Forensic Science International 320 (2021) 110691

Contents lists available at ScienceDirect

Forensic Science International

journal homepage: www.elsevier.com/locate/forsciint

A DWT-SVD based robust digital watermarking for medical image

security
Narima Zermia , Amine Khaldib,* , Redouane Kafib , Fares Kahlessenaneb , Salah Euschib
a
Electronics Department, Faculty of Engineering Sciences, Laboratories of Automation and Signals of Annaba (LASA), Badji Mokhtar Annaba University,
23000, Annaba, Algeria
b
Computer Science Department, Faculty of Sciences and Technology, Artificial Intelligence and Information Technology Laboratory (LINATI), University of
Kasdi Merbah, 30000, Ouargla, Algeria

A R T I C L E I N F O A B S T R A C T

Article history: In this work, we propose a blind watermarking approach for medical image protection. In this approach,
Received 16 September 2020 the watermark will be constituted of the Electronic Patient Record and the image acquisition data. In
Received in revised form 26 November 2020 order to enhance the security and guarantee the data integrity, the Electronic Patient Record hash will be
Accepted 12 January 2021
added to the watermark. The integration process is based on a DWT-SVD combination, a DWT is applied
Available online 13 January 2021
to the retinal image, then, an SVD is applied to the LL sub-band. The watermark will be then integrated
into the least significant bits of the S component obtained by combining the parity of the successive
Keywords:
coefficients. Experimental results for imperceptibility and robustness show that the proposed scheme
Digital watermarking
Color image watermarking
maintains a high quality watermarked image and remains highly robust against several conventional
Medical image attacks.
Electronic patient record © 2021 Elsevier B.V. All rights reserved.
Discrete wavelet transform
Singular value decomposition

1. Introduction Digital image watermarking consists of concealing in visual

information (a digital image), hidden information identifying its
Telemedicine is a remote form of medical practice using owner, or its content [8]. However, traditional watermarking
information and communication technologies [1]. It links one or schemes have some limitations for sensitive images such as
more health professionals to each other or to a patient. The medical images [9]. The inserted mark often alters irreversibly the
implementation of remote medical data visualization interfaces image and can hide subtle details. The proposed approach in this
allows access to patient records containing text and image data [2]. work is a hybrid scheme based on a discrete wavelet transforms
It has many advantages, particularly for remote health care (DWT) and singular value decomposition (SVD). A DWT can be
institutions. However, there are also security and data protection considered as an image decomposition into sub-frequency bands,
risks [3]. Sharing this data, and particularly images on the Internet, which allows developing psychovisual masks easily [10]. In
exposes them to unauthorized manipulation, despite existing addition to the robustness common with DWT, the wavelet
security tools such as access control [4]. Encryption techniques are domain watermark is robust to scaling factor changes. In addition,
the first solution to prevent unauthorized access to digital data. perceptual masks are finer and there are less block effects [11]. In
They meet users' security needs such as confidentiality, integrity this approach, the watermark will consist of the Electronic Patient
and identification. However, these techniques have proven to be Record and the image acquisition data. To enhance security and
insufficient or difficult to use [5]. Indeed, cryptographic tools ensure data integrity, the hash of the Electronic Patient Record will
protect the image only during transmission, but once the image is be added to the watermark. The integration process is based on a
decrypted, there is no control to prevent illegal manipulation [6]. DWT- SVD combination, a DWT is applied to the image for
Watermarking has therefore been proposed to increase the decomposition into four sub-bands, and then, singular value
security of shared images allowing to maintain the confidentiality decomposition is performed to the LL sub-band. The watermark
of patient data and to verify the medical images' integrity [7]. will then be integrated into the least significant bits of the S
component obtained by combining the parity of the successive
coefficients. The proposed integration process will consist of
* Corresponding author.
modifying as few values as possible. For that, a combination of each
E-mail addresses: [email protected] (N. Zermi), three successive coefficients will be calculated. The proposed
[email protected] (A. Khaldi). substitution rules will then allow modifying at most only one

http://dx.doi.org/10.1016/j.forsciint.2021.110691
0379-0738/© 2021 Elsevier B.V. All rights reserved.
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

