Semi Fragile Watermarking Technique using IWT and a Two Level

Tamper Detection Scheme

Nandhini Sivasubramanian and Gunaseelan Konganathan

Department of Electronics and Communication Engineering, College of Engineering, Gunidy, Anna University,

Chennai, Tamil Nadu, India

Keywords: Semi Fragile Watermarking, Integer Wavelet Transform, Image Processing, Tamper Detection, Hamming

Distance.

Abstract: A semi fragile watermarking technique using a two level thresholding scheme for tamper detection is

proposed. The proposed embedding technique uses two level IWT (integer wavelet transform) to embed the

authentication watermark. The authentication watermark generated from the approximate coefficients is

stored in the detail coefficients using least significant substitution to form the watermarked image. The

proposed tamper detection technique for identifying attacks in the watermarked image is a two level

thresholding scheme using normalized hamming similarity (NHS) and a tamper detection map. The

performance of the proposed technique was evaluated for a variety of content preserving manipulations and

malicious attacks. The proposed technique produces a better performance in terms of an increased PSNR

(Peak Signal to Noise Ratio) of the watermarked image and by localizing the malicious attacks when

compared to the existing techniques. The significant performance of the proposed semi fragile

watermarking technique is due to the combined results from both the NHS and the tamper detection map

which helps in localizing the malicious attacks and identifying the incidental manipulations. Also, the

authentication watermark which is a copy of the original image helps in identifying the tampered regions in

the attacked watermarked image.

1 INTRODUCTION

The present digital age of communication calls for a

secured way for communicating the confidential

information from one remote terminal to another.

Watermarking is one of the important techniques for

communication as it authenticates the received data

and also helps in identifying the attacks to the data.

Watermarking can be classified into fragile and semi

fragile. Fragile watermarking is sensitive even to a

single pixel change in the watermarked image and

hence making it unsuitable for watermarking images

in a noisy environment. On the other hand, semi

fragile watermarking is tolerant to incidental

manipulations to the watermarked image which are

called content preserving Manipulations. The

incidental manipulations include addition of noise to

the watermarked image, image compression,

Blurring etc. Semi fragile watermarking techniques

are also sensitive to deliberate malicious attacks to

the watermarked image making it suitable for using

it in noisy environment.

Most of the existing semi fragile watermarking

techniques rely on Discrete Cosine transform (DCT)

or Discrete wavelet transform (DWT) to hide the

watermark. The strategy which is used in semi

fragile watermarking techniques is to embed the

features of an image as a watermark. Some of the

existing DWT based semi fragile watermarking

techniques are discussed in this section. DWT based

watermarking technique (Hang and Park,2003)

embeds the just noticeable feature as a watermark.

Hu and Han(2005) embed the features generated

from the low frequency wavelet coefficients. A

DWT based Zernike moments is used as a feature in

(Liu et al., 2005). Hang and Sun (2003) embed the

semi fragile watermark by combining it with the

human visual model. Some techniques quantize the

wavelet coefficients to embed the watermark. Preda

(2013) embeds the watermark by quantizing the

second level DWT coefficients. Tsai and Chien

(2008) embeds the watermark into the second level

DWT coefficient using two different quantization

parameters. Preda et al., (2015) embeds the

watermark by quantizing the mean of a group of

second level coefficients. The drawback of all these

156

Sivasubramanian, N. and Konganathan, G.

Semi Fragile Watermarking Technique using IWT and a Two Level Tamper Detection Scheme.

DOI: 10.5220/0007759701560164

In Proceedings of the 4th International Conference on Internet of Things, Big Data and Security (IoTBDS 2019), pages 156-164

ISBN: 978-989-758-369-8

Figure 1: Flow chart of the proposed embedding technique.

schemes is that they are tolerant only to JPEG

compression and the effect of other content

preserving manipulations is not discussed

thoroughly. There are only a few semi fragile

watermarking techniques that are tolerant to a

variety of incidental manipulations. Tiwari et al.,

(2017) proposed a novel watermarking technique

based on vector quantization and modified index key

modulation. Benrhouma et al., (2015) proposed a

technique based on cat map and DWT. Qi and Xin

(2011) used a non traditional quantization method to

modify one chosen approximation coefficient.

