Steganalysis of Semi-fragile Watermarking Systems Resistant to

JPEG Compression

Anna Egorova

and Victor Fedoseev

1,2

Samara National Research University, Samara, Russia

Image Processing Systems Institute, Branch of the Federal Scientific Research Centre “Crystallography and Photonics” of

Russian Academy of Sciences, Samara, Russia

Keywords: Image Protection, JPEG, Semi-fragile Watermarking, Targeted Steganalysis, LSB, QIM.

Abstract: Recently, dozens of semi-fragile digital watermarking systems have been designed to protect JPEG images

from unauthorized changes. Their principle is to embed an invisible protective watermark into the image.

Such a watermark is destroyed by any image editing operations, except for JPEG compression with the quality

level in a given range of values. Watermarking systems of this type have been assessed in terms of watermark

extraction accuracy and visual quality of the protected image. However, their steganographic security (i.e.,

robustness against detecting protective information traces by a third party) has not been sufficiently studied.

Meanwhile, if an attacker detects the presence of a watermark in the image, he can get valuable information

on the used image protection technique. It can let him develop a data modification method that alters the

content of the protected image without destroying the embedded watermark. In this paper, we propose a

specific attack to analyze steganographic security of known semi-fragile watermarking algorithms for JPEG

images. We also investigate the efficiency of the proposed attack. In addition, we propose an approach to

counter the attack that can be applied in the existing watermarking systems to enforce their steganographic

security.

1 INTRODUCTION

At present, the role of visual information (in

particular images) has increased in various areas of

the digital economy: e-commerce, medicine,

education, etc. Consequently, image authentication

and malicious change detection in images have

become the tasks of great importance. Note that

images are mostly compressed in practice. For this

reason, distortions caused by JPEG, JPEG 2000, and

other lossy compression formats are often considered

as legal modifications. For compressed images

authenticating, semi-fragile watermarking systems

can be used (Cox, 2008). In this paper, we consider

the watermarking systems that are robust only to

JPEG compression. They embed a watermark

(security information) into the image immediately

after image registration. Such the watermark has the

property to be preserved after JPEG compression, but

it is destroyed after any other modifications of the

image. The performance of such watermarking

systems has been proven (Egorova and Fedoseev,

2019), but their steganographic security (the ability to

detect watermark traces by a third party) has not

previously been examined. Meanwhile, if an intruder

detects the presence of such an embedded watermark

in the image, he can get information on the used

image protection system. Thereby, by using this

information, he can develop a data alteration method

that does not change the protective watermark but

distorts the image content.

In this study, we model a specific attack to analyze

steganographic security of various semi-fragile

watermarking systems designed for the JPEG

compression standard (Lin and Chang, 2000; Mursi et

al., 2009; Preda and Vizireanu, 2015; Fallahpour and

Megias, 2016; Egorova and Fedoseev, 2019). The

need for a new attack is caused by the fact that the

existing targeted attacks for JPEG steganography

methods (JSteg, F5, Model-based, etc.) (Fridrich,

2010) do not fit the semi-fragile embedding concept.

The essence of the proposed attack is that the

number and the distribution of both odd and nonzero

quantized DCT coefficients can be used as significant

features to detect a watermark, i.e., to separate

original images from watermarked ones.

Egorova, A. and Fedoseev, V.

Steganalysis of Semi-fragile Watermarking Systems Resistant to JPEG Compression.

DOI: 10.5220/0009129708210828

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages

821-828

ISBN: 978-989-758-402-2; ISSN: 2184-4321

821

The study also answers the following questions:

 A) How efficient is the proposed attack?

 B) Which of the existing JPEG semi-fragile

watermarking systems is more resistant to the

attack?

 C) How to adjust the watermark embedding

procedures, which are commonly used in existing

JPEG semi-fragile watermarking systems to

protect these systems against the specific attack?

The rest of the paper is organized as follows. Section

2 provides a brief description of the data embedding

techniques that are commonly utilized in the JPEG-

resistant watermarking systems. Section 3 introduces

a method for targeted steganalysis of these systems.

Section 4 presents the results of conducted numerical

experiments and answers questions A and B, while

Section 5 responds to question C and suggests a

modification of watermark embedding procedure that

protects the considered watermarking systems against

the proposed method for targeted steganalysis.

