- Email: [email protected]
Watermarking based image authentication and tamper detection algorithm using vector quantization approach
Accepted Manuscript Regular paper Watermarking based image authentication and tamper detection algorithm using vector quantization approach Archana Tiwari, Manisha Sharma, Raunak Kumar Tamrakar PII: DOI: Reference:
S1434-8411(16)31029-9 http://dx.doi.org/10.1016/j.aeue.2017.05.027 AEUE 51898
To appear in:
International Journal of Electronics and Communications
Received Date: Revised Date: Accepted Date:
19 October 2016 14 March 2017 17 May 2017
Please cite this article as: A. Tiwari, M. Sharma, R.K. Tamrakar, Watermarking based image authentication and tamper detection algorithm using vector quantization approach, International Journal of Electronics and Communications (2017), doi: http://dx.doi.org/10.1016/j.aeue.2017.05.027
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1 2 3 4 5 6
Watermarking based image authentication and tamper detection algorithm using vector quantization approach Archana Tiwari1, Manisha Sharma2* and Raunak kumar Tamrakar3 1,2
Department of Electronics & Telecommunication, Bhilai Institute of Technology, Durg 3
7 8
Department of Applied Physics, Bhilai Institute of Technology, Durg *Corresponding Author: [email protected]
9 10
Abstract-
11
quantization (VQ) approach is presented for digital image authentication. Watermarks are
12
embedded in two successive stages for image integrity verification and authentication. In the
13
first stage, a key based approach is used to embed robust zero level watermark using
14
properties of indices of vector quantized image. In the second stage,semifragile watermark is
15
embedded by using modified index key based (MIKB) method. Random keys are used to
16
improve the integrity and security of the designed system. Further, to classify an attack
17
quantitatively as acceptable or as a malicious attack, pixel neighborhood clustering approach
18
is introduced. Proposed approach is evaluated on 250 standard test images using performance
19
measures such as peak signal to noise ratio (PSNR) and normalized hamming similarity
20
(NHS). The experimental results shows that propose approach achieve average false positive
21
rate 0.00024 and the average false negative rate 0.0012. Further, the average PSNR and
22
tamper detection/localization accuracy of watermarked image is 42 dB
23
respectively; while tamper localization sensitivity is very high. The proposed model is found
24
to be robust to common content preserving attacks while fragile to content altering attacks.
25
Index Terms -Attack classification; Biometric verification watermark; Image authentication;
26
Robust watermarking; Semifragile watermarking; Vector quantization.
27
In the present work, a novel image watermarking algorithm using vector
and 99.8%
1. INTRODUCTION
28
With development of the digital technology recreation of digitally generated information has
29
become very easy, and can be transmitted by digital media with ease among other medias of
30
message conveyance images are most common. Therefore protection of these images and its
31
content authentication is important. To authenticate the integrity and authenticity of a digital
32
image, digital watermarking techniques have been considered an effective technique [1-
33
5].The watermarking technique for authentication purpose can be classified as Robust, fragile
34
and semifragile watermarking techniques based on their level of security. A robust watermark
1
35
finds application in the area of ownership verification and copyright protection [6-9]. A
36
fragile watermark destroys itself even for small accidental or intentional changes [10-12].On
37
the other hand, semi-fragile watermarking is a practical technique which can protect integrity
38
of digital images and can locate the tampered areas [13-16]. Thus fragile or semi-fragile
39
watermarking are those image watermarking techniques which find application in images and
40
video content authentication and tamper detection [7, 13, 16, 17]. Given the limitations of
41
robust and fragile watermarking schemes, multiple-watermarking is applied to digital data
42
where copyright protection and content authentication is required to be addressed
43
simultaneously [7]. Lu [7], pioneered the concept of VQ based image watermarking
44
technique. Various VQ (vector quantization) based watermarking techniques are
45
recommended by researchers in the past decade for image watermarking [17-20] with the
46
objective of image compression and authentication.
47
Most of the methods suggested in above literatures need improvement in issues
48
related to watermarked image quality, identification of attacks as malicious or content
49
preserving manipulation, and efficiency of the system in authentication of images. Therefore,
50
an efficient image authentication algorithm having a good visual quality of watermarked
51
image, having the advantage of comparatively less bandwidth and storage space, with high
52
accuracy in attacks classification, tolerant to acceptable manipulations, and high tamper
53
detection accuracy needs to be designed. Thus in proposed work, watermarks are embedded
54
in two successive stages. In the first phase robust watermark is used for ‘zero level
55
watermarking’ [21] utilizing the statistical properties of VQ indices for enhancing the
56
security of watermarked image by biometric verification. The advantage of first stage of
57
watermarking is, it does not affect VQ reconstructed image. In the second step, the
58
semifragile watermark is embedded by using a Modified Index Key based method (MIKB)
59
for image authentication. A ‘quantitative threshold-based approach’ is suggested for
60
classifying attacks and localizing tampered area. Further statistical analysis is done to find
61
false positive rate, false negative rate, % of correctly identified authenticated images and %
62
of falsely verified images to validate the correctness of the system. Thus, present image
63
authentication fulfils desired characteristics of an effective image authentication technique.
64
The significant features of proposed improvised scheme are-
65 66
Watermarks are embedded using VQ indices properties; it enhances security and saves bandwidth requirements.
2
67
Random keys are used for embedding robust zero level watermark and semi-fragile
68
watermark respectively; random nature of key improves the integrity and security of
69
designed system.
70
71 72
preserving attack.
73 74
A quantitative analysis is suggested to classify the attack as malicious or content
Tamper detection accuracy of proposed system is very high; it can detect tampering upto 10% as malicious attack.
Biometric verification watermark is used to enhance reliability of designed system.
75
The structure of the paper is organized into three sections namely proposed scheme,
76
Experimental results and discussion of deigned algorithm, and lastly conclusion of the paper
77
is addressed.
78
2. PROPOSED SCHEME
79
In this section, the detail of proposed image authentication algorithm is presented. The
80
proposed method uses two-stage VQ, to divide the encoding task into successive stages;
81
wherein initial stage performs a relatively rough quantization of the input vector using a
82
codebook. Then, a second stage quantizer performs on the error vector between the original
83
and quantized first-stage output. The quantized error vector then gives a second
84
approximation to the original input vector thereby provides a more accurate or refined
85
depiction of the input. The binary watermark images are embedded in two successive stages
86
using random key based approach and properties of indices, the output of both stages are
87
combined to get watermarked image. Watermarked image is sent over, an open
88
communication channel, where accidental or malicious changes may affect contents of
89
watermarked image. In the verification and authentication step, the watermark sequence is
90
extracted independently from received watermarked image, using passkey1 and passkey2 in
91
stage 1 and passkey3 in stage 2. The authenticity of the received image is analyzed using the
92
extracted watermark images. Details of tamper detection, localization, and classification of
93
attacks are described in subsequent sections. Figure 1 shows two stage vector quantization
94
procedure followed in proposed method.
95
The proposed technique is described in four subsections; subsection 2.1 describes
96
codebook generation , 2.2 explains proposed watermark embedding technique, subsection 2.3
3
97
discusses watermark extraction in two consecutive stages. Finally, authentication and
98
verification of received image are discussed in subsection 2.4.
99 100
2.1 Codebook generation
101
To create a codebook of M codewords, input training images are selected. Each input image
102
is partitioned into non-overlapping image blocks of m×m pixels. Thereafter, M image blocks
103
from the input image blocks are selected as the initial codewords. This set of initial
104
codewords can be termed as the initial codebook. After generation of the initial codebook, an
105
iterative process for several rounds is executed to generate the codebook using LBG
106
algorithm [22]. In each round, the data clustering procedure and the centroid updating
107
procedure are performed. In the data clustering procedure, the nearest codeword of each input
108
image block in the codebook is to be found. To determine the nearest codeword of each input
109
image block in the codebook, a total number of M squared Euclidean distances are computed.
110
Then, the closest codeword is determined by finding the minimal squared Euclidean distance
111
from itself to the input image block. After the data clustering process is executed, all these
112
input image blocks are classified into M groups. In the centroid updating procedure, the mean
113 114
Figure1: Proposed two stage vector quantization procedure.
115 116
vector of the training image blocks in each group is calculated. These M computed image
117
vectors generate new codebook. By repeatedly executing the same process in each round, a
118
representative set of codewords can be generated for first stage. In second stage error image
119
is used as input training image for generation of codebook2 by repeating same process. In
120
proposed work a new codebook is generated each time algorithm runs, for same input image,
121
to enhance security of algorithm. Randomness is achieved by randomly generating a set of
4
122
input training vectors for codebook generation in the range of minimum to maximum pixel
123
values.
124
2.2 Proposed watermark embedding technique
125
Embedding of the watermark is done in two stages: robust zero level watermark embedding
126
and semifragile watermark embedding. The detail of watermark embedding process is
127
described in two subsections namely 2.2.1 discusses watermark embedding in the first stage,
128
and 2.2.2 explains the second stage of watermark embedding and final watermarked image
129
construction.
130
2.2.1. First stage of watermark embedding: In the first stage, a robust binary fingerprint
131
watermark is embedded. Robust zero level watermark embedding is done in two steps:
132
Step 1: Robust zero level watermark embedding: After generation of codebook1 (C1), the
133
robust binary watermark
134
passkey1, where
135
original image S with size is m
136
(a, b) denotes the image block at the position of (a, b). After that, each vector s (a, b) finds its
137
best codeword in the codebook C1 and a index ‘g’ is assigned to s (a, b), we can then obtain
138
the indices matrix T with elements t (a, b). , which can be represented by:
139
T=VQ(S)=
140
For most images, the VQ indices in the neighboring blocks may be same; this characteristic is
141
used to calculate parameter ‘E,' called embedding function.
142
, with pixel of size (p × q) is scrambled with randomly generated
passkey1
((p × q), resultant scrambled robust watermark is then divided into vectors s (a, b) with size 4
=
, where s
(1)
of adjacent blocks.