Lai(2011) used singular value decomposition and

Tiny GA for semi fragile watermarking purpose. In

some of these approaches the PSNR value of the

watermarked image is very less and some

approaches do not discuss the effect of geometric

attacks on the watermarked images.

The proposed technique tries to address the above

drawbacks by proposing an embedding technique that

preserves the visual quality of the watermarked image

and by proposing a tamper detection technique for

testing the watermarked image to different content

preserving manipulations including geometric attacks.

2 FRAMEWORK OF THE

PROPOSED TECHNIQUE

2.1 Proposed Embedding Technique

A flow chart of the proposed embedding Technique

is shown in Figure 1. Let the size of the cover image,

I used in the proposed technique be MxM. In order

to obtain the watermark and embed it, integer

wavelet transform (IWT) is used to decompose the

cover image. Equation (1) represents the first level

decomposition of the cover image using IWT results

in one approximation coefficient CA and three

details coefficient CH, CV, CD which are of size

(M/2)x(M/2). The detail coefficient CH is again

decomposed according to equation (2) to obtain four

sub bands AA, AH, AV and AD which are of size

(M/4)x(M/4).







, , , 2 CA CH CV CD iwt I

(1)







, , , 2

A AH AV AD iwt CH

(2)

In the proposed technique the approximation

coefficient sub band AA is used to generate the

watermark which will be used for authentication at



Cover Image

Apply two level IWT

Generate the watermark from the

approximation coefficient

Store the watermark in the detail coefficient

Apply two level inverse IWT

Watermarked Image

Semi Fragile Watermarking Technique using IWT and a Two Level Tamper Detection Scheme

157

Figure 2: Flow chart of the proposed tamper detection technique.

the receiver end. The watermark, W which is of size

(M/4) x (M/4) is obtained using equation (3).



(, ) ,2()W i j xor dec n AAiibj

1, (/4)ij M

(3)

In equation (3) dec2bin represents the decimal to 8-

bit binary conversion of a pixel at the position (i, j)

and xor represents the logical exoring of the

resultant bits to obtain the watermark at the position

(i, j).

WW K





(4)

In order to improve the security of the generated

watermark, W is exored with shared secret key

matrix K to form RW which is shown in equation

(4). The shared secret key is a randomly generated

matrix of ones and zeros which is of size

(M/4)x(M/4) .The resultant watermark RW is

embedded into the detail coefficient sub band AV

using least significant bit substitution which are

shown by equations (5),(6) and (7).



Received Watermarked Image

Apply two level IWT

Extract the watermark from the

detail coefficients

Calculate the

watermark from the

approximate

coefficient

Is NHS>

0.99

Plot the tamper detection map

Are False

Positives

Present

Incidental attack

Image is tampered

with malicious

attack

Calculate NHS

Image is authenticated

0.50 ≤

NHS ≤0.99

Image cannot be

authenticated

IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security

158

(1 : 8) 2 ( ( , ))

dec bin AV i j

(5)

(8) ( , )BRWij

(6)

(, ) 2 ( )

Vij bindecB





(7)

In equation (5), every element of AV is converted

into binary bits and the watermark at its

corresponding position is embedded into the least

significant bit of AV which is B(8) to obtain AVˈ.

The final step shown in equations (8) and (9) is the

image reconstruction through inverse IWT to obtain

the watermarked image WI.



2,,, AA AHCH iiwt

VAD





(8)



2,,, CA CHWI iiwt CV CD



(9)

2.2 Proposed Tamper Detecting

Technique

A flow chart of the proposed Tamper detection

Technique is shown in Figure 2.Suppose the

received image WI is tampered via incidental

manipulations or malicious attacks. The proposed

tamper detection technique to differentiate an

incidental/content preserving manipulation from a

malicious attack is explained below:

The first step is the two level decomposition of

the received image WI using IWT which are shown

in equations (10) and (11).







