=Paper=
{{Paper
|id=Vol-2391/paper51
|storemode=property
|title=Evaluation of different embedding methods for JPEG authentication watermarking
|pdfUrl=https://ceur-ws.org/Vol-2391/paper51.pdf
|volume=Vol-2391
|authors=Anna Egorova,Victor Fedoseev
}}
==Evaluation of different embedding methods for JPEG authentication watermarking ==
https://ceur-ws.org/Vol-2391/paper51.pdf
Evaluation of different embedding methods for JPEG
authentication watermarking
A A Egorova1, V A Fedoseev1,2
1
Samara National Research University, Moskovskoe Shosse 34А, Samara, Russia, 443086
2
Image Processing Systems Institute of RAS - Branch of the FSRC "Crystallography and
Photonics" RAS, Molodogvardejskaya street 151, Samara, Russia, 443001
e-mail: varlamova.anna.95@mail.ru
Abstract. This paper considers the applicability of different data embedding methods for semi-
fragile watermarking systems used for JPEG image authentication. The methods include Least
Significant Bit watermarking and various versions of Quantization Index Modulation. In our
investigations, we tested the semi-fragility property against JPEG and compared the visual
quality of the watermarked images. We also checked the watermark fragility to unacceptable
modifications like median filtering, blurring, and adding Gaussian noise. Finally, we analyzed
the provided tampering localization error.
1. Introduction
One of the ways to protect an image from tampering is to embed a fragile or a semi-fragile digital
watermark, a barely visible and removable component, whose presence in the image may testify its
authenticity [1]. Fragile watermarks are destroyed after any image modifications and are usually used
for the data integrity verification. If a specific set of modifications is considered to be acceptable,
semi-fragile watermarks are applied to authenticate the data. They are robust against permitted
transformations and fragile to any other. As a rule, these permitted transformations include
modifications that do not affect image content and structure, for example, weak distortions caused by
lossy compression.
The most common standard for lossy image compression is JPEG. More than 20 semi-fragile JPEG
watermarking systems have been developed since 2000. The most widespread among them are those
that embed a watermark in the frequency domain, namely in the Discrete Cosine Transform (DCT)
coefficients before or after quantization [2-15]. The watermarks embedded by such systems are
visually imperceptible and JPEG-resistant even at low values of the quality factor.
The effectiveness of a particular semi-fragile system depends mostly on its data embedding
method. For this reason, in this paper, we investigate the influence of different embedding methods on
the performance of the JPEG semi-fragile watermarking. We consider and compare the methods that
are commonly applied in JPEG-resistant watermarking. They include Least Significant Bit (LSB)
watermarking [1], Quantization Index Modulation (QIM) [16], and its versions (Sign-QIM [17],
MOD-QIM [18], and DM-QIM [16]). In the experimental part, we test their applicability to semi-
fragile JPEG watermarking and compare the Peak Signal-to-Noise (PSNR) values of the obtained
watermarked images. We also verify the fragility of the considered embedding methods to
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018)
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
unacceptable distortions (exemplified by median filtering, blurring, and adding white Gaussian noise)
and analyze the tampering localization error.
The rest of this paper is organized as follows. In Section 2, the lossy JPEG compression scheme is
described. In Section 3, the description of considered data embedding methods is given. Section 4
presents the experimental results.
2. Lossy JPEG compression scheme
The JPEG lossy compression algorithm consists of the following key steps (also shown in Figure 1)
[17].
Figure 1. Lossy JPEG compression scheme.
1. Division of the original N1 N2 image I into 8×8 nonoverlapping blocks I i , where
i 1,.., N and N N1 N2 / 64 is the total number of nonoverlapping blocks in the image.
2. Calculating blockwise DCT. DCT decomposes the image values into different frequencies.
We denote each obtained block of DCT coefficients as Bi m1 , m2 . The coefficients in the
upper left corner (Figure 2) characterize the low frequency component.
3. Quantization of each block Bi using the quantization matrix QQF of size 8×8, corresponding
to the predetermined compression quality factor QF (from 1 to 100).
B m ,m
Di m1 , m2 round i 1 2 (1)
QQF m1 , m2
The smaller QF is, the higher the values of coefficients of the quantization matrix QQF , more
zeros among quantized DCT coefficients Di m1 , m2 , and the smaller the size of the resulting
archive.