2 SEMI-FRAGILE

WATERMARKING SYSTEMS

FOR JPEG

Let us consider the main steps of the lossy JPEG

compression standard for grayscale images (see

Figure 1). First, a source image

is transformed by

8×8 block discrete cosine transform (DCT), resulting

in coefficients



, where i is an index of 8×8

block and

1..64j 

is a DCT coefficient index in zig-

zag scanning. Then each



is divided by a 8×8

quantization matrix



element-wise, where

the index

is the compression quality factor. Then

the quotients are quantized, resulting in values



. Finally, entropy coding of



performed.

Figure 1: JPEG image compression scheme.

Existing JPEG semi-fragile watermarking

systems embed the watermark by modifying either

DCT





or quantized DCT



coefficients.

Let

be a number of watermark bits to be

embedded in each image block, and

be the

position of the DCT coefficients in zig-zag scanning

to be watermarked, where

1..

kN

The simplest way to embed the watermark during

JPEG compression process is to change the least

significant bits of the quantized DCT coefficients

(LSB method) (Barni and Bartolini, 2004):









ik ik ik

Dj Dj W







(1)

where

,ik

is the k-th bit of information embedded in

the i-th block. This procedure is implemented in the

system proposed in (Lin and Chang, 2000).

Another embedding approach is quantization

index modulation (QIM) (Chen, 2001) that

simultaneously quantizes DCT components and

embeds the watermark. QIM may be implied in

various forms. For example, in (Preda and Vizireanu,

2015), the authors use the following embedding rule:



ik ik QFk

QF k

ik QF k

Bj round W Q j

WQ j













(2)

Another QIM-based system is “Sign-QIM” proposed

in (Egorova and Fedoseev, 2019). In this system, the

watermark embedding is performed according to

these equations:

   

ik ik ikikQFk

Bj Brj SjWQ j

(3)









ik QFk

QF k

Br j round Q j









(4)



 

1, ,

1, .

ik ik

Bj Brj

else













(5)

There are many other JPEG semi-fragile

watermarking systems based on LSB (Ho and Li,

2004; Huang, 2013) or QIM (Wang et al., 2011; Fan

et al., 2011; Ye et al., 2003). They are described in

detail in (Egorova and Fedoseev, 2019).

A system using another embedding algorithm is

proposed in (Mursi et al., 2009). It is based on the

mapping table approach. Such a table is randomly

generated using a secret key. It defines a mapping

between values of



and the set

{0,1}.

Suppose

a bit “1” needs to be embedded in some coefficient. If

its value corresponds to “1” in the mapping table, it

does not change. Otherwise, the value is replaced by

the nearest number corresponding to “1” in the table.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

822

The system (Fallahpour and Megias, 2016)

associates a watermark bit with the parity of the last

nonzero (LNZ)



coefficient position. That is,

this system cannot embed more than one bit per

block. However, in our study, we use a generalized

version of this system that embeds up to 4 bits per

block.

Thus, in this study, we investigate the five

watermarking systems (Lin and Chang, 2000; Mursi

et al., 2009; Preda and Vizireanu, 2015; Fallahpour

and Megias, 2016; Egorova and Fedoseev, 2019)

implementing the four most common embedding

approaches used in semi-fragile watermarking: LSB,

QIM, LNZ, and mapping tables. In the experimental

part, when we select the DCT coefficients in positions

specified in “original” papers for watermarking, we

call the selecting mode “original”. To check different

embedding approaches in the same conditions, we

also test a “sequential” mode, where the first

coefficients in each block are modified. This mode

reduces distortions in the watermarked image and

increases the PSNR metric (Egorova and Fedoseev,

2019). In all the experiments, we use

values

from the set

{1,2,4}

and

50QF 

3 PROPOSED TARGETED

ATTACK AGAINST JPEG

SEMI-FRAGILE

WATERMARKING

Most of the quantized DCT coefficients

generated at the JPEG compression process are equal

to zero. When the image is watermarked by using the

system based on LSB, QIM, or the mapping table, the

statistics of even and odd DCT coefficients of a block

are leveled. Given this, we supposed that the number

of odd and the number of nonzero coefficients per the

quantized DCT block might be the significant

features and can be used to detect the embedded

information in JPEG images.