143 144
. The
(2) The embedding function can be computed as follows:
145
E=
146
Where
(3) ;
147
Where ‘Th’ is a threshold, i.e. standard deviation of one-dimensional indices matrix T. Now
148
embedded robust watermark or passkey2 can be generated by EX-ORing of the scrambled
149
watermark with the embedding function. The watermark is termed as ‘robust zero level
150
watermark,' as it is robust to almost all image manipulations and it does not change VQ
151
compressed image at all.
152
passkey2=
(4)
5
153
Step2: First stage VQ image reconstruction: After robust watermark embedding, the
154
reconstructed image S’ (m, n) and error image Q (m, n) can be obtained as follows. In first
155
stage compressed image is watermarked without actually inserting it explicitly .This concept
156
of Zero-watermark algorithm was suggested by Wen et al. [21], which does not modify the
157
original image. And the most important step is it is constructed from characteristics of the
158
original image. VQ-1
159
(5)
160
Q(m,n)=S (m, n)-S’ (m, n)
161
Equation 5 shows reconstructed image after the first stage. While equation 6 represents the
162
output of the first stage, this is treated as an input to the subsequent stage.
163
2.2.2 Second stage of watermark embedding: In the second stage, a semifragile
164
authentication watermark is embedded. For a codebook2 generation, the error image, Q (m,n)
165
serves as input training sequence to improved LBG algorithm. Further, like the original
166
image, Q (m, n), is also divided into ‘L’ vectors ‘q (a, b),' each of size 4×4, where q (a, b)
167
denotes the image block at the position of (a, b) .After codebook (C2) generation, each vector
168
of input training vector finds its most suitable code vector C2x, in codebook2 (C2) and
169
assigns indices
170
elements x (a, b) which can be represented as follows.
171
X=VQ (Q) =
172
Now each element of indices matrix, X is modified according to binary semifragile
173
watermark Ws .For semifragile watermark embedding; a single bit of watermark Ws is
174
inserted into each element of X. Let K be the size of codebook2, where K=2v then each
175
index has got ‘v’ binary bits. The embedding position of each semifragile watermark bit is
176
decided using randomly generated passkey3, where
177
significant bit) end watermark bit Ws replaces binary bit of each index of matrix X at the
178
position defined by passkey3. For example let, X11=22, its binary equivalent is
179
(000000000010110)2, assuming passkey3=3 and Ws (11) =0, then after watermark
180
embedding, X11 is modified to (000000000010010)2 i.e. 18. Using these modified indices
181
semi fragile watermarked image is reconstructed at VQ decoder, thus method is termed as
182
modified indices key based method (MIKB). Let reconstructed image be
183
stage output, which is represented as follows:
184
=
(6)
to q(a, b). Now using these indices, indices matrix X is created, with
=
(7)
passkey3
. From LSB (least
i.e. second
(8)
6
185
Each time algorithm runs a unique random generated key is used for semifragile watermark
186
embedding. Finally, watermarked image is obtained by combining first stage and second
187
stage outputs.
188
(9)
189
Thus, in the proposed scheme the watermarks are embedded independently in successive
190
stages for verifying integrity and authenticity of the image.
191
2.3 Watermark extraction in two stages
192
This subsection explains watermark extraction procedure of proposed algorithm. It is
193
understood here that both sender and receiver share same codebooks. The procedure adopted
194
for watermark extraction is given below.
195
Step1- Robust watermark extraction in the first stage: At the receiver side, the reverse
196
process of the watermark embedding process is carried out. The received image
197
similar way while encoding, is divided into non-overlapping blocks of size 4×4. The VQ
198
encoder as discussed before is used along with codebook1 to find encoded indices of first
199
stage. Obtained VQ indices are used to compute embedding function ‘E’. Then XOR
200
operation is performed between the embedding function ‘E’ and passkey2 to get back
201
permuted robust watermark
202
passkey1 to get the extracted robust watermark
203
reconstruct the first stage approximate output
204
Step 2-Semifragile watermark extraction: In the second stage of extraction, watermarked
205
image
206 207 208 209 210
. Further, these indices are used to
Codebook 1
Watermarked Image
Partitioned into 16384 blocks each of size 4×4
Indices
First stage VQ encoder
Passkey1 Inverse permutation
212
Extracted robust watermark
213
217 218 219 220 221
. Finally inverse permutation process is carried out with
is segmented using the previous phase output
211
214 215 216
, in
Permuted robust watermark
Embedding function (E) computation
E
Codebook 2 Second stage VQ encoder
Watermarked image segmentation
First stage o/p reconstruction using VQ decoder
Indices Passkey3
Extract bit positions
Restore extracted bits to reconstruct semifragile watermark
Extracted semifragile watermark
Figure 2: Illustration of steps involved in the extraction of the watermark in two subsequent stages. 7
222
To obtain the second stage approximated image Q’’ (m, n) which is further divided into L
223
vectors, each of size 4×4.Standard VQ encoder executes the nearest neighbourhood search to
224
obtain encoded indices (
225
are recovered from indices
226
embedding order to form extracted semi fragile watermark
227
extraction is given in figure 2.
228
2.4 Verification and authentication of received image
229
The verification and authentication procedure is divided into five steps. After extraction of
230
the robust and the semifragile watermarks, the received image is to be verified and
231
authenticated. In next step, details of various stages involved including tamper detection
232
localization procedure and performance evaluation parameters are explained.
233
Step 1: Verification of received image: In the first step the integrity of received
234
image
235
the embedded fingerprint robust watermark, using parameter NHS. NHS [7] is defined in
236
equation 10, where
237
watermark.
238
NHS
). Then using passkey3 semifragile authentication watermark bits . These extracted watermarking bits are then grouped as per . Detail process of watermark
is to be verified, the extracted fingerprint robust watermark is compared with
represents the embedded watermark and
denotes the extracted
(10)
239
Here HD (.,.) stands for the Hamming distance between two binary images, i.e., the number
240
of bits different in two binary images. If the value of NHS is equal to unity, that means
241
embedded watermark and extracted watermark are identical. Thus, the integrity of source is
242
verified, otherwise not.
243 244
(a)
(b)
245
Figure 3 (a) Depicts an error pixel surrounded by five pixels which are error pixel (b) Depicts
246
an error pixel surrounded by two error pixels.
247 248
Step 2: Authentication of received image: In next phase, authentication of received image
249
is performed, the similarity between the embedded and the extracted semifragile watermark is
250
checked. If the value of
251
received image
then there is a possibility of some security attack on . As semifragile watermark is used as authentication watermark in 8
252
proposed algorithm, all authentication analysis is performed using the semifragile watermark.
253
Further difference image is obtained as
254
.If any pixel of
is one it is considered as an error pixel; otherwise, it is not an error pixel. A pixel
255
neighbourhood criterion is used for content authentication of image, as explained in step 3.
256
Step3: Tamper detection, localization, and attack classification: If any error pixel is
257
identified in
258
neighbourhood approach. A pixel is considered as malicious pixel if more than four pixels in
259
its surrounding eight pixels are having value one otherwise it, is classified as a nonmalicious
260
pixel. Figure 3 (a) shows malicious pixel, whereas Figure 3 (b) shows nonmalicious pixel.
261
The threshold (TR) value is obtained after testing of 250 images for different nonmalicious
262
and malicious attacks. It was found malicious attacks contains more than 57 percentage
263
malicious error pixel, so the threshold is set to 57 percentages accordingly. Table 1, shows
264
notations used in the algorithm for authenticating images.
265
Table 1 Notations used in algorithm for authenticating image Notations R(m,n) N N1 N2 T1 T2
266 267 268
, it is checked for malicious or nonmalicious attack by using 3×3 pixel
Descriptions Difference image of size m×n. ). Total no of error pixels in No of pixels having four or more than four surrounding error pixel. , percentage of error pixels. , percentage of malicious pixels.
The algorithmic structure of proposed system for attacks classification is given as:
275
Compute N1 from image . If T1=0, then image is not attacked. The image is considered as authentic. o End. If, T1≠0, then compute the value of NHS. If NHS> 0.7, then extracted image is recognizable. Further, check for authenticity of the biometric image. o If T2TR, thenotations image isused nonauthentic. Else for authenticating images. The scheme uses If T1>0 and approach NHS< 0.7for and T2
276
White pixel areas on the difference image broadly identify the tampered location, while
277
tampering is localized by finding coordinates of rows and columns of all tampered pixels
278
locations. Table 2 explain various evaluation parameters used for tamper detection and
279
classification of attack and Table 3 represents pixels classification statistics.
269 270 271 272 273 274
9
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296
Figure 4 Verification and authentication process Table 2 Performance parameters used to test the effectiveness of proposed scheme. Parameter M
Description Total pixels in received image . Total tampered pixels in received image . Denote no of tampered pixels which are accurately identified as tampered. Denote no of pixels which are not tampered but falsely detected as tampered pixels. Denote no of pixels which are tampered but not identified as tampered pixels.
True positive: TP False negative: FP True negative :(TN)
-
False positive: (FN) False positive rate: False negative rate: % Accuracy Mean square error (MSE)
% Accuracy= (TP+TN) / (TP+TN+FP+FN)
PSNR 10
297
Table no 3 Pixel classification based on statistical results Pixel classification No. of Tampered pixels No of non-tampered pixels
Test positive
Test negative
298 299
3. EXPERIMENTAL RESULT AND DISCUSSION
300
The performance of the proposed watermarking based technique for digital image verification
301
and authentication using vector quantization approach is investigated. The experiment is
302
performed using MATLAB version 7.10.0.499 (R2010a), 64-bit (win64) software. In the first
303
phase, the experiment is conducted on 300 different 8-bit images of resolution 600 dpi each
304
of size 512 ×512, the binary watermark image of size 128×128 as a robust watermark image
305
and semifragile watermark image. Figure-5 shows five test images whose experimental
306
results are given in the subsequent subsection. The work is carried out employing various
307
combinations of codebook1 and codebook2 under the condition codebook size is less than
308
512.The visual quality of the watermarked images is evaluated by PSNR ,which comes to be
309
42 dB (average). It is observed that maximum PSNR is achieved for all images for codebook
310
size 256, for both codebooks.