4. Scanning each block Di m1 , m2 in zigzag order, as shown in Figure 2, and entropy coding.
Further, we denote DCT coefficients shortly as Di j , where j 1..64 is the index of an
element in zigzag order.
Figure 2. Zigzag scanning of a DCT block.
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 372
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
3. Data embedding methods used in JPEG-resistant watermarking
Data embedding is a key step of any JPEG semi-fragile watermarking system. It determines the way of
modifying the spectral components. The most common data embedding methods actively used in
JPEG-resistant watermarking are the LSB [1] and QIM variations [16-18].
It is worth mentioning the loss of information in the JPEG algorithm occurs at the DCT coefficients
quantization stage. For this reason, frequency domain watermarking systems embed the watermark
either at the quantization step or immediately after it.
LSB embedding in JPEG semi-fragile systems is performed after the quantization by replacing one
or more of the quantized DCT image coefficients with watermark bits [4, 5]. Let us assume that the
number of modified DCT coefficients in each block is equal to the number of bits to be embedded per
block NW . We also assume that inter-coefficient relationships are not taken into account during the
watermark embedding process. We denote the positions of the modified DCT coefficients in zigzag
order as jk , where k 1..NW . In general, they are defined by the secret key. Then the LSB method
embeds the watermark by changing the quantized DCT coefficients located in the jk positions:
DiW jk 2 Di jk 2 Wi ,k (2)
th th
where Wi ,k is the k bit of information that is embedded in the i quantized DCT block. All
coefficients excluding jk remain unchanged. The watermark extraction procedure for this method is
obvious.
LSB is actively used [3-5] because it does not require high computational cost, simple to
implement, and makes possible to hide a sufficiently large amount of information. However, its
application in the DCT frequency domain may cause significant distortions.
The QIM-based methods usually lead to smaller distortions of the watermarked image. Unlike
LSB, QIM-based techniques embed a watermark while quantizing DCT coefficients. They modulate
the DCT coefficients by the watermark bits [16]. In the JPEG semi-fragile systems, various versions of
QIM are in use [6-9]. At first, we consider the method applied in the Preda & Vizireanu watermarking
system [7]:
Bj
BiW jk round i k Wi ,k 2QQF jk Wi ,k QQF jk (3)
2QQF jk
Wi ,k round BiW jk QQF jk (mod 2) (4)
Note that (3) is similar to LSB. As in the case of LSB, the components BiW jk are multiples of
the quantization steps QQF jk . The second QIM-based method we consider in this paper is Sign-
QIM – a simple modification of the method by Preda & Vizireanu. Its distinctive feature lies in the
fact that the watermark component sign depends on the direction to which the modified DCT
coefficient is rounded off at the quantization stage. Due to this, the error in the coefficient jk caused
by information embedding does not exceed QQF jk :
BiW jk Bri jk Si jk Wi ,k QQF jk (5)
where
Bj
Bri jk round i k 2QQF jk (6)
2QQF jk
The third method is DM-QIM, which is the most known QIM version [16]. It subtracts the noise-
like component, which is previously added to the host image components, to avoid a mean value shift
instead of adding the remainder of dividing by the quantization step:
Bi jk dWi ,k jk QQF jk
BiW jk round 2QQF jk dWi ,k jk QQF jk (7)
2Q QF jk
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 373
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
where d0 j , d1 j R 1;1 are two pseudorandom arrays used to modulate watermark bits, and
d1 j d0 j sign d0 j
One more QIM-based embedding method was proposed in paper [18] by Glumov & Mitekin. It has
a wider range of obtained values than Preda & Vizireanu. This method is not intended to provide
robustness against JPEG compression, so the embedding is performed in the spatial domain. Another
distinction of [18] from [7] is that it conducts the floor operation x instead of round ( x) . Thus, the
embedding a single bit w into a single image component x by [18] is as follows:
xW x 2 2 w x mod (8)
where is the quantization step. In (8), the last summand provides exactly the extension of the range
of xW values. We denote this QIM-based method as MOD-QIM.
To incorporate MOD-QIM method with the JPEG compression procedure, we bring in
modifications to the rounding function keeping the remainder as follows:
BiW jk Bri jk Si jk Wi ,k QQF jk M i jk (9)
where
Bj
Bri jk round i k 2QQF jk ,
2QQF jk
1, Bri jk Bi jk
Si jk sign Bi jk Bri jk ,
1, else
and M i jk is the value Bri jk mod QQF jk shifted to the range QQF jk / 2, QQF jk / 2 1 .