1, 1, 1, 1 2 CA CH CV CD iwt WI

(10)







1, 1, 1, 1 2 1

A AH AV AD iwt CH

(11)

In order to identify the tampered portions of the

received watermarked image, watermarks CW and

EW are to be obtained. EW which is of size (M/4) x

(M/4) is the extracted watermark from the least

significant bits of AV1 as shown in equations (12)

and (13). CW which is of size (M/4) x (M/4) is the

calculated watermark from AA1 using equation (14).

(1 : 8) 2 ( 1( , ))

dec bin AV i j

(12)

(, ) (8) (, )

Wij B Kij

(13)



(, ) 1(, )2CW i j xor dec AAin ibj

1, (/4)ij M

(14)

Normalized hamming similarity (NHS) (Lu et

al,2005) is calculated between CW and EW using

equation (15) in order to know the effectiveness of

the attack on the watermarked image.

(, )

DCW EW

NHS





(15)

In equation (15), HD is the hamming distance

between CW & EW and NxN is their corresponding

size. Hamming distance represents the number of

positions at which CW and EW differs and this

variation is shown using the tamper detection map.

Using HD,NHS is calculated whose value ranges

from 0 to 1.NHS value of 1 indicates that both CW

and EW are identical and there is no attack on the

watermarked image. Therefore, higher values of

NHS signify that the calculated watermark is more

similar to that of the embedded watermark. In order

to distinguish the incidental manipulations from that

of the malicious attacks a threshold of 0.99 is fixed

on the NHS value (Tiwari et al., 2017). The

significance of this threshold is that a value of NHS

higher than 0.99 implies that the watermarked image

is free from malicious attacks and it is automatically

authenticated. If the value of NHS is less than 0.99

and greater than 0.50, then a tamper detection map is

plotted to ascertain the nature of attacks.

In order to plot the tamper detection map, at first

the tampered regions have to be identified. The

tampered regions are obtained from the hamming

distance calculated between CW and EW. Hamming

distance represents the corresponding positions

where the calculated and embedded watermarks

mismatch. In other words as CW is obtained from

AA1, the corresponding positions from HD can be

directly mapped onto AA1.At this stage, the

elements of AA1 will be labeled either as

authenticated or tampered. In order to refine the

tamper detection process neighbourhood

approximation is used.

Example 1: Illustration when a tampered pixel is identified

as authenticated.

Tampered Authenticated Authenticated

Authenticated Tampered Tampered

Tampered Authenticated Authenticated

Example 2: Illustration when a tampered pixel is identified

as tampered.

Tampered Tampered Authenticated

Authenticated Tampered Tampered

Tampered Authenticated Authenticated

Semi Fragile Watermarking Technique using IWT and a Two Level Tamper Detection Scheme

159

Figure 3: Some of the cover images used for testing (a) Baboon (b) Peppers (c) Lena (d) Goldhill (e) Fishing Boat (f)

Barbara.

The labeling of the eight neighbours of an

element in AA1 is taken into account to finalize

whether an element is tampered or not. As shown in

example 1, if the number of tampered neighbours

surrounding a tampered element is less than three

then the corresponding element is identified as

authenticated. As shown in example 2, if the number

of tampered neighbours surrounding a tampered

element is more than three then the corresponding

element is identified as tampered. By this way the

labeling of the elements in AA1 is fine tuned to plot

the tamper detection map. The tamper detection map

shows the spread of tampered and authenticated

elements in AA1. In order to detect malicious

attacks from the incidental manipulations, it is

important to identify any pattern in the tamper

detection map. An identification of a well defined

pattern outlining an area in the tamper detection map

clearly indicates that the attack is malicious

(Benrhouma et al., 2015). If the potentially tampered

elements are scattered all over the detection map like

a random noise and if it does not contain any

isolated tampered coefficients then the elements are

false positives and should be considered as authentic

(Preda et al., 2015). The final step is the

reconstruction of the received image using equations

(16) and (17).



1, 1, 11 , 12 AA AHCH i Ait

VDw





(16)



,21,, RCACHIiit CwCVD





(17)

3 RESULTS AND DISCUSSION

The cover images used for testing the proposed

tamper detection technique are of size

512x512.Some of the cover images used are shown

in Figure 3. PSNR (peak signal to noise ratio) is

calculated using equation (18) between the cover

and the watermarked images to access the visual

quality of the visual quality of the watermarked

images.