To test this hypothesis, we plotted two graphs (see

Figure 2). The top plot shows the number of odd

coefficients among the first j coefficients of a block

in the zig-zag scanning. The lower plot illustrates the

same statistics of the nonzero coefficients. The plots

show an average result for 1000 halftone images

obtained by each of the five selected watermarking

systems using the “original” mode of coefficients

selection. Black lines in Figure 2 correspond to 50

different host images for clarity. It can be seen that

the scatter of the nonzero and odd statistics is quite

significant. It means that the total numbers of nonzero

and odd values in each block (the values on the graphs

corresponding to j=64) are not sufficiently reliable

signs of the watermark presence. However, the

protected images generally have a higher growth rate

at low j indices, since low-frequency coefficients are

usually used for watermarking in order to reduce

visual distortions.

Figure 2: The average number of odd (up) and nonzero

(down) DCT coefficients among the first j positions in 8×8

watermarked image block obtained by using “original”

coefficient selection mode, compared with the same plots

for 50 host images (black lines marked as I).

The same results are more clearly shown in Figure

3. Here instead of 50 lines for host images, only the

one – averaged over 1000 images is shown (black

line). To save space, in Figure 3, we show only graphs

of odd statistics. Figure 3 also presents the result for

the “sequential” mode of coefficients selection. The

figure shows that the watermarked images have a

higher growth rate at low j. Note that the “sequential”

mode decreases this feature. Figure 3 illustrates that

the line for the system (Fallahpour and Megias, 2016)

is indistinguishable from the line for the empty

containers. This is because the system does not

change the values of the coefficients. Instead, it swaps

some values. Therefore, in relation to our attack, this

Steganalysis of Semi-fragile Watermarking Systems Resistant to JPEG Compression

823

system is secure. However, there is a much simpler

and more effective attack for this system: testing the

hypothesis that odd and even positions of LNZ are

equally probable. It can be done by applying the

statistical methods, for example, by calculating chi-

square statistics (Cox, 2008).

Figure 3: The average number of odd DCT coefficients

among the first j positions in 8×8 source (I) and

watermarked image block obtained by using “original” (up)

and “sequential” (down) coefficients selection modes.

4 EXPERIMENTAL

INVESTIGATION OF THE

PROPOSED ATTACK

4.1 Model and Feature Selection

Based on the analysis of the results shown in Figure

3, we developed a targeted attack in the form of a

linear SVM classifier, which determines whether a

given image contains a watermark or not. This

classifier operates with the following feature set:

((4) (1),NZN NZN (9) (4),NZN NZN (64),NZN

UNZN

(4) (1),ON ON (9) (4),ON ON

(64),ON )

UON ,

where

()NZN j is the number of nonzero

coefficients among the first j,

()ON j is the

corresponding number of odd coefficients,

AUNZN

is the area under the curve of nonzero values, and

AUON

is the area under the curve of odd values.

Differential features reflect the growth dynamics

()NZN j and ()ON j among the most important

DCT coefficients.

(64)NZN and (64)ON show the

total values of the measured statistics. The area

features represent both dynamics and total numbers.

Figure 4: Importance of different features in detecting the

watermarked images (obtained using the “original”

coefficients selection mode).

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

824

In the preliminary tests, we analyzed the

significance of single features from the selected

feature set using the sequential forward selection

procedure (Marcano-Cedeño et al., 2010). These tests

showed that the differential features significantly

outperform all others at all tested

values.

Diagrams in Figure 4 confirm this property and show

that the nonzero features and the odd features vary in

importance for different watermarking algorithms.

However, in the experiments described below, we

use the full 8-feature set, as well as two 4-feature

subsets of nonzero and odd features for comparison.

This was done in order to prevent the exclusion of

potentially significant data.

4.2 Performance Evaluation of the

Developed Attack

To investigate the effectiveness of the developed

attack, we used 1000 images of size 512×512 from

the BOWS-2 dataset (Westfeld, 2009). By applying

each of the five selected systems, we generated 1000

images with watermarks. 70% of the resulting 2000

images were used for training the classifier, and 30%

for testing.

