well. With signal of matched filtering. The optimum selection of depends on the
interest and total disturbance in (16), the minimum eigenvalue of , the noise variance , and
linear filter that operates on and offers maximum SINR at its the target distortion level . The optimum pair
output is allows the output SINR to attain its maximum possible value
. From (21), it is interesting
to note that as the allowed distortion and
the embedding scheme converges to the conventional scheme
of Section II-B. As , the linear operator
converges to the orthogonal projector and the allowable distor-
(18)
tion is budgeted more toward host-data modification and less
The exact maximum output SINR value attained is toward the embedding bit amplitude.
If we assume that is Gaussian, , then
is the optimum bit detector with probability
of error
(19)

In the following, we look at as a function of both the


embedding signature and the parameter , and (24)
identify the signature and parameter values that jointly maxi-
mize the SINR at the output of the maximum SINR filter. Our As in the plain additive SS embedding scenario, if is Gaussian
findings are presented in the form of a proposition below. The then the probability of error is a monotonically decreasing func-
proof is given in the Appendix. tion of . The pair which maximizes
Proposition 3: Consider additive SS embedding in linearly the output SINR of the maximum-SINR filter is also minimizing
transformed host data by (16). Let be eigenvec- the probability of error of the optimum detector to
tors of with corresponding eigenvalues
. For any hidden message-induced distortion level , a pair
(25)
3If c = 0, we revert to the developments of Section II-B. If c = 1; I 0
css becomes the projector orthogonal to s. In this work, as in [22], we confine
ourselves within this class of scalar parametrized linear operators merely for
mathematical convenience and tractability. and optimum detection reduces to .
GKIZELI et al.: OPTIMAL SIGNATURE DESIGN 395

We can show that for Gaussian transform-domain host data where denotes the exclude- data autocorrelation matrix,
, the covert channel capacity is given by that is the autocorrelation matrix of the disturbance to message
defined as

(26)

Then, we can prove that the assignment of Proposi-


tion 3 is also optimal in terms of capacity. This result is summa-
rized in the following proposition whose proof is given in the
Appendix.
Proposition 4: Consider additive SS embedding by (16) with
. Let be eigenvectors of with
corresponding eigenvalues . For any (32)
hidden message-induced distortion level , the pair
of (20) and (21) maximizes the covert channel capacity to The exact maximum output SINR value attained is

(33)

(27) As in Section II for single-message embedding, we propose


to view as a function of the embedding signature
, and identify the signature vector that max-
bits per symbol embedding. imizes the SINR value. Our findings are presented in the form
In the following section, we consider the problem of data em- of Proposition 5 below and parallel the developments of Propo-
bedding in one host with multiple signatures. sition 1 for the single-message case.
Proposition 5: Consider additive SS embedding by (28).
Let be eigenvectors of in (32) with cor-
III. MULTISIGNATURE EMBEDDING responding eigenvalues . For any
message-induced distortion level , a signature that maxi-
We may generalize the signal model in (2) to cover multisig- mizes the output SINR of the maximum SINR filter
nature/multimessage embedding of the form is

(34)
(28)
When , the output SINR is maximized to
where bits coming potentially from distinct
(35)
messages, are embedded simultaneously in with corre-
sponding amplitudes and embedding signatures
. Thus, the contribution and maximum SINR data filtering simplifies to
of each individual embedded message bit to the composite .
watermarked signal is and the mean-squared distortion In summary, Proposition 5 says that the minimum eigen-
to the original host data due to the embedded message alone vector of the disturbance autocorrelation matrix when used as
is the embedding signature allows the output SINR to attain its
maximum possible value . At the same time,
(29)
maximum SINR filtering becomes plain signature (eigenvector)
Under a statistical independence assumption across message matched filtering.
bits, the mean-squared distortion of the original image due to In the context of multisignature optimization, for fixed-bit
the total multimessage insertion is amplitude values and arbitrary signature initializa-
tion , consider repeated
applications of Proposition 5 for . In such an
(30) eigen-update signature cycle each signature is replaced by the
minimum-eigenvalue eigenvector of the disturbance autocorre-
lation matrix as seen by the message corresponding to that sig-
With signal of interest and total disturbance nature. Once all signatures are updated, a second update cycle
in (28), the linear filter that oper- may begin. The whole procedure may continue for a predeter-
ates on and offers maximum SINR at its output is mined number of cycles or until convergence