311
(2)
(1)
(3)
(4)
(5)
312 313
Figure 5: Different input test images used in the proposed method (1) Lena (2) Baboon(3)
314
Pepper(4)Cameraman(5)F-16.
315
(Image source: http://sipi.usc.edu/database/database.php?volume=misc.)[25].
316
0
317
In Table 4, performance of the proposed method is evaluated for ‘Lena’ image against
318
different size of watermark has been evaluated, when no attack is applied. For embedding
319
watermarks the original image is partitioned into a number of blocks according to the size of
320
watermark. Moreover, codebook sizes are half of original image size. However, overall
321
performance of proposed method highly depends on size of semifragile watermark, as
322
watermarked image is not affected by robust watermark.
323 11
324 325
Table 4 Performance of proposed method for different size of Lena image against varying watermarks size. Original image
Size of watermarks Robust Semifragile
Biometric robust Binary semifragile PSNR(dB) watermark watermark
256 256
64 64
64 64
40.11
256 256
32 32
32 32
41.98
512 512
64 64
128 128
42.81
128 128
32 32
64 64
39.55
512 512
256 256
64 64
43.94
326 327
3.1 Verification of Received Image using Robust Watermark
328 329
Figure 6.Robust watermark as verification watermark
330 331
Once watermarked image is received, extracted watermark at first stage is checked, for
332
verification of integrity of image. In present method fingerprint verification method is used
333
for finding integrity of image. Referring to figure 4, under no attack condition NHS is one
334
whereas, when any of passkey1 or passkey 2 changes value NHS≠1, that is considered as
335
threat to security of system. In figure 6, row1 shows results when same key is used both at
336
transmitter and receiver end. Whereas, row2 demonstrate results when passkey1 is changed
337
and passkey2 is not changed.
338
3.2. Robustness to Common Image Processing Manipulations
339
The performance of the method is examined by applying six types of image processing
340
attacks such as Median filtering, blurring, low-pass filtering using a Gaussian filter, salt and 12
341
pepper noise, rotation at various angles, and JPEG compression on 250 watermarked images.
342
Table 5 to 10, demonstrate various results obtained for robustness by processing the four test
343
images by applying different content preserving attacks as enumerated above.
344
terminologies used in the Table are: Tw show number of error pixels obtained from extracted
345
watermark image, T1 is a percentage of error pixels in the difference image, T2 is a
346
percentage of malicious pixels in the difference image. Here blur attack is applied for
347
different radius from 0.0 to 0.9; median filtering is applied to various window sizes from 1×1
348
onwards, Gaussian filtering for different variances, salt, and pepper noise for various
349
percentages from one percent onwards, rotation in both clockwise and anti-clockwise
350
directions and JPEG compression for varying quality factor. The degradation in quality is
351
evaluated using PSNR for received watermarked image and NHS for extracted semifragile
352
watermark images.
353
The
Table 5 Experimental results illustrating blurring the image at different radius. Image Pepper Baboon Lena F-16
Tw 3043 4757 3225 2343
%T 52.4 79.92 87.73 56.33
% T2 52.3 50.06 55.12 41.24
PSNR(dB) 38.49 35.56 38.86 37.28
NHS(semi) 0.8112 0.7067 0.8082 0.8570
Blur radius Radius=1.0 Radius=1.1 Radius=1.1 Radius=1.0
Attack Incidental Incidental Incidental Incidental
354 355
Table 6 Experimental results for median filtering at different window sizes. Image Pepper Baboon Lena F-16
356 357 358
%T 46.98 72.14 70.89 60.55
% T2 23.38 56 29.7 20.55
PSNR(dB) 43.04 36.32 41.72 39.16
NHS(semi) 0.8048 0.8304 0.7150 0.7472
Window size 3×3 filter 2×2 filter 3×3 filter 3×3 filter
Attack Incidental Incidental Incidental Incidental
Table 7 Robust performance for a Gaussian filter for 3×3 window using four different values of variance at mean=0. Image Pepper Baboon Lena F-16
359 360
Tw 3798 3491 4665 4142
Tw 3026 4763 815 749
%T 71.67 80.13 75.48 67.06
% T2 54.1 45.81 51.8 37.37
PSNR 44.31 36.49 40.46 37.68
NHS(semi) 0.8163 0.9136 0.9503 0.9543
Standard deviation Sigma=0.8 Sigma=0.6 Sigma=0.6 Sigma=0.6
Attack Incidental Incidental Incidental Incidental
Table 8 Experimental results for salt and pepper noise attack at various noise densities. Image Pepper Baboon Lena F-16
Tw 4986 4712 4924 4881
%T 6.97 6.94 6.04 6.97
T2 0 0 0 0.00005
PSNR(dB) 41.92 42.52 48.76 40.18
NHS(semi) 0.7136 0.7075 0.7679 0.7021
Noise density 0.07 0.07 0.06 0.07
Attack Incidental Incidental Incidental Incidental
361 13
362
Table 9 Robustness performance at different rotation angles Image Pepper Baboon Lena F-16
363 364
Tw 3242 2387 4498 4641
%T 72.44 72.04 85.90 74.39
% T2 50.5 27.00 54.56 54.3
PSNR 40.41 37.24 40.32 40.95
NHS(semi) 0.8021 0.8543 0.7255 0.7167
Rotation angle 0.5 0.4 0.5 0.5
Attack Incidental Incidental Incidental Incidental
Table 10 Robustness performance of JPEG compression at different quality factor Quality factor Tw %T % T2 PSNR (dB) NHS(semi) Attack 70 2819 87.0 57 40.9 0.85 Incidental 80 1454 88.0 56 43.0 0.91 Incidental 90 171 83.9 51 42.0 0.98 Incidental 99 00 23.0 00 43.0 1.00 Incidental
365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392
Figure 7. Illustration of Robustness nature of the scheme: row-wise (1) Gaussian filtering for window size 3×3 and a Sigma 0.6.(2) Median Filtering for a 2×2 window. (3) Salt and pepper noise addition to noise density 0.07. (4) Blurring with radius 1. (5) Rotation by 0.5 clockwise angles.(6) JPEG compression for quality factor 70. The performance of attacks on the image is evaluated based on the computed
393
percentage of malicious pixels. If the percentage is less than threshold i.e. 57% and the NHS
394
is higher than 0.7 the attack is classified as non-malicious otherwise, it is a malicious attack.
395
Experimental results show that each attack is robust to a certain value of parameters as
396
depicted in Tables 5-10. On comparing the results of the Tables 5-10, it is seen there is a
14
397
small variation in robustness feature of images. For robustness feature, taking minimum
398
values of parameters for different robustness attacks.It is observed that the proposed method
399
is robust for blurring attack for one radius, median filtering up to 2×2 window size, Gaussian
400
filtering for variance 0.6 for the 3×3 filter, for rotation to 0.45 angles, salt and pepper noise
401
density 0.06 and JPEG compression quality factor 70.These results show that there is a slight
402
difference in the level of robustness for different images.
403
Figure 7 illustrates the robust nature of the algorithm; here some results are presented
404
for visualization of experimental results. The algorithm suggests improvement in work
405
proposed by Lu [7]. Lu suggested attacks classification based on NHS value, but the
406
algorithm [7] is unable to classify whether attacks is incidental or malicious. Robustness
407
performance of proposed method is compared with [11, 12, 14, 15, 16]; it is seen none of
408
these algorithms report any method for classifying attacks.
409
3.3. Fragileness characteristics for malicious attacks
410
Further, this algorithm is tested for some content altering attacks such as cutting a portion of
411
watermarked image and cropping the image to different window sizes. Here results of four
412
test images are shown for the cut attack, in Table 11. Table 12, shows crop attack for varying
413
window sizes(window sizes shows dimension of cropped area), the Table shows average
414
results obtained by simulating 250 images. Table 11 and 12 demonstrate that the suggested
415
scheme can identify these attacks as malicious because the percentage of the malicious pixel
416
is more than the threshold in all cases. It is found from experimental results that the proposed
417
method can distinguish a single pixel change in received image and identify tampering
418
greater than 10% as a malicious attempt.
419
Table 11 Simulation results of watermarked test images for Cut attack Image Pepper Baboon Lena F-16
420 421
Tw 29 982 871 128
%T 2.67 2.82 2.77 2.68
% T2 82.33 91.72 89.01 89.46
PSNR(dB) 41.27 39.17 39.51 37.18
NHS(semi) 0.9982 0.9401 0.9468 0.9948
Attack classification Malicious Malicious Malicious Malicious
Table 12 Simulation results of for crop attack results show average values of all images Window Tw %T % T2 PSNR(dB) NHS(semi) Attack classification 8×8 54 0.031 60.49 37.70 0.9967 Malicious 40×40 360 0.640 90.48 38.96 0.9800 Malicious 130×130 1878 6.541 96.96 34.67 0.8853 Malicious 200×200 3149 15.410 98.00 31.48 0.8075 Malicious 300×300 4639 34.600 98.67 28.41 0.7076 Malicious
422
15
423
The sensitivity of given algorithm is clearly shown in Table 12; it can detect and
424
locate cropping as the malicious attack for the 8×8 window. However, the given algorithm
425
can detect a single pixel change in received image. Various performance measures are
426
evaluated for finding tamper detection capacity of algorithm and results are shown in Table-
427
13. The performance evaluation parameters are - TP, TN, FP, FN, FPR, FNR and accuracy of
428
tampered images for different window sizes (tampered area). Values in the Table show
429
average results obtained for various test images. Referring to the same table, results clearly
430
demonstrate that the algorithm shows high accuracy in tamper detection, and is having very
431
low false negative and positive rates. Figure 8, presents a visual presentation of experimental
432
results obtained for various fragile attacks.