coefficient for the integration of two bits of the watermark. The information, a WPT is applied to segment the image and find the
watermarked image must provide the same clinical interpretation minimum energy band where the watermark text is embedded. After
as the original image, so it should not be damaged in any way that the embedding process, a cryptographic key is generated and the
affects the diagnosis. Modifying few information for the water- watermarked image is encrypted. Anand & al proposes a multi-
mark integration will thus allow to obtain a watermarked image watermark embedding method for medical image [19]. The
similar to the original. The watermark extraction will allow embedding watermark is composed of a text and an image. A four
patients authentication using the Electronic Patient Record sub-band decomposition of the host image is first performed by
obtained. The integrity of this information can be verified through applying a DWT. An SVD is applied to the HH sub-band and the text is
the extracted hash, and the image acquisition data will guarantee embedded in the singular value S. The image watermark is divided
the copyright. In order to verify the imperceptibility of our into two sub images, these two parts are then integrated (after
approach, distortion measurements will be calculated between the applying an SVD to both sub bands) into the singular S values of the
original and watermarked images. Robustness tests will be HL and LH sub bands of the cover image. To enhance security, the
performed by applying the most commonly used attacks (filtering, watermarked image is then encrypted using HyperChaotic encryp-
compression, adding noise . . . etc.). This paper is organized as tion algorithm. Devi & al [20] present a blind watermarking scheme
follows: In Section 2 will be presented some recent related work, that combines DWT and SVD. A DWT is applied to the blue
the proposed approach will be presented in Section 3, Section 4 component of the cover image; the resulting LL sub-band is then
will be dedicated to experiments and imperceptibility and decomposed into 8*8 blocks. For each block, a fast Walsh Hadamard
robustness tests. The conclusion of this work and perspectives transformation is applied. Singular value decomposition is then
will be presented in Section 5. applied to the blocks of the LL sub-band formed by the Hadamard
transform. The watermark bits are then integrated into the singular
2. Related works values obtained. In order to increase security, the watermark image is
scrambled using the Arnold Transform. Khare & Srivastava propose a
The various watermarking images methods available in the combination of a discrete wavelet transform, homomorphic
literature are related to the insertion domain of the mark (spatial, transform; singular value decomposition and Arnold transform
frequential, multi-resolution, etc). Each image representation space [21]. In this work, a DWT is first performed, and then homomorphic
brings diverse possibilities in terms of performance [12]. The choice transform is applied to the HL subband to decompose it in
of the insertion domain is therefore a very delicate step in the design illumination and reflectance components. Using Arnold transform,
of watermarking systems. Spatial domain insertion is performed by the watermark is scrambled. The integration process is then
directly modifying the pixel values of the image. The main advantage performed using the singular values of the reflectance component.
of this domain is that the insertion and detection operations are Zhou & al propose a robust and hybrid watermarking scheme [22]. In
inexpensive in computing time, which allows it to be used in real- this work, discrete wavelet transform is performed. Singular value
time watermarking applications [13]. But the main disadvantage is decomposition is then applied to the HL and LH subbands and the
the visual quality degradation of the image after the integration. watermark is embedded by modifying the direct current coefficients.
Moreover, these techniques are not very robust to attacks. Frequency Benoraira & al present a blind and robust image watermarking
domain insertion consists of inserting the mark not directly into the scheme based on discrete wavelet transform and discrete cosine
image but into the transformed domain. The frequency representa- transform [23]. In this approach, discrete cosine transform is first
tion of the image is obtained by applying a Discrete Fourier applied. Then, the two DCT-transformed sub-vectors are used to
Transform or a Discrete Cosine Transform [14]. The main difficulty of embed the bits of the watermark bits. The original sub-vectors are
this type of scheme is the choice of the marked coefficients. The best obtained by the sub-sampling of the approximation coefficients of
schemes offering a good robustness-invisibility compromise gener- the DWT transform of the host image.
ally opt for an insertion in the middle frequencies of the image in
order to maximize imperceptibility by avoiding the low frequencies, 3. Proposed watermarking scheme
without risking that the mark could be removed by compressing or
noise addition if it was inserted in the high frequencies [15]. Despite We can classify watermarking schemes according to their
the main advantage of this domain, which is the invariance to robustness. The robust watermark is designed to protect docu-
translation and scaling, this technique has the disadvantage of being ments. It must be as resistant as possible against attacks. The
expensive in terms of computation time and is not resistant to fragile watermark allows verifying whether the protected image
windowing and small local deformations [16]. The multiresolution has been modified (integrity check). The mark must be fragile, if
domain is a very interesting watermarking space since it is used in the document is modified the mark must disappear [24]. In the
recent JPEG2000 or MPEG4 compression standards. It is character- proposed approach we try to guarantee robustness to attacks while
ized by the decomposition of the image into sub-bands allowing the including a data integrity checker. The proposed watermarking
isolation of low frequency components, which constitute an scheme consists of three main steps: watermark generation,
insertion space less sensitive to modifications [17]. On the other integration and extraction processes.
hand, the spatial content of the image remains preserved after a
multi-resolution transformation; this same content can then be used 3.1. Watermark generation
to locate the mark after a geometric transformation. Mohti & al [18]
propose a hybrid watermarking approach that combines visual The watermark used in this approach is a combination of three
cryptography and Wavelet Packet Transform (WPT). For the parts (Fig. 1). The first part consists of the information that
integration of a watermark containing the medical image owner's identifies the patient during the extraction process (surname, first

Fig. 1. Watermark generation process.

2
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

Fig. 2. Integration process.

name, date of birth, etc.). The second part contains information Table 1
related to the image acquisition (date, time, place, etc.). This Energy distribution of medical image after 2D DWT.
information will allow identifying the source and the author of the Red componenent Green componenent Blue componenent
image.
Sub-band Entropy Energy Entropy Energy Entropy Energy
In order to guarantee the integrity, the information of the first
two parts are concatenated and a hash (by applying the MD5 LL 2,47 % 99,92 % 2,35 % 99,89 % 2,64 % 99,49 %
LH 32,23 % 0,04 % 30,81 % 0,03 % 31,08 % 0,18 %
algorithm) is performed, this 128-bit hash will constitute the third
HL 32,13 % 0,04 % 33,54 % 0,06 % 32,53 % 0,31 %
part of the watermark. All this concatenated information will be HH 33,18 % 0,00 % 33,30 % 0,02 % 33,75 % 0,03 %
put in binary form for insertion.

