Differential Energy Based Watermarking Algorithm Using Wavelet ...

IEICE

TRANS.

FUNDAMENTALS,

VOL.E91-A,

NO.8

AUGUST

2008

1961

PAPER

Special Section on Signal Processing

Differential Energy Based Watermarking Algorithm Using Wavelet Tree Group Modulation (WTGM) and Human Visual System Min-Jen

SUMMARY Wavelettree basedwatermarkingalgorithmsare usingthe waveletcoefficientenergy differencefor copyrightprotectionand ownershipverification.WTQ(WaveletTreeQuantization)algorithmis therepresentativetechniqueusing energydifferencefor watermarking.According to the cryptanalysison WTQ,the watermarkembeddedin the protected imagecan be removedsuccessfully.In thispaper,we presenta noveldifferentialenergywatermarkingalgorithmbased on the wavelettree group modulationstructure,i.e. WTGM(WaveletTreeGroupModulation).The waveletcoefficientsof hostimageare dividedintodisjointsupertrees(each super tree containingtwo sub-supertrees). The watermarkis embedded in the relativelyhigh-frequencycomponentsusingthe groupstrategysuch that energiesof sub-supertreesare close.The employmentof wavelettree structure,sum-of-subsetsandpositive/negativemodulationeffectivelyimprovethe drawbacksof the WTQ schemefor its insecurity.The integration of the HVS(HumanVisual System)for WTGMprovidesa better visual effectof thewatermarkedimage.Theexperimentalresultsdemonstratethe effectiveness of our algorithmin termsof robustnessand imperceptibi li ty. keywords: copyrightprotection,HumanVisualSystem,imagewatermarking,wavelet,wavelettree quantization 1.

Introduction

Digital media files can be easily copied and distributed without any reduction in quality. As a result, digital media files are being widely distributed on the Internet today, through both authorized and unauthorized distribution channels. Piracy is a concern when security measures are not in place to protect content. Conventional cryptographic

systems permit only valid

principals (key holders) access to encrypted data. Once such digital data are decrypted, there is no way to track their reproductions or retransmissions. Over the last decade, digital watermarking has been presented to complement cryptographic protection mechanisms. Invisible watermarks can be broadly classified into three types, i.e. robust, fragile (or semi-fragile) and captioning watermarks [1], [2]. Robust watermarks are generally used for copyright protection and ownership verification as they are robust to nearly all kinds of image processing attacks. Fragile or semi-fragile watermarks are mainly applied to content authentication and integrity attestation as they are fragile to most modifications. Captioning watermarks are usually used for side information conveyance, which are required to convey more information than robust watermarks do. Manuscript received November 12, 2007. Manuscript revised March 10, 2008. The authors are with Institute of Information Management, National Chiao Tung University, Hsing-Chu, 300 Taiwan. a) E-mail: [email protected] DOI: 10.1093/ietfec/e91-a.8.1961 Copyright (c)

2008

The

Institute

of Electronics,

TSAI†a),

Member

and

Chang-Hsing

SHEN†,

Nonmember

Cox et al. [1] proposed a global DCT-based spread spectrum approach to hide watermarks. The frequency domain of the image or sound is viewed as a communication channel, and correspondingly, the watermark is viewed as a signal that is transmitted through it. The watermark is spread over very many frequency bins so that the energy in any one bin is very small and certainly undetectable. Langelaar and Lagendijk [3] introduced the DEW (Differential Energy Watermarking) algorithm for JPEG/MPEG streams in the DCT domain. The DEW algorithm embeds label bits (the watermark) by selectively discarding high frequency DCT coefficients in certain image regions. Wang and Lin [4] introduced the philosophy of WTQ (Wavelet Tree Quantization) in the DWT domain. The wavelet coefficients are grouped into so-called super trees. The wavelet-tree-based watermarking algorithm embeds watermark bits by selectively quantizing super trees. Whether the security of watermarking algorithms can be preserved if the details about algorithms are released is always a controversial issue among watermarking researchers. However, the algorithm will be known to the attacker as it is accepted in the field of cryptology [5]. Das, Maitra and Mitra had presented a successful cryptanalysis against the DEW scheme in [5]. There is a need to analyze each of the popular watermarking algorithms individually and to check whether customized attacks can be mounted to highlight the weakness of the individual watermarking algorithm itself. In this paper, we first introduce the WTQ scheme and then explain how this watermarking algorithm can be attacked by cryptanalysis. Based on the motivation to improve the security robustness of WTQ, we present a differential energy watermarking algorithm based on the wavelet tree group modulation structure, i.e. WTGM (Wavelet Tree Group Modulation). The usage of group modulation makes the proposed watermarking algorithm robust against common signal processing attacks and results in a better detector response. With the characteristic of the wavelet tree structure throughout large spatial regions, it is more robust against geometric distortions. The employment of sum-ofsubsets makes the proposed watermarking algorithm more robust against general cryptanalysis. In addition, the consideration to the CSF (Contrast Sensitive Function) and NVF (Noise Visibility Function) of the HVS (Human Visual System) provides a better visual effect of the watermarked image. The remainder of this paper is organized as follows. In Information

and Communication Engineers

IEICE TRANS. FUNDAMENTALS,

VOL.E91-A,

NO.8 AUGUST 2008

1962

Sect.

2,

the

WTQ

nerability

is

WTGM

scheme

reviewed

in

watermarking

experimental

6, respectively.

2.

WTQ

Scheme

The

WTQ

scheme

ing

scheme.

and

is

and

we

2.1

the

Super

the

one

WTQ

quence

of

should

to

{1,-1}.

the

1 coefficient 16

While bed

the

of

will

be

a pair

the rent in

trees

are

is

detailed

to record

a binary

PN

watermark in Fig.

i={2,3,4} A group

has

we

and

col-

j={1,2,3}, coefficients:

from

level-3, two

tree

(a)

se-

bits,

1(a),

and 21

grouping,

and

groups

and

are

there

WTQ

pair

two

operation, for

are

42

(b)

will

pair.

with

one

super

tree

information

example,

quantize

if

the

Otherwise,

the

to em-

trees

the

For

starts

super so

recording bit.

1, we

corresponding

all

we

watermark bit

for

is used is

grouped,

Here

quantized

watermark the

trees

a super

quantization

corresponding

5

tree.

bits.

for

Sects.

watermark-

[4]

shown

After

super

super

seed

The

in

operations:

4 coefficients

to become

watermark

a secret

as

level-2.

every

all

super

recording

DWT

combined in

the

details.

blind Ref.

vul-

proposed

given

based refer

watermark

level-4,

from

coefficients

of

in Ci ,j, where in Fig. 1(b).

from

randomly

tree

its

the

in

are

can

the

Before

coefficients

4,

Tree

4-level

groups

and

Sect.

described

explain

and

coefficients

form

is

a pair

bit,

perform

locate

briefly

scheme,

watermark

explained In

conclusions

reader

information

In

3.

a wavelet

Interested

Group

briefly

method

results

and

is Sect.

left

of

the

cur-

super

right

tree

one

will

be

trees

will

be

quantized. In

order

transformed

to perform

quantization,

to •gbit-plane•h

form

all

super

first.