433 434 435
Table 13 Statistical analysis of tamper detection capacity of images results show average values Window 40×40 80×80 130×130 200×200 300×300
TP 0.997 0.999 0.999 0.999 0.999
FP 0.00031 0.00063 0.00106 0.00180 0.00350
TN 0.9997 0.9990 0.9989 0.9981 0.9965
FN 0.00250 0.00063 0.00024 0.00010 0.00670
FPR 0.0003 0.0006 0.0002 0.0001 0.0067
FNR 0.0025 0.0006 0.0010 0.0018 0.0035
Accuracy% 99.68 99.94 99.89 99.85 99.77
436
437 438 439 440 441
Figure 8 Illustration of malicious attack and tamper detection capacity of the present scheme under various attacks. Row-wise, (1) shows crop attack for 10 by 10 window .(2) cropping ¼ area from the top of the image. (3) cutting a small portion of the image. (4) text addition to the image. 16
442 443
Table 14 shows performance comparison of the proposed method with the other
444
image authentication algorithms. In the table, first column shows different watermark
445
embedding technique which are, either fragile or semifragile watermarking method.
446
Subsequently performance evaluation matrices such as PSNR, similarity between extracted
447
and embedded watermark and tamper detection and localization capacity are compared.
448
Proposed method is having average PSNR=42 dB, NHS= 1, very high efficiency of tamper
449
detection /localization. Methods suggested by other authors have good PSNR, but exact
450
extraction of watermarks is not reported, all algorithms are able to locate tamper but detection
451
accuracy is either low or not calculated. None of reported technique suggested any method to
452
classify attacks quantitatively. Overall proposed algorithm is having superiority in terms of
453
imperceptibility, exact extraction of hidden watermark, high efficiency in tamper detection
454
/localizing, and quantitative criteria for attack classification shows very low FPR and FNR.
455 456
Table 14 Comparison of recent image authentication algorithms with proposed method. Author/year Jun Chou (2011)[20] Hong Shen (2012)[2] Radu (2013)[16] Hozem (2014)[14] Chunlei (2015)[15] Ming Li(2016)[10] Xiao (2016)[23] Hamid (2017)[24] D.Singh(2017)[11] Proposed work
457 458 459 460 461 462 463 464 465
Technique
PSNR (dB) VQ technique Max 34 DWT technique 30 DWT based 40 approach DWT 41 quantization Two level DWT 36 VQ technique DWT based approach DWT and Zernike moments DCT based technique Two stage VQ technique
Similarity factor NC=0.98 -
31.3 40
Tamper detection/ localizing Possible Not discussed Possible ,resolution min 4 Possible, Can detect 8×8 region NC=0.8 Possible, localization accuracy medium. SSIM=0.88 Possible NC=0.98 Not discussed
40.9
-
Possible
39.3
NC=0.98
Possible
42
NHS=1.0
Possible, accuracy is very high.
Table 15 shows comparison of different VQ image watermarking methods in terms of codebook sharing between transmitter and receiver, watermarking technique employed ,PSNR, Subjective evaluation of tamper detection localization, and FPR and FNR is presented. The ‘-’ in table shows the corresponding data is not available in paper. It can be seen only in Lu [7] method codebook is not shared which resulted in poor quality of extracted semifragile watermark, average PSNR reported is 29 dB. Robust watermarking techniques suggested by [8, 18, 23] show good robustness characteristics but unable to detect tamper. The ‘*’ sign in table shows watermarks which are named as per their characteristics (whose 17
466 467 468 469 470 471 472 473
type is not mentioned in the respective papers). Whereas fragile watermarking techniques gives good tamper detection capacity and can localize tampers accurately, but does not show robustness to content preserving operations. Thus, proposed method shows better PSNR, having tamper detection accuracy =99.8% with low FNR and FPR. It also suggests robustness to various image processing attacks and exact extraction of hidden watermark. Thus, proposed scheme outperforms recent algorithms. Table15 Performance comparison of VQ based image watermarking algorithms. Author/year Lu (2005)[11]
Codebook
Not shared Wu(2005)[21] Shared Feng(2007)[22] Shared Jau (2010)[12] Shared Jun(2011)[15] Shared Joshi(2013)[4] Shared Qin(2016)[16] Shared Ming(2016)[8] Proposed work Shared
Watermarking Robust+ Fragile Robust * Robust * Robust Fragile* Fragile Fragile Fragile Robust+ Semifragile
PSNR Tamper detection(dB) localization 29.31 √ 31.16 30.45 30.23 34.00 38.21 40.11 31.34 42
FPR -
FNR -
x x x √ accuracy-75% √ 0.02 √ √ √ accuracy- 99.8% 0.0002 0.001
474 475
4. CONCLUSIONS
476
In present work novel vector quantization based watermarking scheme for digital image
477
authentication application has been designed using two stages of watermark embedding
478
technique with an additional feature of qualitative analysis of the received signal. In the first
479
stage robust ‘zero level’ watermark is embedded while another image is used as a semifragile
480
watermark for embedding in the second stage using modified index key based (MIKB)
481
method.
482
In proposed algorithm robust watermark is used as verification watermark which
483
destroys when the key is changed. Semifragile watermark is embedded as
484
authentication watermark which supports most of the content-preserving attacks. This
485
supplies an additional layer of security with acceptable performance regarding
486
imperceptibility and robustness. Although vector quantization is a lossy compression
487
technique, designed system uses codebook length equal to half of the cover image
488
size. Using optimum codebook size affords a fuller quality of watermarked image and
489
hence reduces compression distortion.
490
The novelty of proposed work is in the randomness of embedding key generation.
491
Independent watermark embedding is performed in two phases, and threshold based
492
quantitative analysis is executed for attacks classifications. 18
493
Experimental results demonstrate that proposed algorithm can identify cut, crop, and
494
text addition attacks as malicious attacks. The efficiency of the proposed work in
495
detecting tamper is 99.8%. The average false positive rate is achieved 0.00021, and
496
the average false negative rate is achieved 0.0014.
497
Future work can be carried out to evaluate the proposed work for reversibility of
498
original image.
499 500
REFERENCES
501
[1] Cox IJ, Miller ML. Review of watermarking and the importance of perceptual
502
modelling. Electronic Imaging. 1997: 92-99.
503
[2] Shen H, Chen B. From single watermark to dual watermark: a new approach for
504
image watermarking. Computers & Electrical Engineering. 2012; 38(5):1310-24.
505
[3] Wu HC, Chang CC. A novel digital image watermarking scheme based on the vector
506
quantization technique. Computers & Security . 2005; 24(6):460-71.
507
[4] Nguyen TS, Chang CC, Yang XQ. A reversible image authentication scheme based
508
on fragile watermarking in discrete wavelet transform domain. AEU-International
509
Journal of Electronics and Communications. 2016; 70(8):1055-61.
510
[5] Al-Otum HM, Al-Sowayan SS. Color image watermarking based on self-embedded
511
color permissibility with preserved high image quality and enhanced robustness. AEU-
512
International Journal of Electronics and Communications. 2011; 65 (7):619-29.
513
[6] Lu
514
quantisation. Electronics Letters. 2000; 36(4):303-5.
515
[7] Lu ZM, Xu DG, Sun SH. Multipurpose image watermarking algorithm based on
516
multistage vector quantization. IEEE Transactions on Image Processing. 2005;
517
14(6):822-31.
518
[8] Shen JJ, Ren JM. A robust associative watermarking technique based on vector
519
quantization. Digital Signal Processing. 2010; 20(5):1408-23.
520
[9] Huang HC, Feng-Hsing WA, Jeng-Shyang PA. A VQ-based robust multi-
521
watermarking algorithm. IEICE Transactions on Fundamentals of Electronics,
522
Communications and Computer Sciences. 2002; 85(7):1719-26.
523
[10] Li M, Xiao D, Zhang Y. Attack and Improvement of the Fidelity Preserved Fragile
524
Watermarking of Digital Images. Arabian Journal for Science and Engineering. 2016;
525
41(3):941-50.
ZM, Sun SH. Digital image watermarking technique based on vector
19
526
[11] Singh, D., & Singh, S. K. DCT based efficient fragile watermarking scheme for
527
image authentication and restoration. Multimedia Tools and Applications. 2017; 76(1):
528
953-977.
529
[12] Qin C, Ji P, Wang J, Chang CC. Fragile image watermarking scheme based on VQ
530
index sharing and self-embedding. Multimedia Tools and Applications. 2017; 76(2):22
531
[13] Lin CY, Chang SF. Semifragile watermarking for authenticating JPEG visual
532
content. Electronic Imaging .2000: 140-151.
533
[14] Al-Otum HM. Semi-fragile watermarking for grayscale image authentication and
534
tamper detection based on an adjusted expanded-bit multiscale quantization-based
535
technique. Journal of Visual Communication and Image Representation. 2014;
536
25(5):1064-81.
537
[15] Li C, Zhang A, Liu Z, Liao L, Huang D. Semi-fragile self-recoverable
538
watermarking algorithm based on wavelet group quantization and double authentication.
539
Multimedia tools and applications. 2015; 74(23):10581-604.
540
[16] Preda, R. O. Semi-fragile watermarking for image authentication with sensitive
541
tamper localization in the wavelet domain. Measurement.2013; 46(1): 367-373.
542
[17] Yang CW, Shen JJ. Recover the tampered image based on VQ indexing. Signal
543
Processing. 2010; 90(1):331-43.
544
[18] Makur A, Selvi SS. Variable dimension vector quantization based image
545
watermarking. Signal Processing. 2001; 81(4):889-93.
546
[19] Joshi B, Raval MS, Rege PP, Parulkar SK. Multi-stage VQ based exact
547
authentication for biometric images. Journal of Computing. 2013; 2(1–2):R3-25.
548
[20] Chuang JC, Hu YC. An adaptive image authentication scheme for vector
549
quantization compressed image. Journal of Visual Communication and Image
550
Representation. 2011; 22(5):440-9.
551
[21] Wen Q, Sun TF, Wang SX. Concept and application of zero-watermark. Acta
552
Electronica Sinica. 2003; 31(2):214-6.
553
[22] Linde Y, Buzo A, Gray R. An algorithm for vector quantizer design. IEEE
554
Transactions on communications. 1980; 28(1):84-95.
555
[23] Wang, F. H., Jain, L. C., & Pan, J. S. VQ-based watermarking scheme with genetic
556
codebook partition. Journal of Network and Computer Applications.2007; 30(1): 4-23.
557
[24] Shojanazeri, H., Adnan, W. A. W. Ahmad, S. M. S., & Rahimipour, S.