3.2. Integration process

energy is very low. The use of the LL sub-band for watermark
To integrate the watermark into the medical image, a DWT is integration therefore seems more interesting and could give better
first applied. Multi-resolution analysis allows the analysis of the results.
image in different frequency bands, to have a view from the finest
to the coarsest. An SVD is then applied to the LL subband, the 3.2.2. Singular value decomposition (SVD)
watermark bits are then integrated into the coefficients of the Singular value decomposition consists in decomposing a matrix
resulting S matrix (Fig. 2). An inverse SVD is then applied to obtain [27] into a product of 3 matrices U, S and V (S is called matrix of
the modified LL subband. Finally an inverse DWT is performed to singular values). Any matrix I of size m  n of rank r can be
obtain the watermarked medical image. decomposed into a pondered sum of unit matrices m  n by
Singular Value Decomposition [28]. The matrices U and V are unit
3.2.1. Discrete wavelet transform (DWT) matrices and I can therefore be written as:
The discrete wavelet transform decomposes an image into four X
n
sub-bands, an approximation sub-band LL and three detail sub- I ¼ USV t ¼ ðs i ui vti Þ ð5Þ
i¼1
bands: LH, HH and HL, corresponding to vertical, diagonal and
horizontal detail, respectively [25]. The H corresponds to the high- Where S is a matrix where the r first diagonal terms are positive, all the
pass filtering and the L to the low-pass filtering applied on the rows others being null. The r non-zero terms s i are called singular values of
and columns. The decomposition of the approximation sub-band I.
LL gives a pyramidal representation. To obtain the coefficients of 8 9
each sub-band, we can calculate with the Haar filter as follows: < si . . . 0 >
> =
S ¼   ... } ... ð6Þ
pðx; yÞ þ pðx; y þ 1Þ þ pðx þ 1; yÞ þ pðx þ 1; y þ 1Þ >
: >
;
LLðx; yÞ ¼ ð1Þ 0 . . . sn
2
In our approach, an SVD is applied to the LL susband, the matrix
of singular S values will then be used for the watermark
pðx; yÞ þ pðx; y þ 1Þ  pðx þ 1; yÞ  pðx þ 1; y þ 1Þ integration.
 LHðx; yÞ ¼ ð2Þ
2
3.2.3. Substitution process
The proposed watermarking scheme in this work is a
pðx; yÞ  pðx; y þ 1Þ þ pðx þ 1; yÞ  pðx þ 1; y þ 1Þ substitutive scheme declined in two variants. The watermark is
HLðx; yÞ ¼ ð3Þ
2 not added but substituted to the components of the image
according to the rules proposed in the two variants. After
decomposing the image into four susbands and applying SVD to
pðx; yÞ  pðx; y þ 1Þ  pðx þ 1; yÞ þ pðx þ 1; y þ 1Þ
HHðx; yÞ ¼ ð4Þ the LL susband, the substitution is performed in the singular values
2
matrix S.
This representation of the image's DWT coefficients allows to The goal of our approach is to match the result of the
distinguish very particular regions around the image contours and subtraction of two successive coefficients to the watermark bit,
textured areas. These regions correspond to areas of high density of in case of equality nothing is changed, otherwise, the parity of a
significant (or dominant) wavelet coefficients [26]. coefficient is modified, this will systematically modify the result of
For watermark integration, the choice of the sub-band to be the subtraction and thus this result will correspond to the
used depends on the type of images and the type of applications. integrated bit. That will make it possible to dissimulate the bit
After decomposing our medical image into four sub-bands, we can of the watermark, but also will facilitate the extraction, the receiver
see in Table 1 that the energy is concentrated and very high in the will just have to apply a subtraction to the successive coefficients to
LL sub-band. In contrast to the LH, HL and HH bands where the obtain the watermark bits. In the first variant, the parity of the

3
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

calculated. Then, according to the watermark bits X and Y, one of

these coefficients will be modified according to the diagram's rules
of the Fig. 5. In order to integrate all the watermark bits, this
operation is repeated by combining each three successive singular
values.
In order to illustrate the proposed substitution process, we have
used the same matrix as for the example in Fig. 4. As we can see in
the Fig. 6, a matrix of 48 coefficients allows to conceal 32 bits
contrary to the first variant (24 bits). In this example, the 32-bit
substitution only modified eight coefficients.

