CONTEXT BASED WATERMARKING OF SECURE JPEG-LS

IMAGES

A. V. Subramanyam and Sabu Emmanuel

School of Computer Engineering, Nanyang Technological University, Singapore, Republic of Singapore

Keywords:

Compressed domain watermarking, Context based watermarking, JPEG-LS watermarking.

Abstract:

JPEG-LS is generally used to compress bio-medical or high dynamic range images. These compressed images

sometime needs to be encrypted for conﬁdentiality. In addition, the secured JPEG-LS compressed images

may need to be watermarked to detect copyright violation, track different users handling the image, prove

ownership or for authentication purpose. In the proposed technique, watermark is embedded in the context of

the compressed image while the Golomb coded bit stream is encrypted. The extraction of watermark can be

done during JPEG-LS decoding. The advantage of this watermarking scheme is that the media need not be

decompressed or decrypted for embedding watermark thus saving computational complexity while preserving

the conﬁdentiality of the media.

1 INTRODUCTION

The digital media is often embedded with watermarks

for various purposes such as ﬁngerprinting, copyright

violation detection, proof of ownership, authentica-

tion and tamper prooﬁng. And is often distributed in

a secured manner, for e.g in Digital Rights Manage-

ment (DRM) Systems or Clinical Information Sys-

tems (CIS). High dynamic range images in DRM

systems, and Biomedical images in CIS systems are

generally compressed using JPEG-LS. JPEG-LS is

commonly used to compress these images as the

compression efﬁciency is better and also facilitates

near-lossless compression than JPEG and JPEG2000

(Fang, 2009).

In DRM systems, the media providers often

compress the media and distribute it after encrypt-

ing the compressed media via multilevel distributors

(Thomas et al., 2009). In the distribution process the

media is transmitted from owners to the consumers

through different level of distributors. In this sce-

nario, the distributors are entitled only to distribute

the compressed encrypted media to the end user and

as such cannot access the plain content (un-encrypted

content). Distributors request the license server in

the DRM system to distribute the associated licence

containing the decryption keys to open the encrypted

content to the consumers. However, each distributor

sometime needs to watermark the content for traitor

tracing or proving the distributorship. Thus they need

to watermark in the compressed encrypted domain it-

self.

In biomedical ﬁeld, CIS often manages the pa-

tient related media in a compressed and conﬁdential

way in different Healthcare Establishments (Blobel,

2004). In such a communication, different patient

records can be shared between different profession-

als either for subject’s treatment purpose or for a case

study. Thus it becomes necessary to preserve the con-

ﬁdentiality of the distributed media, to prove the own-

ership and also track the user who is dealing with the

record. Therefore the watermark needs to be inserted

in the secured compressed media itself.

In this paper we focus on robust watermarking for

though a robust irreversible watermark can be used

for this purpose but it degrades the quality of the im-

age which is not desirable in the applications spec-

iﬁed above. Therefore it should be watermarked in

such a domain which does not create any undesir-

able distortion in the image. In literature some of

the irreversible algorithms have been proposed for en-

crypted domain watermarking in general, and also for

watermarking of JPEG-LS. In (Caldelli et al., 2006),

Caldelli et. al. proposed a watermarking scheme for

authentication using the prediction error of the pixels

during JPEG-LS encoding . Some techniques have

been proposed where watermarking is done on cer-

161

Subramanyam A. and Emmanuel S..

CONTEXT BASED WATERMARKING OF SECURE JPEG-LS IMAGES.

DOI: 10.5220/0003446201610166

In Proceedings of the International Conference on Signal Processing and Multimedia Applications (SIGMAP-2011), pages 161-166

ISBN: 978-989-8425-72-0

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

tain subbands/bitplanes while encrypting certain other

subbands/bitplanes (Cancellaro et al., 2008), (Lian

et al., 2006). Some of the encrypted domain algo-

rithms are proposed in (Bianchi et al., 2010), (Piva

et al., 2010), (Zhao et al., 2010), (Katzenbeisser et al.,

2008). However, these algorithms cause irreversible

distortion which might not be desirable especially in

case of biomedical images. Thus, here we propose

a technique for watermarking of encrypted JPEG-LS

compressed images, where the compressed bitstream

is encrypted while watermark is embedded using con-

texts (explained in section 2). This paper is organized

as follows. Section 2 gives the preliminaries. In sec-

tion 3, we present the embedding and detection pro-

cess. We discuss the experimental results in section 4.