1 0log

PSNR

MSE

  









(18)

Where

()

i1 j1

ij ij

MSE











It can be shown from table 1 that the average PSNR

value using the proposed embedding technique

exceeds the acceptable value of 38 dB

(Voloshynovsiky et al., 2001). The efficiency of the

proposed technique was tested for a variety of

content preserving manipulations and malicious

attacks.

Table 1: PSNR of the watermarked images.

Cover Image PSNR of the watermarked image

Lena 41.80 dB

Baboon 31.96 dB

Barbara 39.63 dB

Peppers 42.01 dB

Gold Hill 41.21 dB

Airplane(F-16) 41.92 dB

Sailboat on Lake 38.75 dB

Fishing boa

41.39 dB

Elaine 40.05 dB

Table 2: NHS values for various watermarked image with

salt and pepper noise (sigma: 0.01).

Cover Image NHS value

Lena 0.9261

Barbara 0.9249

Elaine 0.9247

Airplane(F-16) 0.9244

Fishing Boat 0.9261

Peppers 0.9230

Sailboat on Lake 0.9268

IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security

160

Table 3: NHS values for various watermarked image with

rotation (degree: 45).

Cover Image NHS value

Lena 0.7414

Barbara 0.7433

Elaine 0.7347

Airplane(F-16) 0.7381

Fishing Boat 0.7374

Peppers 0.7476

Sailboat on Lake 0.7375

Table 4: NHS values for various watermarked image with

image brightening (Contrast Limits: 0.1 & 0.6).

Cover Image NHS value

Lena 0.6135

Barbara 0.5554

Elaine 0.6342

Airplane(F-16) 0.8339

Fishing Boat 0.5806

Peppers 0.6274

Sailboat on Lake 0.6818

3.1 Evaluation of the Proposed

Technique in Terms of Incidental

Manipulations

In order to prove the efficiency of the proposed

tamper detection technique in terms of incidental

manipulations, a variety of content preserving

attacks were considered. An attack is classified as

incidental if the NHS value is greater than 0.99.If the

NHS value is between 0.50 and 0.99 then the tamper

detection map is to be considered for identifying it.

Initially, the salt and pepper noise was added to the

watermarked image and the corresponding NHS

value was calculated. It can be inferred from table 2

that for various images the average NHS value after

adding salt and pepper noise comes to 0.90. It can

also be inferred from table 3 and 4 that rotating an

watermarked image by 45 degrees and adjusting the

contrast parameters produces an average NHS value

between 0.5 and 0.9.So,in order to correctly identify

it as an incidental manipulation tamper detection

map was plotted as can be shown in table 7. The first

row of table 7 shows the tamper detection map when

salt and pepper noise is added to the watermarked

image of ‘peppers’. The second row of table 7 shows

the tamper detection map when the contrast of the

watermarked airplane image was adjusted to 0.1 and

0.6.It can be inferred from the tamper detection map

that the tampered pixels are scattered all over the

image and it does not produce an defined pattern.

Due to the above reasons, the contrast adjustment

manipulation and the addition of salt and pepper

noise is identified as incidental. In the same way

when the watermarked image was attacked by

various incidental manipulations like speckled noise,

gamma correction, wiener filtering and motion

blurring the proposed tamper detection technique

produced NHS value between 0.5 and 0.9 as shown

in table 5.The fifth row of table 7 shows the tamper

detection map when the watermarked ‘cameraman’

image was manipulated by using wiener filtering

(with sigma :0.01).As the tampered pixels are not

isolated and are scattered all over the image, the

wiener filtering attack can be conclusively identified

as incidental. The Possible parameter values for the

content preserving manipulations for which the

Table 5: NHS values for various values of content preserving manipulations for ‘lena’.