558
Authentication of images using Zernike moment watermarking. Multimedia Tools and
559
Applications. 2017; 76(1): 577-606. 20
560
[25] http://sipi.usc.edu/database/database.php?volume=misc. [accessed on 11/01/2016].
561
21
S1434-8411(16)31029-9 http://dx.doi.org/10.1016/j.aeue.2017.05.027 AEUE 51898
To appear in:
International Journal of Electronics and Communications
Received Date: Revised Date: Accepted Date:
19 October 2016 14 March 2017 17 May 2017
Please cite this article as: A. Tiwari, M. Sharma, R.K. Tamrakar, Watermarking based image authentication and tamper detection algorithm using vector quantization approach, International Journal of Electronics and Communications (2017), doi: http://dx.doi.org/10.1016/j.aeue.2017.05.027
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1 2 3 4 5 6
Watermarking based image authentication and tamper detection algorithm using vector quantization approach Archana Tiwari1, Manisha Sharma2* and Raunak kumar Tamrakar3 1,2
Department of Electronics & Telecommunication, Bhilai Institute of Technology, Durg 3
7 8
Department of Applied Physics, Bhilai Institute of Technology, Durg *Corresponding Author: [email protected]
9 10
Abstract-
11
quantization (VQ) approach is presented for digital image authentication. Watermarks are
12
embedded in two successive stages for image integrity verification and authentication. In the
13
first stage, a key based approach is used to embed robust zero level watermark using
14
properties of indices of vector quantized image. In the second stage,semifragile watermark is
15
embedded by using modified index key based (MIKB) method. Random keys are used to
16
improve the integrity and security of the designed system. Further, to classify an attack
17
quantitatively as acceptable or as a malicious attack, pixel neighborhood clustering approach
18
is introduced. Proposed approach is evaluated on 250 standard test images using performance
19
measures such as peak signal to noise ratio (PSNR) and normalized hamming similarity
20
(NHS). The experimental results shows that propose approach achieve average false positive
21
rate 0.00024 and the average false negative rate 0.0012. Further, the average PSNR and
22
tamper detection/localization accuracy of watermarked image is 42 dB
23
respectively; while tamper localization sensitivity is very high. The proposed model is found
24
to be robust to common content preserving attacks while fragile to content altering attacks.
25
Index Terms -Attack classification; Biometric verification watermark; Image authentication;
26
Robust watermarking; Semifragile watermarking; Vector quantization.
27
In the present work, a novel image watermarking algorithm using vector
and 99.8%
1. INTRODUCTION
28
With development of the digital technology recreation of digitally generated information has
29
become very easy, and can be transmitted by digital media with ease among other medias of
30
message conveyance images are most common. Therefore protection of these images and its
31
content authentication is important. To authenticate the integrity and authenticity of a digital
32
image, digital watermarking techniques have been considered an effective technique [1-
33
5].The watermarking technique for authentication purpose can be classified as Robust, fragile
34
and semifragile watermarking techniques based on their level of security. A robust watermark
1
35
finds application in the area of ownership verification and copyright protection [6-9]. A
36
fragile watermark destroys itself even for small accidental or intentional changes [10-12].On
37
the other hand, semi-fragile watermarking is a practical technique which can protect integrity
38
of digital images and can locate the tampered areas [13-16]. Thus fragile or semi-fragile
39
watermarking are those image watermarking techniques which find application in images and
40
video content authentication and tamper detection [7, 13, 16, 17]. Given the limitations of
41
robust and fragile watermarking schemes, multiple-watermarking is applied to digital data
42
where copyright protection and content authentication is required to be addressed
43
simultaneously [7]. Lu [7], pioneered the concept of VQ based image watermarking
44
technique. Various VQ (vector quantization) based watermarking techniques are
45
recommended by researchers in the past decade for image watermarking [17-20] with the
46
objective of image compression and authentication.
47
Most of the methods suggested in above literatures need improvement in issues
48
related to watermarked image quality, identification of attacks as malicious or content
49
preserving manipulation, and efficiency of the system in authentication of images. Therefore,
50
an efficient image authentication algorithm having a good visual quality of watermarked
51
image, having the advantage of comparatively less bandwidth and storage space, with high
52
accuracy in attacks classification, tolerant to acceptable manipulations, and high tamper
53
detection accuracy needs to be designed. Thus in proposed work, watermarks are embedded
54
in two successive stages. In the first phase robust watermark is used for ‘zero level
55
watermarking’ [21] utilizing the statistical properties of VQ indices for enhancing the
56
security of watermarked image by biometric verification. The advantage of first stage of
57
watermarking is, it does not affect VQ reconstructed image. In the second step, the
58
semifragile watermark is embedded by using a Modified Index Key based method (MIKB)
59
for image authentication. A ‘quantitative threshold-based approach’ is suggested for
60
classifying attacks and localizing tampered area. Further statistical analysis is done to find
61
false positive rate, false negative rate, % of correctly identified authenticated images and %
62
of falsely verified images to validate the correctness of the system. Thus, present image
63
authentication fulfils desired characteristics of an effective image authentication technique.
64
The significant features of proposed improvised scheme are-
65 66
Watermarks are embedded using VQ indices properties; it enhances security and saves bandwidth requirements.
2
67
Random keys are used for embedding robust zero level watermark and semi-fragile
68
watermark respectively; random nature of key improves the integrity and security of
69
designed system.
70
71 72
preserving attack.
73 74
A quantitative analysis is suggested to classify the attack as malicious or content
Tamper detection accuracy of proposed system is very high; it can detect tampering upto 10% as malicious attack.
Biometric verification watermark is used to enhance reliability of designed system.
75
The structure of the paper is organized into three sections namely proposed scheme,
76
Experimental results and discussion of deigned algorithm, and lastly conclusion of the paper
77
is addressed.
78
2. PROPOSED SCHEME
79
In this section, the detail of proposed image authentication algorithm is presented. The
80
proposed method uses two-stage VQ, to divide the encoding task into successive stages;
81
wherein initial stage performs a relatively rough quantization of the input vector using a
82
codebook. Then, a second stage quantizer performs on the error vector between the original
83
and quantized first-stage output. The quantized error vector then gives a second
84
approximation to the original input vector thereby provides a more accurate or refined
85
depiction of the input. The binary watermark images are embedded in two successive stages
86
using random key based approach and properties of indices, the output of both stages are
87
combined to get watermarked image. Watermarked image is sent over, an open
88
communication channel, where accidental or malicious changes may affect contents of
89
watermarked image. In the verification and authentication step, the watermark sequence is
90
extracted independently from received watermarked image, using passkey1 and passkey2 in
91
stage 1 and passkey3 in stage 2. The authenticity of the received image is analyzed using the
92
extracted watermark images. Details of tamper detection, localization, and classification of
93
attacks are described in subsequent sections. Figure 1 shows two stage vector quantization
94
procedure followed in proposed method.
95
The proposed technique is described in four subsections; subsection 2.1 describes
96
codebook generation , 2.2 explains proposed watermark embedding technique, subsection 2.3
3
97
discusses watermark extraction in two consecutive stages. Finally, authentication and
98
verification of received image are discussed in subsection 2.4.
99 100
2.1 Codebook generation
101
To create a codebook of M codewords, input training images are selected. Each input image
102
is partitioned into non-overlapping image blocks of m×m pixels. Thereafter, M image blocks
103
from the input image blocks are selected as the initial codewords. This set of initial
104
codewords can be termed as the initial codebook. After generation of the initial codebook, an
105
iterative process for several rounds is executed to generate the codebook using LBG
106
algorithm [22]. In each round, the data clustering procedure and the centroid updating
107
procedure are performed. In the data clustering procedure, the nearest codeword of each input
108
image block in the codebook is to be found. To determine the nearest codeword of each input
109
image block in the codebook, a total number of M squared Euclidean distances are computed.
110
Then, the closest codeword is determined by finding the minimal squared Euclidean distance
111
from itself to the input image block. After the data clustering process is executed, all these
112
input image blocks are classified into M groups. In the centroid updating procedure, the mean
113 114
Figure1: Proposed two stage vector quantization procedure.
115 116
vector of the training image blocks in each group is calculated. These M computed image
117
vectors generate new codebook. By repeatedly executing the same process in each round, a
118
representative set of codewords can be generated for first stage. In second stage error image
119
is used as input training image for generation of codebook2 by repeating same process. In
120
proposed work a new codebook is generated each time algorithm runs, for same input image,
121
to enhance security of algorithm. Randomness is achieved by randomly generating a set of
4
122
input training vectors for codebook generation in the range of minimum to maximum pixel
123
values.
124
2.2 Proposed watermark embedding technique
125
Embedding of the watermark is done in two stages: robust zero level watermark embedding
126
and semifragile watermark embedding. The detail of watermark embedding process is
127
described in two subsections namely 2.2.1 discusses watermark embedding in the first stage,
128
and 2.2.2 explains the second stage of watermark embedding and final watermarked image
129
construction.
130
2.2.1. First stage of watermark embedding: In the first stage, a robust binary fingerprint
131
watermark is embedded. Robust zero level watermark embedding is done in two steps:
132
Step 1: Robust zero level watermark embedding: After generation of codebook1 (C1), the
133
robust binary watermark
134
passkey1, where
135
original image S with size is m
136
(a, b) denotes the image block at the position of (a, b). After that, each vector s (a, b) finds its
137
best codeword in the codebook C1 and a index ‘g’ is assigned to s (a, b), we can then obtain
138
the indices matrix T with elements t (a, b). , which can be represented by:
139
T=VQ(S)=
140
For most images, the VQ indices in the neighboring blocks may be same; this characteristic is
141
used to calculate parameter ‘E,' called embedding function.
142
, with pixel of size (p × q) is scrambled with randomly generated
passkey1
((p × q), resultant scrambled robust watermark is then divided into vectors s (a, b) with size 4
=
, where s
(1)
of adjacent blocks.
143 144
. The
(2) The embedding function can be computed as follows:
145
E=
146
Where
(3) ;
147
Where ‘Th’ is a threshold, i.e. standard deviation of one-dimensional indices matrix T. Now
148
embedded robust watermark or passkey2 can be generated by EX-ORing of the scrambled
149
watermark with the embedding function. The watermark is termed as ‘robust zero level
150
watermark,' as it is robust to almost all image manipulations and it does not change VQ
151
compressed image at all.
