REVERSIBLE DATA HIDING METHODS: A COMPARITIVE STUDY ON

TECHNIQUES AND PERFORMANCE

A PROJECT REPORT

SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE

AWARD OF THE DEGREE

OF

BACHELOR OF TECHNOLOGY
IN
COMPUTER ENGINEERING

Submitted by:

Ritik Jain (2K20/CO/365)

Sachin Mishra (2K20/CO/383)

Saurav Sharma (2K20/CO/417)

Under the supervision of

Dr. Rajeev Kumar

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

DELHI TECHNOLOGICAL UNIVERSITY

(Formerly Delhi College of Engineering)
Bawana Road, Delhi-110042
MAY, 2024
 ii

DELHI TECHNOLOGICAL UNIVERSITY

(Formerly Delhi College of Engineering)
Bawana Road, Delhi-110042

CANDIDATE’S DECLARATION

We, Ritik Jain (2K20/CO/365), Sachin Mishra (2K20/CO/383), Saurav Sharma

(2K20/CO/417), students of [Link]. Computer Engineering, hereby declare that the project

Dissertation titled “REVERSIBLE DATA HIDING METHODS: A COMPARITIVE

STUDY ON TECHNIQUES AND PERFORMANCE” which is submitted by us to the

Department of Computer Science And Engineering, Delhi Technological University, Delhi

in partial fulfillment of the requirement for the award of the degree of Bachelor of

Technology, is original and not copied from any source without proper citation. This work

has not previously formed the basis for the award of any Degree, Diploma Associateship,

Fellowship, or other similar title or recognition.

Place: New Delhi

Date:
Ritik Jain
(2K20/CO/365)

Sachin Mishra
(2K20/CO/383)

Saurav Sharma
(2K20/CO/417)
 iii

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

DELHI TECHNOLOGICAL UNIVERSITY
(Formerly Delhi College of Engineering)
Bawana Road, Delhi-110042

CERTIFICATE

I hereby certify that the Project Dissertation titled “REVERSIBLE DATA HIDING

METHODS: A COMPARITIVE STUDY ON TECHNIQUES AND

PERFORMANCE” which is submitted by Ritik Jain (2K20/CO/365), Sachin Mishra

(2K20/CO/383), Saurav Sharma (2K20/CO/417), students of [Link]. (Computer

Engineering), Delhi Technological University, Delhi in partial fulfillment of the requirement

for the award of the degree of Bachelor of Technology, is a record of the project work carried

out by the students under my supervision. To the best of my knowledge this work has not

been submitted in part or full for any Degree of Diploma to this University or elsewhere.

Place: New Delhi

Date:
Dr. Rajeev Kumar
(Supervisor)
Department of Computer Science and Engineering
DELHI TECHNOLOGICAL UNIVERSITY
 iv

DELHI TECHNOLOGICAL UNIVERSITY

(Formerly Delhi College of Engineering)
Bawana Road, Delhi-110042

ACKNOWLEDGEMENT

A special thanks to Dr. Rajeev Kumar for guiding us through the course curriculum of

”[Link] Project - II (CO402)”. He mentored and motivated us through the entirety of this

innovative project titled “REVERSIBLE DATA HIDING METHODS: A

COMPARITIVE STUDY ON TECHNIQUES AND PERFORMANCE”. We would

also like to extend our sincerest gratitude towards him for giving us his valuable suggestions

and recommendations for the betterment and enhancement of our work, for which we are

forever grateful.

Place: New Delhi

Date:
Ritik Jain
(2K20/CO/365)

Sachin Mishra
(2K20/CO/383)

Saurav Sharma
(2K20/CO/417)
 v

ABSTRACT

Reversible data hiding (RDH), sometimes referred to as lossless or invertible data hiding,
has become a burgeoning subject of study in data hiding throughout the past 20 years. This
trend is highlighted by the growing number of papers on various RDH issues. However,
researchers, particularly those who are new to the topic, may have difficulties due to the
complexity of these breakthroughs. We hope to give readers a basic grasp of RDH by giving
a brief synopsis of the main RDH algorithms used in image processing and emphasising
their distinctive features. We begin by investigating RDH approaches in a number of areas,
including compressed (like JPEG), contrast enhancement, encrypted images (which are
relevant to cloud computing), and expansion to video and audio data. Subsequently, we delve
into commonly used RDH frameworks and their modifications, including traditional
methods like lossless compression, histogram shifting, difference expansion, prediction-
error expansion (PEE), and integer transform (IT). Additionally, we review contemporary
methods such as pairwise PEE, pixel-value-ordering, and multiple histograms modification,
providing detailed assessments of each technique. Our evaluation covers advancements in
embedding frameworks, technological features, extensions, and the current state-of-the-art.
Finally, we discuss potential future research directions rooted in the initial motivations of
the field.
 vi

CONTENTS

CANDIDATE’S DECLARATION ii
CERTIFICATE iii
ACKNOWLEDGEMENT iv
ABSTRACT v
LIST OF FIGURES viii
LIST OF TABLES x
CHAPTER – 1 INTRODUCTION 1
1.1 Reversible Data Hiding 1
1.2 Applications 4
CHAPTER – 2 REVERSIBLE DATA HIDING WITHIN JPEG IMAGES 6
2.1 RDH with Modification of Quantized DCT Coefficients 7
2.2 DE-based RDH Technique 7
2.3 RDH with Modification of the Huffman Table 8
CHAPTER – 3 ROBUST REVERSIBLE DATA HIDING 9
3.1 Prominent techniques for Robust RDH 9
CHAPTER – 4 RDH WITH CONTRAST ENHANCEMENT 11
4.1 Motivation 11
4.2 Four RDH Methods with Contrast Enhancement 12
4.2.1 Method by Histogram Bin Expansion 12
4.2.2 Method with Contrast Enhancement for Medical Images 13
4.2.3 Method with the controlled Contrast Enhancement 14
4.2.4 Automatic Contrast Enhancement method 14
4.3 Image Quality Assessment in RDH with Contrast Enhancement 14
4.3.1 Using SSIM for Image Quality Assessment 15
4.3.2 Using the Methods for Contrast changed images 15
4.3.3 Using the No-Reference Methods 15
CHAPTER – 5 REVERSIBLE DATA HIDING IN ENCRYPTED DOMAIN 16
5.1 Motivation 16
5.2 Vacating Room Before Encryption (VRBE) 17
 vii

5.3 Vacating Room After Encryption (VRAE) 18

5.3.1 VRAE: Data Extraction in the Plaintext Domain 19
5.3.2 VRAE: Data Extraction in the Cipher Domain 20
5.3.3 VRAE: Data Extraction in Both Domains 22
CHAPTER – 6 REVERSIBLE DATA HIDING IN AUDIO AND VIDEO 24
6.1 Video Reversible Data Hiding 24
6.2 Audio Reversible Data Hiding 24
CHAPTER – 7 REVERSIBLE DATA HIDING FOR UNCOMPRESSED IMAGES 26
7.1 Compression-based RDH Technique 27
7.2 DE-based RDH Technique 27
7.2.1 Illustration of Difference Expansion 28
7.3 IT-based RDH Technique 29
7.4 HS-based RDH Technique 29
7.5 PEE-based RDH Technique 31
7.6 PEE-based RDH Technique Extensions 34
CHAPTER – 8 RDH TECHNIQUES FOR ADAPTIVE EMBEDDING 37
8.1 Multiple Histograms Modification 38
8.2 Pairwise PEE 42
8.3 Pixel value ordering 47
8.3.1 IPVO 48
8.3.2 PVO-k 49
8.3.3 k-pass PVO 49
8.3.4 PPVO 50
8.3.5 Pairwise PVO 51
8.3.6 MHM-based PVO 52
8.4 Discussion 52
CHAPTER – 9 THE STATES OF THE ART 55
CHAPTER – 10 FUTURE WORK 60
CHAPTER – 11 SUMMARY 64
CHAPTER – 12 REFERENCES 65
 viii

LIST OF FIGURES

Figure 1.1: The function and characteristics of RDH 1

Figure 2.1: Embedded JPEG compression methods for 3 different data types 6

Figure 2.2: JPEG decompression methods for 3 different data types 6

Figure 3.1: Different cases of distribution of Block histogram 9

Figure 4.1: Poor illumination image and the image achieved by contrast enhancement 11

Figure 5.1: General framework of Reversible Data Hiding in Encrypted domain 16

Figure 5.2: Vacating Room Before Encryption (VRBE) 17

Figure 5.3: Vacating Room After Encryption (VRAE) 18

Figure 5.4: Extraction and recovery of image in domain of plain text 19

Figure 5.5: Extraction of data in the domain of cipher 20

Figure 5.6: Extraction of data in both the domains 22

Figure 7.1: Flowchart explaining the embedding process in histogram shifting in HS-based

RDH 30

Figure 7.2: Example of the modification of the pixels and shifting of histograms in HS

based RDH Technique 30

Figure 7.3: The comparison of pixel histogram and Prediction-error Histogram for Lena

image 32

Figure 7.4: Example of modification of pixels and the histogram-shifting in the PEE RDH

technique 33

Figure 7.5: Embedding modules of PEE, and the used enhancements 33

Figure 7.6: An example of the asymmetric PEH’s histogram adjustment 36

Figure 8.1: The data embedding example of MHM 39

 ix

Figure 8.2: Two-dimensional mapping comparison, with the standard PEE 43

Figure 8.3: The alteration of pixels using pairwise PEE 44

Figure 8.4: The 2D mapping generated using the adaptive approach for the Lena and

Airplane test image with an EC value of 10,000 bits 44

Figure 8.5: Diagram depicting the process of embedding and extracting data in a PVO-

based PEE system 47

Figure 8.6: Comparing the alteration of the PEH for PVO and IPVO during the embedding

process 48

Figure 8.7: Embedding comparison of PVO-k and k-pass PVO 50

Figure 8.8: Two-dimensional mapping comparison 51

Figure 10.1: Future work: High dimensional mapping 60

Figure 10.2: RDH for the 3D mesh model 60

Figure 10.3: RDH for DNN 61

Figure 10.4: Example for reversible adversarial 61

 x

LIST OF TABLES

Table 1.1: Overview of the Research Topic of RDH and their maximal citations 3

Table 7.1: Summary of the RDH Techniques 26

Table 7.2: Differentiation for components of Existing Predictors of PEE 34

Table 8.1: An examination of the existing methodologies for adaptive embedding 37

Table 8.2: Overview of the current MHM-based Methods 40

Table 8.3: The Conventional PEE’s Nine Potential PE Pair Mappings 42

Table 8.4: Summary of the Adaptive Optimization Algorithms for the Pairwise PEE 45