(31) (36)
396 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 2, FEBRUARY 2007

It can be proven that convergence of (36) is guaranteed and, as and maximum SINR data filtering simplifies to
shown in a CDMA literature context [28], at each cycle the gen-
(41)
eralized total-squared correlation of the signature set

where , is nonincreasing.4 If, in addition to the assumptions of Proposition 6, we are al-


If we consider channel coding the message bits before embed- lowed to assume that , then the optimum detector
ding and assume that the host in (28) is Gaussian, the stegano- for the bit of interest is with probability
graphic system determined by the signature matrix is a special
of error . As a simple il-
case of the -user Gaussian multiple access channel with av-
lustration of the use of Proposition 6 for conditionally optimal
erage input distortion constraints [31], [33]. For such a channel,
multisignature design, suppose that we want to embed in the
the sum capacity (defined as the maximum sum of mes-
host data vector message bits
sage coding rates at which messages can be recovered reliably
with fixed corresponding amplitudes (mean-
[33], [34]) is a reasonable criterion of quality for the signature
squared distortions ) and
set and equals
signatures to be chosen. By Proposition 1 of Sec-
tion II, we set equal to the bottom eigenvector of
. Given and under the constraint that we search for
an orthogonal to , by Proposition 6, we design
(37) which is the next available eigenvector of from the bottom.
Given and under the constraint that we
Minimization of the metric translates to maximization search for an orthogonal to both and , by Proposition
of [34]. However, decrease in at each cycle of the 6, we assign . We continue this way until the final
algorithm in (36) does not necessarily imply increase of assignment . Once again, a welcome side ef-
as seen for instance in [35][38] via binary signature examples. fect of this conditionally SINR optimal signature design proce-
Apart from global optimality limitations, the multicycle mul- dure is that the maximum SINR receiver for each message bit
tisignature optimization procedure in (36) requires recalculation , simplifies to a matched filter and requires
of the disturbance autocorrelation matrix and eigen decompo- no knowledge of other system parameters.
sition at each step of each cycle. A simple low-cost alterna- For (fixed) unequal embedding amplitude values
tive to (36) could be a conditionally optimal single-cycle design , the exact order by which the eigenvectors of
method based on the following proposition. are drawn to become signatures is important if we consider
Proposition 6: Consider additive SS embedding according to the sum capacity of the steganographic system. We can show
that a necessary condition for a maximum sum capacity
(38) solution under the constraint of eigenvector assignment is that
the ordering of the bit amplitudes be inversely proportional to
the ordering of the eigenvalues of the corresponding signature
Let be eigenvectors of in (1) with corre- eigenvectors. This statement is given below in the form of a
sponding eigenvalues . If the signatures lemma. The proof can be found in the Appendix.
are eigenvectors of , then for Lemma 1: Consider additive SS embedding by (28) with
any message-induced distortion level , a signa- and let be eigenvectors of with cor-
ture that maximizes the output SINR of the maximum responding eigenvalues . Without loss of
SINR filter for the bit of interest subject to the constraint generality assume that . Then
is

(39)
If we generalize our approach and view the individual am-
plitudes/distortions as design parameters themselves, then we
where is the minimum-eigenvalue eigenvector of avail- can search for the optimal amplitude assignment that maximizes
able. sum capacity subject to a total allowable distortion constraint
When , the output SINR is (conditionally) maxi- . We derive the optimal amplitude values in the
mized to lemma below. The proof is given in the Appendix.
Lemma 2: Consider additive SS embedding by (28). Let
(40) be eigenvectors of with corresponding
eigenvalues . If the signatures as-
4Yet, there is no guarantee that TSC (S) will converge to its minimum pos- sociated with the message bits are distinct eigenvectors of
sible value (global minimum) [29], [32]. then the sum
GKIZELI et al.: OPTIMAL SIGNATURE DESIGN 397

capacity is maximized subject to an expected total distortion Let be eigenvectors of in (1) with corre-
constraint if sponding eigenvalues . Assume that
, are all eigenvectors of and is the index of
the minimum-eigenvalue eigenvector of available. For any
(42) given message-induced distortion level , an
pair that maximizes the output SINR of the maximum SINR
where and is the KuhnTucker coefficient filter subject to the constraint is
[31] chosen such that the distortion constraint
is met.
To find the necessary parameter value in (42), we suggest
to arrange the participating eigenvalues in
ascending order . Then