The results on classification accuracy obtained

using both “original” and “sequential” modes of

coefficients selection at the different

are

presented in Table 1. The higher the accuracy value,

the higher the probability of a successful attack.

As predicted, the system (Fallahpour and Megias,

2016) is completely secure against the attack at any

N . Other systems can be detected. If the “original”

mode is used for coefficients selection, the least

resistant system is (Lin and Chang, 2000) based on

LSB. The detection accuracy for this system is 73 –

100%, depending on

. Three other systems show

very similar results and can be detected with accuracy

higher than 69% for

N  . When using the

“sequential” mode, all the accuracy values become

lower, and the detection accuracy could be rated as

acceptable only at



. The highest values are

gained by the system (Mursi et al., 2009).

Table 1: Accuracy of different systems detection by the developed attack (full feature set). The higher the accuracy value,

the higher the probability of a successful attack.

1 2 4

Positions selection method Original Sequential

Adaptiv

-2

Origi

Sequential

Adaptiv

-2

Origi

Sequential

Adaptiv

-2

Lin & Chang, 200

0.73 0.56 0.54 0.97 0.61 0.53 1.00 0.69 0.63

Preda & Vizireanu, 2015 0.62 0.57 0.54 0.69 0.64 0.58 0.85 0.67 0.77

Sign-QIM 0.65 0.57 0.55 0.72 0.65 0.67 0.85 0.72 0.76

Fallahpour & Megias, 201

0.50 0.51 0.51 0.50 0.51 0.50 0.50 0.51 0.51

Mursi et al., 2009 0.64 0.63 0.58 0.78 0.52 0.56 0.84 0.72 0.62

Table 2: Accuracy of different systems detection by the developed attack (odd features only). The higher the accuracy

value, the higher the probability of a successful attack.

1 2 4

Positions selection method Original Sequential

Adaptiv

-1

Origi

Sequential

Adaptiv

-1

Origi

Sequential

Adaptiv

-1

Lin & Chang, 200

0.79 0.54 0.50 0.90 0.60 0.50 0.95 0.80 0.50

Preda & Vizireanu, 2015 0.66 0.54 0.51 0.78 0.59 0.51 0.83 0.79 0.50

Sign-QIM 0.69 0.53 0.52 0.71 0.60 0.52 0.86 0.74 0.55

Fallahpour & Megias, 201

0.51 0.51 0.51 0.51 0.51 0.51 0.51 0.51 0.51

Mursi et al., 2009 0.57 0.51 0.51 0.59 0.54 0.51 0.60 0.57 0.50

Table 3: Accuracy of different systems detection by the developed attack (nonzero features only). The higher the accuracy

value, the higher the probability of a successful attack.

1 2 4

Positions selection method Original Sequential

Adaptiv

-1

Origi

Sequential

Adaptiv

-1

Origi

Sequential

Adaptiv

-1

Lin & Chang, 200

0.72 0.57 0.51 0.93 0.59 0.52 0.99 0.67 0.58

Preda & Vizireanu, 2015 0.61 0.54 0.50 0.65 0.58 0.52 0.82 0.64 0.55

Sign-QIM 0.61 0.53 0.51 0.65 0.57 0.53 0.81 0.65 0.55

Fallahpour & Megias, 201

0.51 0.50 0.51 0.51 0.51 0.50 0.51 0.51 0.50

Mursi et al., 2009 0.62 0.53 0.51 0.75 0.57 0.50 0.78 0.65 0.55

Steganalysis of Semi-fragile Watermarking Systems Resistant to JPEG Compression

825

Table 4: Average PSNR of secured images after watermark embedding by the investigated systems.

1 2 4

Positions selection method Original Sequential

Adaptiv

-2

Origi

Sequential

Adaptiv

-2

Origi

Sequential

Adaptiv

-2

Lin & Chang, 200

41.6 44.0 39.9 39.2 41.0 36.9 36.6 38.0 33.8

Preda & Vizireanu, 2015 45.7 46.0 40.5 42.3 43.0 37.4 39.1 40.0 34.3

Sign-QIM 49.3 49.5 42.1 45.9 46.7 39.1 42.6 43.7 35.9

Fallahpour & Megias, 201

35.4 35.3 35.3 35.4 35.2 35.2 35.4 35.3 35.3

Mursi et al., 2009 47.6 47.5 36.1 44.3 42.0 38.6 40.8 41.6 30.3

Tables 2-3 show that the detection accuracy

reduces if either nonzero only or odd only features are

selected for classification in the attack. Overall, the

obtained results claim the effectiveness of the

proposed attack in some range of conditions.