Table 8.5: Comparitive analysis of prediction accuracy between commonly utilized

predictors and PVO 52

Table 8.6: Results for the Proportion of the Better Predictions achieved with the use of

PVO 54

Table 9.1: Summary of the Datasets 55

Table 9.2: Summary of Evaluation Criteria in RDH 56

Table 9.3: Comparison between the current representative methods of RDH 56

 1

CHAPTER-1 INTRODUCTION

1.1 Reversible Data Hiding (RDH)

RDH, also known as Reversible Data Hiding, is a well respected method employed to protect
multimedia data [91]. It provides multiple benefits, as depicted in Figure 1.1. RDH facilitates
covert communication across a shared channel. Embedding has a negligible effect on the
quality, but it presents difficulties for adversaries to identify the marked carrier within large
data sets. The receiver end of RDH performs two functions: it retrieves the data embedded
and then precisely restores the original content or information of the carrier. RDH
demonstrates exceptional proficiency in retrieving the carrier without any errors. RDH is a
highly promising method or technique that can be employed in situations where maintaining
data integrity is of utmost importance, such as in engineering images, business documents,
and legal evidence. RDH is particularly advantageous as it can prevent irreversible
deformation. Presently, RDH processes commonly employ digital images as a means of
research because they are simple to use and can be easily transmitted. The methods related to
storage formats can be classified into three groups: RDH in encrypted images [6, 19, 37, 59,
76, 79, 81, 83, 87, 111, 117, 125, 127–129, 136, 137], RDH in uncompressed images [1, 2, 9,
14, 16, 30, 38, 42, 45, 47, 48, 51, 57, 62, 72, 86, 90, 94, 95, 101, 135, 138], and RDH in
compressed images [13, 20, 25, 31, 32, 80, 85, 107, 113].

Fig 1.1: The function and characteristics of RDH.

Among the three categories, uncompressed photographs have received the most significant
focus from the academic community regarding RDH techniques. This is mostly due to the
presence of excessive redundancy in the information representation of uncompressed
pictures, which facilitates the creation of efficient reversible embedding systems. Many
approaches undergo initial testing on uncompressed images before being extended to various
carrier formats. Table 1.1 is a comprehensive summary that includes reference information,
proportions, and concise explanations. The citations are obtained from Google Scholar
searches. Regarding RDH in uncompressed images, the earlier algorithms primarily
 2

emphasized the development of efficient methods for changing pixels in a way that can be
undone. The subsequent advancements in the field have been profoundly influenced by these
fundamental algorithms. There are essentially five fundamental embedding frameworks.
RDH is performed using lossless compression with the following parameters: [2, 16, 135,
138]. Difference expansion (DE) is applied with the parameters [30, 38, 95]. Integer
transform (IT) is utilized with the parameters [1, 9, 72, 86, 101]. Histogram shifting (HS) is
performed with the parameters [42, 45, 51, 62]. Prediction-error expansion (PEE) is applied
with the parameters [14, 47, 48, 90, 94]. These systems share a fundamental notion, albeit
having distinct technological characteristics: they utilize image redundancy to achieve
reversible data embedding. Scholars have conducted additional research and development on
embedding theories that are built upon standard RDH algorithms, with the aim of enhancing
performance.

The primary criteria for assessing the efficacy of a reversible data hiding (RDH) system are
generally the Peak Signal-to-Noise Ratio(PSNR) and the embedding capacity(EC). Method’s
carrying capacity, often measured in bits per pixel(bpp), is indicated by the EC, which is
often evaluated using the embedding rate (ER). A higher error rate (ER) implies that the
cover picture has the capacity to accommodate a greater number of hidden message bits,
potentially alleviating limitations on database storage. PSNR signifies level of distortion
caused due to the embedding by assessing the similarity of each pixel between an annotated
image and an original image. A higher PSNR indicates better preserved quality and implies a
less perceptible hidden message. To achieve better optimal balance between embedding
capacity and PSNR, RDH algorithms must adapt their data embedding strategies according to
the content of the images. In response to this need, several adaptive RDH methods have been
created, including pairwise Prediction Error Expansion(PEE) [68], Pixel Value
Ordering(PVO) [46], and multiple histograms modification (MHM) [49]. These methods
have been developed with distinct objectives: enhancing embedding theory and addressing
adaptive embedding challenges. However, in practical implementation, these systems often
require customization to suit specific requirements and uses. This customization enhances the
algorithm's practical utility and ability to be applied to many scenarios. Several efficient
algorithms have been published in various fields, including RDH in data encryption, RDH
for images of JPEG format, RDH in two-dimensional vector graphics, RDH for
three-dimensional mesh models, RDH with the Contrast Enhancement, RDH for colored
images, robust RDH, visible reversible watermarking, RDH for videos, and RDH for
human-visual system (HVS). Table 1.1 contains detailed information regarding these plans.
These studies have explored pertinent concepts of RDH and its extensions to tackle
pragmatic concerns.
 3

Research Topic Maximum Citations Description

Lossless compression 1463 Initial RDH technique for

uncompressed pictures.

DE 3637 It is the basic method of the RDH in

the uncompressed pictures.

IT 3369 The extension of Difference

Expansion(DE).

HS 1703 A RDH technique designed for

uncompressed photos that is very
effective. It is also widely applicable
in other media, such as JPEG pictures
and videos.

PEE 1594 The widely used method derived from

DE and HS.

MHM 342 The standard technique for adaptive

embedding.

Pairwise PEE 494 An improved technique for changing

pixels based on PEE.

PVO-based RDH 485 1. An effective extension of PEE.

2. One of the hottest RDH
technique.

RDH in encrypted 1066 1. A useful framework that takes

images into account the cloud storage and
data sharing environment.
2. To protect plaintext, conceal
the secret message inside the
encrypted pictures.

RDH in JPEG images 413 1. Conduct embedding in

transformed coefficients.
2. The preservation of file size is
an objective.

RDH in 2D vector 68 The various aspects of 2D vector

graphics graphics, such as the floating-point
values.

RDH in 3D mesh 99 The many mesh topologies and vertex

models structures found in 3D mesh models.

RDH with the 179 After embedding, improve the carrier

contrast-enhancement image's contrast.
 4

Research Topic Maximum Citations Description

RDH in color images 97 1. Embedding of many channels.

2. An additional goal is grayscale
invariance.

Robust RDH 280 After the carrier is assaulted, be sure

to carry out the proper extraction,
such as JPEG compression, cropping,
rotation, etc.

Visible reversible 149 The hidden information is no longer

watermarking deceptive, meaning it can now be seen
from the marked carrier in order to
meet the required goals.

RDH in videos 94 1. Incorporate the message

within the video's frames.
2. Accomplish additional duties,
such as hiding intra-frame errors.

RDH considering the 124 Instead of just adjusting the PSNR,

HVS lessen the quality distortion that is
visible to the human eye.
Table 1.1: Overview of the Research Topic of RDH and their maximal citations

1.2 Applications

1. Buyer and vendor Protocol: To prevent unauthorized access, a seller of the content
encrypts content before transmitting it. The seller guarantees the uniqueness of each piece of
content they sell by watermarking it and integrating the buyer's information (fingerprinting)
to deter unauthorized distribution and enable monitoring and tracing. Nevertheless, by
incorporating their thumbprint into the content and distributing it without charge, there is a
possibility of falsely incriminating a genuine consumer. In the case of buyer-seller protocol,
it is recommended that content seller incorporates an unidentifiable watermark of some sort
into the final form to prevent this from happening.

2. Cloud Storage: In order to safeguard their privacy, customers employ encryption to secure
their private photographs, videos, and audio files before transferring them to the cloud. Cloud
administrators incorporate encrypted data into various materials for administrative needs,
including the data which is required for collection of statistics and indexing.

3. Confidentiality of the information of a patient: Images and documents of diagnosis that

contain an information of the patient are encoded for identification and encrypted to ensure
privacy. Nurses can extract this embedded information, which facilitates content linking.
Physicians and other medical professionals have the ability to access original unaltered
diagnostic images and documents, as well as patient identity information.
 5

4. Confidential Data: In military environments, subordinate personnel such as clerks and

lower-ranking officers are able to retrieve hidden tags from encrypted documents, audio files,
and video files for the purpose of organizing them. Personnel with higher level clearance,
such as generals, are able to read labels which are embedded, and original files.

5. Encryption of audio, picture, or video content for field reporters before delivering it to
headquarters. This ensures that only authorized personnel, excluding competitors, may access
it for exclusive coverage. In order to address the issue of content counterfeiting, it is possible
to incorporate the GPS position and sender ID of the field reporter for the purpose of
verification.

6. Surveillance video is recorded and to ensure privacy, certain techniques such as

region-of-interest masking are used to encrypt the film selectively. Moreover, crucial
information for identification, such as camera ID, time, and date, is included, particularly
when the footage is used as a proof in a court.

7. Labelling data is essential for ensuring privacy and confidentiality in the current era of
extensive data development and collection, often driven by advancements in data-intensive
scientific discoveries. The joint adaptation methodology is a practical method for effectively
managing such data.
 6

CHAPTER-2 REVERSIBLE DATA HIDING WITHIN JPEG IMAGES

JPEG, or Joint Photographic Experts Group standard, is the preferred image format for digital
cameras and other photographic equipment because it provides the best possible balance
between visual fidelity and compression ratio [101]. As such, JPEG images are a perfect
format for hiding items using the Reversible Data Hiding (RDH) method. Nonetheless, more
attention has been paid to RDH implementation in uncompressed photos than in compressed
images. This disparity partially results from JPEG compression's propensity to reduce image
size by removing specific high-frequency elements, while RDH tries to preserve the quality
of the original image.

Fig 2.1: Embedded JPEG compression methods for 3 different data types

Fig 2.2: JPEG decompression methods for 3 different data types

For JPEG images, lot of research have been done and several RDH methods have been
published in the literature; But compared to RDH methods for uncompressed photos, their
number is much less. Three main strategies have been investigated for RDH of JPEG images:
adjusting quantized DCT coefficients, modifying quantization tables, and modifying Huffman
codes. Each of these strategies has its own characteristics and achieves variable performance.
The separate operations of the three approaches during image compression and
decompression are illustrated in Figures 2.1 and 2.2.
 7

2.1 RDH WITH MODIFICATION OF QUANTIZED DCT COEFFICIENTS

We consider quantized DCT coefficients in the first class of RDH for JPEG images.
According to [102], Friedrich et al. were the first to propose the idea of losslessly
compressing the least significant bit (LSB) level of particular DCT coefficients inside a JPEG
image in order to make room for RDH. Furthermore, Xuan et al. presented a lossless data
hiding technique using histogram pairs for JPEG images, which is described in [103]. The
quantized DCT coefficient histogram is divided into three parts using this technique: (a) a
segment that can be used to embed data, (b) an unaltered segment where the absolute values
of the coefficients are less than a predefined threshold, and (c) a shifted segment where the
absolute values of the coefficients are greater than the threshold. An explanation of the
extraction and embedding procedures is given below,

In this case, the selected threshold (T), stop point (S) represents the quantized DCT
coefficient values ​are shown by (x) before the data embedding and (x0) after, (u) is a
single-phase function ( where (u(S) = 1 ) if (S geq 0 ) and (u(S) = 0 ) if (S < 0 ) ) and (b in {0
, 1} ) is the message bit. In addition, this method suggests minimizing image distortions
through an ideal search strategy. We select frequencies that are low and medium to make sure
that the results are undistributed for data embedding.