Section 5 concludes the paper.

2 PRELIMINARIES

In this section we brieﬂy discuss the JPEG-LS com-

pression algorithm (Weinberger et al., 2000), and the

challenges in watermarking a compressed-encrypted

JPEG-LS image. In ﬁgure 1-a, an initialcontext is

computed in the Gradient block using the pixels a,

b, c and d. This initialcontext is a positive quantity

and is mapped into equal negative and positive values

using a classmap. For e.g., let the initialcontext be

denoted as Q ∈ [0,728] (Weinberger et al., 2000). Let

us denote Q

neg

∈ [−1, −364] and Q

pos

∈ [1, 364] and

let classmap f (.) denote the function which maps Q

to Q

neg

or Q

pos

or 0, i.e, f : Q → {Q

pos

neg

,0}. In

the regular mode, the current sample x is predicted in

Predictor block using a ﬁnite past pixel set compris-

ing of pixels a, b and c. Generally the ﬁxed prediction

has some bias in it which is canceled out using the

context in which pixel occurs. Finally, the prediction

residual between the corrected (bias canceled) pre-

dicted value and the original pixel is Golomb coded.

Further each pixel (of a M x N image) occurs under

px[ j]

∈ {Q

pos

neg

,0} while encoding, where px[ j]

denotes the pixel at position j, ∀ j = 1, 2, ..., M x

N and Q

px[ j]

denotes its context. Also, the Golomb

coded bit stream can be encrypted using a secure ci-

pher scheme such as RC4 (Schneier, 1996) for conﬁ-

dentiality, which can then be distributed.

Now, modifying such a randomized bitstream

to insert watermark faces certain limitations. The

Golomb coded bitstream is highly sensitive to bit er-

rors. Even if one bit gets corrupted, the rest of the

decoding can be compromised. Therefore, we choose

the context of the encoded image for embedding wa-

termark without requiring any decryption or decom-

pression. Next we discuss the proposed algorithm.

3 PROPOSED ALGORITHM

The proposed algorithm involves encryption of

Golomb coded bit stream while embedding water-

mark through context. Encryption is performed on the

output bit stream of JPEG-LS compression (ﬁgure 1-

a). The encryption algorithm that we propose to use

is RC4 cipher (Schneier, 1996). The encryption does

not result in any increase in compressed ﬁle size as

the encryption is done using a stream cipher. The en-

crypted bit stream along with classmap f and context

frequencies is sent to the watermark embedder (ﬁgure

1-b) where the watermark embedding takes place.

The ﬁrst step in watermarking process involves

ﬁnding the frequency or number of occurrences of

each context Q ∈ Qpos. Then according to the dy-

namic range of highest frequency contexts, the wa-

termark signal is divided into smaller length seg-

ments. Further Euclidean distance between the se-

lected contexts against the watermark segments is

computed. The contexts which give minimum dis-

tance will represent watermark information. Finally

contexts involved in computing Euclidean distance

are exchanged with the other contexts which does not

belong to highest occurring contexts. The main pur-

pose behind selecting highest occurring contexts is

that the higher the dynamic range, the more number of

watermark signal bits can be embedded. We describe

the watermark embedding procedure next.