5 MODIFYING THE

WATERMARKING SYSTEMS

TO MAKE THE ATTACK LESS

EFFECTIVE

In this section, we suggest two methods for adjusting

the watermark embedding procedures in the existing

systems that allow enforcing the security of the

systems against the proposed attack. To make an

adjustment method universal, we modify the mode

for selecting positions of DCT coefficients. The idea

is to select a coefficient in a way that makes the

nonzero and odd curves after the embedding (see

Figure 3) close to graphs of host images. To get this

effect, we generate these positions randomly. The

probability of each position selection is proportional

to the numerical derivative of the host image lines

shown in Figure 3. This mode we name “Adaptive-

1”. It is investigated in the same conditions as other

positions selection modes. Tables 2-3 present the

classification accuracy when only nonzero or only

odd features were used. It can be seen that the

“Adaptive-1” mode is more secure than other modes

because it provides lower accuracy. However, the full

feature set does not provide lower classification

accuracy. To save space, these results are not

specified in Table 1.

To overcome the disadvantage of the “Adaptive-

1”, we modify this method slightly. In addition to

watermark insertion in some positions, we replace

with zero some existing nonzero coefficients. Their

positions are also selected according to the numerical

derivative of the host image nonzero line. The number

of such coefficients is proportional to

. The

experimental results obtained for this mode are shown

in Table 2 in the column “Adaptive-2”. These data

show that the accuracy becomes lower for all cases

except for the two QIM-based systems for



Thus, this mode has reasonable potential.

It is clear that zeroing some DCT coefficients

reduces the watermarked image quality. To estimate

the quality loss, we measured PSNR in all the cases

tested earlier. Table 4 contains the obtained numerical

results, while Figure 5 shows some examples of images

protected using all the analyzed watermarking systems

in combination with “Adaptive-2” at

N 

. The

numbers and shown images prove that visual quality is

reduced. However, the loss is not so dramatic, and the

quality of the watermarked images can be recognized

as acceptable in many applications.

6 CONCLUSIONS

In this paper, at first, we proposed a new targeted

steganographic attack against semi-fragile

watermarking systems designed for the JPEG

compression standard. The goal was to analyze the

steganographic security of such systems. The attack

consists of the calculation of the nonzero and the odd

DCT coefficients statistics, selection of significant

features, and training an SVM classifier. We showed

logically and experimentally the significance of the

selected features for the given problem. To investigate

the efficiency of the developed attack, we applied it for

five different watermarking systems and measured the

classification accuracy for different watermark length.

The systems were tested in both “original” and

“sequential” modes of coefficients selection. The

results showed that the attack is effective for all the

systems except one (Fallahpour and Megias, 2016)

(but this system has more crucial security problems)

when more than one bit per block is embedded.

“Sequential” mode has been found more secure than

“original”, in addition to its higher quality investigated

in (Egorova and Fedoseev, 2019) and also shown in

Table 4. In addition, in this paper, we proposed and

investigated two methods to counter the developed

attack, i.e., ways to make the systems more secure.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

826

Source image

Lin & Chang, 2000

PSNR = 33.79

Preda & Vizireanu, 2015

PSNR = 34.51

Sign-QIM

PSNR = 36.18

Fallahpour & Megias, 2016

PSNR = 36.46

Mursi et al., 2009

PSNR = 35.50

Figure 5: Examples of secured images protected using the proposed “Adaptive-2” positions selection method at



These methods consist in specific modes of

coefficients selection, “Adaptive-1” and “Adaptive-2”,

and allow obtaining more “natural” DCT coefficients

statistics. The second method, “Adaptive-2”, in

addition to adaptive coefficients selection, zeroes some

values. The experiments show that “Adaptive-1”

improves security when the attack is applied using

shortened feature sets (nonzero only features, or odd

only features). However, for the full set, its effect is not

so clear. The other method undoubtedly makes the

systems more secure but reduces the visual quality of

the protected images (by 5 dB on average in terms of

PSNR). However, if we use the Sign-QIM system

(Egorova and Fedoseev, 2019), which is the best in

visual quality, we can still obtain 36 dB in the case of

4 bit per block watermarking and 42 dB in the case of

1 bit per block watermarking.