2.2 QUANTIZATION OF RDH TABLE ADJUSTMENTS

Another method involves making changes to the JPEG quantization tables. The idea of
​modifying the quantization table to add a lossless bit with the DCT coefficient was first
proposed by Fridrich et al. in [109]. In particular, the quantization coefficient Q(i, j) is
divided by two if it is even, and the equivalent coefficient D(i, j) is multiplied by two
(designated D0(i, j)) without modifying the image value. Visual appearance. Since D0(i, j) is
even, it is possible to embed a bit in the LSB of D0(i, j), ensuring that data embedding is fully
reversible. An example of the idea is:
 8

where the message bit is represented by 𝑏∈{0,1}. There are two problems with this strategy,
though. First of all, it uses an unconventional quantization table, which means that the hidden
data must be included in the JPEG picture header. Second, the Huffman coding method
modifies the stream of quantized coefficients, making it less compressible overall. This
results in a JPEG image that is tagged, meaning it contains hidden data. An RDH model for
altering particular entries in a quantization table and inserting secret messages with quantized
DCT table mean frequency coefficients within a cover image was presented by Chang et al.
in [110]. The DCT coefficients of every block are expanded in both the original and updated
quantization tables, which makes data input easier. By increasing the tagged JPEG file's
storage capacity, this technique improves at both information hiding speed and image quality.
A method for creating room for data embedding and attaining high delivery power and
accuracy was developed by Wang et al. [113]. We divide the quantized table entries by an
integer and then with the same integer is multiplied with DCT coefficients. However, as the
JPEG files' initial quantization table balances file size and visual quality, the RDH technique
for JPEG photos, which is based on altering JPEG quantization tables, may inevitably upset
this equilibrium.

2.3 RDH WITH MODIFICATION OF THE HUFFMAN TABLE

Changing the Huffman table is the third approach. The finding that current encoders only use
a part of the JPEG code space was exploited by Mobasseri et al. [114]. By merging usable
and unneeded variable-length codes (VLC), they were able to integrate data. This method can
maintain or even reduce file size while transporting the payload. When the data is hidden,
parts of the bitstream code are replaced with code that matches the secret bits. By applying a
mapping technique designed to improve integration, Hu et al. [116] effectively utilised the
underutilised variable length codes (VLC), leading to a significant increase in data capacity.
For JPEG Huffman sets, Wu and Deng [117] presented a unique integer vector transformation
technique that guarantees synchronous bitstream modification with the encoded message.
Any JPEG bitstream can use this technique, including ones with optimised Huffman tables.
These techniques preserve the precision and recorded dimensions of the tagged JPEG picture,
but they provide only a restricted range of embedding choices. Therefore, they are best suited
for applications such as image authentication, where minimal integration is sufficient.
 9

CHAPTER-3 ROBUST REVERSIBLE DATA HIDING

In some cases, if an image containing hidden data is irreversibly processed, it may not be
possible to completely restore the source picture. However, this is viable that much of our
hidden information can still be retrieved from the processed image. Therefore, RDH
applications often request robustness. In recent decades, several robust reversible data hiding
(RRDH) techniques have been published for digital images. Below we look at some related
methods.

3.1 PROMINENT TECHNIQUES FOR ROBUST RDH

Vleeschouwer et al. [141], [142] presented an original RRDH method based on the
relationships between adjacent pixels. The technique is said to withstand moderate JPEG
compression. Addition and subtraction calculations are performed using the modulo-256
system to solve overflow/underflow problems. However, adding modulo-256 can cause white
pixels to become black pixels and vice versa, causing "salt and pepper" noises & degrading
picture pixels. We et al. [143], [144] introduced the RRDH scheme to reduce salt and pepper
noise after identifying this limitation. They discovered and used a robust statistical metric
derived from patch theory to encrypt [Link] scheme involves dividing the original
image into parts, each of dimensions 8 × 8. Use of Errors Correctioning Code &
combinations techniques can lead revertibility unless picture has undergone an irreversible
procedure, & robustedness can be obtained if the picture underwent processes such as JPEG
compression. Every 8×8 unit is divided into 2 groups: A, which consists of all items denoted
with a plus, & B, which consists of all items denoted with "-". Hence, contrast between the
two sets, specifically A and B, is determined by calculating the difference value.

Fig 3.1: Different cases of distribution of Block histogram

 10

In this plan, the pixels of table A and B are indicated by ai and bi , respectively. Since, here is
a powerful correlation among grayscale values in the pixels in the blocks, the differencing
value α is thought to be near to zero. Furthermore, since α reflects the stats of each & every
pixel in the block, this statistic is likely to remain constant despite small changes in certain
pixels after JPEG compression. The effectiveness of the statistical metric comes from its
inherent robustness to JPEG2000/JPEG compressioning & little changes. In the given
method, one bit is embedded in every block, so data inserting capability is determined by the
block size. Larger block sizes generally improve the durability of embedded pieces.
Therefore, a balance between the amount of information to be entered and the level of
robustness must be found according to the needs of the specific [Link] solve
overflow/underflow problems, blocks are divided into four groups, each with a special way of
inserting bits. In the case I, if the block gray values ​are far enough from the histogram limits,
the difference value α to hide the value '0' remains the same and is moved above the
threshold K to the value hides "1". In cases II and III, the histogram is moved by K to the
right or left to insert a '1', while the insertion of a '0' has no effect. In the situation of Case IV,
when certain pixel values ​are close to the histogram limits, the value '0' is not changed, but an
error occurs at '1' and is corrected by ECC. Many subsequent RRDH systems were built using
the same framework.

Tsai et al. [145], [146] recently developed a zero-watermark RRDH technique using
α-trimmed avg algorithm and SVM. In this approach, initial cover picture remains the same
when the trained SVM learns the embedded message, allowing data extraction through
evaluation of Support Vector Machine algo. The α-trimmed averaging method is used to
combat noise interference. Yin et al. established a method for selecting and evaluating
different feature regions in RRDH.
 11

CHAPTER-4 REVERSIBLE DATA HIDING WITH CONTRAST

ENHANCEMENT

This section describes contrast enhancement techniques in reversible data hiding (RDH). The
starting point is to explain why it is important for RDH to create image contrast enhancement
methods. The RDH techniques discussed in references [110, 147, 148, 149] are carefully
analyzed to emphasize the priority of contrast enhancement over peak signal-to-noise ratio
(PSNR) maximization to preserve the quality of host images. The next section focuses on
evaluating the quality of the images compared to the contrast RDH image. Future research in
this area is highlighted at the end of the section..

4.1 MOTIVATION
Previously , Reversible Data Hiding (RDH) was designed for authentication in applications
that are sensitive to distortions such as satellite images, military and medical applications. In
these areas, it is important to maintain the perceived quality of the host image. The peak
signal-to-noise ratio (PSNR) is often used to estimate the difference between the original
image and the image containing the hidden data. When data embedding processes add
distortion to the image, PSNR falls short as an objective metric. Since there is always a
trade-off between PSNR and embedding power, the performance of image RDH approaches
is sometimes analyzed using a plot of PSNR versus hiding rate.

Fig 4.1 Dark image and image achieved by increasing contrast. (a) Original (b) Image after
contrast enhancement. PSNR: 18.66dB.

Although the quality of a picture including hidden data may be preserved by employing
sophisticated RDH techniques to maintain a high PSNR , the unavoidable distortions brought
about by data hiding frequently prevent the visual quality from improving. Merely
maintaining picture quality is insufficient for low-quality photographs, such as those taken in
poorly lit environments; improving visual quality is more important. For instance, in order to
show features for visual examination in medical or satellite photos, contrast augmentation is
frequently required. Contrast enhancement may dramatically increase the visibility of picture
 12

features, as seen in Fig. 4.1, even if the improved image's PSNR is just 18.66 dB. This
demonstrates that in some cases, maintaining a high PSNR is less important than enhancing
visual quality.

Traditional RDH approaches often place a higher priority on maintaining a high PSNR since
it is not always appropriate to use PSNR to judge image quality in certain instances. A new
kind of RDH that improves the host image's visual quality is thus required. For some
applications, contrast augmentation is especially desired, and this has been accomplished in
[110] by use of histogram bin expansion procedures, which result in a histogram equalisation
effect [140]. Similar to this, further techniques suggested in [147]–[149] put less emphasis on
a high PSNR in RDH and more emphasis on maintaining visual quality through contrast
augmentation. The outcomes of the experiments indicate that these techniques can really
enhance visual quality. The techniques covered in each of the next parts [147]–[149] will be
reviewed separately.

4.2 IV RDH METHODS WITH CONTRAST INTENSIFICATION

IV Contrast-enhanced RDH techniques are reviewed in this section. These techniques all
have the capacity to boost visual quality in specific scenarios by utilising contrast
augmentation.

4.2.1 METHOD BY HISTOGRAM BIN EXPANSION [136]

In theory, histogram equalization can be used to improve image contrast. The technique [110]
modifies the histogram of pixel values ​to mask data and increase contrast at the same time.
The first step in creating a histogram of a grayscale image is to count the number of times
each pixel value (between 0 and 255) occurs in the host image. The two highest frequency
non-empty bins are selected and labeled as fL & fR, respectively. The following describes
how the histogram bin is expanded for each scanned pixel value f in raster order:.

where b is the binary value (0 or 1) of the bitstream to be buried and f₀ is the new value that
replaces f. The hidden data is entered by expanding the top two boxes to four boxes after
applying the above equation to each pixel of the host image. Then, the two highest bands of
the readjusted histogram are selected for expansion by adjusting the values ​of fL and fR.
Histogram equalization is an effect comparable to the continuous extension of the two highest
cells in the transformed histogram.
 13

Preprocessing of the master image is necessary to avoid overflow and underflow problems
due to histogram bin expansion. In preprocessing, pixel values ​from 0 to S-1 are incremented
by S, while values ​from 256 to S to 255 are decremented by S when S pairs of histogram bins
are enlarged together (S < 64). A binary location map of the size of the host image is made to
store the locations of these modified pixels; The number 1 is used to indicate the location of
the edited pixels, and zeros for all other locations. This local map is losslessly compressed
using the JBIG standard and integrated into the host image with other data to restore the
original image. An extended histogram bin must be understood to obtain the original image
during data extraction. If the pixel value f0 is in the set {fLL-1, fLL, fLR, fLR+1}, the binary
value b0 can be extracted from it by identifying the last two augmented boxes, fLL and fLR,
as shown below:.