3.1 Embedding Process

Let Q

∀ j = 0, 1, ....,L − 1 be the L highest occur-

ring contexts which are selected from Q

pos

for Eu-

clidean distance criteria. Here, L is chosen such that

L ≤ max(Q

pos

)/2, for e.g. in our case max(Q

pos

) =

364. Let freq

represent the frequency of Q

∀ j =

0,1,....,L − 1. Now, let us arrange freq

∀ j =

0,1,....,L − 1 in ascending order as shown in ﬁg-

ure 2 and different freq

’s are then grouped in range

, where range R

∀ i = 0, 1,....,K − 1 represent

the i

range, such that K ≤ L. Although it would

be easy to range the frequencies in uniform inter-

vals, this may not be a good idea as it will increase

detection error rate as the dynamic range of fre-

quencies in range will vary too much. Therefore,

we choose unequal range lengths, with range limits

[freq

,freq

i+1

− 1] ∈ R

∀ i = 0, 1, ....,K − 1. Let

∀ i = 0,1,....,K − 1 be the minimum number of

bits used to represent the elements present in i

range

∀ i = 0,1,....,K − 1. Let their be E

number of fre-

quency elements (and hence number of contexts) in

range R

∀ i = 0,1,....,K − 1. The values of the

parameters K and b

used for simulation purpose are

SIGMAP 2011 - International Conference on Signal Processing and Multimedia Applications

162

Gradients

Flat?

Predictor/

bias

cancellation

Context

Modeler

Golomb

coder

Run

Counter/

Coder

regular

run

Mode

Context

Input Image

(I)

Samples

c b d

a x

Image Samples

Classmap f

Qpos, Qneg, 0

Encrypted compressed

bitstream

Watermarked

context (Q

w

)

Decoder

S

d

Detection

Classmap f

w

Watermark

Detected (Y/N)

Watermark

embedding

Classmap f

Encrypted compressed bitstream

+ Classmap f

w

Image Samples

Predicted

Values

Prediction Errors

Classmap f +

(a) (b) (c)

Contexts (frequencies) +

I

w

JPEG-LS encoder Watermark embedder JPEG-LS decoder

EK

Compressed

stream

Encryption

Figure 1: (a)JPEG-LS encoding (b) Watermark embedding/detection (c) Decoding (watermarked context generation).

given in section 4.

Further watermark signal W of length L

is also

segmented. Corresponding to the number of fre-

quency elements in i

range R

, the length E

watermark bits is divided into E

number of segments,

denoted by w

i j

∀ j = 0,1,...,E

−1, with each segment

of length b

bits. This is done corresponding to each

range of ﬁgure 2. Also the watermark length L

chosen such that,

K−1

∑

i=1

(1)

Now the watermark can be represented as W =

||...||w

i j

∀ i = 0,1,....,K − 1; j = 0,1,...,E

− 1.

Then the Euclidean distance between each watermark

segment w

i j

∀ i = 0,1,....,K − 1; j = 0,1, ..., E

− 1

is computed against frequencies of the selected con-

texts.

argmin

i j,k

i j

,freq

) ∀ k = 0,1,....,L − 1 (2)

min

Qorder

= Q

(3)

where, min

Qorder

represents the context whose

frequency is closest to the watermark segment. Now,

the contexts represented in Q

are exchanged with the

contexts which are not involved in Euclidean distance

criteria for watermarking. Now without loss of gen-

erality, let Q

′

and Q

′′

denote these contexts such that

∪

′

∪

′′

= Q

pos

. Also let Q

, Q

′

and Q

′′

denote

the corresponding negative contexts of Q

, Q

′

and Q

′′

i.e., Q

= −Q

, Q

′

= −Q

′

and Q

′′

= −Q

′′

respec-

tively, such that Q

∪

′

∪

′′

= Q

neg

. Also Q

and

′

, and, Q

and Q

′

are chosen equilength.

In JPEG-LS encoding, initialcontext Q is mapped

through f as, f : Q → {Q

′

′′

′

′′

,0}.