Incidental Manipulation Parameter Value NHS value

Salt and Pepper Noise Sigma: 0 1

Sigma:0.1 0.5914

Gamma Correction Gamma: 0 1

Gamma: 2 0.5367

Wiener Filtering Filter size:3x3 0.5140

Speckle Noise Sigma: 0 1

Sigma: 0.1 0.5068

Gaussian Blur Sigma:4 0.6178

Image Brightening Contrast Limits:0.3 & 0.7 0.5923

Motion Blur Len: 5, theta: 45 0.5145

Len :20,theta:45 0.5240

Rotation Degree:6 0.5855

Degree:45 0.7433

Degree:80 0.62

Semi Fragile Watermarking Technique using IWT and a Two Level Tamper Detection Scheme

161

proposed tamper detection technique will identify as

incidental and not malicious is given in table 6 .The

better performance of the proposed technique is

because a copy of the image in the form of

watermark is used for tamper detection.

3.2 Evaluation of the Proposed

Technique in Terms of Malicious

Attacks

The efficiency of the proposed tamper detection

technique was also tested for malicious attacks like

object addition and deletion. The main objective of

the proposed two level thresholding is to properly

identify malicious attacks from incidental

manipulations. The malicious attacks was found to

produce a NHS value that was greater than 0.9.Since

the proposed technique is a two level thresholding

process, an object addition or deletion is clearly

outlined in the tamper detection map. This results in

identifying it as a malicious attack. As shown in the

third and the fourth column of the table 7, an object

addition or deletion to the original image clearly

outlines the tampered part which shows where the

malicious attack had taken place. The better

performance of the proposed tamper detection

technique is due to the two level thresholding of

NHS and tamper detection map to identify incidental

manipulations from malicious attacks. Further the

proposed embedding technique almost embeds a

copy of the original image by using a watermark of

size 128x128 which helps in identifying the

tampered elements at the receiver end.

Table 6: List of Incidental manipulations and its

parameters.

Manipulations Parameters

Salt and pepper Noise Sigma:0-0.1

Speckle Noise Sigma:0-0.1

Gaussian Blur Sigma:2-5

Motion Blur Len:5-20,theta:45

Gamma Correction Gamma:0.5-1.5

Rotation Degree:5-80

Wiener Filtering Size:3x3

Image Brightening Contrast Limits:0.3 & 0.7

Table 7: Tamper Detection map for various types of attacks.

Cover Image Attacked Image Tamper Detection Map Classification

Incidental

Malicious

IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security

162

Table 7: Tamper Detection map for various types of attacks (cont.).

Cover Image Attacked Image Tamper Detection Map Classification

Malicious

Incidental

Table 8: Comparison of the characteristics of the proposed technique with various methods.

Paper Technique

Maximum

PSNR(dB)

Tamper

Localization

Attacks Classification

Shen and Chen,

2012

DWT technique 30 ---

JPEG compression, Mean and

Median Filtering, Noise

Preda, 2013

DWT based

approach

40 Yes JPEG compression, Filtering

Li et al., 2015 Two level DWT 36 Yes JPEG compression, Gaussian Noise.

Zhang et al., 2016

DWT based

approach

40 ---

JPEG compression, Salt & Pepper and Gaussian

Noise, Speckle Noise, Image Rescaling

Shojanazeri et al.,

2017

DWT and Zernike

Moments

40.9 Yes

JPEG compression, Rotation, Scaling, Translation,

Additive Noise.

Proposed

IWT based

technique

42 Yes

Salt and pepper Noise, Speckle Noise, Gaussian Blur,

Motion Blur, Gamma Correction, Rotation, Wiener

Filtering, Image Brightening

Finally, table 8 compares the characteristics of

the proposed technique with the existing methods.

4 CONCLUSION

A semi fragile watermarking technique using integer

wavelet transform and a two level thresholding

scheme to identify attacks in the watermarked image

is proposed. Due to the usage of LSB substitution to

embed the authentication watermark, the

degradation in the visual quality of the watermarked

image is reduced. As a result, the PSNR of the

watermarked images using the proposed embedding

technique is greater when compared to the existing

techniques. On analyzing the proposed tamper

detection technique to a variety of content

preserving manipulations like addition of noises,

blurring, filtering, geometric attacks, image

brightening it is found that the image authenticity is

correctly verified. When malicious attacks like

object addition and object deletion was tested on the

watermarked image, the tampered pixels was clearly

outlined in the tamper detection map. The better

Semi Fragile Watermarking Technique using IWT and a Two Level Tamper Detection Scheme

163

performance of the proposed technique was due to

the two level thresholding scheme of NHS and

tamper detection map to identify the tampered

portions in the watermark image.