To obtain the histogram bins, the following operations are performed in parallel while the
data is extracted in raster order:

This equation can be used to recover the hidden data from the forward recovered histogram if
the previously extended bins are known. At the same time, the histogram is repeatedly
reconstructed using the described steps until the pre-processed image is reached. The location
map constructed from the contrast image must be used to transform the preprocessed pixels to
restore the original image. Any pixel value [S, 2S-1] is decremented by S when its associated
value in the local map is 1, because pixel values ​that were previously in the range [0, S-1]
were incremented by S during preprocessing. Similarly, keep it above all at the top if the
location map has an equivalent value of 1, the value of pixel in pixel [256-2S, 255-S] is
increased by S. Similarly, if the location map has an equivalent value of 1, the value of a
pixel in pixel [256-2S, 255-S] is increased by S. It is important to note that in [110], the
contrast image keep safe all the values ​of the extended histogram bin for data hiding, the total
number of compressed position map and extended histogram bin pairs. This allows you to
blindly restore the original image. of.

4.2.2 METHOD WITH CONTRAST ENHANCEMENT FOR MEDICAL IMAGES

Although the technique described in [110] can improve contrast and high-power RDH, it can
cause artificial distortions in photos with strong backgrounds. To overcome this limitation, a
new RDH technique specially designed for medical images is introduced [147]. First, the
 14

image is segmented into background and region of interest (ROI) using Otsu's approach to
background segmentation. Next, background pixels with pixel values ​greater than a given
percentage are segmented into main pixels. Selective contrast enhancement of the ROI can be
achieved by removing the histogram bins from their key values ​of the extension. Two
intervals of length S, indicating the number of pairs of histogram bins to be enlarged, are
adaptively selected in the histogram to reduce pixel inclusions to reduce distortions caused by
preprocessing. Histogram shift is used to overlap these selected periods to reduce pixel value
distortions due to preprocessing. Since the image with contrast hides all the bins of the
enlarged histogram, the hidden data can be extracted in a similar way as described in [110].
However, the original image can be blindly reconstructed using a contrast image, revealing a
hidden location map.

4.2.3 METHOD WITH THE CONTROLLED CONTRAST ENHANCEMENT [148]

An alternative approach to using [110] to correct visual distortions associated with large
volume data hiding is presented in [148]. To avoid over-enhancement, a maximum relative
contrast error (RCE) limit defined in [143] is imposed, which controls the extent of contrast
enhancement. Since an RCE above 0.5 indicates increased contrast, the upper threshold is set
to 0.55 to maintain visual accuracy. Next, data hiding will be extended to the Haar IWT
domain to increase the hiding power. This method generally provides better image quality in
scenarios that require large amounts of masking..

4.2.4 AUTOMATIC CONTRAST ENHANCEMENT METHOD [149]

Another RDH-based technique to achieve reversible contrast enhancement is presented in
[149]. To avoid preprocessing artifacts, their method generates a location map for each
histogram expansion, unlike [110], where the location map is built only once. In particular,
the lower bin is merged with the adjacent bin, while the upper bin of the histogram is
increased. Several iterations are performed to smooth the histogram as smoothly as possible.
Automatic contrast enhancement continues until the built-in capacity can no longer keep up
with the increasing total information. Experimental results show that the approach of [149]
provides better visual quality even for low exposure photos. Furthermore, the resulting
contrast-enhancing effects are similar to those produced by Matlab's global histogram
smoothing function "histeq"..

4.3 IMAGE QUALITY ASSESSMENT IN RDH WITH CONTRAST ENHANCEMENT

The relative contrast error (RCE) is determined using methods described in [110,147,148].
However, because it can be high in an enhanced image with visual defects, the RCE score
could not accurately reflect the visual quality. Although the methods presented in
[110,147,148,149] can improve visual quality in some situations, high contrast does not
always produce high-quality images. Furthermore, PSNR [110,147,148,149] is still used as
the comeonimage quality standard, as there is currently no proper image quality estimator.
Since PSNR is not always the best metric to evaluate image quality, it is important to find
other standards that can better control RDH processes. This is followed by a review of the use
of reference-free techniques for image quality and Structural SIMilarity (SSIM) evaluation.
These techniques are specifically designed for images with altered contrast. of.
 15

4.3.1 USING SSIM FOR IMAGE QUALITY ASSESSMENT

The common features/similarity between two images is calculated/measured by the SSIM
index, a commonly used tool for evaluating image quality. As opposed to objective
techniques that measure variations between a destroyed/distorted image and a reference
image, the SSIM index concentrates on assessing quality by measuring the rate at which
structural information is lost. Its scale goes from 0 to 1, where 1 represents the same images.
In [147], the SSIM index is used in addition to PSNR to evaluate image quality. On test
photos from USI-CIPI and Kodak, experimental results show that the SSIM index between
the real and contrast-intensified images often drops with increasing concealing rate, while the
corresponding RCE value increases. Interestingly, regardless of whether visual distortions are
present in the image content, the SSIM index tends to go down with increased contrast. The
SSIM index's susceptibility to potential visual distortions as well as increased contrast
emphasises the limits of using it as the only metric to evaluate image quality in RDH with
contrast augmentation. As a result, it might not be adequate to evaluate image quality in this
situation by using SSIM and PSNR.

4.3.2 USING THE METHODS FOR CONTRAST CHANGED IMAGES

Techniques created especially for photos with contrast modifications can be applied when the
real image or its randomness is available. As an illustration, [148] provides a patch-structure
representation method for evaluating the calibre of images with contrast modifications. There
are several ways to assess perceptual distortions by breaking down a picture patch into its
average/mean intensity, signal power/strength, and signal structural components. By creating
a local contrast quality map, this technique makes it possible for the patch-based contrast
quality index (PCQI) approach to give exact estimates of how humans would perceive
variations in contrast.

4.3.3 USING THE NO-REFERENCE METHODS

As described in [150]–[155], more no-reference techniques have been added to SSIM and
PCQI for the assessment of picture quality. In general, a good no-reference technique for
contrast intensified situations should be high resilient to increased contrast as long as the
image quality is maintained. A no-reference quality assessment method based upon natural
scene statistics (NSS) is presented in [155] for contrast-distorted images. A large-scale image
database is used to develop NSS models utilising moment & entropy features. As such, the
degree of unnaturalness, or variation from the NSS models, can be used to evaluate the
quality of a contrast-enhanced image.
 16

CHAPTER-5 REVERSIBLE DATA HIDING IN DOMAIN

5.1 Motivation
Most individuals concur that using encrypting to preserve privacy is a good idea. The needs
of clouds computing systems and other privacy-preserving technologies have spurred an
increasing curiosity in signal processing within encrypted domains. Initial focus has also been
paid to the combination of data concealing and cryptography. Only some of the cover data is
protected by encryption in some of the joint methods that are now in use, enabling the rest of
the information to contain extra messages. In a similar vein, cover data in the transform
domain's lower and upper bit-planes are watermarked and encrypted, respectively.

Nevertheless, there is a chance that some information will escape through the cover in these
joint methods since just part of encryption is used. Furthermore, the data embedding
procedure is irreversible and does not handle the separating of the initial cover & encrypted
info in a watermarked one.

The majority of the studies on RDH covered throughout previous chapters is intended for the
plaintext domain, in which extra bits are inserted in initial one, unencrypted media content.
But as interest in signal processing inside encrypted domains increases, so does the need to
reversibly incorporate more data inside the encrypted domain. This results from the
requirement to safely transfer or keep multimedia files with third parties, in which case the
media owner may encrypt the data before sending it. The study of reversible embedding
becomes relevant in some scenarios, such as when a helper or channel administrator who isn't
aware of the initial material tries to add extra messages—like image annotations,
authentication data, or origin information—to the hidden media.
For instance, in the field of medical image management, sensitive data may be embedded into
encrypted medical photographs by a database administrator in an attempt to protect patient
privacy. Furthermore, it is anticipated that following data extraction and decryption at the
recipient's end, the original content will be accurately retrieved. A summary of the most
recent RDH-ED reversible embedding techniques is given in the following paragraphs.

Fig 5.1: General framework of Reversible Data Hiding in Encrypted Domain

Figure 5.1 presents the overall framework of RDH-ED. The data proprieter, info hiding agent,
and getter are three main players in the workflow, and they each have distinct tasks that are
explained below:
• Content Owner: Whether processing is used or not, the source material is encrypted to hide
the primary data. The owner of the material chooses the encryption key.
 17

• Data Hider: Inserts extra information into material that is encrypted. For security, the data
hider makes use of a data concealing key.
• Receiver: Three possibilities are available to receivers who grip distinct keys. In order to
retrieve a rough approximation of the material, Option 1 entails decrypting the designated
encoded content. Extraction of the excess hidden bits is the second option. The creation of
retrieved material that is exact replicas of the source is made possible by option 3.

"Vacating room before encryption (VRBE)" and "vacating room after encryption (VRAE)"
are the two primary categories into which algos are divided. We provide a thorough rundown
of every division below.

5.2 VACATING ROOM BEFORE ENCRYPTION (VRBE)

Fig 5.2: VRBE

Because embedding area is generated in the plaintext realm by the VRBE architecture, it is
formed prior to encryption [156]–[159]. As such, before encrypting, the material's proprietor
must perform further preprocessing. Using pictures as a demonstration (and in a similar
manner going forward), Figure 5.2 shows the VRBE framework concept. In order to provide
embedding capacity within digital images, the Least Significant Bits of some pixels are
merged into other pixels using a traditional RDH technique [156]. The owner then creates an
encrypted image by encrypting the processed image. As a result, the data hider can use these
freed LSB positions in the encrypted image to carry a large payload of up to 0.5 bits per pixel
(bpp).

A different strategy that makes use of a prediction technique is presented [157], adhering to a
similar idea. Predicted mistakes are produced in this approach by estimating certain pixels
based on the pixels that remain before encryption. Then, a particular encryption strategy is
designed to encrypt these anticipated faults, and the remaining pixels can be encrypted using
a normal encryption process (such as AES). Rather than embedding data directly into
encrypted images, more information can be integrated by adjusting the encrypted histogram
of predicted mistakes.