Watermark extraction is carried out by (4).
4. Experimental part
In the experimental research, we implemented and tested the selected embedding methods using
different criteria. All experiments were carried out using the images from the University of Waterloo
repository [20].
4.1. Efficiency of the embedding methods in JPEG semi-fragile watermarking
The first experiment assesses the efficiency of the considered methods in JPEG-resistant semi-fragile
watermarking. In this experiment, we embedded NW 4 bits into the DCT coefficients in the fixed
positions (low, medium and high frequency coefficients were modified). For data embedding, we used
QF 50 . Then the watermarked images were compressed to JPEG using various quality factors QF * ,
both lower and higher than QF .
After that, we extracted the hidden bits from each obtained image and estimated the bit error rate
(BER) as:
1 N NW
BER XOR Wi ,k ,Wi ,Rk
N NW i 1 k 1
(10)
The results of the experiment averaged by the dataset are presented in Figure 3 and Table 1.
Table 1. Integral BER deviations from theoretical values (after JPEG compression with all possible
QF* values).
Embedding method errFN errFP
LSB 8.486 0.042
Preda-QIM 8.543 0.041
Sign-QIM 8.540 0.037
DM-QIM 5.446 0.066
MOD-QIM 8.282 1.286
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 374
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
0,55
0,50 LSB
0,45 QIM by Preda & Vizireanu
0,40 MOD-QIM
0,35 DM-QIM
0,30 Sign-QIM
BER
0,25
0,20
0,15
0,10
0,05
0,00
0 20 QF* 60 40 80 100
Figure 3. BER after JPEG compressions with different QF*.
In the ideal case, BER should be close to 0 for QF * QF . For smaller QF * , BER should be
close to 0.5 that corresponds to random guessing. In practice, there is an inevitable transition phase
where BER gradually decreases from 0.5 to 0. Figure 3 shows that it is true for all the considered
methods. The shorter the transition phase, the better the method.
Besides, there may be nonzero values of BER at QF * QF due to the rounding of pixel values
after the inverse DCT that cause distortions of the spectral components.
To assess the deviation of the obtained curves from the ideal case (a step function), we used the
following heuristic measures:
0.5 BER QF
QF 1
errFN *
(11)
QF QF 25
*
QF 24
errFP
QF QF
*
BER QF * (12)
These two expressions characterize the integral BER deviation from their theoretical values. The
obtained errFN and errFP values are presented in Table 1. The table demonstrates that in terms of
errFN measure, DM-QIM considerably outperforms the rest methods. However, DM-QIM provides a
high errFP value. LSB, Preda-QIM, and Sign-QIM are very close in errFN values, while MOD-QIM
has a large number of errors at QF * QF .
4.2. Investigation of introduced distortion level
In the second experiment, we estimated how the quality of the resulting image depends on the number
of embedded bits and the positions of the modified coefficients. For this purpose, we calculated the
Peak Signal-to-Noise (PSNR) measure.
The numbers of modified coefficients in each frequency domain were predetermined, but their
positions were random. We considered coefficients 2-14 in the zigzag scan as the low frequency
domain, 15-35 coefficients as the medium frequency domain, and 36-64 coefficients as the high
frequency domain. As in the first experiment, QF was equal to 50 . The results of the second
experiment are presented in Table 2.
Table 2 shows that LSB, MOD-QIM and Preda-QIM provide quite close quality of the
watermarked images. DM-QIM showed the best results. Analysis of various configurations of
modified frequency domains showed that it is better to embed information into low frequency
components. For instance, if 10 bits are embedded in the low frequency coefficients using DM-QIM,
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 375
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
the quality of the resulting image is higher than if we embed one bit in the high frequency coefficients
by any method.
Table 2. Averaged PSNR of watermarked images after watermark embedding by different methods.