3.3. Extraction process

Fig. 3. Substitution rules for the first variant.
Depending on the extraction process, the watermark is qualified
subtraction of two successive coefficients (S1 and S2) is calculated as non-blind if the original image is used to generate the mark. If
and the result is compared to the bit to be integrated (X). Then, the detection of the hidden information is done directly from the
according to the watermark bits X, one coefficient will be modified watermarked image without knowledge of either the original
according to the diagram's rules of the Fig. 3. image or the mark, the watermark is considered as blind. In our
This operation is repeated to integrate all the watermark bits. approach, the extraction of the mark is performed in blind mode.
This is done using all available singular values by combining every This is the only mode where we can really talk about watermark
two successive values. In case of inequality, the parity of the first extraction since we do not presuppose knowledge of the
coefficient (S1) will be modified to satisfy the equation: watermark or of the original image. This is the most interesting
mode of extraction, but also the most difficult to implement. For
X ¼ ðS1   S2Þ% 2 ð7Þ
the two proposed variants, the extraction will be done by applying
As we can see in the Fig. 4, the 24-bit integration is performed a DWT to the watermarked medical image (Fig. 7), then, a singular
using 48coeficient, in this example the bits are integrated by value decomposition will be applied to the LL susband.
modifying only nine coefficients. The successive coefficients (in the diagonal) of the S singular
In the second variant, three successive coefficients (in the value matrix will then be used to extract the watermark. For the
diagonal of the S matrix) are used to integrate two watermark bits. first variant, the bits of the watermark will correspond to the parity
The parity of three successive coefficients (S1, S2 and S3) is of the subtraction of two successive coefficients (Eq. (7)). As we can

Fig. 4. Substitution example for the first variant.

Fig. 5. Substitution rules for the second variant.

4
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

Fig. 6. Substitution example for the second variant.

Fig. 7. Extraction process.

see in the example of Fig. 8, using Eq. (7) on a matrix of 48 The resulting binary sequence will then be divided into three
coefficients allows the extraction of 24 watermark bits. parts, one part will constitute the patient information, the second
For the second variant, the bits (X and Y) of the watermark will part will contain the information related to the image acquisition.
correspond to the parity of the subtraction of three successive The last part represents the hash of the first two parts. It will then
coefficients S1, S2 and S3: be possible to verify the integrity of the extracted data by
comparing the extracted hash with the hash calculated from the
X ¼ ðS1   S2Þ% 2 ð8Þ
information contained in the first two parts.

4. Experimental results and discussions

Y ¼ ðS1   S3Þ% 2 ð9Þ
Fig. 9 presents an example of the application of equations (Eqs. For our experiments, we used a medical image database ODIR
(8) and (9)) to a matrix of 48 coefficients. As we can see it allows (Ocular Disease Intelligent Recognition) which is a structured
the extraction of 32 watermark bits. ophthalmic database of 5,000 patients collected by Shanggong

Fig. 8. Extraction example for the first variant.

5
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

Fig. 9. Extraction example for the second variant.

Medical Technology Co., Ltd [28]. from different hospitals/medical Robustness indicates the ability to detect the mark after
centers in China (Fig. 10). The experiments were performed on 200 modifications due to some involuntary attacks (treatments that
images from this database.In this experiment, all methods are do not necessarily aim to remove the mark) or voluntary attacks
performed on the same platform of 2.27GHZ CPU, 2.00GB RAM, that have the sole purpose of removing the watermark.
Win 7 and MATLAB 8.5.0 (R2015a). In order to evaluate our
approach and determine the efficiency of the decomposition in
4.1. Imperceptibility and capacity test
singular values, we conducted two experiments. The first is to
apply both variants by integrating the watermark in the least
The watermarking process should allow the patient's informa-
significant bits of the LL subband coefficients (without applying
tion integration while maintaining good image quality. Many
SVD). DWT-V1 and DWT-V2 will correspond to the watermarking
distortions can be generated during the integration process. To
results obtained by applying the first and second variant
evaluate the imperceptibility of the proposed scheme, different
respectively. In the second experiment, the watermark is
distortion measurements are calculated:
integrated into the coefficients of the Singular values matrix S
of the LL sub-band according to the two variants proposed in
 The PSNR allows to evaluate the degradation in dB caused by the
Section 3. DWT-SVD-V1 and DWT-SVD-V2 will correspond to the
integration process to the watermarked image. When PSNR is
watermarking results obtained by the application of the first and
high, the distortion becomes less significant [29]. Generally a
second variants respectively.
watermarking scheme is considered as imperceptible when the
As can be seen in Fig. 11, the two images are visually similar and
PSNR is higher than 36 dB:
the distortions caused by the integration process are visually
undetectable. Whatever the watermark scheme, the developer has
" #
to make a compromise between imperceptibility, capacity and maxI2 ði; jÞ
robustness to attack. Capacity represents the amount of informa- PSNR dB ¼ 10 log10 M    N P 2
ð10Þ
i; j ½ Iði; jÞ  J ði; jÞ
tion embedded in proportion to the amount of information
contained in the digital image. Imperceptibility consists in Where I (i, j) is the value of the luminance of the reference pixel
reducing the visual impact of the watermark so that the water- (i, j) and J (i, j) that of the image to be tested, the two images being
marked image is visually equivalent to the original image. of size [M x N].

Fig. 10. Medical images used for the analysis of the proposed scheme.

