Discrete Wavelet Transform based Watermarking for Image Content

Authentication

Obaid Ur-Rehman and Natasa Zivic

Chair for Data Communications Systems, University of Siegen, Hoelderlinstrasse 3, 57076 Siegen, Germany

{obaid.ur-rehman, natasa.zivic}@uni-siegen.de

Keywords: Watermarking, Content Authentication, Wavelet Decomposition, Security Analysis.

Abstract: A watermarking scheme based on discrete wavelet transform for content based image authentication is

proposed in this paper. The proposed scheme is tolerant to minor modifications which could be due to

legitimate image processing operations. The tolerance is obtained by protecting the low frequency data of

the wavelet transform using approximate message authentication codes. Major modifications in the image

content are identified as forgery attacks. Simulation results are given for unintentional modifications, such

as channel noise, and for intentional modifications such as the object insertion and deletion. Security

analysis is given at the end to analyze the security strength of the proposed image authentication scheme.

1 INTRODUCTION

With the rapid growth of Internet and

communications technologies, and the widespread

availability of multimedia generation and editing

tools, images can be easily generated and shared

over the Internet. However, due to this advancement,

image content can be conveniently edited and

reconstructed. As a consequence, the significance of

the techniques for image integrity verification and

content authentication is ever increasing. A digital

watermark is typically a visible or invisible signature

inserted inside the image to proof its authenticity or

ownership at a later stage. Digital watermarking

(Cox, et al., 2007), as opposed to digital signatures,

do not require extra bandwidth for transmission and

are designed to be prone to minor modifications in

the image data. However, they should be sensitive to

modifications in the image content. Digital

signatures, on the other hand, are very sensitive to

any modifications in the image data (or content).

With the standard authentication mechanisms, a well

known phenomenon called Avalanche Effect

(Fiestel, 1973) will result in failed authentication

even in the presence of a single bit error. There is a

new class of authentication mechanisms emerging

recently called the soft authentication mechanisms

(Ur-Rehman, 2013) or noise tolerant authentication

mechanisms. These mechanisms are designed to be

tolerant to minor modifications in the data, i.e., the

authentication will succeed even if the data protected

by these mechanisms is a little different than the data

on which the authentication tag was computed. A

watermarking scheme for image authentication is

proposed in this paper which is based on the

approximate message authentication code (AMAC)

(Graveman and Fu, 1999). AMAC is tolerant to

minor changes in data, whereas the standard

message authentication code (MAC) does not

tolerate any modification of the data.

This paper is organized as follows. Section II

discusses some related work. Section III discusses

the building blocks of the proposed watermarking

scheme. Section IV presents the watermark

generation, embedding and the watermark extraction

mechanisms. Simulation results are given in Section

V. Security analysis of the proposed scheme is

presented in Section VI. Finally, the paper is

concluded in Section VII.

2 RELATED WORK

Amongst the many methods for noise tolerant data

authentication, approximate message authentication

code (AMAC) is used in this work. Other techniques

for noise tolerant data authentication include, noise

tolerant message authentication code (NTMAC)

(Boncelet, 2006) and soft input decryption (Zivic,

2008). AMAC is based on majority logic, in which

620

Ur-Rehman, O. and Zivic, N.

Discrete Wavelet Transform based Watermarking for Image Content Authentication.

DOI: 10.5220/0006232306200625

In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 620-625

ISBN: 978-989-758-222-6

the authenticator tag is generated by arranging the

data in rows and columns. Then after XORing with

pseudorandom bits, the majority logic is used to

obtain the authentication tag. In AIMAC

(Graveman, Xe and Arce, 2000), which is a variation

of AMAC, the AMAC is adapted to image data,

such that it is tolerant to minor changes in the image

data but still able to differentiate intentional

forgeries. The results in the presence of image

modification scenarios including JPEG compression,

image forgery and additive Gaussian noise are given

in (Graveman, Xe and Arce, 2000).