(43)

where the cutoff index is the greatest integer in


and . (47)
The optimal message amplitude/distortion allocation solution of and
Lemma 2 can be viewed as a power waterfilling procedure [31]
in the eigen domain of the host.
Finally, as the last technical development in this paper, we ex-
amine the possibility of carrying out multisignature embedding
in linearly transformed host data. We assume that the host data
vector is linearly transformed by an operator of the form (48)
where and When and , the output SINR is (condi-
, are the parameters and signatures to be designed. tionally) maximized to
The final composite signal is
(49)
(44)

and maximum SINR data filtering simplifies to .


and the mean-squared distortion induced by each individual If in addition to the assumptions of Proposition 7, we are al-
message , is lowed to assume that , then the optimum detector
for the bit of interest is with probability
of error
(45)
(50)
With signal of interest , the autocorrelation ma-
trix of the disturbance is
As an illustrative example of the use of Proposition 7, suppose
. Unfortunately, that we would like to embed in the host data vector
in contrast to (38) for multisignature embedding in nontrans- message bits with individual corresponding
formed data, remains a function of (as well as ). In mean-squared host distortion . We first design
this context, unconditionally optimal multisignature multicycle the system parameters , and for message bit in the
optimization along the lines of (36) is practically an unrealistic absence of any other message in the host image. According to
objective. Instead, we suggest to design sequentially the am- Proposition 7 , the optimal parameter selection is
plitudes , parameters , and signatures of the embedded
messages such that the output SINR
is conditionally maximized given all past fixed embeddings
(single-cycle optimization). Our developments are presented in and by (45) . Next, we proceed with the
the form of Proposition 7 below whose proof is given in the second message bit and optimize and subject to the
Appendix. desired distortion level and the constraint . Since
Proposition 7: Consider additive SS embedding according to is already an eigenvector of , Proposition 7 offers
the equation shown at the bottom of the next page, and by (45)
. We continue calculating signatures ,
parameters , and amplitudes as above, always subject to the
(46) desired distortion and orthogonality between the signature to
be designed and all other prior signatures. Provided that ,
398 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 2, FEBRUARY 2007

the final set of designed parameters is shown in the equation at


the bottom of the page, and .
Optimal distortion allocation for sum capacity maximization
in multimessage embedding in linearly transformed host data
subject to the constraint that all messages are assigned distinct
eigenvectors of , and subject to a total distor-
tion constraint
is a joint optimization problem with
respect to and . We suggest an iterative solution approach
based on Proposition 7 above and Lemma 2 presented earlier in
this section. We initially fix the distortions induced by each mes-
sage and find the optimum parameters according to Propo-
sition 7. Then, we perform optimum amplitude allocation ac-
cording to Lemma 2. Based on this allocation, we reevaluate all
by Proposition 7. We continue until convergence is observed.
In the following, we present extensive experimental results
that we obtained from the implementation of the developed
steganographic algorithms.

IV. EXPERIMENTAL STUDIES

To carry out an experimental study of the developments


presented in the previous sections, we consider first as a host
example the familiar grayscale 512 512 Boat image in
Fig. 2(a) that has been widely used in the pertinent literature.
We perform 8 8 block DCT single-signature embedding
over all bins except the dc coefficient. Hence, our signature
length is and we embed bits. For
the sake of generality, we also incorporate white Gaussian
noise of variance dB. Fig. 3 shows the recovery BER
under signature matched filter detection as a function of the
distortion created by the embedded message over the 0- to 2
Fig. 2. (a) Boat image example (512 512 grayscale). (b) Boat image after
20-dB (s ; c ) embedding of 4 Kbits and additive white Gaussian noise of
20-dB range for four different embedding schemes: a) SS variance 3 dB.
embedding with an arbitrary signature, b) SS embedding with
an arbitrary signature and optimized selection of the host data
transformation parameter as in [22] (known as improved additive white Gaussian noise disturbance in the block DCT
spread-spectrum or ISS), c) SS embedding with an optimal domain.
signature according to Proposition 1, and d) SS embedding In Fig. 4, we repeat the same experiment of Fig. 3 on the
with a jointly optimal signature and host data transforma- 256 256 grayscale Baboon image in Fig. 1(a) (signature
tion parameter according to Proposition 3. The length , hidden message of bits, and
demonstrated BER improvement of our joint signature and additive white Gaussian noise disturbance of variance 3 dB).
parameter optimization procedure, in particular, measures in Comparatively speaking (Boat versus Baboon host or Fig. 3
orders of magnitude. Fig. 2(b) shows the Boat image after 20 versus Fig. 4 results), message recovery from the Baboon host
dB -embedding of the 4 Kbit message and 3-dB appears to be a somewhat more difficult problem. Yet, the
GKIZELI et al.: OPTIMAL SIGNATURE DESIGN 399

Fig. 3. Bit-error rate versus host distortion (Boat image,  = 3 dB). Fig. 5. Bit-error rate versus host distortion (average findings over USC-SIPI
image database [41],  = 3 dB).