(c) 2.2

A as

Form

super

tree

shown

which Fig. ger bits

the

will 1(c),

Bit-Plane

will

be

in

Fig.

be

removed

the

pair.

If we

pair

will

be

quantized. can

get

the

error. will be

bit

we to

of

is the

same

is

the

with

found

find

the

trees

out

In

a tree-

tree

is we

information

we

know order

tree, of

In

Cryptanalysis

[6],

Das

termarking model

and

which

Maitra

algorithm by

weakness. is image

that

be

the

dependent,

on the paper

every

tested

Therefore,

techniques As

claimed

as

The

Das

WTQ

and

scheme,

existing a

and

says, •gKnowledge but

not

dependent

put found of

on

secret

every

super

destroy

can

use

removal

of

constructing

the

correlation.•h

a super tree

is

tree

obtained

watermarking

indirect

Although

is

unknown,

with

two

information

technique

through

but groups.

in

the

super

knowledge

cryptanalysis

on

identification

WTQ

of quantized

estimate

of

quantized

groups.

reference

error,

can and and

be

divided

into

non-quantized

tree

groups,

quantization

of

non-

wa3.1

cryptographic Maitra

of

groups.

bit.

WTQ

should

cryptanalysis.

cryptanalytic the

on

successful

that to

we

steps: 3.

for

the

In

in a tree-

pattern,

watermark

is enough

scope,

which

quantization

current

bits

or big-

with

both

will

the

of error.

recorded

watermark,

then

according

information

reference

have

format

scope

the

When we discarded.

the

and

bit-plane the

area

watermark

checked

the

to

gray

to extract

Thus,

to

calculating

of the

every

want

for

according

energy

WTQ,

Quantization

transformed

1(c)

than the reference in the gray area In

for

Fig. 1 Illustrations of the wavelet transform and groups of wavelet coefficients. (a) A four-level wavelet transform and its subbands. The coefficients correspond to the same spatial location are grouped together. (b) The 21 coefficients of a wavelet tree. (c) Binary representation of the coefficients in the nth super tree. The shaded area denotes the bits to be discarded.

the

Identification

of

Quantized

and

Non-quantized

Groups

out

groups, seed,

All

groups

identification.

will The

be

transformed principle

into is calculating

bit-plane the

format energy

for of

last

TSAI

and SHEN:

DIFFERENTIAL

ENERGY

BASED

WATERMARKING

ALGORITHM

1963

rows.

In our

last

rows

this

group

simulation,

in current as

a quantized

non-quantized

group.

3.2

Estimate

of

After

identification,

estimate. groups every

last group

First,

we

are

used.

empty,

we

Otherwise,

this

If the

bits

can

assume

group

of

4.1

is

a

take

the

set the

of

quantized

quantization

find out that Thus, ƒÃ' is

the the

energy estimated

groups error

for of

removed reference

all in

error.

3.3

Quantization of Non-quantized Groups

When reference error has been estimated, the set of nonquantized groups will be quantized using this estimated reference error. After this step, all groups are almost quantized. In WTQ, every watermark bit is recorded by quantizing only one tree in a pair. Making all groups quantized means making all super trees quantized because a super tree is merged with two groups. Thus, if all trees are quantized, the difference caused by quantization between two trees in a pair will be eliminated. As the difference between both trees declines, it is difficult for the detector to extract the watermark bit accurately. According to our simulation of the cryptanalysis attack for WTQ, the unquantized bitplane could be successfully identified and the last two rows could be removed. Therefore, the watermark will be removed even without the reference error estimation. Therefore, WTQ is not secure enough for digital watermarking in principle. 4.

Designs of WTGM Algorithm

There are several issues need to be addressed if the energy difference will be applied for the wavelet based watermarking scheme. The first is the choice of the tree structure. How many levels should the image to be decomposed and achieve the robustness and the scalability of the watermark? The second is how to balance the robustness and the fidelity of the image on a designed energy differential watermarking algorithm? The third is how to maximize the detector response in order to render a better performance. The fourth is the security of the watermarking algorithm, the most important issue to be addressed. To resolve those mentioned issues, the same pyramidal decomposition is applied in WTGM. In addition, the idea of sum-of-subsets [5] for selecting supertrees is adopted to securely embed the watermark. Instead of the bit plane quantization for watermark embedding, the usage of positive/negative modulation will effectively render a better detector response. Also, the consideration to the CSF and NVF of the HVS provides a better visual effect and imperceptibility. The details will be explained next.

PM (Positive Modulation) and NM (Negative Modulation)

Lu, et al. [7] had analyzed the behaviors of transformed coefficients under attacks. In principle, there are four

Error

calculate

in the set, and group is almost ƒÃ'.

rows

almost

group.

Reference

we

two

are

possible types of modulations: Modu(+, +), Modu(+, -), Modu(-, +), and Modu(-, -), where Modu(+/-, -/+) denotes a positive/negative transformed coefficient modulated with a negative/positive watermark quantity. No matter whether the DCT or the wavelet domain is employed, the probabilities of occurrence of the four types of modulations are all very close to 0.25. They further classified the behaviors of attacks into two categories. The first category contains those attacks like compression and blurring, which tend to decrease the magnitudes of most of the transformed coefficients. Under these circumstances, it is hoped that every transformed coefficient can be modulated with a quantity that has different sign. The reason is that it can adapt to compression-style attacks and enables more than 50% of the modulated targets to contribute a bigger positive value to the detector response. Only Modu(+, -) and Modu(-, +) will contribute positively to the detector response. The second category contains those attacks such as sharpening and histogram equalization, which have the tendency of increasing most of the magnitudes of transformed coefficients. Only Modu(+, +) and Modu(-, -) will contribute positively to the detector response. Lu, et al. emphasized that the random modulation strategy does not help the detector response. Scenarios in the attacking process are illustrated in Fig. 2. No matter whether the positive modulation or the negative modulation is employed, the modulated wavelet coefficient can effectively resist the attack in scenario 1 and 2. However, the modulated wavelet coefficient is unable to resist the attack alone in scenario 3 if the strength of the attack is larger than that of the modulation. If a watermarking algorithm simultaneously employs the positive modulation and the negative modulation in embedding a watermark bit, it can succeed in resisting the attack in scenario 3 (as cocktail watermarking did in [7], which simultaneously embedded two watermarks in complementary roles). Since the DEW scheme and the WTQ scheme only employed the philosophy of negative modulation, the detector was unable to bring the brilliant results under any kind of attacks mentioned above. Moreover, the scheme can be easily defeated by the attacker if it only employs unilateral modulation, regardless of the positive modulation or the negative modulation. Thus, a good differential energy watermarking algorithm should take both modulated methods into account for higher detector response and better security. 4.2

Wavelet Tree Structure

We employ the same wavelet tree structure as depicted in the WTQ scheme. However, each tree can be extended to involve high-frequency components as illustrated in Fig. 4.