6
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

integration process [33]. It is calculated by comparing the values

(R, G and B) of the original images (P) and the watermarked
images (Q) as follows:

M X
X N
NCVC   ¼ ðPr ði; jÞ  Q r ði; jÞÞ þ  ðPg ði; jÞ  Q g ði; jÞÞ
i¼1 j¼1
þ  ðQ b ði; jÞ  Q b ði; jÞÞ ð15Þ

 The average rate of intensity change [34] is analyzed using the

UACI parameter by comparing the original image (CI1 ) and the
watermarked image (CI2 ). The two images being of size [M x N].
Fig. 11. Watermarking results, a: Original image, b: Image watermarked using
DWT-SVD-V2. The UACI is defined by the following formula:

XM X N
CI1 ði; jÞ   CI2 ði; jÞ
 The MSE represents the mean square error between the UACI  ¼ 100% ð16Þ
watermarked and original image [30]. This measure evaluates i¼1 j¼1
FT
the influence of the mark on the image and is defined as follows:
Where F is the largest pixel value supported by the image format

1 X X
MSE  ¼   i ð Iði; jÞ  J ði; jÞÞ2 ð11Þ  The bits per pixel (BPP) allows to measure the payload capacity
MN j [35], it represents the number of bits integrated compared to the
number of pixels of the image.
Where I (i, j) is the value of the luminance of the reference pixel (i,
j) and J (i, j) that of the image to be tested, the two images being of
size [M x N].
Number of secret bits embedded
BPP ¼ ð17Þ
 SSIM (Structural SIMilarity) is a measure of similarity between Total pixels in the cover image
two images [31]. It was developed to measure the visual quality According to the results obtained in Tables 2 and 3, the two
of a compressed image compared with the original image. The variants combining DWT and SVD offer better imperceptibility
idea of SSIM is to measure the structural similarity between the results. The integration of the watermark in the least significant bit
two images, rather than a pixel-to-pixel difference as is done for of the singular values generates less distortion than the integration
example by PSNR. The SSIM metric is calculated over multiple in the DWT image coefficients. The combination of DWT and SVD
windows of an image. The measurement between two x and y during the integration process generates watermarked images
windows of size N x N is defined by the following formula: similar to the original images with SSIM rates of 0.9997 and 0.9998.
However, the number of pixels and values changed by the DWT-
   SVD combination are significantly higher than the integration in
2mx my þ  C 1 2dxy þ  C 2
SSIMðx; yÞ  ¼    ð12Þ the DWT coefficients. Furthermore, the integration in singular
m2x þ  m2y þ C 1 d2x þ  d2y þ C 2 (non-zero) values considerably reduces the integration capacity of
our approach.
2 2
Where: mx - Average of x. my - average of y. dx - variance of x. dy - In order to evaluate the capacity of the proposed approach as
variance of y. dxy - covariance of x y. C1=k1L2-Constant to avoid well as the impact on imperceptibility, we experimented several
instability when m2x þ  m2y is close to zero; k1 = 0.01, = 255. integration rates and calculated the average value of PSNR
2 2
obtained. As we can see in Table 3, the more the capacity of the
C2=k2L2- Constant to avoid instability when dx þ  dy is close to approach increases, the more the PSNR obtained decreases. The
zero; k2 = 0.03, = 255.

 The number of pixels changed ratio (NPCR) indicates the number

Table 2
of pixels modified by the integration process [32], it is calculated Imperceptibility results.
by comparing the pixels of the original image ðCI1 Þ and
watermarked ðCI2 Þ.The two images being of size [M * N]. The PSNR MSE SSIM NPCR UACI NCVC BPP

NPCR is defined by the following formula: DWT-V1 27.34 2.32 0.8989 15.5% 0.00017 65.90% 10.52
DWT-V2 27.19 2.12 0.8919 11.9% 0.00004 75.12% 7.89
DWT-SVD-V1 56.12 0.10 0.9998 4.2% 0.00006 16.72% 0.10
XM X N
Dði; jÞ
DWT-SVD-V2 57.41 0.13 0.9998 4.0% 0.00016 12.81 % 0.13
NPCR  ¼ 100% ð13Þ
i¼1 j¼1
T

Table 3
 Imperceptibility and capacity results for different integration rate.
0;  if  CI1 ði; jÞ ¼ CI2 ði; jÞ
Dði; jÞ ¼   ð14Þ
1;  if  CI1 ði; jÞ 6¼ CI2 ði; jÞ DWT-V1 DWT-V2 DWT-SVD-V1 DWT-SVD-V2

Integration rate PSNR BPP PSNR BPP PSNR BPP PSNR BPP
25 % 64.71 3.15 69.81 0.78 69.42 0.02 69.23 0.03
50% 44.00 5.26 45.44 3.94 55.90 0.04 63.87 0.06
 The number of colorimetric values changed (NCVC) is used to 100 % 27.34 10.5 27.19 7.89 56.12 0.095 57.41 0.13
calculate the number of colorimetric values modified by the

7
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

Table 4
Imperceptibility results comparison with related work.