152
passkey2=
(4)
5
153
Step2: First stage VQ image reconstruction: After robust watermark embedding, the
154
reconstructed image S’ (m, n) and error image Q (m, n) can be obtained as follows. In first
155
stage compressed image is watermarked without actually inserting it explicitly .This concept
156
of Zero-watermark algorithm was suggested by Wen et al. [21], which does not modify the
157
original image. And the most important step is it is constructed from characteristics of the
158
original image. VQ-1
159
(5)
160
Q(m,n)=S (m, n)-S’ (m, n)
161
Equation 5 shows reconstructed image after the first stage. While equation 6 represents the
162
output of the first stage, this is treated as an input to the subsequent stage.
163
2.2.2 Second stage of watermark embedding: In the second stage, a semifragile
164
authentication watermark is embedded. For a codebook2 generation, the error image, Q (m,n)
165
serves as input training sequence to improved LBG algorithm. Further, like the original
166
image, Q (m, n), is also divided into ‘L’ vectors ‘q (a, b),' each of size 4×4, where q (a, b)
167
denotes the image block at the position of (a, b) .After codebook (C2) generation, each vector
168
of input training vector finds its most suitable code vector C2x, in codebook2 (C2) and
169
assigns indices
170
elements x (a, b) which can be represented as follows.
171
X=VQ (Q) =
172
Now each element of indices matrix, X is modified according to binary semifragile
173
watermark Ws .For semifragile watermark embedding; a single bit of watermark Ws is
174
inserted into each element of X. Let K be the size of codebook2, where K=2v then each
175
index has got ‘v’ binary bits. The embedding position of each semifragile watermark bit is
176
decided using randomly generated passkey3, where
177
significant bit) end watermark bit Ws replaces binary bit of each index of matrix X at the
178
position defined by passkey3. For example let, X11=22, its binary equivalent is
179
(000000000010110)2, assuming passkey3=3 and Ws (11) =0, then after watermark
180
embedding, X11 is modified to (000000000010010)2 i.e. 18. Using these modified indices
181
semi fragile watermarked image is reconstructed at VQ decoder, thus method is termed as
182
modified indices key based method (MIKB). Let reconstructed image be
183
stage output, which is represented as follows:
184
=
(6)
to q(a, b). Now using these indices, indices matrix X is created, with
=
(7)
passkey3
. From LSB (least
i.e. second
(8)
6
185
Each time algorithm runs a unique random generated key is used for semifragile watermark
186
embedding. Finally, watermarked image is obtained by combining first stage and second
187
stage outputs.
188
(9)
189
Thus, in the proposed scheme the watermarks are embedded independently in successive
190
stages for verifying integrity and authenticity of the image.
191
2.3 Watermark extraction in two stages
192
This subsection explains watermark extraction procedure of proposed algorithm. It is
193
understood here that both sender and receiver share same codebooks. The procedure adopted
194
for watermark extraction is given below.
195
Step1- Robust watermark extraction in the first stage: At the receiver side, the reverse
196
process of the watermark embedding process is carried out. The received image
197
similar way while encoding, is divided into non-overlapping blocks of size 4×4. The VQ
198
encoder as discussed before is used along with codebook1 to find encoded indices of first
199
stage. Obtained VQ indices are used to compute embedding function ‘E’. Then XOR
200
operation is performed between the embedding function ‘E’ and passkey2 to get back
201
permuted robust watermark
202
passkey1 to get the extracted robust watermark
203
reconstruct the first stage approximate output
204
Step 2-Semifragile watermark extraction: In the second stage of extraction, watermarked
205
image
206 207 208 209 210
. Further, these indices are used to
Codebook 1
Watermarked Image
Partitioned into 16384 blocks each of size 4×4
Indices
First stage VQ encoder
Passkey1 Inverse permutation
212
Extracted robust watermark
213
217 218 219 220 221
. Finally inverse permutation process is carried out with
is segmented using the previous phase output
211
214 215 216
, in
Permuted robust watermark
Embedding function (E) computation
E
Codebook 2 Second stage VQ encoder
Watermarked image segmentation
First stage o/p reconstruction using VQ decoder
Indices Passkey3
Extract bit positions
Restore extracted bits to reconstruct semifragile watermark
Extracted semifragile watermark
Figure 2: Illustration of steps involved in the extraction of the watermark in two subsequent stages. 7
222
To obtain the second stage approximated image Q’’ (m, n) which is further divided into L
223
vectors, each of size 4×4.Standard VQ encoder executes the nearest neighbourhood search to
224
obtain encoded indices (
225
are recovered from indices
226
embedding order to form extracted semi fragile watermark
227
extraction is given in figure 2.
228
2.4 Verification and authentication of received image
229
The verification and authentication procedure is divided into five steps. After extraction of
230
the robust and the semifragile watermarks, the received image is to be verified and
231
authenticated. In next step, details of various stages involved including tamper detection
232
localization procedure and performance evaluation parameters are explained.
233
Step 1: Verification of received image: In the first step the integrity of received
234
image
235
the embedded fingerprint robust watermark, using parameter NHS. NHS [7] is defined in
236
equation 10, where
237
watermark.
238
NHS
). Then using passkey3 semifragile authentication watermark bits . These extracted watermarking bits are then grouped as per . Detail process of watermark
is to be verified, the extracted fingerprint robust watermark is compared with
represents the embedded watermark and
denotes the extracted
(10)
239
Here HD (.,.) stands for the Hamming distance between two binary images, i.e., the number
240
of bits different in two binary images. If the value of NHS is equal to unity, that means
241
embedded watermark and extracted watermark are identical. Thus, the integrity of source is
242
verified, otherwise not.
243 244
(a)
(b)
245
Figure 3 (a) Depicts an error pixel surrounded by five pixels which are error pixel (b) Depicts
246
an error pixel surrounded by two error pixels.
247 248
Step 2: Authentication of received image: In next phase, authentication of received image
249
is performed, the similarity between the embedded and the extracted semifragile watermark is
250
checked. If the value of
251
received image
then there is a possibility of some security attack on . As semifragile watermark is used as authentication watermark in 8
252
proposed algorithm, all authentication analysis is performed using the semifragile watermark.
253
Further difference image is obtained as
254
.If any pixel of
is one it is considered as an error pixel; otherwise, it is not an error pixel. A pixel
255
neighbourhood criterion is used for content authentication of image, as explained in step 3.
256
Step3: Tamper detection, localization, and attack classification: If any error pixel is
257
identified in
258
neighbourhood approach. A pixel is considered as malicious pixel if more than four pixels in
259
its surrounding eight pixels are having value one otherwise it, is classified as a nonmalicious
260
pixel. Figure 3 (a) shows malicious pixel, whereas Figure 3 (b) shows nonmalicious pixel.
261
The threshold (TR) value is obtained after testing of 250 images for different nonmalicious
262
and malicious attacks. It was found malicious attacks contains more than 57 percentage
263
malicious error pixel, so the threshold is set to 57 percentages accordingly. Table 1, shows
264
notations used in the algorithm for authenticating images.
265
Table 1 Notations used in algorithm for authenticating image Notations R(m,n) N N1 N2 T1 T2
266 267 268
, it is checked for malicious or nonmalicious attack by using 3×3 pixel
Descriptions Difference image of size m×n. ). Total no of error pixels in No of pixels having four or more than four surrounding error pixel. , percentage of error pixels. , percentage of malicious pixels.
The algorithmic structure of proposed system for attacks classification is given as:
275
Compute N1 from image . If T1=0, then image is not attacked. The image is considered as authentic. o End. If, T1≠0, then compute the value of NHS. If NHS> 0.7, then extracted image is recognizable. Further, check for authenticity of the biometric image. o If T2
276
White pixel areas on the difference image broadly identify the tampered location, while
277
tampering is localized by finding coordinates of rows and columns of all tampered pixels
278
locations. Table 2 explain various evaluation parameters used for tamper detection and
279
classification of attack and Table 3 represents pixels classification statistics.
269 270 271 272 273 274
9
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296
Figure 4 Verification and authentication process Table 2 Performance parameters used to test the effectiveness of proposed scheme. Parameter M
Description Total pixels in received image . Total tampered pixels in received image . Denote no of tampered pixels which are accurately identified as tampered. Denote no of pixels which are not tampered but falsely detected as tampered pixels. Denote no of pixels which are tampered but not identified as tampered pixels.
True positive: TP False negative: FP True negative :(TN)
-
False positive: (FN) False positive rate: False negative rate: % Accuracy Mean square error (MSE)
% Accuracy= (TP+TN) / (TP+TN+FP+FN)
PSNR 10
297
Table no 3 Pixel classification based on statistical results Pixel classification No. of Tampered pixels No of non-tampered pixels
Test positive
Test negative
298 299
3. EXPERIMENTAL RESULT AND DISCUSSION
300
The performance of the proposed watermarking based technique for digital image verification
301
and authentication using vector quantization approach is investigated. The experiment is
302
performed using MATLAB version 7.10.0.499 (R2010a), 64-bit (win64) software. In the first
303
phase, the experiment is conducted on 300 different 8-bit images of resolution 600 dpi each
304
of size 512 ×512, the binary watermark image of size 128×128 as a robust watermark image
305
and semifragile watermark image. Figure-5 shows five test images whose experimental
306
results are given in the subsequent subsection. The work is carried out employing various
307
combinations of codebook1 and codebook2 under the condition codebook size is less than
308
512.The visual quality of the watermarked images is evaluated by PSNR ,which comes to be
309
42 dB (average). It is observed that maximum PSNR is achieved for all images for codebook
310
size 256, for both codebooks.
311
(2)
(1)
(3)
(4)
(5)
312 313
Figure 5: Different input test images used in the proposed method (1) Lena (2) Baboon(3)
314
Pepper(4)Cameraman(5)F-16.
315
(Image source: http://sipi.usc.edu/database/database.php?volume=misc.)[25].