Fig. 4. Bit-error rate versus host distortion (Baboon image,  = 3 dB). Fig. 6. Capacity versus distortion (Baboon image,  = 3 dB).

proposed joint signature and host transformation parameter op- Next, we consider the problem of multisignature embedding.
timization scheme maintains a better than BER at 20-dB We keep the Baboon image as the host and wish to hide
host distortion and outperforms the proposed signature-only data blocks/messages of length 1024 bits each with each
optimization scheme by about eight orders of magnitude. block/message having its own individual embedding signature.
To address the need for experimental verification of highest Each message is allowed to cause the same expected distor-
credibility, we carried out the experiments of Figs. 3 and 4 over tion to the host . Therefore, for sta-
the whole USC-SIPI database [41] of 44 miscellaneous images. tistically independent messages, the total distortion to the host
Fig. 5 shows the average BER versus distortion for the database. is . As before, for the sake of generality, we
The average database findings are quite similar to the individual add to the host white Gaussian noise of variance 3 dB. We
Baboon (or Boat) results. study five different multisignature embedding schemes: a) Em-
In Fig. 6, we return and continue the work with the Ba- bedding with arbitrary signatures, b) ISS embedding [22], c)
boon host and plot the capacity versus distortion performance multicycle eigen-signature design by (36), d) conditional opti-
curves for the four embedders under consideration. We see, for mization by Proposition 6 (sequential assignment,
example, that at 20dB host distortion the jointly optimized ), and e) conditional optimization by Proposition
embedder offers 2.7 information bits payload per embedded 7 (sequential assignment, ). The results
symbol (suggesting implicitly the suitability of a higher than in Fig. 7 reiterate the importance of optimized host-data ma-
binary message alphabet). The bit payload number goes down nipulation in conjunction with signature optimization. In fact,
to 1.2 for signature-only optimization, 0.6 for ISS embedding under joint sequential optimization even the least fa-
[22], and 0.5 for arbitrary signature embedding. vored message outperforms in recovery BER the most
400 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 2, FEBRUARY 2007

Fig. 7. BER as a function of the per-message distortion D (Baboon image, Fig. 9. Sum capacity versus distortion (Baboon image, K = 15;  = 3 dB).
K = 15 messages of size 1024 bits each,  = 3 dB).

Fig. 8. Baboon image after multisignature embedding via Proposition 7 (K =


15 messages of size 1 024 bits each, per-message distortion 20 dB,  = 3 dB).

favored message under sequential signature-only Fig. 10. Sum capacity versus distortion (average findings over USC-SIPI
image database [41], K = 15;  = 3 dB).
optimization or multicycle signature-only design for
per-message distortion values above 18 dB.5 Fig. 8 shows the
Baboon image after embedding the fifteen messages (15 1024
numbers, for all embedders the database images have on the av-
bits) via joint sequential , optimiza-
erage about ten more information bits payload per 15-symbol
tion with 20-dB per-message distortion (31.8-dB total distor-
embedding than the Baboon image at 32-dB total distortion.
tion) and 3-dB variance additive white Gaussian noise.
Finally, in Fig. 9, we present sum capacity results when the V. CONCLUSION
two proposed schemes, sequential design and sequential
joint design, , employ waterfilling power We considered the problem of hiding digital data in a dig-
allocation (use of Lemma 2 alone or coupled use of Proposition ital host image via SS embedding in an arbitrary transform do-
7 and Lemma 2, correspondingly). We see, for example, that main. We showed that use of the minimum-eigenvalue eigen-
at 32-dB total distortion the waterfilled design offers vector of the transform domain host data autocorrelation ma-
information bit payload of about 36 bits per 15 symbols em- trix as the embedding signature offers the maximum possible
bedded, while the waterfilled design offers only SINR under linear filter message recovery and, conveniently,
about 12 bits per 15 symbols embedded. In Fig. 10, we repeat does so under plain signature correlation (signature matched
the experiment of Fig. 9 over the 44 images of the USC-SIPI filtering). If we allow ourselves the added assumption of (col-
database [41]. The relative average sum capacity behavior of ored) Gaussian transform-domain host data, then we see that the
the embedders remains the same over the database. In absolute above described system as a whole becomes minimum proba-
bility of error and maximum Shannon capacity optimal, as well.
5The performance of the most favored message (i = 1) for both proposed
conditional schemes, s ; c and s ; c = 0, remains the same as in the
single message case (see Fig. 4) since the (orthogonal) eigenvector signature To take these findings one step further, we examined SS em-
assignment completely avoids multimessage interference. bedding in transform-domain host data that are modified by a
GKIZELI et al.: OPTIMAL SIGNATURE DESIGN 401