The NTMAC algorithm is also tolerant to slight

modifications in data. The idea is based on splitting

the data into blocks, calculating standard MAC on

each block and retaining a portion of the whole

MAC for each block. This portion is used to detect

changes in the block. The concept of partitions is

used to introduce tolerance. Again certain variations

and improvements on NTMAC have been proposed

in literature. These include weighted noise tolerant

message authentication code (WNTMAC) (Ur-

Rehman, et al., 2011). WNTMAC is based on

NTMAC but introduces the concept of weights to

differentiate the relatively more important parts of

data from the lesser important parts. EC-WNTMAC

is an extension of WNTMAC, where the error

localization and correction capability is introduced

aside from error tolerance. However, all of the above

mentioned approaches are based on image data

authentication. They need to be used together with

image features for authentication of image content.

NTMAC was used for image content

authentication in (Ur-Rehman and Zivic, 2012).

Features of the image were generated based on

discrete cosine transform (DCT) and they were

protected using NTMAC. If error correction is

desired in addition to authentication, then error

correcting codes have been used together with

content authentication. This helps in error

localization and correction. In (Lee and Won, 2000),

Reed-Solomon (RS) codes are used to calculate

parity symbols for each row and column of an

image. These parity symbols are embedded as a

watermark in the two least significant bit (LSB)

planes of the image. RS decoder is used to “correct”

the modifications in the watermarked image. In

(Tabatabaei, et al., 2015), a two phase authentication

scheme is proposed which performs image

authentication in two stages. In one stage, the error

correcting codes are used to (partially) correct the

image and in the second stage, a tolerant

authentication is performed. The threshold is totally

flexible and can be adjusted to achieve the desired

level of flexibility.

Amongst the other interesting techniques, two

methods for self-embedding an image in itself were

proposed in (Fridrich and Goljan, 1999). This helps

in recovering those portions of the image which are

somehow damaged, e.g., through cropping, or

tampering. In the first method, the 8 × 8 blocks of an

image are transformed into frequency domain using

discrete cosine transform (DCT) and the coefficients

are embedded in the least significant bits of other

distant blocks. This method has a good quality of

reconstruction but it is very fragile. The second

method is based on the principle similar to

differential encoding where a circular shift of the

original image with decreased colour depth is

embedded into the original image.

3 BUILDING BLOCKS OF THE

PROPOSED WATERMARKING

SCHEME

3.1 Digital Watermarking

Digital watermarking is a technique of covertly

embedding digital data with secret information that

can be extracted by the recipient (Zivic, 2015). The

watermark should be unique, so that it can be later

used for authentication. Additionally, the watermark

should also be complex making it difficult for an

attacker to extract and damage or replace it. An ideal

watermark should be such that extracting it damages

the cover object. Applications of digital

watermarking include owner identification,

name a few. Watermarks are typically based on

image features. The features of the cover image are

extracted at first as,

   (1)

where Feature(·) is a feature extraction function,

applied on the cover image Image to obtain the

image feature f. The features uniquely identify the

cover image and two different images will have

completely different features. However, images with

the same content as the cover image will have more

or less the same features. The image feature, f, is

then used to generate a watermark, by protecting it

using a secret key, k, as,

  ,  (2)

Discrete Wavelet Transform based Watermarking for Image Content Authentication

621

where w is the watermark and

GenerateWatermark(·) is a watermark generation

function. Only the intended recipient(s) with the

shared key can extract the watermark and

authenticate the image. The watermark is then

embedded into the cover image. Two methods are

typically used for watermark embedding, i.e., either

in the spatial domain or in the frequency domain.

The watermark can be extracted at any later stage to

verify the authenticity of the protected data.

3.2 Discrete Wavelet Transform

Discrete wavelet transform (DWT) is used to

decompose an image hierarchically. Wavelet

transform decomposes the image into band limited,

low and high frequency components, which can be

reassembled to reconstruct the original image. A

DWT operation decomposes an image into four

components represented as LL, LH, HL and HH and

as shown in Fig. 1. Where L represents applying a

low pass operation and H represents applying a high

pass operation. Here LL is the low resolution

approximation image and it closely resembles the

original image. The other sub bands, LH, HL, and

HH represent other details such as edges etc. An

example DWT of the Lena image is shown in Fig. 2.