Patch-level sparse representation is incorporated to further improve the situation [158]. A

subset of atoms in an overcomplete dictionary may represent an image patch linearly, as
demonstrated by the widely used sparse coding approach. Mostly residual mistakes are stored
 18

and self-embedded in the cover picture because sparse coding yields an approximate solution.
Furthermore, the encrypted picture contains the obtained dictionary contained in it. By taking
use of sparse coding's strong representation, a sizable amount of space is freed up, which
enables the data hider to insert additional hidden messages into the encrypted picture.

To create embedding space, [159] combines Paillier homomorphic encryption with difference
expansion, a traditional plaintext RDH approach. With this method, preparation is required.
The original image must be adjusted before encryption, and the owner of the image uses the
difference expansion technique to do so. The processed image is then sent to the data hider
encrypted using Paillier homomorphic encryption. The chosen encrypted image is then
created by the data hider by adding an extra bit to each pair of neighbouring encrypted pixels.
Using the homomorphic characteristic of Paillier encryption, the recipient decrypts the bits
that were inserted and uses this information to rebuild the original cover image by comparing
every pair of decrypted pixels. The VRBE technology can be deemed unworkable since it
necessitates more preprocessing by the content owner before encrypting the material. This is
where the more realistic VRAE structure, which we will discuss next, comes into play.

5.3 VACATING ROOM AFTER ENCRYPTION (VRAE)

Fig 5.3: Vacating Room After Encryption Framework

In VRAE techniques, the data hider modifies specific bits of the encrypted data to incorporate
extra bits, while the information proprietor encrypts the signal itself directly. Figure 5.3
shows a model of the VRAE framework using photographs as a demonstration.

Depending on the field in which extra information may be retrieved, VRAE techniques can
be categorised into three basic groups: data retrieval in the cypher domain, information
extraction in the plaintext domain, and data retrieval in both domains. First, we will discuss
extracting data in plaintext.

5.3.1 VRAE: DATA EXTRACTION IN THE PLAINTEXT DOMAIN

 19

Fig 5.4: Extraction and recovery of image in domain of plain text

In these methods, the data-hiding key is used to recover the initial picture after the encryption
key has been used to decode an encrypted media carrying additional data and retrieve the
embedded data. Figure 5.4 shows the schematic that shows data extraction and picture
restoration in the plaintext domain.

The first method for encrypted pictures was suggested by Zhang [162]. It entails the data
hider dividing the secret picture into blocks and adding a single bit to every unit by changing
the 3 least significant bits (LSB) of half of the blocks' pixels. The indicated encrypted picture
is transformed into a rough image upon decryption by the recipient. After building a new
block by adjusting the three LSBs of pixels, the receiver uses a function to estimate the
picture texture of each block. The original block is assumed to be substantially smoother than
the disturbed block by taking use of the spatial correlation found in natural pictures.
Therefore, it is possible to take out the contained bits and reconstruct the initial image
together.

This method's embedding rate depends on the block size that is chosen. Errors might occur in
the data extraction and picture recovery procedures if an incorrect block size is used. Qian et
al. [163] present a similar method for JPEG pictures in which data is hidden in the encoded
JPEG bit-stream and the initial bit-stream is recovered by examining the blocking artefacts
caused by data hiding. This strategy has garnered a lot of attention. To improve performance,
for example, [164] uses a side-match technique and spatial correlation between adjacent
blocks to obtain greater embedding payloads with lower error rates in picture recovery. By
using imbalanced bit flipping and adding a flipping ratio, more gains are achieved. To further
improve the accuracy of data extraction and picture recovery, a more precise function for
estimating the image texture of each image block is included.

Moreover, just a smaller portion of the encrypted image's LSBs are flipped during the data
embedding procedure outlined, as opposed to flipping the LSBs of half of the image's pixels.
The estimated image's visual quality is much improved by this modification. Furthermore,
throughout the data extraction and picture recovery processes, a unique adaptive judgement
function is utilised to estimate the image texture of each block depending on the distribution
 20

features of local image material. To some extent, this method aids in reducing the mistakes in
the reconstructed picture and the retrieved bits.

The previously listed techniques mostly focus on data extraction following decryption. Stated
otherwise, the primary content must be revealed prior to data extraction by extracting the
supplementary data from the textual domain. As a result, no information can be extracted
from the indicated encrypted media containing extra bits by someone who has the data-hiding
key but not the encryption key. This category is also known as a non-separable approach
since the retrieval of the initial media and the extraction of the embedded bits are frequently
linked. On the other hand, there is a different sort of method called a separable solution,
which allows information extraction, or extraction of data in the cypher domain, to be done
independently before picture decryption.

5.3.2 VRAE: DATA EXTRACTION IN THE CIPHER DOMAIN

Fig 5.5: Extraction of data in the domain of cipher

In these techniques, a legitimate recipient with a data embedding key can extract extra pieces
of the encrypted media directly in the encryption region. Similarly, a data receiver with an
encryption key can decrypt the received data, resulting in a representation of the original data
carrier, called an approximate data carrier. Both data hiding and encryption keys allow the
receiver to pick up the extra bits and return the original content intact. Figure 5.5 describes
the data extraction process in the encryption area. This concept was first proposed by [166],
where the owner encrypts the original image using Advanced Encryption Standard (AES) and
the cache adds one bit to each block of n pixels, resulting in embedding. rate of 1/n bits per
pixel (bpp). Data extraction and image recovery at the receiving end involves local standard
deviation analysis after decrypting the encrypted image with [Link] said, this
approach has two main disadvantages for the receiver. Originally, a hacker could use
statistical analysis to gather information about the encrypted bits, since each block is
encrypted separately using AES with a unique encryption key. Repeated encrypted blocks
due to redundancy in the image can be a security risk. Second, if the receiver decodes the
tagged encrypted image directly, the quality of the decoded image may be insufficient and not
meet the requirements of human vision..
 21

Zhang proposed a technique called RDH-ED for encrypted images to overcome the problems
of the previous method and enable the extraction of additional information. The process
involves compressing the encoded bits to make room for the extra bits. The cache divides the
encoded image into groups, and the least significant bit levels (LSBs) of each group are
compressed, allowing data to be embedded. After decoding, the receiver-side LSBs are
determined using the MSBs of the neighboring pixels, allowing the original LSBs to be
retrieved. Since the extra bits are added to the least significant bits (LSB), they can be
removed before [Link] techniques also create embedding space by compressing
coded bits [169], [172], [173]. Zhang and colleagues developed a lossless compression-based
RDH technique for encrypted data using LDPC codes to compress specific LSB files [169].
The authors of [174] encode certain bits as syndrome bits to make room for more data. In
reference [172], the least significant bits of the encoded pixels are losslessly compressed
using the Hamming distance calculation.

Moreover, although most RDH-ED techniques are based on stream encryption, Qian et al.
presents a suitable method for block-encrypted images [171]. In this process, the owner
encrypts the image using a block cipher, and then the server adds extra bits to the encrypted
image to create a tagged encrypted image. Individuals with the embed key can obtain
additional bits from the encrypted region. Compared to current block encryption-based
RDH-ED techniques [166], this method improves both the security and quality of images.

Besides, in [170] discusses a universal reversible data embedding method that can be used in
any encrypted domain. It uses Golomb-Rice (GRC) code words to exploit code redundancy in
encrypted signals. By dividing the signal into parts and encoding them with GRC, each part
can be swapped two bits in the GRC code in a reversible way. The test results show that the
input load of the encrypted signal is on average 0.169 bits per [Link], data decryption
can only be done in an encrypted domain before decrypted or in a decrypted domain after
decrypted using the aforementioned methods. This is a big problem: a legitimate recipient
who has the data hiding key, but no decryption key, cannot directly recover the hidden pieces
of the encrypted domain. On the other hand, if the receiver has the key to hide the data and it
is decrypted with extra bits, it cannot retrieve the hidden [Link], a new RDH-ED
framework is needed, which allows extracting embedded bits from both plaintext and cipher
fields. This means that only with the data hiding key, recipients can extract the hidden bits
from the encryption region, improving flexibility and usability.
 22

5.3.3 VRAE: DATA EXTRACTION IN BOTH DOMAINS

Fig 5.6: Extraction of data in both the domains

In this context, a legitimate recipient can extract the hidden data from the encrypted media
using the data secret key of the encryption domain. Additionally, with the encryption key, the
recipient can decrypt the encrypted information to obtain a copy that closely resembles the
original version. If there is an average approximation and a key to hide the data, more bits
can be extracted and the original image can be completely restored. Figure 5.6 illustrates the
data extraction process in both fields.

Reference [176] presents an original solution that uses pseudorandom sequential modulation.
In this method, part of the data in the least significant bit layers of the coded image is
replaced by additional data, while the data in the remaining LSB layers is modified by
pseudo-random sequences affected by the replaced bits and embedded data. A user who has
the key to hide the data in the additional data can easily retrieve it from the encrypted
domain. Direct decryption can result in an image that shows the original key content because
the data embedding only affects the least significant bit. The detection of modulated
sequences showing minimal variation allows extracting the embedded data from the decoded
image, ensuring an error-free recovery of the original content as long as the embedded rate is
low..

In certain RDH-ED techniques, such as those described in [176]–[180], information is added

to regions that remain unchanged by encryption. In [176], gray scales of adjacent pixels are
hidden using identical pseudo-random bits. Additional information is added to the different
bit levels in an interpretable way using a method that optimizes the parameters to achieve a
good balance between payload and distortion. Because encryption does not affect the usable
 23

data space, both plaintext and encrypted data can be appended or decrypted using the same
methods [177] presents a fully separable RDHEI approach based on block segmentation, RC4
coding and block histogram adjustment. The original image is divided into separate blocks,
each encrypted with RC4 and a shared key that preserves the redundancy of the structure
needed to embed the extra bits using the block histogram transfer. This method allows
accurate separation of labeled encoded and directly decoded images, but is not suitable for
images with saturated pixels. Similarly, [178] uses cross-division to classify image pixels and
encrypts all pixels in each class with a common RC4 key. This preserves the difference
between the histograms after encryption, making it easy to enter the difference in a reversible
way by changing the histogram.