parametrized projection-like linear operator. We found the joint Proof of Proposition 3: For a target distortion value , the
signature and parameter values under the optimality scenaria term in (19) equals and is maximized for
mentioned above. Conveniently, the jointly optimal signature is . We will show that the second term in (19),
still the minimum-eigenvalue eigenvector and the SINR optimal , is maximized by , as
linear filter at the receiver side is still the signature correlator. well. By the matrix inversion lemma
Yet, joint signature and parameter optimization was seen to offer
dramatic improvements in SINR, probability of error, and ca-
pacity.
Finally, we extended our effort to cover multisignature/mul-
timessage embedding. First, under signature-only optimization
we developed a computationally costly multicycle eigen-sig-
nature design scheme based on the disturbance autocorrelation and
matrices. The alternative suggestion based on the host data auto-
correlation matrix alone and sequential (conditional) eigen-sig- (53)
nature optimization is practically much more appealing. A wa-
terfilling amplitude assignment algorithm was developed as well
to maximize sum capacity under eigen-signature designs. All
multisignature findings were generalized to cover parametrized
projection-like modification of the host data with, once again,
dramatic improvements in probabilty of error or sum capacity.
As a brief concluding remark, image-adaptive signature(s) or (54)
signature(s)/parameter(s) optimization as described in this work
can be carried out over a set of host images (frames) if desired. Combining (53) and (54), we obtain
The only technical difference is the calculation of the host data
autocorrelation matrix which now has to extend over the whole
host set. As long as the cumulative host autocorrelation matrix is
not constant-value diagonal , significant gains are to be
collected over standard nonadaptive SS embedding techniques. (55)

APPENDIX
Proof of Proposition 2: We need to find that maximizes The derivative of the righthandside of (55) with respect to
in (14) subject to . Since is a strictly monotonic gives
function

(56)

Hence, is
(51) a decreasing function of . Yet,

Using the rank-one update rule [40] . Therefore

(57)

Since for any

Therefore

(58)
(52)
By direct differentiation of the final expression in (58) and root
selection, we obtain in (21).
402 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 2, FEBRUARY 2007

Proof of Proposition 4: For the signal model in (16), the with respect to and obtain
channel capacity is (59), shown at the bottom of the page. Ap-
plying a rank-one update [40] to the determinant in the numer- (63)
ator, we obtain
where is the KuhnTucker coefficient [31] chosen
such that . Substitution of the optimal ampli-
tude allocation of (63) in (61) gives the maximum attainable
sum capacity value
(60)
where , and are as in (43).
The result follows from the proof of Proposition 3. Proof of Proposition 7: The output SINR of the maximum
Proof of Lemma 1: The sum capacity of the channel in SINR filter for the signal model in (46) is
(28) provided that all signatures associated with the mes-
sage bits are distinct eigenvectors of
is given by

(61)

Then,
(64)
Let be the matrix with columns the eigen-
vectors of the diagonal matrix
with the eigenvalues of , and the number of available eigen-
vectors of that do not correspond to any .
Since , are eigenvectors of we can write
but

and

(65)

for and where and are diagonal matrices of dimension .