REFERENCES

H. Kang, and J. H. Park, “ A semi-fragile watermarking

using JND ,” in Proceedings of the Pacific Rim

Workshop on Digital Steganography (STEG 2003),

2003,pp. 127–131.

Y. P. Hu, and D. Z. Han, “ Using two semi-fragile

watermarks for image authentication,” in Proceedings

of the Fourth International Conference on Machine

Learning and Cybernetics, Guangzhou, China, 2005,

pp. 5484–5489.

H. Liu, J. Lin, and J. Huang, “Image authentication using

content based watermark,” in Proceedings of IEEE

International Symposium on Circuits and Systems,

Kobe, Japan, 2005, pp. 4014–4017.

H. Yang, and X. Sun, “ Semi-fragile watermarking for

image authentication and tamper detection using HVS

model,” in Proceedings of International Conference

on Multimedia and Ubiquitous Engineering, Seoul,

South Korea ,2007,pp. 1112–1117.

R. O. Preda, “Semi-fragile watermarking for image

authentication with sensitive tamper localization in the

wavelet domain,” Measurement, vol.46, no.1, pp.367–

373, January 2013.

M. J. Tsai, and C. C. Chien, “A wavelet-based semi-

fragile watermarking with recovery mechanism,” in

Proceedings of the IEEE International Symposium on

Circuits and Systems, Seattle, WA, USA, 2008, pp.

3033–3036.

R. O. Preda, I. Marcu, and A. Ciobanu, “Image

authentication and recovery using wavelet-based dual

watermarking ,” UPB Science Bulletin, Series C,

vol.77, no.4,pp.199–212,2015.

A. Tiwari, M. Sharma, and R.K. Tamrakar,

“Watermarking based image authentication and

tamper detection algorithm using vector quantization

approach,” AEU -International Journal of Electronics

and Communications, vol.78, pp.114–123, August

2017.

O. Benrhouma, H. Hermassi, and S. Belghith, “Tamper

detection and self-recovery scheme by DWT

watermarking,” Nonlinear Dynamics, vol.79, no.3,

pp.1817-1833,February 2015.

X. Qi, and X. Xin, “A quantization-based semi-fragile

watermarking scheme for image content

authentication,” Journal of Visual Communication and

Image Representation, vol.22, no.2, pp.187–200,

February 2011.

C. C. Lai, “A digital watermarking scheme based on

singular value decomposition and tiny genetic

algorithm,” Digital Signal Processing, vol.21, no. 4,

pp.522–527, July 2011.

Z. M. Lu, D. G. Xu, and S. H. Sun, “ Multipurpose image

watermarking algorithm based on multistage vector

quantization,” IEEE Transactions on Image

Processing, vol.14, no.6, pp. 822–831, June 2005.

S. Voloshynovskiy, S. Pereira, T. Pun, J.J. Eggers, and

J.K.Su, “Attacks on digital watermarks: classification,

estimation based attacks and benchmarks,” IEEE

Communications Magazine., 2001, 39 (8), 118–126,

August 2001.

H. Shen, and B. Chen, “From single watermark to dual

watermark: a new approach for image watermarking,”

Computers & Electrical Engineering, vol.38, no.5,

pp.1310–1324, September 2012.

C. Li, A. Zhang, Z. Liu, L. Liao, and D. Huang, “Semi-

fragile self-recoverable watermarking algorithm based

on wavelet group quantization and double

authentication,”

Multimedia Tools and Applications,

vol.74, no.23, pp.10581–10604, December 2015.

Z. Zhang, C. Wang, and X. Zhou, “Image watermarking

scheme based on DWT-DCT and SSVD,”

International Journal of security and its applications,

vol.10, no.10, pp. 191-206, October 2016.

H. Shojanazeri, W. A. W. Adnan, S. M. S. Ahmad, and S.

Rahimipour, “Authentication of images using Zernike

moment watermarking,” Multimedia Tools and

Applications,vol.76 , no.1, pp.577–606, January 2017.

IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security

164