Then, to embed watermark, this mapping is changed

by changing f to f

such that, f

: Q →

′

′′

′

′′

,0}. The mapping can be

done as, for e.g., Q

′

↔ min

Qorder

∀ j = 0,1,...., L −

1, where ’↔’ denotes one-to-one mapping i.e., each

element of Q

′

is mapped to an element of min

Qorder

However, this mapping is secret, can be randomized

using a secret embedding key EK, and does not affect

the watermarked image quality (explained in section

4.2). Thus Q

occurring during encoding is replaced

with Q

′

for watermarking. Q

and Q

′

should also be

exchanged correspondingly in order to maintain the

relation between Q

pos

and Q

neg

((Weinberger et al.,

2000)). Since only the classmap f and context fre-

quencies are used for embedding, the encrypted bit

stream remains intact. The compressed-encrypted bit

stream along with f

is then sent to the consumer (ﬁg-

ure 1-c). Next we discuss the detection algorithm.

3.2 Detection Process

The detection is done while decoding the Golomb

coded bit stream. The context map f

which is sent

to the end user generates watermarked context while

decoding. The extraction involves f

, EK, min

Qorder

and the decryption key. Then the frequencies of the

exchanged contexts in place of contexts occurring in

min

Qorder

are computed, which are then correlated

against different watermarks. The original watermark

gives the highest correlation, in case if it is present.

Now, min

Qorder

gives us the order in which the

contexts represent the watermark segments. Using

EK the context mapping is retrieved and, f

gives

us the corresponding context which should be looked

for in place of the contexts present in min

Qorder

. Us-

ing f

(which gives the mapping of the contexts),

the watermarked contexts can be retrieved as, Q

min

Qorder

→ Q

′

∀ i = 0,1,....,L − 1 i.e, searching

CONTEXT BASED WATERMARKING OF SECURE JPEG-LS IMAGES

163

R

K-1

Frequencies in ascending order

freq

R0

freq

R1

freq

RK-2

freq

RK-1

freq

RK

R

1

R

0

R

k-2

freq

R2

Figure 2: Frequencies arranged in ascending order and grouped.

for the contexts in Q

′

which are mapped to contexts

in min

Qorder

. The frequencies of Q

in the water-

marked contexts will then represent the frequencies of

min

Qorder

in the original contexts. Let the estimated

watermark be

W = ˆw

||...|| ˆw

∀ k = 0, 1, ....,L − 1,

where ˆw

is given by,

ˆw

= freq

∀ k = 0,1, ...., L − 1 (4)

The detection involves correlating the estimated

watermark

W against different watermarks

∀ i =

1,2,....,n

, where n

denotes the number of water-

marks. Correlation is given by,

corr(

W ,

) =

E[(

W − µ

)(

− µ

)]

∀ i = 1,....,n

(5)

where corr(.,.) denote the correlation measure, E[.]

denote the expectation operator, µ denote the mean,

σ denote the variance. The correlation value is then

subjected to a threshold T (explained in section 4) to

detect the presence of watermark.

4 EXPERIMENTAL RESULTS

AND DISCUSSIONS

Experiments are carried out on gray scale biomedical

images. The parameters K = 5, b

= 4, b

= 5, b

6, b

= 8 and b

= 10 are used. These parameters are

selected based on the frequency range and number of

contexts occurring in each group. The b

is chosen

to maximize the capacity and minimize the detection

error rate. In Table I, the embedding capacity, PSNR,

SSIM, and side information for images of different

resolutions is given.

4.1 Embedding Capacity

It can be seen from the Table I that the capacity in-

creases with increasing image size. This happens be-

cause the larger the image, the higher the frequency

of occurrence of contexts. And the higher the fre-

quency, more number of bits are required to represent

as compared to lesser frequency. However, the image

Brain gives lesser capacity, this is because the occur-

rence of all the contexts is uniform as compared to our

assumption of highest occurring frequencies used for

watermarking. This leads to lesser number of bits re-

quired for representing the frequencies of the contexts

involved in Euclidean distance criteria and hence less

capacity.