Overall, the paper draws the attention of

researchers to the problem of steganographic security

of semi-fragile watermarking systems and gives some

practical methods to improve it.

ACKNOWLEDGMENTS

The work was funded by the RSF grant #18-71-

00052.

REFERENCES

Barni, M. and Bartolini, F. (2004). Watermarking systems

engineering: Enabling digital assets security and other

applications, CRC Press.

BOWS-2. Image Repository. (n.d.). Retrieved from

http://bows2.ec-lille.fr/.

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(4). 1423–

1443. https://doi.org/10.1109/18.923725.

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

(2008). Digital watermarking and steganography,

Elsevier.

Egorova, A. A and Fedoseev, V. A. (2019). A classification

of semi-fragile watermarking systems for JPEG images.

Computer Optics, 43(3), 419–433. https://doi.org/

10.18287/2412-6179-2019-43-3-419-433.

Fallahpour, M. and Megias, D. (2016). Flexible image

watermarking in JPEG domain. In 2016 IEEE

International Symposium on Signal Processing and

Information Technology (ISSPIT), 311–316.

https://doi.org/10.1109/ISSPIT.2016.7886055.

Fan, C. H., Huang, H. Y., and Hsu, W. H. (2011). A robust

watermarking technique resistant JPEG compression.

In Journal of Information Science and Engineering,

27(1), 163–180.

Steganalysis of Semi-fragile Watermarking Systems Resistant to JPEG Compression

827

Fridrich, J. (2009). Steganography in digital media:

Principles, algorithms, and applications. Cambridge

University Press.

Ho, C. K. and Li C. T. (2004). Semi-fragile watermarking

scheme for authentication of JPEG images. In

Proceedings of the International Conference on

Information Technology: Coding and Computing, 1, 7–

11. https://doi.org/10.1109/ITCC.2004.1286417.

Huang, L. Y. (2013). Authentication watermarking

algorithm resisting JPEG compression based on

preliminary quantization. In Information Technology

Journal, 12(6), 3723–3728. https://doi.org/10.3923/

itj.2013.3723.3728.

Lin, C. Y. and Chang, S. F. (2000). Semifragile

watermarking for authenticating JPEG visual content.

In Electronic Imaging. 140–151.

https://doi.org/10.1117/12.384968.

Marcano-Cedeño, A., Quintanilla-Domínguez, J., Cortina-

Januchs, M. G., & Andina, D. (2010). Feature selection

using Sequential Forward Selection and classification

applying Artificial Metaplasticity Neural Network.

IECON 2010 - 36th Annual Conference on IEEE

Industrial Electronics Society, 2845–2850.

https://doi.org/10.1109/IECON.2010.5675075

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, 681–687.

https://doi.org/10.5281/zenodo.1070359.

Preda, R. O. and Vizireanu, D. N. (2015). Watermarking-

based image authentication robust to JPEG

compression. Electronics Letters, 51(23), 1873–1875.

https://doi.org/10.1049/el.2015.2522

Wang, H., Ho, A., and Zhao, X. (2011). A novel fast self-

restoration semifragile watermarking algorithm for

image content authentication resistant to JPEG

compression. Digital Forensics and Watermarking,

7128. 72–85. https://doi.org/10.1007/978-3-642-

32205-1_8.

Westfeld, A. (2009). Fast Determination of Sensitivity in

the Presence of Countermeasures in BOWS-2.

Information Hiding (pp. 89–101). Springer.

https://doi.org/10.1007/978-3-642-04431-1_7.

Ye, S., Zhou, Z., Sun, Q., Chang, E., and Tian, Q. (2003).

A quantization based image authentication system. In

Fourth International Conference on Information,

Communications and Signal Processing, 2003 and the

Fourth Pacific Rim Conference on Multimedia, 955–

959. https://doi.org/10.1109/ICICS.2003.1292599.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

828