Figure 1: Single level DWT decomposition.

3.3 Approximate Message

Authentication Code

As already said, AMAC is an algorithm from the

class of noise tolerant authentication algorithms,

designed to tolerate minor modifications in a

message/image. This is different from the standard

MAC algorithms, which do not tolerate even a single

bit modification. The threshold on the acceptable

number of bit modifications is adaptable and the

tolerance exhibited by AMAC is due to the majority

logic. AMAC tag generation on a message M is

shown in Fig. 3, where L is the tag length and R and

S are positive integers. R is usually chosen to be

equal to S for simplicity (Graveman and Fu, 1999).

A pseudorandom number generator (PRNG) is used

to generate a stream of pseudorandom bits in the

AMAC using a secret key, k

, shared between the

sender and the intended receiver. As long as the bit

changes in the data are below the threshold, the data

is declared authentic. If the changes exceed the

threshold, the data is declared unauthentic.

Figure 2: The single level DWT decomposition of Lena

image.

4 WATERMARK GENERATION,

EMBEDDING AND

EXTRACTION

4.1 Watermark Generation

A source image is taken and DWT is computed on it.

The LL sub band of the DWT is taken and passed

through the AMAC algorithm. The AMAC tag is

taken as the watermark of the image. If there are

minor changes in the image, the LL sub band will

not change much and thus the AMAC will remain

the same. For changes beyond a threshold, such as in

case of forgery attacks, e.g., object insertion or

object removal, the AMAC will change. The

threshold is adjustable as discussed in the section on

AMAC. The length of AMAC tag is chosen to be

256 bits. The watermark generation for an example

Lena image is shown in Fig. 4.

4.2 Watermark Embedding

The watermark is embedded in the cover image. In

this work, the watermark is self embedded in the

source image. The source image is split into 8 × 8

pixel non overlapping blocks. The length of the

AMAC tag is 256 bits, which is split into 32 sub-

AMACs of 8 bits each. One sub-AMAC is taken at a



Singlelevel

DWT

decomposition

Image

LL HL

LH

HH

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

622

time and inserted into the LSBs of the first 8 pixel

values of the next image block which is obtained via

a secret permutation. The next image block is chosen

at random, using a secret key, k

, as the seed value.

Thus the AMAC tag is scrambled in the LSB of

image blocks. This makes it hard for an attacker, to

extract and replace the watermark without the

knowledge of the secret key.

Figure 3: AMAC tag generation (Graveman and Fu,

1999).

Figure 4: Watermark generation.

4.3 Watermark Extraction

When the authenticity of the image has to be proven,

the watermark is extracted back from the image. The

watermarked image is taken and split into 8 × 8

pixel non overlapping blocks. The LSBs of the first

8 pixel values of each next block is taken and

appended to the existing watermark to obtain the

complete watermark (the AMAC tag). The next

block is chosen again using the pseudorandom

permutation based on the shared secret key, k

, to

obtain the same sequence as obtained in the

watermark embedding procedure.

4.4 Image Authentication

An image is verified by comparing the extracted

watermark with the recomputed watermark. As the

watermark is embedded in the spatial domain, a part

of the cover image which is the source image as

well, is distorted. However, since AMAC is tolerant

to modifications below the chosen threshold, the

authentication will succeed even if there are other

deviations from the original.

5 SIMULATION RESULTS

Resolution of images used in these simulations is

256 × 256 pixels. The input image is first converted

to a grayscale image before being process further.

Single level DWT transform is applied on the

grayscale image. The length of AMAC is chosen to

be 256 bits and the length of a sub-AMAC is chosen

to be 8-bits. Simulations results are given in this

section for authentication in the presence of

intentional and unintentional modifications. Results

for unintentional modifications are based on “Salt &

Pepper” noise of varying magnitudes. Object

insertion is performed to test the proposed method in

the presence of intentional noise / forgery attacks.