4.2 Watermarked Image Quality

The image quality does not get affected by the water-

marking process which is explained as follows. Since

the context is used for bias cancelation, section 2, we

analyze the affect of context exchange on bias estima-

tion and establish the fact that, changing the context

according to the proposed scheme does not affect bias

estimation and hence decoded image quality. The bias

is estimated based on prediction errors accumulated

in the context. Let at position px

a pixel x occurs in

context A with accumulated error A

, and B

be the

accumulated error in context B. Let the contexts A and

B occur A

and B

times respectively until this point.

Now, the bias is given as

Bias

= ⌈A

⌉ (6)

Bias

= ⌈B

⌉ (7)

The bias for the predicted value of pixel x is can-

celed as

ˆx = ˆx + Bias

(8)

where ˆx represent predicted value of pixel x.

Let us now calculate the bias in case when the con-

texts are exchanged i.e., A is exchanged with B. In this

case, B occurs A

times while, A occurs B

times. The

prediction error accumulated remains same as that in

case of no exchange. Now, the bias is given as,

Bias

= ⌈A

⌉ (9)

Bias

= ⌈B

⌉ (10)

The bias cancelation is now given as,

ˆx = ˆx + Bias

(11)

SIGMAP 2011 - International Conference on Signal Processing and Multimedia Applications

164

Table 1: Image, Resolution, Embedding capacity, PSNR, SSIM, Side information.

(percentage of (in bits) compressed ﬁlesize)

Chest1 373 x 387 1147 53.42 .9981 .148

Abdominal 636 x 614 1735 53.33 .9913 .038

Brain 800 x 600 1426 53.59 .9862 .078

Chest2 2048 x 2494 2448 52.84 .9992 .003

Figure 3: First and second row : original image, decompressed image , watermarked-decompressed image respectively.

From equations 8 and 11, it is clear that watermark-

ing through contexts does not affect the prediction of

pixels and hence the image quality.

Figure 3 gives the original image, decompressed

image and watermarked decompressed image respec-

tively.

From this ﬁgure 3 it is clear that the watermark

does not effect the quality of the image as watermark-

ing is performed on the context of the image. PSNR

is given in Table I, and is computed as,

MSE =

m−1

∑

i=0

n−1

∑

j=0

(I(i, j) − I

(i, j))

(12)

PSNR = 10 log

(255

/MSE) (13)

Mean square error (MSE), is the sum of squares of

difference between the original image I and water-

marked decompressed image I

. The Structural Sim-

ilarity Index (SSIM) measures the similarity between

I and I

and is given in Table I. SSIM is given as,

SSIM(I, I

) =

(2µ

+ c

)(2σ

+ c

)

(µ

+ µ

+ c

)(σ

+ σ

+ c

)

(14)

where, µ

(.)

denotes mean, σ

(.)

denotes standard devia-

tion, c

= 6.5 and c

= 58.5 are constants. PSNR and

SSIM measures show that the decompressed image

quality remains intact even after being watermarked.

0 2 4 6 8 10

x 10

−0.4

−0.3

−0.2

−0.1

Figure 4: Detection against random watermarks (positive

upward spike at center corresponds to embedded water-

mark)

4.3 Detection Performance

The detection performance is given in ﬁgure 4. For

simulation 100000 watermarks are generated ran-

domly and the average is reported for 10 randomly

generated context set. It is evident from the ﬁg-

ure 4 that the correct embedded watermark gives the

highest correlation against the extracted watermark.

On the experimental basis we ﬁnd that a threshold

T > 0.8 does not give any error in detection.

The watermark embedding takes place through

context mapping. Since context values can only be

CONTEXT BASED WATERMARKING OF SECURE JPEG-LS IMAGES

165

exchanged with other contexts, attacks like additive

noise, scaling, cropping, ﬁltering or other attacks can-

not be performed as these attacks will change the con-

text value randomly further not rendering decoding.

The watermark can be attacked by changing the

mapping f

. Here for detection purpose, the water-

mark embedder can request the owner for the decryp-

tion key. Further decrypting the bit stream, the orig-

inal contexts can be generated. Since the position of

original context is known, the corresponding water-

marked context position is also known. Thus the wa-

termarked context can be retrieved and the detection

follows similar to the process described in section 3.2.