IEICE

TRANS.

FUNDAMENTALS,

VOL.E91-A,

NO.8

AUGUST

2008

1964

Suppose that each watermark bit is embedded using one super tree, half of a super tree is used for PM and the other is used for NM. We use the term super tree to refer to the collection of n trees (i.e. 1 super tree=n trees). A particular super tree can be divided into two sub-super trees, each containing n/2 trees. The energy of a tree t is defined as the sum of absolute values of q-p+1 wavelet coefficients. The energies of sub-super tree A and sub-super tree B are given by: EA(p,q,n)=Σ

Σ￨θi,t￨

(1)

EB(p,q,n)=Σ

Σ￨θi,t￨

(2)

(a)

where Įi

,t denotes q denotes the

and lation

from

p

sub-super suitable

for

judging B differ

4.4

(b) Fig.

2

Scenarios

Negative

in

the

attacking

modulation. •go•h

indicates

the

wavelet

process.

denotes

modulated

the

wavelet

(a)

original

coefficient,

Positive

modulation.

wavelet

coefficient, •gm•h

and •ga•h

means

the

For

attacked

by

be

a collection

from

level

level

2,

and

64

idea

of

adopted

in

sum

to

there

The

W.

are In

For

As

to

a key

factor,

if we

to form modulate bedded.

the

can

for

[8]

and

coefficient

be

they

all

Das,

two

will

be

out

coefficients

supertrees

the

as

wi

subsets

and

of

in-

integers

that

W=31,

and

how

trees tree,

according

we

can can the

is

that

NP-

used

this

[5]

the

DEW

the

coefficients

be use

scheme. are

aggregation. in the set

grouped this

watermark

of

HVS

visibility

various

the

quency

and

variation

of grating

has

of

measurements were

spatial

contrast

is

given

on and

thresholds

of of

for by

as

periodic the

for

good of

are

the

such

thresholds. gratings by

study by

a measure pattern

given

was the

such

the

studied

sinusoidal

orientations

gratings

Contrast

luminance

signals

to determine

purpose

orientation.

a need properties

visual

performed

frequencies The

been

incorporates

thresholds

conditions.

termine

that

to

given of as

defre-

relative a sinu-

equation

(3)

where Lmaxand Lminare maximal and minimal luminance of a grating. Reciprocal values of contrast thresholds express the contrast sensitivity (CS), and Mannos and Sakrison [9] originally presented a model of the contrast sensitive function (CSF) for luminance (or grayscale) images is given as follows:

then

an

Mitra of

idea

of

be

a criterion

H(f)=2.6*(0.0192+0.114*f)*e-(0.114*f)1.1 itself

a closed value of energy of a tree as an element

super

The

following:

a positive

the

problem

the

Function)

there

quality

measurements

with

{24,7}.

Maitra

which

is

watermark.

formulated

vulnerability

it renders with energy

selecting

S={11,13,24,7},

the

is just

of sub-super tree A and to some small quantity ƒÂ.

images,

image

The

viewing

from

will

energy or equal

two

will

C=(Lmax-Lmin)/(Lmax+Lmin)

embed be

{11,13,7},

so-called the

There

will

WTGM

sum-of-subsets

find

tree

coefficients

1.

(weights)

find

example,

resolve

grouped together If we treat the and

integers is to

problem.

method

WTGM

[5]

can

the

3, 16

level

for

securely

subsets:

fact,

one

Any

i.e. |EA-EB|•…ƒÂ,

Sensitive

psychovisual

soidal

to

goal

two

complete

for

problem n positive

W.

level

from

sum-of-subsets WTGM

are

teger

from

S2

each

coefficients,

Selection

sum-of-subsets There

and 5.

Trees

is transformed,

wavelet

coefficients

set S1 in Sect.

Super

The

85

4, 4 coefficients

parameter discussed

4.3

of

image

(Contrast

for

These a 512•~512

the than

watermarked

metrics HVS.

that

0•…q•…84).

EA=EB,

modulation. |EA-EB|•…ƒÂ

whether by less

CSF

q (0•…p•…84, with

in the tree t, p do the modu-

to

(b)

coefficient.

Suppose

to

trees

the ith wavelet coefficient coefficient number used

S,

together principle bit

to em-

where

f=√

f x2 + f y2

is

the

spatial

frequency

(4) in

cy-

cles/degree of visual angle (fx and fy are the spatial frequencies in the horizontal and vertical directions, respectively). Figure 3 depicts the CSF curve which characterizes luminance sensitivity of the HVS as a function of normalized spatial frequency. According to the CSF curve, we can see that the HVS is most sensitive to normalized spatial frequencies between 0.025 and 0.125 and less sensitive to low and high frequencies [10]. Therefore, this knowledge from CSF

TSAI and SHEN: DIFFERENTIAL

ENERGY BASED WATERMARKING

ALGORITHM

1965

(5)

βk=0.01+(7.20-rk)2/7.202

where k denotes the decomposed level and rk represents the wavelet coefficient CSF of the perceptual importance weight as Fig. 4 shows. The level 1 has the largest rate for modulation, which corresponds to high-frequency components. The level 3 has the smallest rate for modulation. Under the circumstances the sum-of-subsets is employed, the actual modulation quantity of low-frequency components will be relatively small since they have larger energies. Contrarily, the actual modulation quantity of highfrequency components will be relatively large since they have smaller energies. In our study, low-frequency comFig. 3

ponents can tolerate more common signal processing while high-frequency components can tolerate more geometric attacks. The usage of high-frequency components is pretty different from the WTQ scheme for its nature of watermarking.

Luminance CSF (Courtesy of [10]).