Fig. 5 shows the 4 sub-bands of the single level

DWT decomposition of Lena image in the presence

of “Salt & Pepper” noise of magnitude 0.001. As it

can also be observed from Fig. 5, the LL sub-band

of the Lena image in the presence of “Salt &

Pepper” noise resembles the LL sub-band of the

original Lena image. The Hamming distance

between the two is 215. However, their AMAC tags

are similar based on the chosen value of threshold to

allow for bit differences of up to 300 bits in the data.

Therefore the Lena image in the presence of “Salt &

Pepper” noise passes the authentication test of the

proposed method and is declared authentic.



XOR

with

PRNG

Majority

LRows

ata

time

Permutation

00111001110110111….01111

… … … …

Message(M)orImage(I)

00111001110110111….01111000000(Zeropadding)

M(0) M(1) ….. M(L‐1)

M(L) …. ….. ….

…. …. ….. ….

…. …. ….

M(L(RS‐1))….M(LRS‐1)

(0) T

(1) ….. T

(L‐1)

(L) …. ….. ….

…. …. ….. ….

…. …. ….

(L(RS‐1))….T

(LRS‐1)

M(k

) M(k

+1)…..M(k

+L‐1)

M(k

) …. ….. ….

…. …. ….. ….

…. …. ….

M(k

RS‐1

)…. M(k

RS‐1

+L‐1)

(0) T

(1) …. T

(L‐1)

(L) …. …. ….

…. …. …. ….

(L(R‐1))…. T

(LR‐1)

(LR) …. …T

(LR+L‐1)

…. …. …. ….

(

2R‐1

))

….T

(

2LR‐1

)



(L(RS‐2))…..T

(L(RS‐2)+L‐1)

… …. ….. ….

(

RS‐1

))

….T

(

LRS‐1

)



(0,0)

(1,0)

….

(S‐1,0)

(0,1)

(1,1)

.…

.…

(0,L‐1)

.…

.…

(S‐1,L‐1)

A(0) A(1) ….. A(L‐1)

AMAC

Discrete Wavelet Transform based Watermarking for Image Content Authentication

623

Figure 5: Authentication in the presence of “Salt &

Pepper” noise of magnitude 0.001.

Figure 6: Authentication in the presence of “Salt &

Pepper” noise of magnitude 0.1.

However, if the noise level exceeds the

threshold, then the image is declared unauthentic. In

Fig. 6, the “Salt & Pepper” noise of magnitude 0.1

results in a Hamming distance of 22443 between the

LL sub bands of the original and the modified

images. It can be noticed from the figure that the

other sub bands are also severely affected by the

high magnitude of noise, though they are not used in

the authentication. Thus the authentication test fails

as the AMAC tags are different for both the images.

The test case of forged Lena image is shown in Fig.

7, with extra hair on the forehead. The LL band of

the forged image has a Hamming distance of 16315

with the LL band of the original Lena image of

similar dimension. Thus the AMAC tags are

different, resulting in a failed authentication.

Figure 7: Authentication in the presence of forgery attack,

with extra hair on the forehead.

6 SECURITY ANALYSIS

The security analysis of AMAC is given in (Onien,

Safavi-Naini and Nickolas, 2011), where it is proven

that if Hamming distance is used for distance

measurement, then it might not be secure for large

messages. The general security analysis of the

proposed authentication scheme can be done by

considering the key recovery and substitution

attacks. In key recovery attack, the secret key of the

scheme is disclosed using a sufficient number of

authenticated image-hash pairs. An attacker can then

use the recovered secret key to generate a watermark

of his own image to deceive the receiver. In the

substitution attack, the attacker tries to substitute a

valid image and its tag with another authentic image

and its tag. This attack is successful when the

substituted image is perceptually different than the

original image whereas the difference between their

watermarks or tags is below the threshold value.

6.1 Key Recovery Attack

The watermark generation and insertion uses two

secret keys in order to generation and embed the

watermark. An attacker must recover two secret

keys, one for watermark generation and another one

for watermark embedding. In AMAC, the reshaped

matrix of size R × L × S bits is used, which means

the attack complexity is about 2

R×L×S

function (tag

generation and verification) operations, which is

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

624

very high complexity even for small images, such as

56 x 56 pixels.