In this case the detection performance is similar as

with the case of unattacked watermarked copy. This is

because the absolute

context

value cannot be changed

but only replaced with other context.

Another possible attack is collusion attack. In this

case different watermarked copy holders may collude

to extract or destroy the watermark. Colluders may

get the highest frequency contexts and thus the cor-

responding watermarked contexts. However, this nei-

ther gives any information about the mapping used

for watermarking nor the order in which context fre-

quencies should be used (as the watermark is embed-

ded using Euclidean distance criteria). Thus the at-

tacker has to perform brute force attack to extract wa-

termark.

4.4 Side Information

The side information i.e., the contexts and its frequen-

cies occurring in original image is also given in Ta-

ble I. It is clearly evident that the percentage of side

information required is far less than the compressed

ﬁlesize.

5 CONCLUSIONS

In this paper we propose a novel technique to em-

bed a robust watermark in the JPEG-LS compressed

and encrypted images. The algorithm is simple to

implement as it is performed on the context of re-

ceived compressed-encrypted media and does not re-

quire any decompression or decryption. Also the wa-

termark is detected correctly in case of different at-

tacks. The quality of the media is preserved as the wa-

termark does not affect the pixel values itself, rather

only the context classmap is changed. The side infor-

mation required for watermarking is also very low as

compared to the compressed ﬁlesize.

REFERENCES

Bianchi, T., Piva, A., and Barni, M. (2010). Composite

signal representation for fast and storage-efﬁcient pro-

cessing of encrypted signals. IEEE Transactions on

Information Forensics and Security, 5(1):180–187.

Blobel, B. (2004). Authorisation and access control for

electronic health record systems. International Jour-

nal of Medical Informatics, 73(3):251–257.

Caldelli, R., Filippini, F., and Barni, M. (2006). Joint near-

lossless compression and watermarking of still im-

ages for authentication and tamper localization. Signal

Processing: Image Communication, 21(10):890–903.

Cancellaro, M., Battisti, F., Carli, M., Boato, G., De Na-

tale, F., and Neri, A. (2008). A joint digital water-

marking and encryption method. In Security, Foren-

sics, Steganography, and Watermarking of Multime-

dia Contents X. Proc of the SPIE, volume 6819, pages

68191C–68191C, 2008.

Fang, T. (2009). On performance of lossless compression

for HDR image quantized in color space. Signal Pro-

cessing: Image Communication, 24(5):397–404.

Katzenbeisser, S., Lemma, A., Celik, M., van der Veen,

M., and Maas, M. (2008). A buyer–seller water-

marking protocol based on secure embedding. IEEE

Transactions on Information Forensics and Security,

3(4):783–786.

Lian, S., Liu, Z., Zhen, R., and Wang, H. (2006). Com-

mutative watermarking and encryption for media data.

Optical Engineering, 45:1–3.

Piva, A., Bianchi, T., and De Rosa, A. (2010). Secure

client-side ST-DM watermark embedding. Informa-

tion Forensics and Security, IEEE Transactions on,

5(1):13–26.

Schneier, B. (1996). Applied Cryptography. John Wiley

and Sons, New York.

Thomas, T., Emmanuel, S., Subramanyam, A., and Kankan-

halli, M. (2009). Joint watermarking scheme for

multiparty multilevel DRM architecture. IEEE

Transactions on Information Forensics and Security,

4(4):758–767.

Weinberger, M., Seroussi, G., and Sapiro, G. (2000). The

LOCO-I lossless image compression algorithm: prin-

ciples andstandardization into JPEG-LS. IEEE Trans-

actions on Image Processing, 9(8):1309–1324.

Zhao, B., Kou, W., Li, H., Dang, L., and Zhang, J. (2010).

Effective watermarking scheme in the encrypted do-

main for buyer-seller watermarking protocol. Infor-

mation Sciences, pages 4672–4684.

SIGMAP 2011 - International Conference on Signal Processing and Multimedia Applications

166