4.5

NVF (Noise Visibility Function) of HVS

S. Voloshynovskiy et al. [14] presented a stochastic approach based on the computation of a NVF (Noise Visibility Function) that characterizes the local image properties and identifies texture and edge regions. This allows us to determine the optimal watermark locations and strength for the watermark embedding stage. Their argument: the channel capacity is not uniform, i.e. the noise is more visible in flat areas and less visible in regions with edges and textures. Accordingly, when the local variance is small, the image is flat, and a large enough variance indicates the presence of edges or highly texture areas. The adaptive scheme based on NVF calculated from stationary GG model is the best model in our simulation, which is defined as follows: Fig. 4 A four-level wavelet tree structure. The coefficients correspond to the same spatial location are grouped together. Each tree consists of one coefficient

from

level

4, 4 coefficients

from

level 2, and 64 coefficients from level indicated at the center of each band.

level

3, 16 coefficients

1. rk(ƒÀk) values

for each

NVF(x,y)=w(x,y)/ w(x,y)+σ2I

from

level k are

where

w(x,y)=γ[η(γ)]γ/‖r(x,y)‖2-γ

variance.

η(γ)=√

(gamma can

be

used

to

develop

a

simple

[10]-[13]

is

image

independent

HVS

the

model. CSF in

the

the to

masking

discrete

method their

wavelet of

masks

which

ceptual

transforms

importance

mask

in the

five-level

12-weight

DWT

each

CSF

the

same

are

wavelet CSF

in in

to

with

[11].

12-weight

transform.

mask

Fig.

Figure the

weights

CSF

refers

into

of

use CSF

subband

the

square

masking. is

determined

function The

in

adequate by:

[10]

to

parameter

and

rate

which

DWT

CSF

time

the

rate ƒÀk

variance.

watermark.

Since

a better

which

For

is

most

very

CSF visual

renders

a better

4.6

WTGM

Algorithm

We

summarize

the

ing

algorithms,

which

real

local

images,

the

close,

the

image

CSF

and

NVF

the

energy

enhance

constraints

watermark

regions

of

the

of the and

the

edge

is by

0.3•…ƒÁ•…1.

visibility the

enhance

global

∞0e-uus-1du

determined

combination the

retains

is

range

value

The

decrease

NVF

r(x,y)

is in the PSNR

σ2I is the

Γ(s)=∫

r(x,y)=(I(x,y)-I(x,y))/ƒÐI. ƒÁ and

local

the

effectively

Huang

shown

and

different.

the

4 illustrates

the

Even

per-

approximate

modulation

and

and

Γ(3/γ)/Γ(1/γ),

parameter

shape

is quite

CSF 3

function)

shape

mean

to

according

well-designed

curve

led

the

masking

presented

method

apply

coefficients

Some CSF

to

effect,

the

detector

can of

modulation

while

strength

quality

at the in texture

same and

response.

for

subband. We

effect

use

way

wavelet

the

weight

[10]

each

the

importance.

Tang

the

domain.

weighting

perceptual

one

(6)

the for

ideas

mentioned

integrate

above the

advantages

in

the

follow-

of

wavelet

IEICE

TRANS.

FUNDAMENTALS,

VOL.E91-A,

NO.8

AUGUST

2008

1966

tree structure, sum-of-subsets for supertree selection, positive/negative modulation for watermark embedding and the CSF and NVF of the HVS into the WTGM. To quantify the existence of the watermark, the normalized correlation coefficient (NC) will be examined in order to identify the existence of the watermark. The formula of normalized correlation coefficient is as follows:

Nw

equals

be

to

3)

The by

Huffman

the

size

would

data

Since

the

The coefficient value is within -1 and 1. The complete design of the proposed algorithm is summarized as following:

4)

For

each

watermark

a)

Select

b)

Choose ƒ¿.

c)

If

bit

wi

(i=0

to

Nw-1)

do

to

for

is

originally

denotes

the

we

ployed, 5) ƒÀk is

EA,

ith

consisting

of

n trees.

EB

,t=ƒÆi,t*(1+ƒ¿*ƒÀk*ƒÁkx,y) i=p,...,q.

ii) Įi,t=Įi

for (PM

for

IfƒÀk=1

i=p,...,q.

for

(NM

for

tree

7)

A)

ing

t=(n/2),...,n-1,

sub-super

The

tree

B)

the

readers

could

detailed

infor-

use,

there like

PGP

[16]

that

lets

indi-

extremely

are

free

strong

en-

change

difference,

required

i.e.,

after

to enthe

modand

modification).

If

the

HVS

is em-

for the strength of the embedding parameter the

NVF

watermark. and ƒÁkx ,y=1parameter where

embedding

the

is used.

the

CSF

the

the is not

NVF

HVS

employed.

employed.

super

embedding

is not

employed.

is not of

tree

list

procedure

generator algorithm Most pseudo-random quences

,t=ƒÆi,t*(1-ƒ¿*ƒÀk*ƒÁkx,y) and

ii) Įi

for

i=p,...,q.

(NM

for

i=p,...,q.

for (PM

for

mon

t=0,...,(n/2)-1,

sub-super

,t=ƒÆi,t*(1+ƒ¿*ƒÀk*ƒÁkx,y) and

tree

A)

will

but

the

be

stored

random

is not necessary to generator algorithms

tree

B)

Blum

Shub, it is very

tions.

dom

number

Pass

various

back

the

the

modulated

modified

verse

DWT

The

watermark

The

length

of

The

max

wavelet

to obtain

trees

to their

coefficients

original

posi-

through

a watermarked

are of

uniformly

these

distributed

algorithms

generators, lagged Fibonacci shift registers, generalized

t=(n/2),...,n-1,

sub-super

which classes

fore,

Arrange

6)

the

information the

data

have

dur-

number

be recorded. produce se-

else i) Įi

5)

after

and ƒÁkx,y=1,

If ƒÁkx,y=1,

t=0,...,(n/2)-1,

sub-super

,t*(1-ƒ¿*ƒÀk*ƒÁkx,y)

and

is (5)

If ƒÀk=1,

con-

the

need |(E'A-E'B)/(EA+EB)|•†ƒ¿(E'A

it stands the CSF

formula

then

and

d)

tree

6)

(wi=-1)

i) Įi

super

the

keys.

fractional

energy

of

files system

with

long

the

required

ification, E'B are

and

more

decrypt

de-

encryp-

doesn't

Interest

documents

and

authenticity

encode

for

encryptions

their

cryptography

change

practical

and

inissue

Data

who

[15]

For

to encrypt

and

secret

encryption and

them.

application.

algorithms

of

of

efficiently

study

a public-key

tool

and

research

anyone

unscramble

secure

cryption

force

by

cryptography

NVF(x,y) the

read

any further

is a critical

systematically can

will

small.

communication.

data

be

available

viduals

4) ƒ¿

secret

key

which

secure

be

compression

important

The

which

the

the

tools

the

K

without

K can

transmission

extraction.

can't

mation

is

confidentiality

they

refer

K data

guarantee

of

key

zip

to

algorithms

proper

WTGM Watermark Embedding:

or

key

bytes

comparatively

key

for

image

4608

image

algorithms

tion

so

of

advanced

is needed

the

is

the

data

tents

1536,

coding be

secure

watermark

cryption

generator and group them in various super trees. Each super tree should be divided in two sub-super trees such that EA=EB. Store this group information which we call the image key K.

image the

studies

of which

amount

reduced

formation,

(7)

value bits

compression.

for

1) Generate a seed by mapping a signature/text through a one-way deterministic function. Obtain a PN sequence Wof length Nw using the seed. 2) Compute wavelet coefficients of a host image. Group the coefficients to form trees. 3) Randomly arrange the trees using some pseudorandom

the

12•~2•~1536

the

in-

image.

super

Fortuna,

and

flexible

for

generator trees

under

to

are

[17] linear

and

com-

congruential

generators, feedback shift

linear feedback registers, Blum

the

twister.