DWT-V1 DWT-V2 DWT-SVD-V1 DWT-SVD-V2 Mothi & al Anand & al Devi & al
PSNR 27.34 27.19 56.12 57.41 49.62 44.19 54.64

Table 5
Robustness comparison with related work.

Mothi & al Devi & al Anand & al DWT-V1 DWT-V2 DWT-SVD-V1 DWT-SVD-V2
Salt & Pepper Noise 0.9671 / 0.9251 0.9296 0.9250 0.9882 0.9782
Gaussian noise 0.9819 0.9953 0.9803 0.9741 0.9681 0.9622 0.9707
Rotation 0.8713 0.85933 0.8817 0.8215 0.8391 0.8709 0.8611
JPEG compression 0.9873 0.9965 0.9388 0.9464 0.9570 0.9204 0.9238
Speckle noise 0.8191 / 0.9687 0.8707 0.8529 0.9012 0.9197
Cropping 0.9473 0.8141 0.3586 0.5613 0.6925 0.8101 0.7913
Average filtering 0.9603 0.7690 0.9860 0.9708 0.9661 0.9613 0.9595
Histogram equalization 0.9688 0.8764 0.7223 0.8191 0.8034 0.9877 0.9873
Resizing / 0.9057 / 0.8992 0.8879 0.8976 0.9103
Gamma correction 0.9347 1 / 0.9664 0.9711 0.9688 0.9881
Sharpening 0.9732 0.9253 0.6506 0.8479 0.7632 0.9690 0.9736
Scaling / / 0.7157 0.8215 0.8937 0.8982 0.9519

choice of the amount of information to be integrated will allow for robustness tests are satisfactory. The proposed watermarking
a trade-off between capacity and imperceptibility. The images used scheme generates a watermarked image that is robust to geometric
in our experiments are of size 512 * 512, the application of a DWT to and destructive attacks. This is due to the small amount of
this image will generate 82636 L L coefficient for each color plane information required for the watermark integration. In the second
(R, G and B). The first variant of our integration process requires proposed variant, only one bit (of only one coefficient out of three)
two coefficients for the integration of two bits of the watermark, so is probably modified for the integration of two watermark bits.
this variant allows the integration of 123954 watermark bits which Based on the obtained results, we found that in terms of
corresponds approximately to 15494 characters. The second imperceptibility, all the four schemes offer the same PSNR rates
variant of our approach allows the integration of two watermark for the concealment of a small amount of information. However,
bits by combining three coefficients. This makes it possible to hide for the concealment of a large number of bits, the scheme
165272 bits (approximately 20659 characters). The combination of combining DWT and SVD is more efficient. Robustness tests
a DWT and an SVD considerably reduces the capacity of our concluded that the two variants combining DWT and SVD are
approach, a singular value decomposition of the LL sub-band almost similar. We can conclude from the capacity tests that the
coefficients allows to generate 3820 embeddable coefficients. The second variant combining DWT and SVD is significantly better.
first variant of the integration process will allow to integrate 1910 Indeed, the first variant of the integration process will allow
bits of the watermark which corresponds to approximately 238 integrating 1910 bits of the watermark which corresponds to
characters. The second variant of the integration process will allow approximately 238 characters. The second variant of the integra-
to conceal 2546 bits (approximately 318 characters). tion process will allow concealing 2546 bits (approximately 318
In order to evaluate our approach, we compared the imper- characters).
ceptibility results of our approach with related works. As we can
see in Table 4, the PSNR rates obtained by the two variants 5. Conclusion
combining a DWT and an SVD are higher than those obtained in
recent related works that make the imperceptibility of our scheme Medical information protection ensures the security of medical
acceptable. Indeed, in the proposed integration scheme, the parity information records which are, for a patient, a complex set of
of the coefficients is combined for the watermark integration clinical examinations, diagnostic annotations and other findings
which considerably reduces the modifications made to the image. and images, contained in the Electronic Patient Records (EPR). The
importance of the security of medical images is due to the serious
4.2. Robustness test consequences that can be caused by malicious alterations or
attacks, such as misdiagnosis, invalidity of computerized data in
We evaluated the robustness of the proposed system against forensic medicine, possibility of misleading medical insurance
some attacks on the watermarked image such as noise addition, agencies, etc. In this work, we proposed an approach to ensure the
and JPEG compression, for each attack the different correlations are reliability of retinal images. The watermark composed of Patient
calculated. The correlation variation for each attack is given in the information allows to authenticate the patient as well as the data
Table 5. related to the image acquisition. In addition, the hash embedded in
As we can see in Table 5, both approaches have the same the watermark guarantees that the data has not been altered by
resistance to resizing and scaling attacks, but, the approach using unauthorized persons. After applying a DWT to the medical image,
only DWT is more resistant to average filtering and JPEG a SVD is performed in the LL sub-band. The bits of the watermark
compression attacks. We can also conclude from the obtained are then integrated into the least significant bits of the single
results that the approach combining a DWT and SVD offers better values obtained by combining the parity of the successive
resistance to rotation and Salt & Pepper noise attacks. The obtained coefficients. The wavelet transform allows overcoming some
results of this approach are also significantly better for Histogram weaknesses of the Fourier transform, essentially the space-
equalization and Sharpening attacks. From the normalized frequency localization. The representations at different resolutions
correlation values obtained, it can be concluded that the make possible to extract the main trends of the image in a small