316
0
317
In Table 4, performance of the proposed method is evaluated for ‘Lena’ image against
318
different size of watermark has been evaluated, when no attack is applied. For embedding
319
watermarks the original image is partitioned into a number of blocks according to the size of
320
watermark. Moreover, codebook sizes are half of original image size. However, overall
321
performance of proposed method highly depends on size of semifragile watermark, as
322
watermarked image is not affected by robust watermark.
323 11
324 325
Table 4 Performance of proposed method for different size of Lena image against varying watermarks size. Original image
Size of watermarks Robust Semifragile
Biometric robust Binary semifragile PSNR(dB) watermark watermark
256 256
64 64
64 64
40.11
256 256
32 32
32 32
41.98
512 512
64 64
128 128
42.81
128 128
32 32
64 64
39.55
512 512
256 256
64 64
43.94
326 327
3.1 Verification of Received Image using Robust Watermark
328 329
Figure 6.Robust watermark as verification watermark
330 331
Once watermarked image is received, extracted watermark at first stage is checked, for
332
verification of integrity of image. In present method fingerprint verification method is used
333
for finding integrity of image. Referring to figure 4, under no attack condition NHS is one
334
whereas, when any of passkey1 or passkey 2 changes value NHS≠1, that is considered as
335
threat to security of system. In figure 6, row1 shows results when same key is used both at
336
transmitter and receiver end. Whereas, row2 demonstrate results when passkey1 is changed
337
and passkey2 is not changed.
338
3.2. Robustness to Common Image Processing Manipulations
339
The performance of the method is examined by applying six types of image processing
340
attacks such as Median filtering, blurring, low-pass filtering using a Gaussian filter, salt and 12
341
pepper noise, rotation at various angles, and JPEG compression on 250 watermarked images.
342
Table 5 to 10, demonstrate various results obtained for robustness by processing the four test
343
images by applying different content preserving attacks as enumerated above.
344
terminologies used in the Table are: Tw show number of error pixels obtained from extracted
345
watermark image, T1 is a percentage of error pixels in the difference image, T2 is a
346
percentage of malicious pixels in the difference image. Here blur attack is applied for
347
different radius from 0.0 to 0.9; median filtering is applied to various window sizes from 1×1
348
onwards, Gaussian filtering for different variances, salt, and pepper noise for various
349
percentages from one percent onwards, rotation in both clockwise and anti-clockwise
350
directions and JPEG compression for varying quality factor. The degradation in quality is
351
evaluated using PSNR for received watermarked image and NHS for extracted semifragile
352
watermark images.
353
The
Table 5 Experimental results illustrating blurring the image at different radius. Image Pepper Baboon Lena F-16
Tw 3043 4757 3225 2343
%T 52.4 79.92 87.73 56.33
% T2 52.3 50.06 55.12 41.24
PSNR(dB) 38.49 35.56 38.86 37.28
NHS(semi) 0.8112 0.7067 0.8082 0.8570
Blur radius Radius=1.0 Radius=1.1 Radius=1.1 Radius=1.0
Attack Incidental Incidental Incidental Incidental
354 355
Table 6 Experimental results for median filtering at different window sizes. Image Pepper Baboon Lena F-16
356 357 358
%T 46.98 72.14 70.89 60.55
% T2 23.38 56 29.7 20.55
PSNR(dB) 43.04 36.32 41.72 39.16
NHS(semi) 0.8048 0.8304 0.7150 0.7472
Window size 3×3 filter 2×2 filter 3×3 filter 3×3 filter
Attack Incidental Incidental Incidental Incidental
Table 7 Robust performance for a Gaussian filter for 3×3 window using four different values of variance at mean=0. Image Pepper Baboon Lena F-16
359 360
Tw 3798 3491 4665 4142
Tw 3026 4763 815 749
%T 71.67 80.13 75.48 67.06
% T2 54.1 45.81 51.8 37.37
PSNR 44.31 36.49 40.46 37.68
NHS(semi) 0.8163 0.9136 0.9503 0.9543
Standard deviation Sigma=0.8 Sigma=0.6 Sigma=0.6 Sigma=0.6
Attack Incidental Incidental Incidental Incidental
Table 8 Experimental results for salt and pepper noise attack at various noise densities. Image Pepper Baboon Lena F-16
Tw 4986 4712 4924 4881
%T 6.97 6.94 6.04 6.97
T2 0 0 0 0.00005
PSNR(dB) 41.92 42.52 48.76 40.18
NHS(semi) 0.7136 0.7075 0.7679 0.7021
Noise density 0.07 0.07 0.06 0.07
Attack Incidental Incidental Incidental Incidental
361 13
362
Table 9 Robustness performance at different rotation angles Image Pepper Baboon Lena F-16
363 364
Tw 3242 2387 4498 4641
%T 72.44 72.04 85.90 74.39
% T2 50.5 27.00 54.56 54.3
PSNR 40.41 37.24 40.32 40.95
NHS(semi) 0.8021 0.8543 0.7255 0.7167
Rotation angle 0.5 0.4 0.5 0.5
Attack Incidental Incidental Incidental Incidental
Table 10 Robustness performance of JPEG compression at different quality factor Quality factor Tw %T % T2 PSNR (dB) NHS(semi) Attack 70 2819 87.0 57 40.9 0.85 Incidental 80 1454 88.0 56 43.0 0.91 Incidental 90 171 83.9 51 42.0 0.98 Incidental 99 00 23.0 00 43.0 1.00 Incidental
365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392
Figure 7. Illustration of Robustness nature of the scheme: row-wise (1) Gaussian filtering for window size 3×3 and a Sigma 0.6.(2) Median Filtering for a 2×2 window. (3) Salt and pepper noise addition to noise density 0.07. (4) Blurring with radius 1. (5) Rotation by 0.5 clockwise angles.(6) JPEG compression for quality factor 70. The performance of attacks on the image is evaluated based on the computed
393
percentage of malicious pixels. If the percentage is less than threshold i.e. 57% and the NHS
394
is higher than 0.7 the attack is classified as non-malicious otherwise, it is a malicious attack.
395
Experimental results show that each attack is robust to a certain value of parameters as
396
depicted in Tables 5-10. On comparing the results of the Tables 5-10, it is seen there is a
14
397
small variation in robustness feature of images. For robustness feature, taking minimum
398
values of parameters for different robustness attacks.It is observed that the proposed method
399
is robust for blurring attack for one radius, median filtering up to 2×2 window size, Gaussian
400
filtering for variance 0.6 for the 3×3 filter, for rotation to 0.45 angles, salt and pepper noise
401
density 0.06 and JPEG compression quality factor 70.These results show that there is a slight
402
difference in the level of robustness for different images.
403
Figure 7 illustrates the robust nature of the algorithm; here some results are presented
404
for visualization of experimental results. The algorithm suggests improvement in work
405
proposed by Lu [7]. Lu suggested attacks classification based on NHS value, but the
406
algorithm [7] is unable to classify whether attacks is incidental or malicious. Robustness
407
performance of proposed method is compared with [11, 12, 14, 15, 16]; it is seen none of
408
these algorithms report any method for classifying attacks.
409
3.3. Fragileness characteristics for malicious attacks
410
Further, this algorithm is tested for some content altering attacks such as cutting a portion of
411
watermarked image and cropping the image to different window sizes. Here results of four
412
test images are shown for the cut attack, in Table 11. Table 12, shows crop attack for varying
413
window sizes(window sizes shows dimension of cropped area), the Table shows average
414
results obtained by simulating 250 images. Table 11 and 12 demonstrate that the suggested
415
scheme can identify these attacks as malicious because the percentage of the malicious pixel
416
is more than the threshold in all cases. It is found from experimental results that the proposed
417
method can distinguish a single pixel change in received image and identify tampering
418
greater than 10% as a malicious attempt.
419
Table 11 Simulation results of watermarked test images for Cut attack Image Pepper Baboon Lena F-16
420 421
Tw 29 982 871 128
%T 2.67 2.82 2.77 2.68
% T2 82.33 91.72 89.01 89.46
PSNR(dB) 41.27 39.17 39.51 37.18
NHS(semi) 0.9982 0.9401 0.9468 0.9948
Attack classification Malicious Malicious Malicious Malicious
Table 12 Simulation results of for crop attack results show average values of all images Window Tw %T % T2 PSNR(dB) NHS(semi) Attack classification 8×8 54 0.031 60.49 37.70 0.9967 Malicious 40×40 360 0.640 90.48 38.96 0.9800 Malicious 130×130 1878 6.541 96.96 34.67 0.8853 Malicious 200×200 3149 15.410 98.00 31.48 0.8075 Malicious 300×300 4639 34.600 98.67 28.41 0.7076 Malicious
422
15
423
The sensitivity of given algorithm is clearly shown in Table 12; it can detect and
424
locate cropping as the malicious attack for the 8×8 window. However, the given algorithm
425
can detect a single pixel change in received image. Various performance measures are
426
evaluated for finding tamper detection capacity of algorithm and results are shown in Table-
427
13. The performance evaluation parameters are - TP, TN, FP, FN, FPR, FNR and accuracy of
428
tampered images for different window sizes (tampered area). Values in the Table show
429
average results obtained for various test images. Referring to the same table, results clearly
430
demonstrate that the algorithm shows high accuracy in tamper detection, and is having very
431
low false negative and positive rates. Figure 8, presents a visual presentation of experimental
432
results obtained for various fragile attacks.
433 434 435
Table 13 Statistical analysis of tamper detection capacity of images results show average values Window 40×40 80×80 130×130 200×200 300×300
TP 0.997 0.999 0.999 0.999 0.999
FP 0.00031 0.00063 0.00106 0.00180 0.00350
TN 0.9997 0.9990 0.9989 0.9981 0.9965
FN 0.00250 0.00063 0.00024 0.00010 0.00670
FPR 0.0003 0.0006 0.0002 0.0001 0.0067
FNR 0.0025 0.0006 0.0010 0.0018 0.0035
Accuracy% 99.68 99.94 99.89 99.85 99.77
436
437 438 439 440 441
Figure 8 Illustration of malicious attack and tamper detection capacity of the present scheme under various attacks. Row-wise, (1) shows crop attack for 10 by 10 window .(2) cropping ¼ area from the top of the image. (3) cutting a small portion of the image. (4) text addition to the image. 16
442 443
Table 14 shows performance comparison of the proposed method with the other
444
image authentication algorithms. In the table, first column shows different watermark
445
embedding technique which are, either fragile or semifragile watermarking method.