Mersenne

WTGM arrange the

principle

to apply the

any

wavelet of

Theregood

ran-

trees

into

sum-of-subsets.

WTGM Watermark Extraction

Note: 1) 2)

trees.

W

under

4 level

mark

bit

of NM,

is

a super

the

is a binary watermark

value wavelet

WTGM

is

tree

(212=4096•†1536•~2).

key

K

ing

information

needs

using for at

least

number for

one PM

and 12

of •}1.

bits the

Since

water-

each

tree

the

other

is

to

mark

each

recording

sum-of-subsets

super image

super

bits

for

of

a 512•~512

Therefore,

12•~2•~Nw under

Nw=the

decomposition.

used

needs

sequence

of Nw=1536

embedded tree

PN

where

half

used

for super

the

image

the

order-

principle.

1) Generate a seed by mapping a signature/text through a one-way deterministic function. Obtain a PN sequence Wof length Nw using the seed. 2) Compute wavelet coefficients of a host image. Group the coefficients to form trees. 3) Reorganize the trees using the image key K. 4) For each watermark bit wi (i=0 to Nw-1) do

If

a) Select the ith super tree consisting of n trees. b) Calculate EA and EB. c) If (EA>EB) then wi=-1.

TSAI and SHEN: DIFFERENTIAL

ENERGY

BASED WATERMARKING

ALGORITHM 1967

else wi=1. 5)

Compute

6)

If ƒÏ

the

is

normalized

above

the

otherwise,

5.

mark.

Experiment

correlation ƒÏ.

threshold ƒÏT,

it does

not

Without

tion

the

watermark

W

rate

is

artifacts

exists;

ment

exist.

in the of

the

watermarked

(as

shown

Results

evaluate

the

performance

512•~512

Lena,

bits/pixel

resolution

a four-level length the

trees

part

used

to relatively DWT) scheme. 6-85

In

p=5,

order

[4]

of

sentative WTQ meet

we

and

length

Nw=512,

a false

the

used

in

is the

CDF

9/7

5.1

Visual

From

this

5,

study

filters

the

different

image

the

same.

Compared

ages

demonstrate and

the

WTGM(S2) quency

are

responds ure

Table

the

to

6(c)

1

uses

the

of

strength

of

why

the

HVS

watermark

is

im-

so that

the

first

4 of WTQ

coefficient

num-

to relatively

high-

all

the

(a

(b

(c

(d

shown

typically

repre-

compare

and

water-

values

the

with

is

for

the With

of NC

is chosen

as

in

watermark to be The

tree

to Lena,

same

1.

used

the

strength)

39.8dB

0.23

Fig. 5 Watermarked images and error images. (a) WTGM(S1) watermarked Lena image at PSNR=38.2dB. (b) Scaled error image between (a) and the original image. (c) WTGM(S2) watermarked Lena image at PSNR=38.2dB. (d) Scaled error image between (c) and the original image.

wavelet

watermarking

in WTQ.

parameter

errors

1-21

and

will

The

embed used number

settings

5(d),

the

(S1)

in

for Lena,

Goldhill

high

fre-

image.

6(b)

scheme. the

and Peppers

(b

(c

(d

images uses

which

embed

(a water-

Lena

WTQ to

im-

different

watermark, the

im-

different.

more

Figure

6-85

at

error

are

watermarked

the

result kept

watermarked

under

38.2dB.

is

the

(S2)

have

quality

of

subbands

coefficient

and

will

PSNR

watermarked

visual

to

the

between

WTGM(S1)

values

The parameter

5(b)

image

the

setting

even

setting

settings.

PSNR

the

the

wavelet

Figs.

by

than

number

3 and

1.03•~10-7.

while

by

6 shows

with

efficient

different

quality

parameter

all

intensify

example

uses

2,

(watermark

also

another

Comparison

original

signals

the are

watermarked

Figure marking

for

6(d)).

is

NM.

1 (S1))

Table

of

of rate

8

The

as

setting

in

which

two

in

ages

the

shown

quality

modulation

of

same

To

38.7

probability

Quality

Fig.

38.2,

obvious employ-

corresponds

PSNR

it is

of ƒ¿

The visual

larger

modula-

DWT).

same

threshold ƒÏT

positive

filters

value

since as

for

comparison,

approach.

Fig.

the

it has

the are

a rectangle.

improves even

even there

half

(level the

2 of

since

of

Peppers

respectively,

at the

the

values

WTQ,

fair

set

used parts.

uses

1 and

based

set

is

the

6 trees,

Set

corresponds

algorithm tree

PSNR

Goldhill

be

WTQ

wavelet

the

for

make

of

q=20)

WTGM(S2)

(level

will

scheme,

p=0,

q=84)

to

images

in

part

components

marked

two

which

second

(i.e.

frequency

are

into

components

watermarking,

The

others

with

apparently image,

7

We

marked

HVS, 0.378),

employ

sequence

Parameter

(i.e.

low-frequency

for

the

region

and

the with

We

consists

divided

1-21

images

a watermark

tree

(Watermarking

number

method,

watermarking.

and

and

are

WTGM(S1)

ber

PM

proposed

Peppers

for

a super

for

the

and

used

transform

experiments

coefficient

as

are

Therefore,

are

The

of

Goldhill

wavelet

512.

in

Figure

to the

(ƒ¿=0.177

HVS

the

portant. To

consideration small

cocorFig-

water-

images.

Fig. (a)

6

Close-ups

Original

0.177). WTGM(S2)

(c)

image.

for

comparison

(b)

WTGM(S1)

WTGM(S2) watermarked

with

watermarked with

visual

effects

watermarked

HVS

Lena

without

(ƒ¿=2.131).

(PSNR=38.2dB).

Lena HVS

without

HVS

(ƒ¿=0.378).