Consider the permutation matrix which partitions into
Proof of Lemma 2: To identify the set of amplitudes (or that contains the available eigenvectors of
equivalently distortions) that maximizes the concave sum ca- with the corresponding eigenvalues in descending order
pacity function in (61) subject to the distortion constraint , and that contains all used eigenvectors:
we differentiate the Lagrange functional . Let and
be the diagonal matrices that contain the eigenvalues of the
eigenvectors in and , respectively. Then

(62)
and (66)

(59)
GKIZELI et al.: OPTIMAL SIGNATURE DESIGN 403

where and are diagonal matrices of dimension subject to . We will show that both terms in (73),
. Since is an orthonormal basis of , we can write and
as , are maximized by the same . For the first term we
have

(67)
minimum eigenvector of (74)
where and . The constraint
is equivalent to . Hence Consider now the second term. By the matrix inversion lemma

(68)

The unit norm requirement, , implies that


or .
We want to find subject to
and . Substituting (75)
from (68) to the first term in (64), we obtain Using the matrix inversion lemma again
(69)
Using (65) and (68), the second term in (64) becomes

(70)
(76)
which, using (66), reduces to Combining (75) and (76), we have

(77)
(71)

We know that if a matrix is invertible and Differentiation of the righthandside of (77) with respect to
gives
exist, then see the equation shown at the bottom of
the page [27]. Hence, (71) reduces further to

(72)
Using (69) and (72), the initial optimization problem can be
written equivalently as
Hence

(73)
404 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 16, NO. 2, FEBRUARY 2007