Another approach involves RDH-ED, which exploits homomorphic encryption [181]. This
paper proposes lossless and recoverable data hiding methods for encrypted images using
public-key cryptosystems with probabilistic and homomorphic [Link] the lossless
method, the encryption pixels are replaced by different values ​to hide the extra bits of
multiple LSB encryption pixels by multi-layer wet paper coding. The embedded data can be
directly extracted from the encrypted domain without affecting the decryption of the original
plaintext image. When using preprocessing in the inverse model, the histogram of the image
is reduced before encryption to avoid pixel oversaturation of the plaintext region due to data
embedding changes in the encrypted images. Even with a small distortion, it is possible to
extract the embedded data and restore the original image directly from the extracted
[Link] the lossless and recovery models are compatible, data embedding processes can
be performed simultaneously on encrypted [Link] a combined technique, the receiver
can recover part of the hidden information both before and after decryption, and finally
returns the original image as clear text.
 24

CHAPTER-6 REVERSIBLE DATA HIDING IN AUDIO AND VIDEO

6.1 VIDEO REVERSIBLE DATA HIDING

Static pictures are the main focus of RDH analysis. Studies on audio/video RDH, on the other
hand, has received little attention. A video RDH approach that utilised histogram
manipulation was devised by Liu et al. [160]. To increase resilience, the content that has to be
hide is initially encoded utilising the BCH syndrome code prior data concealing.
Subsequently, the encoded information is integrated with the quantized DCT parameters of
the intra-frame's 4 × 4 blocks. Using Multi-Vide Coding (MVC) films as a medium, Song et
al. [161] introduced a reversible video steganography system for encrypting hidden
information into every block's motion vector. To accomplish reversible nature the concept of
the inner product among the vector modulations with the motion vector is put forward. Zhao
et al. [162] presented an RDH approach that utilises two-dimensional (2D) histogram
transformation to boost the ability to be inserted of histogram driven reversible watermarking
approaches. To offer superior capacity-distortion efficiency while integrating info with 2D
histogram alteration, 2 quantized DCT parameters are arbitrarily chosen for every
embeddable 4 × 4 luminance unit. Compared to other reversible video watermarking
approaches now in use, our approach does not modify the motion-compensated prediction
error histogram. Rather, it depends on the interpolation faults of increasing
motion-compensated frames. With this method, high-capacity information may be inserted
into videos with the least amount of distortion. Furthermore, as covered in [164], it presents
an intra-frame error concealing technique in H.264/AVC based on RDH. These techniques
embed the motion vectors (MV) of macroblocks (MB) into other MBs in the same
intra-frame by means of histogram modification. The integrated MV can be recovered from
the appropriate MB in order to restore the damaged data if an MB becomes contaminated
during decoding. This method tackles the requirement for cloud data management
confidentiality and safety, occasionally necessitating the encryption of video material.
Reversible data hiding (RDH) in protected H.264/AVC movies has been proposed lately
[165]. With this technology, even without knowing the initial video material, the data-hider
may incorporate hidden data via an altered histogram shifting approach within the encoded
video.

6.2 AUDIO REVERSIBLE DATA HIDING

Relative with video RDH, there is comparatively little study on audio reversible data hiding
(RDH). An audio RDH technique based on PEE was presented by Yan and Wang [182], who
included confidential data within the increased prediction error. In order to handle the
underflow/overflow problems, authors used a location map approach, which meant that in
order to guarantee reversibility, secret data had to be included with the compressed location
map. In a similar vein, Nishimura [183] increased embedding capacity and SNR by using
Burg's approach to generate anticipated coefficients. Nishimura [184] presented a variable
extension approach in addition to prediction error expansion with regard to capacity
 25

management capabilities. Lower expansion ratio leads to lower deterioration and lower
payload. Studies have shown that stego audio with satisfactory reliability may be obtained by
inserting an arbitrary payload of a maximum of 0.4 bit per instance into CD-format music
transmissions. Speech signals coded with G.711 µ-law are likewise treated using this
approach. The differential evolution approach is used to discover optimal coefficients, as
opposed to [182], which uses a predefined set of linear prediction coefficients.

By drastically lowering the embedding position information, the histogram shifting approach
is used to improve capacity control. Nishimura presented an additional reversible audio data
concealing technique uses quantization index modulation (QIM), as described in [186]. Here,
amplitude expansion within QIM creates gaps in the amplitude histogram that conceal secret
information. The approach's capacity to provide good imperceptibility is confirmed by an
objective perceptual evaluation. However, when applied to altered stego audio, the technique
loses its reversibility. In a subsequent work [187], Nishimura refined the method from [186],
making it semi-reversible for perceptually coded stego audio and reversible for unaltered
stego audio. The secret data is implanted via direct-sequence spread spectrum, or DSSS
modulation. This sequence originates from the amplitude expansion of integer modified
discrete cosine transform (MDCT) coefficients in the time and frequency domains. The
amplitude histogram's apertures are used to insert the reversible payload, in a method akin to
that described in [187]. The evaluation findings show that for the stego audio, the mean
objective difference grade (ODG) is higher than the "perceptible, but not annoying" level.
Achieving semi-recovery from perceptually coded stego audio is indicated by negligible
changes in ODG between the coded and recovered audio.
 26

CHAPTER-7 REVERSIBLE DATA HIDING FOR UNCOMPRESSED

IMAGES

The process of creating RDH for uncompressed pictures typically involves two parts. In the
initial stages, we focus on constructing several RDH frameworks and devised five
conventional techniques:
- lossless compression,
- Prediction Error Expansion,
- Integer Transform,
- Histogram Shifting,
- Difference Expansion.

The second step involves the development of the theory and the exploration of more efficient
methods. It aims to enhance performance by addressing the issue of unnecessary utilization.
The classical PEE framework is utilized as the main development framework for
implementing efficient algorithms in this stage. These algorithms include
- Pixel Value Ordering,
- Multiple Histograms Modification,
- pairwise PEE.

In this section, we will focus on the conventional RDH techniques implemented in the early
phase and briefly discuss their expansions. An elementary illustration is visible in Table 7.1.

RDH Description Expectable ER Embedding Reference

Techniques Distortion

Lossless After making 1.5 bpp High [2, 16, 135,

compression some room in the 138]
image by
compressing it,
embed the
pieces.

DE Use difference 0.9 bpp Medium [30, 38, 95]

expansion for
data embedding.

IT Apply the integer 2.17 bpp Medium [1, 9, 72, 101]

transform for
data embedding.

HS According to the 1 bpp Low [17, 42, 62]

histogram,
embed the bits.
 27

RDH Description Expectable ER Embedding Reference

Techniques Distortion

PEE For data 1.8 bpp Low [34, 47, 90,

embedding, alter 94]
the
prediction-error
histogram.
Table 7.1: Summary of the RDH Techniques

7.1 Compression-based RDH Technique

RDH algorithms were initially developed primarily through the use of lossless compression
methods [2, 16, 135, 138]. To utilize the available space for data embedding, the partial
image material undergoes lossless compression to provide a compact representation with
reduced size. Because compression is lossless, it is possible to restore the image. The
compression method and the embedding performance are closely related. One objective is to
enhance the balance between performance and compression efficiency. Celik et al. [2]
proposed the classic approach of least significant bit (LSB) based reversible data hiding
(RDH). This method employs quantization to determine the minimum value of the cover
pixels, which is then utilized for embedding. The information is recursively incorporated in
[135] through the decompression and compression methods of an entropy coder. Zhang [138]
proposed partitioning the cover pixels into many segments and applying the optimal value
transfer rule to change them. This method effectively accomplishes reversible embedding
using a range of compression algorithms. The RDH approach, which relies on compression,
has demonstrated its effectiveness. However, a drawback of attempting to incorporate
additional hidden information into the image is that it can become difficult to preserve its
intended meaning. When there are enormous amounts of something, it is common to see
clear visual distortion.

7.2 DE-based RDH Technique

The DE-based RDH algorithm applies the Haar transformation to transformation each of the
two adjacent pixels for data embedding. The reversible embedding and extraction of data is
facilitated by the commonalities between the two pixels. The designated pixels for a pair of
pixels (p1,p2) are p1 = 2p1 − [1/2 (p1 + p2)] and p2 = 2p2 − [1/2 (p1 + p2)] + m, where m
represents the binary message bit. To put it simply, the disparity between the two pixels is
increased by 2(p2 − p1) + m in order to include one additional bit. The least significant bit
(LSB) of the magnified difference can be utilized by the recipient to compute the embedded
bit, whereas the inverse transformation is employed to recover the original cover pixels.

In the context of data embedding (DE) in digital images, the embedding rate (ER) can reach a
maximum of 0.5 bits per pixel (bpp) when using a single layer embedding technique.
Additionally, the average values of two adjacent pixels before and after the embedding
 28

process stay unchanged. Several variations of RDH have been evolved from this fundamental
method. However, the DE technique in the initial stage sometimes fails to include the
discriminating processing for the pixel pairings during embedding. Hence, the utilization of
multi-layer embedding may lead to substantial distortion while aiming for a higher ER.
Furthermore, to reduce the amount of embedding capacity used, DE-based approaches
typically rely on the location map to identify pixels that are not available and then compress
them without losing any data.

7.2.1 Illustration of Difference Expansion

Given two numbers, x=206 and y=201, we aim to securely and embed a bit b = 1 reversibly.

We calculate the average of the integer, denoted as L, and the difference, denoted as H,
between the values of x, and y.

L = (206+201) / 2 = 203

H = 206-201 = 5

Then, we convert the difference-values (H) into its binary equivalent, which is (101)2. By
appending bit b binary representation to the LSB of the binary representation of (H), the
resulting difference-value will be
H’ = (101b)2 = (1011)2 = 11.

It can also be written as follows: H’ = 2h + b = 2(5) + 1 = 11

Ultimately, we calculate the updated values by utilizing the new difference-value (h') and the
initial integer average-value (L).
x’ = 203 + (11+1)/2 = 209; y' = 203 - (11)/2 = 198

We will then recover the embedded bit b, and recreate original pair (x, y) from the embedded
pair (x', y'). Then, we calculate integer mean, and subtraction.

L' = (209 +198) / 2 = 203

H' = 209 - 198 = 11

Examine binary form of H', where H'=11=(1011)2. Retrieve the LSB, which is equal to 1 in
this particular scenario, and designate it by the embedded bit b. Consequently, the
original-value of difference remains h = 101, which is equal to 5.

Mathematically,
b = LSB(H') = 1
H = H' / 2 = 5
 29

By utilizing the integer average-value (L') and the restored difference-value (H), it is possible
to precisely recreate original-pair (x, y). We have encoded a single bit b by extending the
size of the difference-value (H) from 3 bits of length (when H=5) to the 4 bits of length
(when H' = 11).

The reversible data-embedding execution, h' = 2h + b, is called DE.

7.3 IT-based RDH Technique

The primary objective of IT based RDH technique is to modify a sequence of pixels at the
same time in order to embed multiple-bits. In the embedding process, it is possible to modify
a pixel sequence (p1,p2,...,pn) as a single entity in order to incorporate n − 1 bits. When n =
2, DE can be considered as a specific case of IT. Based on the citation [9], the specified
pixels (p1, p2, ..., pn) are recognized as

The message bit, represented as mj, belongs to the set {0, 1}. The jth index range starts from
1 and ends at n-1, and the ith index ranges starts from 1 and ends at n. When the value of n is
more than 2, it is evident that an information technology-based resource discovery and
handling (RDH) strategy may offer a higher capacity compared to a data embedding (DE)
approach. For a large value of n, the maximum error rate is around 1 bit per pixel. The
succeeding works [72, 86, 101] significantly improve the performance in terms of ER and
PSNR. Furthermore, considering many pixels simultaneously enhances the stability of the
integer average, rendering it more resistant to alterations.