446
Subsequently performance evaluation matrices such as PSNR, similarity between extracted
447
and embedded watermark and tamper detection and localization capacity are compared.
448
Proposed method is having average PSNR=42 dB, NHS= 1, very high efficiency of tamper
449
detection /localization. Methods suggested by other authors have good PSNR, but exact
450
extraction of watermarks is not reported, all algorithms are able to locate tamper but detection
451
accuracy is either low or not calculated. None of reported technique suggested any method to
452
classify attacks quantitatively. Overall proposed algorithm is having superiority in terms of
453
imperceptibility, exact extraction of hidden watermark, high efficiency in tamper detection
454
/localizing, and quantitative criteria for attack classification shows very low FPR and FNR.
455 456
Table 14 Comparison of recent image authentication algorithms with proposed method. Author/year Jun Chou (2011)[20] Hong Shen (2012)[2] Radu (2013)[16] Hozem (2014)[14] Chunlei (2015)[15] Ming Li(2016)[10] Xiao (2016)[23] Hamid (2017)[24] D.Singh(2017)[11] Proposed work
457 458 459 460 461 462 463 464 465
Technique
PSNR (dB) VQ technique Max 34 DWT technique 30 DWT based 40 approach DWT 41 quantization Two level DWT 36 VQ technique DWT based approach DWT and Zernike moments DCT based technique Two stage VQ technique
Similarity factor NC=0.98 -
31.3 40
Tamper detection/ localizing Possible Not discussed Possible ,resolution min 4 Possible, Can detect 8×8 region NC=0.8 Possible, localization accuracy medium. SSIM=0.88 Possible NC=0.98 Not discussed
40.9
-
Possible
39.3
NC=0.98
Possible
42
NHS=1.0
Possible, accuracy is very high.
Table 15 shows comparison of different VQ image watermarking methods in terms of codebook sharing between transmitter and receiver, watermarking technique employed ,PSNR, Subjective evaluation of tamper detection localization, and FPR and FNR is presented. The ‘-’ in table shows the corresponding data is not available in paper. It can be seen only in Lu [7] method codebook is not shared which resulted in poor quality of extracted semifragile watermark, average PSNR reported is 29 dB. Robust watermarking techniques suggested by [8, 18, 23] show good robustness characteristics but unable to detect tamper. The ‘*’ sign in table shows watermarks which are named as per their characteristics (whose 17
466 467 468 469 470 471 472 473
type is not mentioned in the respective papers). Whereas fragile watermarking techniques gives good tamper detection capacity and can localize tampers accurately, but does not show robustness to content preserving operations. Thus, proposed method shows better PSNR, having tamper detection accuracy =99.8% with low FNR and FPR. It also suggests robustness to various image processing attacks and exact extraction of hidden watermark. Thus, proposed scheme outperforms recent algorithms. Table15 Performance comparison of VQ based image watermarking algorithms. Author/year Lu (2005)[11]
Codebook
Not shared Wu(2005)[21] Shared Feng(2007)[22] Shared Jau (2010)[12] Shared Jun(2011)[15] Shared Joshi(2013)[4] Shared Qin(2016)[16] Shared Ming(2016)[8] Proposed work Shared
Watermarking Robust+ Fragile Robust * Robust * Robust Fragile* Fragile Fragile Fragile Robust+ Semifragile
PSNR Tamper detection(dB) localization 29.31 √ 31.16 30.45 30.23 34.00 38.21 40.11 31.34 42
FPR -
FNR -
x x x √ accuracy-75% √ 0.02 √ √ √ accuracy- 99.8% 0.0002 0.001
474 475
4. CONCLUSIONS
476
In present work novel vector quantization based watermarking scheme for digital image
477
authentication application has been designed using two stages of watermark embedding
478
technique with an additional feature of qualitative analysis of the received signal. In the first
479
stage robust ‘zero level’ watermark is embedded while another image is used as a semifragile
480
watermark for embedding in the second stage using modified index key based (MIKB)
481
method.
482
In proposed algorithm robust watermark is used as verification watermark which
483
destroys when the key is changed. Semifragile watermark is embedded as
484
authentication watermark which supports most of the content-preserving attacks. This
485
supplies an additional layer of security with acceptable performance regarding
486
imperceptibility and robustness. Although vector quantization is a lossy compression
487
technique, designed system uses codebook length equal to half of the cover image
488
size. Using optimum codebook size affords a fuller quality of watermarked image and
489
hence reduces compression distortion.
490
The novelty of proposed work is in the randomness of embedding key generation.
491
Independent watermark embedding is performed in two phases, and threshold based
492
quantitative analysis is executed for attacks classifications. 18
493
Experimental results demonstrate that proposed algorithm can identify cut, crop, and
494
text addition attacks as malicious attacks. The efficiency of the proposed work in
495
detecting tamper is 99.8%. The average false positive rate is achieved 0.00021, and
496
the average false negative rate is achieved 0.0014.
497
Future work can be carried out to evaluate the proposed work for reversibility of
498
original image.
499 500
REFERENCES
501
[1] Cox IJ, Miller ML. Review of watermarking and the importance of perceptual
502
modelling. Electronic Imaging. 1997: 92-99.
503
[2] Shen H, Chen B. From single watermark to dual watermark: a new approach for
504
image watermarking. Computers & Electrical Engineering. 2012; 38(5):1310-24.
505
[3] Wu HC, Chang CC. A novel digital image watermarking scheme based on the vector
506
quantization technique. Computers & Security . 2005; 24(6):460-71.
507
[4] Nguyen TS, Chang CC, Yang XQ. A reversible image authentication scheme based
508
on fragile watermarking in discrete wavelet transform domain. AEU-International
509
Journal of Electronics and Communications. 2016; 70(8):1055-61.
510
[5] Al-Otum HM, Al-Sowayan SS. Color image watermarking based on self-embedded
511
color permissibility with preserved high image quality and enhanced robustness. AEU-
512
International Journal of Electronics and Communications. 2011; 65 (7):619-29.
513
[6] Lu
514
quantisation. Electronics Letters. 2000; 36(4):303-5.
515
[7] Lu ZM, Xu DG, Sun SH. Multipurpose image watermarking algorithm based on
516
multistage vector quantization. IEEE Transactions on Image Processing. 2005;
517
14(6):822-31.
518
[8] Shen JJ, Ren JM. A robust associative watermarking technique based on vector
519
quantization. Digital Signal Processing. 2010; 20(5):1408-23.
520
[9] Huang HC, Feng-Hsing WA, Jeng-Shyang PA. A VQ-based robust multi-
521
watermarking algorithm. IEICE Transactions on Fundamentals of Electronics,
522
Communications and Computer Sciences. 2002; 85(7):1719-26.
523
[10] Li M, Xiao D, Zhang Y. Attack and Improvement of the Fidelity Preserved Fragile
524
Watermarking of Digital Images. Arabian Journal for Science and Engineering. 2016;
525
41(3):941-50.
ZM, Sun SH. Digital image watermarking technique based on vector
19
526
[11] Singh, D., & Singh, S. K. DCT based efficient fragile watermarking scheme for
527
image authentication and restoration. Multimedia Tools and Applications. 2017; 76(1):
528
953-977.
529
[12] Qin C, Ji P, Wang J, Chang CC. Fragile image watermarking scheme based on VQ
530
index sharing and self-embedding. Multimedia Tools and Applications. 2017; 76(2):22
531
[13] Lin CY, Chang SF. Semifragile watermarking for authenticating JPEG visual
532
content. Electronic Imaging .2000: 140-151.
533
[14] Al-Otum HM. Semi-fragile watermarking for grayscale image authentication and
534
tamper detection based on an adjusted expanded-bit multiscale quantization-based
535
technique. Journal of Visual Communication and Image Representation. 2014;
536
25(5):1064-81.
537
[15] Li C, Zhang A, Liu Z, Liao L, Huang D. Semi-fragile self-recoverable
538
watermarking algorithm based on wavelet group quantization and double authentication.
539
Multimedia tools and applications. 2015; 74(23):10581-604.
540
[16] Preda, R. O. Semi-fragile watermarking for image authentication with sensitive
541
tamper localization in the wavelet domain. Measurement.2013; 46(1): 367-373.
542
[17] Yang CW, Shen JJ. Recover the tampered image based on VQ indexing. Signal
543
Processing. 2010; 90(1):331-43.
544
[18] Makur A, Selvi SS. Variable dimension vector quantization based image
545
watermarking. Signal Processing. 2001; 81(4):889-93.
546
[19] Joshi B, Raval MS, Rege PP, Parulkar SK. Multi-stage VQ based exact
547
authentication for biometric images. Journal of Computing. 2013; 2(1–2):R3-25.
548
[20] Chuang JC, Hu YC. An adaptive image authentication scheme for vector
549
quantization compressed image. Journal of Visual Communication and Image
550
Representation. 2011; 22(5):440-9.
551
[21] Wen Q, Sun TF, Wang SX. Concept and application of zero-watermark. Acta
552
Electronica Sinica. 2003; 31(2):214-6.
553
[22] Linde Y, Buzo A, Gray R. An algorithm for vector quantizer design. IEEE
554
Transactions on communications. 1980; 28(1):84-95.
555
[23] Wang, F. H., Jain, L. C., & Pan, J. S. VQ-based watermarking scheme with genetic
556
codebook partition. Journal of Network and Computer Applications.2007; 30(1): 4-23.
557
[24] Shojanazeri, H., Adnan, W. A. W. Ahmad, S. M. S., & Rahimipour, S.
558
Authentication of images using Zernike moment watermarking. Multimedia Tools and
559
Applications. 2017; 76(1): 577-606. 20
560
[25] http://sipi.usc.edu/database/database.php?volume=misc. [accessed on 11/01/2016].
561
21