is a decreasing function of . How- [14] J. Hernandez, M. Amado, and F. Perez-Gonzalez, DCT-domain wa-
termarking techniques for still images: Detector performance analysis
ever, and a new structure, IEEE Trans. Image Process., vol. 9, no. 1, pp.
where and are the minimum 5568, Jan. 2000.
[15] C. Qiang and T. S. Huang, An additive approach to transform-domain
and maximum eigenvalues in . Therefore information hiding and optimum detection structure, IEEE Trans.
Multimedia, vol. 3, no. 3, pp. 273284, Sep. 2001.
[16] C. B. Adsumilli, M. C. Q. Farias, S. K. Mitra, and M. Carli, A robust
error concealment technique using data hiding for image and video
transmission over lossy channels, IEEE Trans. Circuits Syst. Video
minimum eigenvector of (78) Technol., vol. 15, no. 11, pp. 13941406, Nov. 2005.
[17] J. J. Eggers and B. Girod, Quantization effects on digital watermarks,
We conclude [cf. (73), (74), (78), and (68)] that the SINR ex- in Signal Process., Feb. 2001, vol. 81, pp. 239263.
pression in (64) is maximized when is the minimum avail- [18] P. Moulin and M. K. Mihak, A framework for evaluating the data-
able eigenvector of for any . Hence hiding capacity of image sources, IEEE Trans. Image Process., vol.
11, no. 9, pp. 10291042, Sep. 2002.
[19] S. Pereira, S. Voloshynovskiy, and T. Pun, Optimized wavelet domain
watermark embedding strategy using linear programming, in Proc.
SPIE Wavelet Applications Conf., Orlando, FL, Apr. 2000, vol. 4056,
pp. 490498.
[20] P. Moulin and A. Ivanovic , The zero-rate spread-spectrum water-
marking game, IEEE Trans. Signal Process., vol. 51, no. 4, pp.
10981117, Apr. 2003.
[21] X. G. Xia, C. G. Boncelet, and G. R. Arce, A multiresolution water-
mark for digital images, in Proc. IEEE Int. Conf. Image Processing,
Nov. 1998, vol. 1, pp. 548551.
(79) [22] H. S. Malvar and D. A. Florncio, Improved spread spectrum: A new
modulation technique for robust watermarking, IEEE Trans. Signal
Process., vol. 51, no. 4, pp. 898905, Apr. 2003.
where is the minimum available eigenvalue of (equiva- [23] C. Fei, D. Kundur, and R. H. Kwong, Analysis and design of wa-
lently is the bottom element of ). The optimum termarking algorithms for improved resistance to compression, IEEE
value can be computed by setting the derivative of the last Trans. Image Process., vol. 13, no. 2, pp. 126144, Feb. 2004.
[24] M. H. M. Costa, Writing on dirty paper, IEEE Trans. Inf. Theory, vol.
expression in (79) equal to zero. The latter gives two candidate IT-29, no. 3, pp. 439441, May 1983.
values for . We select the value that maximizes (79), which [25] B. Chen and G. Wornell, Quantization index modulation: A class of
is the one in (48). provably good methods for digital watermarking and information em-
bedding, IEEE Trans. Inf. Theory, vol. 47, no. 5, pp. 14231443, May
2001.
REFERENCES [26] P. Moulin and J. A. OSullivan, Information-theoretic analysis of in-
[1] F. Hartung and M. Kutter, Multimedia watermarking techniques, formation hiding, IEEE Trans. Inf. Theory, vol. 49, no. 3, pp. 563593,
Proc. IEEE, vol. 87, no. 7, pp. 10791107, Jul. 1999. Mar. 2003.
[2] G. C. Langelaar, I. Setyawan, and R. L. Lagendijk, Watermarking dig- [27] D. G. Manolakis, V. K. Ingle, and S. M. Kogon, Statistical and Adap-
ital image and video data: A state-of-the-art overview, IEEE Signal tive Signal Processing. New York: McGraw-Hill, 2000.
Process. Mag., vol. 17, no. 9, pp. 2046, Sep. 2000. [28] C. Rose, S. Ulukus, and R. D. Yates, Wireless systems and interference
[3] L. Marvel and C. G. Boncelet, Spread spectrum image steganog- avoidance, IEEE Trans. Wireless Commun., vol. 1, no. 7, pp. 415428,
raphy, IEEE Trans. Image Process., vol. 8, no. 8, pp. 10751083, Jul. 2002.
Aug. 1999. [29] C. Rose, CDMA codeword optimization: Interference avoidance and
[4] M. Kutter and S. Winkler, A vision-based masking model for spread- convergence via class warfare, IEEE Trans. Inf. Theory, vol. 47, no.
spectrum image watermarking, IEEE Trans. Image Process., vol. 11, 9, pp. 23682382, Sep. 2001.
no. 1, pp. 1625, Jan. 2002. [30] H. L. Van Trees, Detection Estimation and Modulation Theory, Part
[5] J. R. Smith and B. O. Comiskey, Modulation and information hiding
I. New York: Wiley, 2001.
in images, Lecture Notes Comput. Sci., vol. 1174, pp. 207226, 1996.
[31] T. M. Cover and J. A. Thomas, Elements of Information Theory. New
[6] M. Wu and B. Liu, Data hiding in binary image for authentication and
York: Wiley, 1991.
annotation, IEEE Trans. Multimedia, vol. 6, no. 4, pp. 528538, Aug.
[32] P. Anigstein and V. Anantharam, Ensuring convergence of the
2004.
MMSE iteration for interference avoidance to the global optimum,
[7] N. Nikolaidis and I. Pitas, Robust image watermarking in the spatial
domain, Signal Process., vol. 66, pp. 385403, May 1998. IEEE Trans. Inf. Theory, vol. 49, no. 4, pp. 873885, Apr. 2003.
[8] I. J. Cox, J. Kilian, F. T. Leighton, and T. Shannon, Secure spread [33] M. Rupf and J. L. Massey, Optimum sequence multisets for syn-
spectrum watermarking for multimedia, IEEE Trans. Image Process., cronous code-division-multiple-access channels, IEEE Trans. Inf.
vol. 6, no. 12, pp. 16731687, Dec. 1997. Theory, vol. 40, no. 4, pp. 12611266, Jul. 1994.
[9] M. Barni, F. Bartolini, A. De Rosa, and A. Piva, Optimum decoding [34] P. Viswanath and V. Anantharam, Optimal sequences for CDMA
and detection of multiplicative watermarks, IEEE Trans. Signal under colored noise: A Schur-Saddle function property, IEEE Trans.
Process., vol. 51, no. 4, pp. 11181123, Apr. 2003. Inf. Theory, vol. 48, no. 6, pp. 12951318, Jun. 2002.
[10] , A new decoder for the optimum recovery of nonadditive water- [35] G. N. Karystinos and D. A. Pados, New bounds on the total squared
marks, IEEE Trans. Image Process., vol. 10, no. 5, pp. 755766, May correlation and optimum design of DS-CDMA binary signature sets,
2001. IEEE Trans. Commun., vol. 51, no. 1, pp. 4851, Jan. 2003.
[11] C. Qiang and T. S. Huang, Robust optimum detection of transform [36] C. Ding, M. Golin, and T. Klve, Meeting the Welch and Karystinos-
domain multiplicative watermarks, IEEE Trans. Signal Process., vol. Pados bounds on DS-CDMA binary signature sets, Des., Codes Cryp-
51, no. 4, pp. 906924, Apr. 2003. togr., vol. 30, pp. 7384, Aug. 2003.
[12] G. Csurka, F. Deguillaume, J. J. K. ORuanaidh, and T. Pun, A [37] P. Ipatov, On the Karystinos-Pados bounds and optimal binary
Bayesian approach to affine transformation resistant image and video DS-CDMA signature ensembles, IEEE Commun. Lett., vol. 8, no. 2,
watermarking, Lecture Notes Comput. Sci., vol. 1768, pp. 270285, pp. 8183, Feb. 2004.
2000. [38] G. N. Karystinos and D. A. Pados, The maximum squared correla-
[13] M. Barni, F. Bartolini, A. De Rosa, and A. Piva, Capacity of full frame tion, sum capacity, and total asymptotic efficiency of minimum total-
DCT image watermarks, IEEE Trans. Image Process., vol. 9, no. 8, pp. squared-correlation binary signature sets, IEEE Trans. Inf. Theory,
14501455, Aug. 2000. vol. 51, no. 1, pp. 348355, Jan. 2005.
GKIZELI et al.: OPTIMAL SIGNATURE DESIGN 405