7.4 HS-based RDH Technique

Ni et al. [62] proposed the RDH with HS technique. This technique utilizes gray-scale
histogram to alter the pixels. In order to include the hidden information, the histogram's
topmost bins are selected for the expansion in HS. This involves grouping pixels based on
their grey values. There will be no ambiguity when reversible alterations are applied to the
pixels from different categories. Pixels can be classified into three broad categories: moving,
expanding, and unchanging pixels. They perform multiple jobs during the process of
embedding. Usually, only the shifting and enlargement of pixels are altered. The expansion
pixels during embedding will either have "±1" attached to them or remain unchanged,
depending on the message bits. Meanwhile, in order to guarantee reversibility, the moving
pixels will be appended with either "1" or "−1". The specific pixels (p1, p2, ..., pN) that are
designated for the cover pixel sequence (p1, p2, ..., pN) are determined.
 30

Fig 7.1: Flowchart explaining the embedding process in histogram shifting in HS-based RDH

Fig 7.2: Example of the modification of the pixels and shifting of histograms in HS based
RDH Technique
 31

Assuming that N is representing the number of pixels, m representing the message bit, and
(a,b) represent the two expansion bins, where a and b are both greater than 0 and less than
255. In order to achieve reversible embedding, it is important to store the parameters (a,b) as
side information. Usually, the bins of expansions, which usually have the highest peaks in the
pixel histogram, are selected to reflect the grey values that correspond to the largest
categories. Expanding pixels are pixels that have values within the range of a and b, while
shifting pixels are pixels that have values greater than b or less than a. This rule stipulates
that while embedding, the distributive nature of the pixel-value histogram will be changing in
a predictable manner. Pixels with values greater than b will have their histogram bins moved
to the right. The bins of histogram corresponding to pixels having values of b will extend to
incorporate an extra bit. The histogram bins are adjusted to the left for pixels with values less
than or equal to a. Figure 7.2 illustrates an instance of HS-based customization. The
expansion bins are chosen as (a,b) = (57, 59) at this location. The three expansion-pixels (57,
59, 59) are adjusted based on the bits with the secret (0, 0, 1), resulting in the pixels mark
(57, 59, 60) as per the HS-modification rule. Pixels highlighted in blue color are the pixels
which might be shifting. The values are incremented or decremented by 1 without encoding
any bits. The HS predictable modification approach enables receiver to detect bits which are
hidden and perform the recovery in a reverse way, meaning it can determine the cover pixels.

The embedded bits extraction depends on category of pixel indicated. Specifically, if the
pixel pi belongs to the set {a - 1, b + 1}, then m = 1. Otherwise, if pi belongs to the set {a,
b}, then m = 0. The approach of embedding of Histogram Shifting is simple, and more
instinctive as compared to the techniques of DE and IT. Alteration methods do not require
intricate mathematical calculations, and the performance may be directly assessed by
simulating Histogram Shifting before embedding. This approach possesses significant
flexibility. In general, the design of embedding of HS is efficient for features of statistics
derived from some other coefficients, such as histogram of quantified discrete
cosine-transform coefficients [32, 113] and histogram of deep convolutional-network
parameters [17], among others.

7.5 PEE-based RDH Technique

Prediction Error Expansion introduces prediction in the RDH, which further enhances the
performance of embedding by using correlation exploitation more effectively. The prediction
errors(PE) are sequentially calculated based on cover pixels and their adjoining pixels. This
process allows for the generation of a prediction error histogram (PEH).
 32

Fig 7.3: The comparison of pixel histogram and Prediction-error Histogram for Lena image.

The PEH is a mathematical function that defines the number of pixels gathered in a given
sequence. It is denoted as h(e) = #{1 < i ≤ N : ei = e}, where the function #{·} represents the
counting of pixels. The expression ei = pi − p∗ i represents the pixel's prediction error, where
pi is the actual pixel value and p∗ i is the predicted pixel value. Within the context of PEE,
the PEs will categorize the pixels based on their attributes rather than their gray values. The
produced Probability Density Function (PDF) closely resembles the Laplacian distribution.
The distribution of the PEH exhibits a higher degree of sharpness and concentration when
compared to the pixel histogram. Consequently, the compression of visual data results in a
more condensed representation, thereby creating additional capacity for embedding data.
Figure 7.3 presents a comparison between the histogram derived from PEs and pixels. PEH
exhibits a clear concentration at zero, with significantly larger values in its bins. The
PEE-based technique utilizes a similar approach to HS for the embedding process. One
distinction lies in the fact that PEE selects the expansion pixels based on the values of PEs.
The highlighted pixels (p1, p2, ..., pN) are obtained in the following manner:

The two expansion bins, a and b, are determined adaptively based on the distribution of PEH
and the provided capacity. The values of a range from -255 to 0, while the values of b range
from 0 to 255. It should be noted that the expansion bins in PEE are now defined as the
values of PEs instead of the pixel values. Just like in HS, the histogram bins of PEH will
 33

undergo predictable modifications during the embedding process. Figure 7.4 displays the
structure of the PEE framework. In contrast to the image depicted in Figure 7.2, the
significant difference is in the categorization of expansion pixels, shifting pixels, and
unaltered pixels based on the PEs. In the context of PEE, it is crucial to maintain the original
prediction for a pixel both before and after embedding in order to guarantee reversibility. The
recipient should acquire identical forecasts as those of the giver. An effective approach is to
ensure that the identical context utilized for prediction may be retrieved following the
embedding process. The local environment for the receiver can be made identical by
scanning and processing the designated pixels in reverse order, as indicated by the popular
predictors [8, 14, 90, 94]. The PEs are thus configured.

Fig 7.4: Example of modification of pixels and the histogram-shifting in the PEE RDH
technique.

Fig 7.5: Embedding modules of PEE, and the used enhancements.

The estimated data obtained from the annotated image can be utilized to accurately
categorize the annotated pixels. There are three instances to consider. In each case, the pixel
pi is noted and its PE (pixel error) is calculated as the difference between pi and p∗i.
Additionally, the parameters (a,b) are acquired from the side information.
The value of the cover pixels is restored to pi − 1 if its designated PE (Pixel Error) is greater
than b.
 34

The user's text is empty. The value of the cover pixels is preserved as pi itself, provided that
its designated PE falls within the range of a and b. No information provided. The value of the
cover pixels is restored as pi + 1, when its designated PE is smaller than a. Similarly, the
embedded bits will be extracted as "1" if ei belongs to the set {a - 1, b + 1}, and extracted as
"0" if ei belongs to the set {a, b}. PEE combines the benefits of both DE and HS. The
computational complexity is deemed acceptable, and the embedding efficiency is higher. The
collection of embedding modules offers researchers numerous chances for exploration and
enhancement. Figure 7.5 depicts the progression of PEE and its associated enhancements.

7.6 PEE-based RDH Technique Extensions

Within PEE-based RDH approaches, primarily two types of enhancements are there: precise
prediction [8, 10, 11, 14, 28, 29, 55, 58, 66, 90, 102, 139, 140] and adaptive embedding
strategy [34, 47, 48, 89, 90]. One way to improve is by increasing the accuracy of the
predictor, resulting in a more precise distribution of the generated PEH. In the PEE-based
architecture, a higher number of embeddable pixels indicates a more precise prediction based
on the derived statistical data.

Predictors Description MSE Var Mean

MED [94] Median Edge Detector(MED) is a widely 182.86 125.13 6.34

utilised model. It is based on the correlation
between the target and its three nearby pixels.

DP [95] The prediction produced by the DE-based 178.07 152.68 4.38

approach. In this case, the pixel next to the
target predicts it.

GAP [8, A widely used predictor is the 355.56 262.76 8.39

14] gradient-adjusted predictor (GAP). In contrast
to MED, it makes use of the more intricate
aspects for forecasting. The utilised
neighbouring pixel count has been raised to
seven. To adaptively modify the prediction,
the local gradients are computed.

RP [90] The widely utilised predictor, known as the 119.17 79.10 5.37
rhombus predictor (RP). It provides a precise
forecast by averaging the values of the four
neighbours.

CNNP-1 The Hu and Xiang proposed CNN-based 81.05 53.52 4.16

[28] predictor. By designating the picture material
as two non-overlapping sets, it creates the
training sets. To forecast the pixels from
another set, all of the pixels from one set are
fed into the model.
 35

Predictors Description MSE Var Mean

CNNP-2 The CNNP-1 enhanced predictor. The target 41.30 27.72 2.80
[29] image is split into four sections by it. The
prediction is carried out taking into account
the worldwide context.
Table 7.2: Differentiation for components of Existing Predictors of PEE

It enhances the effectiveness of embedding and reduces the distortion caused by excessive
shifting. A precise predictor has the ability to provide a histogram that is more focused. This
implies that the initial content gets condensed into a representation with reduced entropy. It
has the potential to counterbalance the rise in distortion caused due to embedding. Gathering
suitable adjacent pixels [10, 140] and formulating the precise frameworks [8, 11, 14, 28, 58,
66, 115] are useful techniques to improve the accuracy of the prediction. Ou et al. [66]
employed the partial differential equation and the pixel gradient to determine more
appropriate weights for the adjacent pixels. Dragoi et al. [10, 11] utilized various prediction
models to generate the target image and applied the least squares method to acquire an
accurate answer. Ridge regression is proposed in Reference [102] to enhance the
generalization capacity of the training weights. Searching for the proper coefficients requires
more time as the price. In recent times, the convolutional neural network (CNN) has emerged
as a potent technique for predictors implementation. Luo et al. [55] initially identified that
the CNN can effectively capture the non-linear mapping between the pixel of the target and
its surrounding pixels in its neighborhood. A model was trained to predict stereo pictures and
its exceptional performance was proven. Subsequently, Hu and Xiang [28, 29] developed a
CNN-based predictor for grayscale images. They extensively researched about the pixels
allocation, the training method, and a lot of other intricate issues. Zhang et al. [139] and Yang
et al. [120] implemented improved predictor based on CNN, each employing distinct
network architectures. Table 7.2 presents the comparative outcomes for components of the
current predictors. Here, we give the average, variance, and mean square error (MSE). The
initial two outcomes are computed using the absolute values of the acquired PEs, whereas the
MSE is determined by comparing the target pixels with the predicted values. The
experimental data, obtained from Reference [29], are derived from the analysis of 500 test
photographs. Furthermore, the inclusion of the asymmetric PEH alteration [7, 40, 41] is also
a significant expansion. In contrast to prior PEE-based approaches, this technique proposes
the generation of a skewed histogram by utilizing multiple extreme predictions. Figure 7.6
depicts the skewed PEH that was developed, along with its modification framework.
 36

Fig 7.6: An example of the asymmetric PEH's histogram adjustment.