(ƒ¿= (d)

IEICE TRANS. FUNDAMENTALS,VOL.E91-A, NO.8 AUGUST 2008 1968

Table

2

Watermarks

extracted

from JPEG

compressed

watermarked

images.

(a)

(b)

5.2 (c) Fig.

7

(a)

WTGM(S2)

Test

WTGM(S2)

visibility

of embedding

watermarked watermarked

watermarked marked

on

(d)

Barbara

Barbara

Fig. 8

Lena Lena

without

with

HVS

with HVS

of watermark without HVS

(PSNR=25.0dB).

HVS

(ƒ¿=1.731).

(ƒ¿=9.741).

(ƒ¿=0.872).

(d)

(c)

(b) WTGM(S2)

WTGM(S2)

water-

(ƒ¿=3.374).

(a)

(b)

(c)

(d)

Close-ups for comparison with Figs. 7(a)-(d).

quality of watermarked image will be as low as 25dB for comparison purpose. From these results, we can see that there are obvious artifacts in the regions near the shoulder in Fig. 7(a) and the foot of the table in Fig. 7(c). The HVS can effectively decrease the visibility of the watermark (as shown in Figs. 7(b) and (d)). Figures 8(a)-(d) are the closeups of the images in Figs. 7(a)-(d).

Common Image Processing Attacks

1) JPEG CompressionAttacks In this experiment, we perform JPEG compression with different quality factors (QF) on the watermarked image. The extracted results and NC values are depicted in Table 2. From these results, we can see that the proposed algorithm is robust to JPEG compression. For all cases, the extracted watermarks are with relatively high-NC values. The result of WTGM(S1) is superior to that of WTGM(S2). Even for the case that QF is equal to 20, we can still detect the embedded watermark. Since the setting for S2 reserves the watermark in the level 1 component, JPEG intentionally removes the high frequency components which make setting S1 perform better than setting S2. Therefore, the results from Table 2 are reasonable. 2) SPIHT CompressionAttacks SPIHT (Set Partitioning in Hierarchical Trees) is an imageecompression algorithm that exploits the inherent similarities across subbands in a wavelet decomposition of an image. It implies uniform quantization and bit allocation applied after wavelet decomposition. Table 3 shows the extracted tracted NC values and corresponding PSNR values between original image and attacked image. From these results, we can see that the proposed algorithm can tolerate the incidental distortions induced by high-quality SPIHT compression. Since SPIHT first removes the high frequency components during the rate reduction, the results of WTGM(S1) is also superior to those of WTGM(S2). 3) JPEG2000 CompressionAttacks JPEG2000 [18] is a new image compression standard which has good performance in high bit rate coding. It adopts wavelet transform instead of discrete cosine transform to utilize the intersubband correlation. Table 4 shows the extracted NC values and corresponding PSNR values between original image and attacked image. Since there is no data from WTQ results under JPEG2000 attack, the results under SPIHT attack are shown for comparison purpose. From these results, we can see that the proposed WTGM al-

TSAI

and SHEN:

DIFFERENTIAL

ENERGY

BASED

WATERMARKING

ALGORITHM

1969

Table 3 Watermarks extracted from SPIHT compressed watermarked images. (a) Lena. (b) Goldhill. (c) Peppers.

Table 5 Watermarks extracted from spatial-domain-attacked watermarked images. (a) Lena. (b) Goldhill. (c) Peppers .

(a)

(b)

(a)

(c)

Table 4 Watermarks extracted from JPEG2000 compressed watermarked images. (a) Lena. (b) Goldhill. (c) Peppers.

(b)

(a)

(b)

(c)

(c)

gorithm can tolerate the incidental distortions induced by JPEG2000 compression. Since JPEG2000 first removes the high frequency components during the rate reduction, the results of WTGM(S1) is also superior to those of WTGM(S2) which has similar performance as shown in Table 3.

4) Spatial-Domain Image Processing Attacks Several spatial-domain image processing techniques, including histogram equalization, image cropping, brightness enhancement, contrast enhancement, median filtering, Gaussian filtering, sharpening, and rescale are performed on the watermarked image. The extracted results are depicted in Table 5. For all cases, the watermark information therein can be successfully recognized. Especially for those cases of histogram equalization, Gaussian filtering and sharpening, the result of WTGM(S2) is superior to that of WTGM(S1). Except for the case of Peppers Gaussian filtered image, the proposed algorithm can outperform the

IEICE

TRANS.

FUNDAMENTALS,

VOL.E91-A,

NO.8

AUGUST

2008

1970

Table

6

Watermarks

extracted

from shifted

watermarked

images.

Table 8 Watermarks extracted from multiple watermarked Lena. (b) Goldhill. (c) Peppers.

images.

(a)

(a)

Table

7

Watermarks

extracted

from rotated

watermarked

images.

(b)

(c)

5.4

Security

Measurement

1) Multiple

Watermarking

For ply

one

group tector WTQ

scheme

with

relatively

high-NC

values.

the

Geometric

Attacks

or

results

of

the

PSNR

Pixel

Shifting

This cularly.

6.

is

for image.

that

of the

the

functions.

attack

for

the

pixels

left.

attacks

as

shown of

13 pixels

to

12

Gold-

rate

up

to

images.

size.

and

can

and

domain.

We

3•‹ for

that

Goldhill

can

be

image

we

is

aptree

the de8 shows

through can

fallen

see

into

mul-

that

even

25dB,

the

detected.

removal WTQ

is

can

resist

WTGM(S2),

ues

of

of

We

perform

this

which

reduces

subbands,

images.

and

one

scheme.

embedded

watermarked rithm

than

rotate

7

Table and

the

9 shows

8 bitplanes

images

are

strategies

that

Under fallen

impact

on

proposed for

26

algo-

WTGM(S1)

the

and

to

designated

the

the

which

into

used

attack

removed

respectively.

attacked

major

PSNR

val-

29dB.

shown

and

is

described

a geometri-

can 2.5•‹

for

The

sis.

watermarked

in Table

Complexity

of

Lena

is

WTGM

with

The

complexity

also

whole

transform,

Human

Vision

System

low

from

of WTGM the

complexity

view

should

sum-of-subsets,

with

of be

CSF

Human

Vision

mathematical discussed

and

NVF

analyfor

wavelet

calculation

re-

spectively.

7. From resist

computation

System

counter-clockwise

WTGM(S2)

image

5.5

scaled

software

is the

a small

the

the

scaling

and are

the

by

[19]

it provides

rotation

results

see

image cropping

StirMark

to 3•‹ in clockwise

extracted

we

the

image,

since

the

the

attacked

to disturb Table

attacked

results,

may

wavelet

Scaling)

rotating

attack

spatial

The

and

by

on

these

attacker

same

attempt watermark.

images

From of

an

the

the

Removal

Bitplane defeat

in Taup

for

modulation

Bitplane

cir-

Apparently

a shift

and

lower

image

This

0.25•‹

results, of

this

in the

from

directions.