8
N. Zermi, A. Khaldi, R. Kafi et al. Forensic Science International 320 (2021) 110691

number of coefficients, while precisely locating the discontinuities. [8] D. Degadwala, M. Kulkarni, D. Vyas, A. Mahajan, Novel image watermarking
This allows enhancement of the most significant areas of the approach against noise and RST attacks, Procedia Comput. Sci. 167 (2020)
213–223.
image. Experimental results of imperceptibility and robustness [9] S. Thakur, A.K. Singh, S.P. Ghrera, A. Mohan, Chaotic based secure water-
show that the proposed scheme maintains a high quality of marking approach for medical images, Multimed. Tools Appl. 79 (2020) 4263–
watermarked images and is very robust against several conven- 4276.
[10] A. Anand, A.K. Singh, Z. Lv, G. Bhatnagar, Compression-then-Encryption based
tional attacks. Experiments performed on the test images allowed secure watermarking technique for smart healthcare system, IEEE Multimed.
to evaluate the performance of this method and to evaluate its (2020), doi:http://dx.doi.org/10.1109/MMUL.2020.2993269 pp. 1–1.
robustness against JPEG compression and the addition of noise. [11] R.M. Hadi, P. Ayubi, Blind digital image watermarking based on CT-SVD and
chaotic cellular automata, 2012 2nd International Econference on Computer
However, instead of using the Singular value decomposition, we and Knowledge Engineering (ICCKE) (2012) 301–306, doi:http://dx.doi.org/
could in future works use the Schur decomposition. The successful 10.1109/ICCKE.2012.6395396 October.
application of singular values decomposition in digital water- [12] A.K. Singh, Data hiding: current trends, innovation and potential challenges,
ACM Trans. Multimedia Comput. Commun. Appl. (2020), doi:http://dx.doi.org/
marking suggests the use of Schur decomposition for the same
10.1145/3382772.
purpose. Schur decomposition represents an alternative to [13] M. Yousefi Valandar, M. Jafari Barani, P. Ayubi, A blind and robust color images
the singular values decomposition. Indeed, singular value decom- watermarking method based on block transform and secured by modified 3-
position requires about 11N3 flops as compared to a Schur dimensional Hénon map, Soft Comput. 24 (2020) 771–794.
[14] M. Jafari Barani, M. Yousefi Valandar, P. Ayubi, A new digital image tamper
decomposition which requires only 8N3/3. This represents about detection algorithm based on integer wavelet transform and secured by
three times less computing time. This might be interesting to encrypted authentication sequence with 3D quantum map, Optik 187 (2019)
reduce the computation time. DWT is a typical multi-scale 205–222.
[15] O.P. Singh, A.K. Singh, G. Srivastava, N. Kumar, Image watermarking using soft
transform (MST) method that does not have the translation- computing techniques: a comprehensive survey, Multimed. Tools Appl. (2020),
invariance property. In addition, DWT decompose the image into doi:http://dx.doi.org/10.1007/s11042-020-09606-x.
one low-frequency sub-band and three high-frequency sub-bands [16] M.Y. Valandar, M.J. Barani, P. Ayubi, M. Aghazadeh, An integer wavelet
transform image steganography method based on 3D sine chaotic map,
(horizontal, vertical, and diagonal). Only three decomposition Multimed. Tools Appl. 78 (2019) 9971–9989.
directions are not enough to represent textures of different scales. [17] A.K. Singh, Robust and distortion control dual watermarking in LWT domain
It might be interesting in future work to use non-subsampled MST using DCT and error correction code for color medical image, Multimed. Tools
Appl. 78 (2019) 30523–30533, doi:http://dx.doi.org/10.1007/s11042-018-
methods, such as non-subsampled contourlet transform (NSCT) or
7115-x.
non-subsampled shearlet transform (NSST). In order to improve [18] R. Mothi, M. Karthikeyan, Protection of bio medical iris image using
the security of the proposed scheme, a watermark encryption watermarking and cryptography with WPT, Measurement 136 (March)
(2019) 67–73.
before integration could be performed by a symmetrical encryp-
[19] A. Anand, A. Kumar Singh, An improved DWT-SVD domain watermarking for
tion algorithm using a pre-shared key. medical information security, Comput. Commun. 152 (February) (2020) 72–80.
[20] H. Devi, K. Singh, Red-cyan anaglyph image watermarking using DWT,
Ethical approval Hadamard transform and singular value decomposition for copyright
protection, J. Inf. Secur. Appl. 50 (February) (2020).
[21] P. Khare, V.K. Srivastava, A reliable and secure image watermarking algorithm
This article does not contain any studies with human using homomorphic transform in DWT domain, Multidimens. Syst. Signal