Number of PSNR
Number of modified AC
bits per coefficients per
LSB Preda-QIM Sign-QIM DM-QIM MOD-QIM
block, NW domains
(LF-MF-HF)
1 1-0-0 43.95 42.92 45.87 47.03 44.74
1 0-1-0 33.85 33.74 34.68 37.72 33.97
1 0-0-1 28.93 28.94 29.23 31.67 28.96
2 2-0-0 41.55 40.55 43.56 44.12 42.17
2 0-2-0 31.84 31.70 32.69 34.68 31.98
2 0-0-2 26.42 26.43 26.74 28.76 26.45
4 4-0-0 38.85 37.87 40.92 41.15 39.42
4 0-4-0 29.32 29.18 30.23 31.71 29.50
4 0-0-4 23.68 23.68 24.02 25.75 23.72
4 1-1-2 25.64 25.62 26.05 28.13 25.70
10 10-0-0 34.90 33.93 36.99 37.22 35.45
10 0-10-0 25.78 25.62 26.69 27.70 25.97
10 0-0-10 20.04 20.03 20.40 21.80 20.09
10 2-3-5 22.12 22.08 22.56 24.15 22.55
10 3-3-4 22.85 22.80 23.31 24.88 22.19
10 2-4-4 22.68 22.62 23.15 24.73 22.76
10 1-3-6 21.49 21.46 21.91 23.50 21.56
Mean 29.05 28.77 29.94 31.45 29.27
4.3. Investigation of watermark fragility to unacceptable distortions
The considered data embedding methods should be fragile to typical distortions corrupting image
content. To verify this property, we performed median filtering and image blurring with a sliding
window of size from 3×3 to 15×15, and additive white Gaussian noise with variance values from 400
to 1000. The results are presented in Figures 4, 5, and 6, respectively ( QF 50 , the number of
embedded bits per block NW 4 ). Since all these distortions are unacceptable, the relative extraction
error ( BER ) should ideally be close to 0.5, which corresponds to the probability of random guessing
of the correct bit.
0,55
0,50
0,45
0,40
0,35
0,30
0,25
BER
0,20
0,15 LSB Preda-QIM MOD-QIM
0,10
0,05
0,00
3 5 7 9 11 13 15
Sliding window size
Figure 4. The effect of median filtering on the extraction error.
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 376
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
0,55
0,50
0,45
0,40
0,35
0,30
0,25
BER
0,20
0,15
0,10 LSB Preda-QIM MOD-QIM
0,05
0,00
3 5 7 9 11 13 15
Sliding window size
Figure 5. The effect of blur on the extraction error.
0,55
0,50
BER
LSB Preda-QIM MOD-QIM
0,45
400 500 600 700 800 900 1000
Noise variance
Figure 6. The effect of additive noise on the extraction error.
Figures 4-5 show that LSB slightly outperforms QIM-based methods. However, QIM-based
methods also provide the BER that exceeds 0.4 after nonlinear and linear filtering even with a
window size of 5×5, which is a very good result. With a 3×3 window, the error is also high enough, so
the considered methods are fragile to the distortions. After adding noise to the watermarked image,
almost all methods behave perfectly.
Thus, according to the experimental results, it can be concluded that the considered data embedding
methods are fragile to these three types of distortions.
4.4. Investigation of tampering localization error
Some watermarking systems used for authentication perform content-based watermark generation
aimed to raise tampering localization accuracy. One of such systems is proposed in paper [7] Preda &
Vizireanu. For each block, it calculates a hash value of a pseudo-random sequence and block
coordinates and uses the obtained code as a watermark. This technique protects the image from copy-
move attacks. However, in this research, we did not apply any technique improving localization
accuracy, because we just aimed to compare the embedding methods.
Therefore, we constructed the watermark in a pseudo-random manner, so the localization error was
overestimated. Theoretically, the probability of skipping a distorted block, in this case, should be close
to 1 / 2 N , where N is the number of bits embedded in each block. Thus, for example, if NW 4 , the
percentage of error should be about 6.25%.
The dependence of the fraction of falsely detected blacks on the fraction of tampered blocks is
presented in Figure 7. It illustrates that the graphs for different methods, as expected, are very close to
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 377
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
each other and correspond to the theoretical estimation. The only exception is MOD-QIM that
provides a high error and does not depend strongly on the number of tampered blocks. Consequently,
this method cannot be applied for JPEG semi-fragile watermarking.
0,10
LSB Preda-QIM MOD-QIM
Fraction of falsely
0,08
detected blocks
0,06
0,04
0,02
0,00
0 0,2 0,4 0,6 0,8 1
Fraction of tampered blocks
Figure 7. The influence of the number of modified blocks on the tampering localization error
(the watermark is randomly generated).