Peppers

the

resist

images has

rotated

original

testing

tation

such

2)

shifting to

can

(Rotation done

the

here

these

former

is

scaling

image

Peppers

the

Attacks

adopted

cal

pixels

to resist

and

attack

to

the

by

latter.

The

image

shift

Shift)

done

WTGM(S2)

For

Rotation

angle,

is

unable

Lena

hill

(Circular

attacks

we

Contrarily,

pixels

2)

of

Here,

WTGM(S1) ble

Attacks

kind

watermarked

value

all,

using

technique in the embedded

the

still

to

watermarks

watermarking.

watermark 1)

well-known

more

modulation or to destroy

tiple 5.3

algorithms

a roand

Suppose (high-pass) and

assume

CDF

9/7

the for

M•†N. filters

synthesis

wavelet

filters

are

transform. The

is 4(N+M)+2

cost

h

(low-pass)

and

g

Take |h|=2N, |g|=2M, of

the

and

standard could

be

algorithm speeded

for up

by

TSAI

and SHEN:

DIFFERENTIAL

ENERGY

BASED

WATERMARKING

ALGORITHM

1971

Table 9 Watermarks extracted from bitplane-removed ages. (a) Lena. (b) Goldhill. (c) Peppers.

watermarked im-

Table 10

Summary of WTGM with WTQ schemes .

(a)

(b)

by

the

the

window

local

mean

is O(l2),

l(=2L+1)

window

size

is

of

use

ance

NVF

since

image

width

tion

of

algorithm

wavelet

The

transform

problem

a real

sum-of-subsets

stead.

Empirical

[21]

get

such

plexity

an

CSF

weighting as

plementation tion from

about

and

the

linear-time

the

other

in the

hand,

DWT

in

the

Fig.

the

domain be

be

applied

by

the

energy

Therefore,

average

the

time

comif

R

quicksort-based

is employed associated

pre-calculated

for

to

apply

each of

well

subband CSF

can parameter

the

complexity

be pre-calculated is

decided.

by r(x,y)

the

NVF,ƒÅ(ƒÁ) look-up in

Eq.

and table

(6)

under

Intel

In

suitable

image

sub-

size

(we

global time

vari-

complex-

2

loop

of

seconds

to

practical

3.0GHz,

complete

the

for

WTGM

applications

simulation

WTGM

Pentium

conclusion,

for and

12

amount

than l. whole

than

comfrom

the

results.

gamma while

is determined

the

WTQ

In

as

other

is

shown

we

general,

WTGM(S2)

the

all

The

it provides WTGM

to

clearly

WTGM

is superior

with

be

compression

resisting

as

geometric

10). not

the

use

quantization

to

cryptanalysis-like

remove

categories

in

with can

SPIHT

ineffective

does

and

can

and

in Table

WTGM

useful

in almost

Also,

methods.

JPEG but

watermarks not

comwhich geomet-

components-WTGM(S1)

in resisting

(as

high-frequency other methods, processing,

cryptanalysis.

than

cryptanalysis,

the

WTQ

well

effect

addition,

relatively

is superior to common signal

resist as

effective

In bed

with

low-frequency

distortions.

in

WTGM

visual

as

im-

the coefficient multiplicacan be efficiently done

of

the

distortions

Therefore,

Regarding

shape

ric

more

linear-time.

function

general,

ponents-WTGM(S2) can effectively

a better

perceptual

complexity

total

(O(n2.l2+n2)=O(n2))

the

less

analysis

are

The

overall

the

variance

there

Thus,

the

larger

images.

study,

Summary

relatively

the

becomes table. This

5.6

In to

[21].

masking

the

in-

quick-

comparisons of

and

its

and

than O(n2)

this global

the

to

variance

of

on

tree

and

to

results).

related

with idea

based

Therefore,

in WTGM look-up

dealing

is low

mathematical

NP-

implementation

(2+2ln2)R

CSF

can 4.

is not

the

complexity on

plexity known

actually

easily.

only

function

shown

can

a

a sum-of-subsets

that

supertrees

arrangement

sorted is

On the

the

requires

selection

WTGM but

shows

in WTGM

is

the

extraction need

and

equals

store

n is much

testing

In the

subband

simulation,

will

512•~512

computa-

mathematics.

itself

However,

study

The

time

problem

problem

to order

are

to 2(N+M+2). is linear

[5].

sum-of-subsets sort

[20]

sum-of-subsets

complete

items

in

and

decomposition.

more

and

1GRAM lifting

no

our

embedding

the

to

is

mean

size.

Besides,

computation is

From

(c)

window

wavelet

array

which

of local

approximately

static

for

variance

L=1.

wavelet

takes O(n2)

ity

for

each

4 level

calculation

can

local

complexity is the

3•~3

for

after

the

The

is

obtained

bands

and

size.

the see from

watermark

that the

for

WTGM detailed

medium-high

in resisting

common

em-

attack

for

WTGM.

outperforms comparsion.

frequency signal

setting process-

IEICE

TRANS.

FUNDAMENTALS,

VOL.E91-A,

NO.8

AUGUST

watermarking

for

2008

1972

ing, geometric distortions as well as cryptanalysis with better visual perception than WTGM(S1). Due to the difference of watermark embedding location for setting S1 and S2, the results are expected compared with other wavelet based approaches. However, the weakness for the WTGM the tree combination information must be kept secret addresses extra storage space. The extended study working on the design to efficiently reduce this extra

is that which should cost

[2]

Lu

no.10, [3]

[4]

Langelaar

Process.,

S.H.

Wang

T.K.

C.S. marking

[8]

Tsai

[9]

J.L.

no.4, [10]

scheme,•h

Trans.

for

Process.,

Im-

copyright

vol.13,

no.2,

2004, Sze,

differscheme,•h

Feb. tree

2005.

quantization

2004.

Liao, •gCocktail

IEEE

tree

Trans.

July

group

IEEE

water-

Multimed.,

vol.2,

modulation

International

Processing,

IEEE

(WTGM)

Conference

vol.II, effects

of images,•h

on

pp. 173-176,

of a visual

Trans.

Inf.

2007.

fidelity

Theory,

cri-

vol.20,

1974. Tang, •gA

IEEE

H.Y.

Sakrison, •gThe

S.X.

pp. 768-775,

pp. 219-230,

and

protection,•h

Signal

encoding

and

robust

2000.

and D.J.

of optimal

a modified

no.2,

Shen, •gWavelet

pp. 525-536,

ing

Image

of wavelet

watermarking,•h

and

Huang

contrast-sensitive

Multimedia,

vol.13,

visible

no.2,

watermark-

pp. 60-66,

April-June

2006. [11]

L. Yong,

L.Z.

based

wavelets

on

vol.27, [12]

A.P.