participants or animals performed by any of the authors Process. (June) (2020), doi:http://dx.doi.org/10.1007/s11045-020-00732-1.
[22] X. Zhou, H. Zhang, C. Wang, A robust image watermarking technique based on
DWT, APDCBT, and SVD, Symmetry 10 (2018), doi:http://dx.doi.org/10.3390/
CRediT authorship contribution statement sym10030077.
[23] A. Benoraira, K. Benmahammed, N. Boucenna, Blind image watermarking
technique based on differential embedding in DWT and DCT domains,
Zermi Narima: Conceptualization, Methodology, Formal anal- EURASIP J. Adv. Signal Process. 1 (2015), doi:http://dx.doi.org/10.1186/s13634-
ysis, Investigation, Software, Writing - original draft. Khaldi 015-0239-5.
Amine: Supervision, Resources, Validation, Writing - original draft. [24] S. Bravo-Solorio, F. Calderon, C. Li, K. Nandi, Fast fragile watermark embedding
and iterative mechanism with high self-restoration performance, Digit. Signal
Kafi Redouane: Methodology, Writing - original draft. Kahlesse- Process. 73 (February) (2018) 83–92.
nane Fares: Investigation, Software, Writing - original draft. [25] M.E. Moghaddam, N. Nemati, A robust color image watermarking technique
Euschi Salah: Methodology, Software, Validation, Writing - using modified Imperialist Competitive Algorithm, Forensic Sci. Int. 233 (2013)
193–200, doi:http://dx.doi.org/10.1016/j.forsciint.2013.09.005.
original draft.
[26] I. Assini, A. Badri, K. Safi, A. Sahel, A. Baghdad, A robust hybrid watermarking
technique for securing medical image, Int. J. Intell. Eng. Syst. 11 (June) (2018)
Declaration of Competing Interest 169–176.
[27] N. Falgun, V.K. Srivastava, A blind medical image watermarking: DWT-SVD
based robust and secure approach for telemedicine applications, Multimed.
The authors declare that they have no known competing Tools Appl. 76 (2017) 3669–3697.
financial interests or personal relationships that could have [28] H. Hu, L. Hsu, H. Chou, An improved SVD-based blind color image
appeared to influence the work reported in this paper. watermarking algorithm with mixed modulation incorporated, Inf. Sci. 519
(May) (2020) 161–182.
[29] Ocular Disease Recognition, Right and Left Eye Fundus Photographs of 5000
References Patients, (2021) . https://odir2019.grand-challenge.org/.
[30] Fares Kahlessenane, Amine Khaldi, Redouane Kafi, Salah Euschi, A DWT
[1] K. Swaraja, K. Meenakshi, P. Kora, An optimized blind dual medical image Based Watermarking Approach for Medical Image Protection, J. Ambient
watermarking framework for tamper localization and content authentication Intell. Humaniz. Comput. (2020), doi:http://dx.doi.org/10.1007/s12652-020-
in secured telemedicine, Biomed. Signal Process. Control 55 (January) (2020). 02450-9.
[2] A.K. Singh, C. Kumar, Encryption-then-Compression-Based copyright protec- [31] Q. Hu, M.D. Abràmoff, M.K. Garvin, Automated separation of binary
tion scheme for E-Governance, IT Prof. 22 (2020) 45–52. overlapping trees in low-contrast color retinal images, Med. Image Comput.
[3] P. Ayubi, M. Jafari Barani, M. Yousefi Valandar, B. Yosefnezhad Irani, R. Comput. Assist. Interv. 16 (2013) 436–443.
Sedagheh Maskan Sadigh, A new chaotic complex map for robust video [32] F. Kahlessenane, A. Khaldi, S. Euschi, A robust blind color image watermarking
watermarking, Artif. Intell. Rev. (2020), doi:http://dx.doi.org/10.1007/s10462- based on Fourier transform domain, Optik 208 (April) (2020).
020-09877-8. [33] R. Dubolia, R. Singh, S. Bhadoria, R. Gupta, Digital image watermarking by
[4] E. Farri, P. Ayubi, A blind and robust video watermarking based on IWT and using discrete wavelet transform and discrete cosine transform and
new 3D generalized chaotic sine map, Nonlinear Dyn. 93 (2018) 1875–1897. comparison based on PSNR, International Conference on Communication
[5] C. Kumar, A.K. Singh, P. Kumar, Dual watermarking: an approach for securing Systems and Network Technologies (2011) 593–596.
digital documents, Multimed. Tools Appl. 79 (2020) 7339–7354. [34] Euschi Salah, Khaldi Amine, Kafi Redouane, Kahlessenane Fares, A Fourier
[6] A. Anand, A.K. Singh, Joint Watermarking-Encryption-ECC for patient record Transform Based Audio Watermarking Algorithm, Appl. Acoust. 172 (2021)
security in wavelet domain, IEEE Multimed. 27 (2020) 66–75. 107652, doi:http://dx.doi.org/10.1016/j.apacoust.2020.107652.
[7] A. Khaldi, Diffie-Hellman Key Exchange through Steganographied Images, [35] S. Ahmadi, G. Zhang, S. Wei, Robust and hybrid SVD-based image water-
Revista de Direito Estado e Telecomunicações 10 (May) (2018) 147–160. marking schemes: a survey, Multimed. Tools Appl. 79 (2020) 1075–1117.