[39] P. Viswanath, D. N. C. Tse, and V. Anantharam, Asymptotically op- Dimitris A. Pados (M95) was born in Athens,
timal waterfilling in multiple antenna multiple access channels, IEEE Greece, on October 22, 1966. He received the
Trans. Inf. Theory, vol. 47, no. 1, pp. 241267, Jan. 2001. Diploma degree in computer science and engi-
[40] C. D. Meyer, Matrix Analysis and Applied Linear Algebra.. Philadel- neering (five-year program) from the University
phia, PA: SIAM, 2000. of Patras, Patras, Greece, in 1989, and the Ph.D.
[41] USC-SIPI Image Database, [Online]. Available: http://sipi.usc.edu/ degree in electrical engineering from the University
database/database.cgi?volume=misc of Virginia, Charlottesville, VA, in 1994.
From 1994 to 1997, he held an Assistant Pro-
fessor position in the Department of Electrical and
Computer Engineering and the Center for Telecom-
munications Studies, University of Louisiana,
Lafayette. Since August 1997, he has been with the Department of Electrical
Engineering, State University of New York at Buffalo, where he is presently an
Associate Professor. His research interests are in the general areas of commu-
nication theory and adaptive signal processing with an emphasis on wireless
multiple access communications, spread-spectrum theory and applications,
coding, and sequences.
Dr. Pados is a member of the IEEE Communications, Information Theory,
Signal Processing, and Computational Intelligence Societies. He served as an
Associate Editor for the IEEE SIGNAL PROCESSING LETTERS from 2001 to 2004
and the IEEE TRANSACTIONS ON NEURAL NETWORKS from 2001 to 2005. He
received a 2001 IEEE International Conference on Telecommunications best
Maria Gkizeli (S00M03) received the Diploma paper award and the 2003 IEEE Transactions on Neural Networks Outstanding
degree in electrical engineering from the Democritus Paper Award for articles that he coauthored with his students.
University of Thrace, Xanthi, Greece, in 1997, and
the M.Sc. and Ph.D. degrees in mobile and satellite
communications from the University of Surrey,
Guildford, U.K., in 1998 and 2002, respectively. Michael J. Medley (S91M95SM02) received
From September 2002 to September 2004, she the B.S., M.S., and Ph.D. degrees in electrical
was a Postdoctoral Fellow with the Communications engineering from Rensselaer Polytechnic Institute,
and Signals Laboratory, Department of Electrical Troy, NY, in 1990, 1991, and 1995, respectively.
Engineering, State University of New York at Since 1991, he has been a Research Engineer
Buffalo, where she also held an appointment as an for the United States Air Force at the Air Force
Adjunct Instructor from August 2003 to May 2004. Since September 2004, she Research Laboratory, Rome, NY, where he has
has been a Visiting Lecturer with the Department of Electronics, Technological been involved in adaptive interference suppression
Education Institute of Crete, Chania, Greece. Her research interests lie in the and spread-spectrum waveform design. In 2002, he
areas of communication theory and systems, covert communications, wireless joined the State University of New York Institute
networks, satellite communications, and image processing. of Technology, Utica, as an Assistant Professor.
Dr. Gkizeli is a member of the IEEE Signal Processing Society and the Tech- His research interests include transform domain signal processing, adaptive
nical Chamber of Greece. filtering, steganography/steganalysis, and spread-spectrum communications.

You might also like