5. Conclusion
In the paper, we investigated the different data embedding methods that are usually used in JPEG
semi-fragile watermarking systems such as LSB and various QIM versions (MOD-QIM, Sign-QIM,
and DM-QIM). The evaluation of their performance showed that all considered methods could be
applied to embed a JPEG semi-fragile digital watermark in the frequency domain apart from MOD-
QIM. The study of the quality of the images formed by the watermark embedding process showed the
superiority of the DM-QIM method over the others. We also showed that visual distortions of the
watermarked image are more visually imperceptible when the watermark is embedded in the low
frequency DCT coefficients. As JPEG semi-fragile watermarks must be destroyed with any image
modifications, apart from JPEG, we checked the fragility of the considered methods to the distortions:
median filtering, blurring and white Gaussian noise. Finally, we carried out that the error of tampering
localization coincides with theoretical value for all considered methods excluding MOD-QIM.
6. References
[1] Cox I 2008 Watermarking and Steganography (Morgan Kaufmann) p 624
[2] Lin C Y and Chang S F 1999 Issues and solutions for authenticating MPEG video Proceedings
of SPIE 54-65
[3] Lin C Y and Chang S F 2000 Semifragile watermarking for authenticating JPEG visual content
Security and Watermarking of Multimedia Contents II 3971 140-151
[4] Ho C K and Li C T 2004 Semi-fragile watermarking scheme for authentication of JPEG images
ITCC 1 7-11
[5] Huang L Y 2013 Authentication watermarking algorithm resisting JPEG compression based on
preliminary quantization Information Technology Journal 12 3723-3728
[6] Ye S, Zhou Z, Sun Q, Chang E and Tian Q 2003 A quantization-based image authentication
system Proceedings of the 2003 Joint 2 955-959
[7] Preda R O and Vizireanu D N 2015 Watermarking-based image authentication robust to JPEG
compression Electronics Letters 51 1873-1875
[8] Wang H, Ho A and Zhao X 2011 Novel fast self-restoration semi-fragile watermarking
algorithm for image content authentication resistant to JPEG compression 72-85
[9] Fan C H, Huang H Y and Hsu W H 2011 A Robust Watermarking Technique Resistant JPEG
compression J. Inf. Sci. Eng 27 163-180
[10] Fallahpour M and Megias D 2016 Flexible image watermarking in JPEG domain ISSPIT 311-
316
[11] Mursi M, Assassa G M R, Aboalsamh H and Alghathbar K 2009 A DCT-based secure JPEG
image authentication scheme World Academy of Science, Engineering and Technology 53 681-
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 378
Image Processing and Earth Remote Sensing
A A Egorova, V A Fedoseev
687
[12] Lin E T, Podilchuk C I and Delp E J 2000 Detection of image alterations using semifragile
watermarks Security and Watermarking of Multimedia Contents II 3971 152-164
[13] Al-Mualla M E 2007 Content-adaptive semi-fragile watermarking for image authentication 14th
IEEE International Conference on Electronics, Circuits and System 1256-1259
[14] Wong P H W, Au O C L and Wong J W C 2001 Data hiding technique in JPEG compressed
domain Security and Watermarking of Multimedia Contents III 4314 309-320
[15] Xiao J, Ma Z, Lin B, Su J and Wang Y 2010 A semi-fragile watermarking distinguishing JPEG
compression and gray-scale-transformation from malicious manipulation IEEE Youth
Conference on Information, Computing and Telecommunications 202-205
[16] Chen B and Wornell G 2001 Quantization index modulation: a class of provably good methods
for digital watermarking and information embedding IEEE Transaction on Information Theory
47 21
[17] Egorova A A and Fedoseev V A 2019 A classification of semi-fragile watermarking systems for
JPEG images Computer Optics 43 (in print)
[18] Glumov N I and Mitekin V A 2011 A new semi-fragile watermarking algorithm for image
authentication and information hiding Computer Optics 35(2) 262-267
[19] Wallace G K 1992 The JPEG still picture compression standard IEEE Transactions on
Consumer Electronics 38 xviii–xxiv
[20] Image repository The waterloo fractal coding and analysis group URL:
http://links.uwaterloo.ca/Repository.html
Acknowledgment
The work was partly funded by the Russian Federation Ministry of Science and Higher Education
within a state contract with the "Crystallography and Photonics" Research Center of the RAS under
agreement 007-ГЗ/Ч3363/26 (in part of JPEG implementation) and by the RFBR grant # 18-71-00052
(in parts of review and investigations).
V International Conference on "Information Technology and Nanotechnology" (ITNT-2018) 379