Cheng,

no.11,

L.R.

perceptual

Z.H.

Xu, •gTranslucent

error-correct

Iyer,

masks

Digital

and

and

pp. 1533-1539.

Beegan,

Signal

for

code,•h

Nov. and

Bell, •gDesign

image

Processing

watermark

J. of Computers,

2004.

A.E.

wavelet

digital

Chinese

and

compression,•h

Workshop,

pp. 88-93,

evaluation

of

Proc.

10th

IEEE

IEEE

CS

Press,

2002. [13]

M.J.

Tsai

and

C.W.

watermarking

by

models,•h 1437, [14]

Trans.

S. Voloshynovskiy, approach Proc.

Dresden, B.

3rd

based

multipurpose

watermarks

with

Fundamentals,

Schneier,

A.

Herrigel,

to

content

Int.

color

human

image

vision

vol.E91-A,

Sept.

Applied

Code

in C,

PGP,

[17]

P.L'Ecuyer, •gUniform

N.

no.6,

Baumgaertner,

adaptive

Workshop

Germany,

[16]

system pp. 1426-

and

digital

Information

T. Pun, •gA

image Hiding,

watermarkpp. 211-236,

1999.

Cryptography:

Second

ed.,

Protocols,

John

Wiley,

Algorithms,

New

York,

and

1996.

http://www.pgp.com/downloads/freeware/index.html

erations [18]

dual

2008.

stochastic

[15]

Lin, •gWavelet

using

IEICE June

Source

1) The group information for trees needs to be kept for watermark extraction, which needs more storage space. 2) The tolerance for geometric attacks is sill insufficient, feature-based or other RST (Rotation, Scaling and Translation) invariant mechanisms can be taken into account for better synchronization.

C.H

the

B.B.

IWDC

Dec.

Speech

on

2001.

and

vol.53,

C.J.

image

image

Mannos

terion

ing,•h

On the other hand, there are still some issues needed to be further studied as following:

digital

Acoustics,

Jan.

energy

IEEE

quantization

Trans.

(DEW)

Huang,

and

digital

video,•h

J. Mitra, •gCryptanalysis

Process.,

pp. 209-224,

M.J. for

and

scheme,•h

for

vol.10,

differential

and

tree

IEEE

S. Maitra, •gCryptanalysis

S.K.

image

Process.,

2004.

Signal

and

Lu,

no.4,

pp. 148-158,

watermarking

Trans. Das

images

Lin, •gWavelet

S. Maitra,

watermarking [7]

Y.P.

Feb.

energy

T.K.

no.1,

watermarking,•h

Das,

IEEE

1) The proposed algorithm can tolerate more common signal processing and geometric attacks. 2) The length of the image key is large, which renders a better confusion/diffusion for security. 3) The human visual characteristics are considered in the wavelet tree based watermarking systems to provide a better visual quality. 4) The watermark can be public for users, and if any malicious user tries to destroy the watermark and sell those attacked copies, the user could be identified.

and

Image

Lagendijk, •gOptimal

encoded

vol.10,

Trans.

2001.

R.L.

of DCT

age

IEEE

Oct.

and

pp. 154-165,

[6]

Liao, •gMultipurpose protection,•h

pp. 1579-1592,

G.C.

ential

provides a better visual quality of the watermarked image. Compared with the WTQ scheme, the advantages of the proposed algorithm are as follows.

H.Y.M. and

protection

Conclusion

An efficient differential energy watermarking algorithm based on wavelet tree group modulation has been presented. In the proposed algorithm, the watermark is embedded in the relatively high-frequency components using the group strategy for each super tree such that energies of sub-super tree A and that of sub-super tree B are close. The employment of wavelet tree structure, sum-of-subsets and positive/negative modulation effectively improve the robustness of the watermark. The consideration to the CSF and NVF of the HVS

and

watermarking

[5]

6.

C.S.

authentication

random

Research,

vol.53,

JPEG

2000

compression,

JPEG

2000)

published

number

pp. 77-120, the from

generation,•h

Annals

of Op-

1994.

International ISO/IEC,

standard [Online]:

(IS

15444-1:

http://www.ece.

uvic.ca/mdadams/hasper/•` [19]

StirMark,

http://www.petitcolas.net/fabien/software/StirMarkBen

chmark_4_0_129.zip [20]

Acknowledgments

I. Daubechies

1 and

lifting

Journal

no.3,

1. This work was supported by the National Science Council in Taiwan, Republic of China, under NSC95-2416H009-027 and NSC96-2416-H009-015. 2. Partial technical background has been presented in the conference ICASSP 2007 [8].

References

[1]

I.J.

Cox,

J. Kilian,

spectrum cess.,

watermarking vol.6,

no.12,

F.T.

Leighton, for

and

multimedia,•h

pp. 1673-1687,

Dec.

T. Shamoon, •gSecure IEEE 1997.

Trans.

spread Image

Pro-

[21]

R.

steps,•h

pp. 247-269, Sedgewick,

Structures, dison

Wesley,

W. Sweldens, •gFactoring of Fourier

May

2001.

wavelet and

transforms

into

Applications,

vol.4,

1998.

Algorithms Sorting,

Analysis

Searching,

in

C, and

Parts Graph

1-5:

Fundamentals,

Algorithms,

3rd

Data ed.,

Ad-

TSAI and SHEN: DIFFERENTIAL

ENERGY BASED WATERMARKING

ALGORITHM 1973

Min-Jen Tsai received the B.S. degree in Electrical Engineering from National Taiwan University in 1987, the M.S. degree in Industrial Engineering and Operations Research from University of California at Berkeley in 1991, the Engineer and Ph.D. degrees in Electrical Engineering from University of California at Los Angeles in 1993 and 1996, respectively. From 1996 to 1997, he was a senior researcher at America Online Inc. In 1997, he joined the Institute of Information Management at the National Chiao Tung

University

search

interests

sic, digital

in Taiwan include

watermarking

puting for electronic Eta Kappa Nu.

and is currently

multimedia

system

and authentication,

commerce.

an associate web

services,

Dr. Tsai is a member

Chang-Hsing

professor.

and applications,

Shen

His re-

digital

foren-

enterprise

com-

of IEEE,

ACM,

and

has received B.S. de-

gree in information management from National Central University in 2000, the M.S. degrees in Institute of Information Management at the National Chiao Tung University in the year 2006. From year 2002 to 2003, he was in the development team of anti-virus service at Trend Micro. He later joined Inventec Besta in 2006 and focuses on digital entertainment and learning of mobile device.