6.2 Substitution Attack

An attacker can execute the substitution attack in

two steps, a forgery attack on the AMAC and then

an attack on the watermark embedding. The

possibility for an attacker to pass the first step can be

calculated as follows. Let T be the threshold value

below which the difference between the AMAC tags

is acceptable and let t indicates the threshold for

difference between DWT’s LL sub-band tolerance.

The probability (P

) of changes in the “majority”

selection round of the AMAC is calculated in

(Onien, Safavi-Naini and Nickolas, 2011) as,

 



































































































tji

122

(3)

Based on Pt, the probability of deceiving the

attacker (P

) is calculated as,





















)1( (4)

can be decreased by increasing the length of

AMAC tag.

7 CONCLUSIONS

The paper proposes a watermarking scheme for

content based image authentication. The scheme

consists of generating the watermark based on image

features using discrete wavelet transform and

protecting them using the noise tolerant AMAC

algorithm. Simulation results show the noise tolerant

authentication capability for unintentional

modifications, such as through channel noise.

However, intentional modifications, such as forgery

attacks can be recognized using the proposed

watermarking scheme.

REFERENCES

Cox, I, Miller, M, Bloom, J, Fridrich, J and Kalker, T,

2007, Digital Watermarking and Steganography,

Morgan Kaufmann.

Fiestel, H 1973, ‘Cryptography and Computer Privacy’,

Scientific American, vol. 228, no. 5, pp. 15-23.

Ur-Rehman, O 2013, Applications of iterative soft

decision decoding, Aachen, Shaker Verlag.

Graveman, R and Fu, K 1999, ‘Approximate message

authentication codes’, Proceedings of 3

Fed. lab

Symposium on Advanced Telecommunications /

Information Distribution.

Boncelet, C 2006, ‘The NTMAC for authentication of

noisy messages’, IEEE Transactions on Information

Forensics and Security, vol. 1, no. 1, pp. 35-42.

Zivic, N 2008, Joint Channel Coding and Cryptography,

Aachen, Shaker Verlag.

Gravemen, R, Xie, L and Arce, GR 2000, ‘Approximate

image message authentication codes’, Proceedings of

Annual Symposium on Advanced

Telecommunications and Information Distribution

Research Program.

Ur-Rehman, O, Zivic, N, Tabatabaei, AE and Ruland, C

2011, ‘Error Correcting and Weighted Noise Tolerant

Message Authentication Codes’, 5

International

Conference on Signal Processing and Communication

Systems, Hawaii, 12-14 December.

Ur-Rehman, O and Zivic, N 2012, ‘Noise Tolerant Image

Authentication with Error Localization and

Correction’, 50

Annual Allerton Conference on

Communication, Control and Computing, Monticello,

Illinois, 1-5 October.

Lee, J and Won, CS 2000, ‘A Watermarking sequence

using parities of error control coding for image

authentication and correction’, IEEE Transactions on

Consumer Electronics, vol. 46, no. 2, pp. 313-317.

Tabatabaei, AE, Ur-Rehman, O, Zivic, N and Ruland, C

2015, ‘Secure and Robust Two-phase Image

Authentication’, IEEE Transactions on Multimedia,

vol. 17, no. 7, pp. 945-956.

Fridrich, J and Goljan, M 1999, ‘Images with self-

correcting capabilities’, Proceedings of the IEEE

International Conference on Image Processing, pp.

Kobe, 25-28 October.

Zivic, N ed. 2015, Robust Image Authentication in the

Presence of Noise, New York, Springer.

Onien, D, Safavi-Naini, R and Nickolas, P 2011,

‘Breaking and repairing an approximate message

authentication scheme’, Discrete Mathematics,

Algorithms and Applications, vol. 3, no. 3, pp. 393-

412.

Discrete Wavelet Transform based Watermarking for Image Content Authentication

625