The histogram referred to be "skewed" exhibits an unbalanced distribution. Specifically, it

comprises a summit and two asymmetrical flanks. The more acute edge is referred to as the
short tail, while the opposite side is known as the long tail. In this approach, the embedding
is completed twice in a like manner. During each layer of embedding, only the side with the
smaller length is altered for the purpose of embedding. An advantage of this strategy is that
any modifications made in the initial stage can be restored in the subsequent step. This
implies that the ED can be dynamically calculated and modified while performing the
embedding process. An further method to enhance PEE is to dynamically choose the suitable
pixels for alteration. It should be noted that this type of enhancement is not limited to
PEE-based extensions. Furthermore, there are other systems that exhibit comparable
enhancements. Here, we shall exclusively utilize PEE as an illustration, as it is presently the
most prevalent RDH technique. Sachnev et al. [90] suggested arranging the pixels according
to the magnitudes of their local variations. The objective of this strategy is to utilize the
smooth pixels for embedding to the maximum extent possible. When it comes to
reversibility, the method uses the local variances instead of the values of PEs to choose
elements. Pixels with lower local variances are prioritized for embedding, resulting in a
considerable reduction in embedding distortion. Empirical evidence has demonstrated the
efficacy of this sorting strategy in reducing distortion, particularly in cases of low-capacity
embedding. Additionally, a current area of focus in research [47–49] is to dynamically
determine the expansion bins or modification method based on the provided image content.
Li et al. [47] suggested categorizing image pixels into two groups based on their smoothness
levels: pixels from smooth regions and pixels from textural regions. Smooth regions will
have their expansion pixels encoded with two bits, as their PEs tend to be less. The texture
pixels will retain a single-bit embedding. This approach inherently classifies pixels in a more
detailed manner, leading to a more efficient utilization of redundancy. Table 8.1 provides a
concise overview of the current adaptive embedding methodologies. In the following section,
we will focus on some popular extensions, namely MHM [27, 49, 60, 69, 78, 96–98, 103,
105, 108], pairwise PEE [5, 12, 15, 65, 68, 84, 114, 131], and PVO [21–23, 44, 46, 61, 63,
64, 67, 70, 73, 88, 99, 104, 112, 130, 133]. Two sample ways of adaptive embedding are
MHM and pairwise PEE. MHM achieves multiple-level embedding, while pairwise PEE
achieves deep-level [Link] helps create a more accurate predictor by locally
organizing the pixel values for prediction.
 37

CHAPTER-8 RDH TECHNIQUES FOR ADAPTIVE EMBEDDING

In the subsequent section three widely used RDH schemes and their extensions: MHM ,
pairwise PEE , and PVO. These schemes represent advancements over the traditional
approaches outlined earlier and possess their own distinct technological features.
Furthermore, the presentation of these typical techniques will be provided in the end.

Method Country Number Year Description

of
Citations

Pixel sorting South Korea 967 2009 The sorting approach

[90] mostly modifies the smooth
pixels.

Pixel selection China 711 2011 Additional covert

[47] information is concealed
inside the fine details of the
pixels.

Multi-level China 94 2014 The pixels are divided

embedding [18] according to their
smoothness levels. Each
level is assigned a distinct
embedding capability.

MHM [49] China 342 2015 First we make different

PEHs to set the suitable
parameters for each
sub-PEHs. Different
approaches are used for
pixels based on the local
context.

Pairwise PEE China 494 2013 The PEH modification is

[68] enhanced in the
two-dimensional space by
leveraging the correlations
of second order between
pixels.

PVO [46] China 485 2013 Choose the starting and

ending points from the
sequence of pixels which
are arranged from every
single block for insertion.

AGM [89] Malaysia 40 2016 The PE with a high

frequency is embedded
 38

Method Country Number Year Description

of
Citations

with more than one bit and

the low-frequency PEs are
discarded for embedding.

Reduce shifting China 135 2019 The excess moving pixels

[34] are removed in an adaptive
manner based on the
surrounding environment.
Table 8.1: An examination of the existing methodologies for adaptive embedding.

8.1 Multiple Histograms Modification

MHM is a comprehensive framework that encompasses the principles and concepts of PEE.
Within the context of MHM, the classification of cover pixels is determined by considering
both the properties of PEs (Prediction Error) and the degrees of texture. In PEE-based
approaches, the smooth pixels typically have PEs with values close to zero. This is due to the
fact that smooth regions typically exhibit strong correlations. Therefore, it is more feasible to
acquire precise forecasts for the seamless pixels. The Prediction Error Histogram (PEH)
derived from those places exhibits a more focused distribution. By including the message
into the seamless pixels, the embedding process will experience reduced distortion. A viable
and logical approach is to assign separate expansion bins to pixels with different amounts of
texture. In Reference [49], the N cover pixels (p1,p2, . . . ,pN ) are classified into M
categories. Each category is individually handled using distinct parameters. Initially, the
pixels that possess an equivalent amount of complexity are gathered and grouped in the
corresponding subset. Now to determine the complexity level, we analyse the specific
circumstances and conditions of the given situation. Furthermore, each subset will generate a
Prediction Error Histogram (PEH) consisting of two expansion bins. To clarify, the original
PEH will be partitioned into M sub-PEHs that fulfill the condition h(e) = Σhc(e) for c ranging
from 1 to M. In this context, "hc" refers to the sub-PEH that counts the PEs with complexity
of "ci", where "c" belongs to the set {1, 2, ..., M}. Consequently, the cover pixels are
categorized based on the pair (ei, ci). Furthermore, by the implementation of exhaustive
search, the most optimal expansion bin set {(ac, bc)}M c=1 is determined adaptively, taking
into account the specific characteristics of each subset. Thus, each subset will have its own
distinct mappings for the pixels. This constitutes the modification model with several tiers.
Figure 3.1 displays an illustration of the embedding process based on MHM, as described in
reference [49]. The cover pixels are assigned varied gray levels based on their histogram
indexes. The alterations for the cover pixels corresponding to the sub-PEH hc can be
summarized into five scenarios, based on the expansion bins {(ac ,bc )}M c=1 for each
subhistogram.
—When the value of ei is greater than bc, the cover pixel is incremented by 1 to shift the
space.
 39

—The cover pixel is enlarged to pi + m in order to encode one hidden bit, when ei is equal to
bc.
— The cover pixel remains intact when the value of ac is less than ei and bc.
—The cover pixel is enlarged to pi −m in order to encode a single secret bit, only when ei is
equal to ac.
—When the value of ei is less than ac, the cover pixel is adjusted by subtracting 1 to create a
space.

Fig 8.1: The data embedding example of MHM.

Alteration of the pixels are done from various levels that is based on their corresponding
embedding parameters. Similarly, it is imperative to assign a minimal storage capacity
capacity for storing the chosen parameters. In the course of the extraction procedure, the
cover pixels are restored in a manner that is on the opposite side. Generally, MHM exhibits
flexibility by modifying the set of expansion bins. The precise alteration can be achieved by
choosing appropriate expansion bins for the pixels from various texture regions.
Subsequently, Ou et al. [69] and Qi et al. [78] made further enhancements to [49] by
extending its framework to a more comprehensive structure. The cth histogram will find Kc
pairs of expansion bins {(ac,k ,bc,k )}Kc k=1 that fulfill the conditions ac,k < ... < ac,1 ≤ 0 ≤
bc,1 < ... < bc,k. During the process of embedding, the alterations made to the cover pixels in
each sub-PEH are categorized into five distinct instances. The distinction lies in the
adjustment of the shifting step size for each pixel. Specifically, the identified pixels are
determined in the following manner:
—Case 1: If the value of ei falls between bc,k and bc,k+1, the cover pixel should be
relocated to the right by k steps, meaning that pi = pi + k.
—Case 2: If the value of ei is equal to bc,k, then the pixel is enlarged to obtain the designated
one by adding (k − 1) + m to pi.
—Case 3: If the value of ei falls between ac,1 and bc,1, the cover pixel remains unaltered.
—Case 4: If the value of ei is equal to ac,k, then the marked pixel is calculated as pi minus (k
minus 1) minus m. The user's text is empty.
—Case 5: If the value of ei is between ac,k+1 and ac,k,then the pixel in the cover should be
moved to the left by k steps, so that pi = pi − k.
 40

In the References [69, 78], it is possible to include additional hidden messages in each
histogram. The maximum achievable error rate (ER) reported in Reference [69] is 0.612 bits
per pixel (bpp), which is considerably greater than the ER of the approach proposed by Li et
al. [49] (0.176 bpp). Several other extensions have been suggested because MHM works so
well. Table 8.2 provides a concise overview of the current MHM-based methods. The
enhancements for MHM can be primarily categorized into two types: pixel division and
parameter optimization. Initially, employing a more efficient approach to division could
guarantee that the recurring elements are consolidated together to the greatest extent feasible.
Choosing appropriate expansion bins is advantageous. Furthermore, a meticulously crafted
expansion bin selection procedure has the potential to discover a more efficient method of
change. In addition, optimizing the parameter search technique is also driven by the desire to
decrease computational complexity. The heuristic pixel division method employed in [49] is
not adaptable. They propose utilizing fuzzy c-means clustering to generate the PEH instead.
It effectively constructs numerous more precise distributed PEHs. After that, Wang et al. [96]
pointed out that bin selection in MHM is the process of finding the right payload for each
suboptimal performance improvement.

Research Generation of Selection of Complexit Maximum PSNR

Work PEH Bins y of ER
method

MHM [49] Determine the Exhaustive O(218) 0.25 bpp 61.02 dB

computational Finding with
complexity by limitations in a
considering the specified
12 neighboring range.
elements.

Wang et al. Use fuzzy Same as in O(218) 0.18 bpp 61.15 dB

[97] c-means Reference
clustering. [49].

Ou et al. [69] Same as in Discover O(220) 0.8 bpp 61.01 dB

Reference [49]. numerous sets
of containers
by conducting
a thorough
search for
extensive
solutions.

Qi et al. [78] Same as in Get series of O(227) 0.64 bpp 61.04 dB

Reference [49]. bins for each
PEH using
dynamic
programming.

Wang et al. Allocate the Assign the - 0.7 bpp 60.92 dB

 41

Research Generation of Selection of Complexit Maximum PSNR

Work PEH Bins y of ER
method

[96] acquired payload to

complexity each sub-PEH
levels in equal using an
proportions. evolutionary
process.

Hou et al. Use a deep Expand the O(226) 0.24 bpp 61.46 dB
[27] neural network search area
to learn a and utilize
suitable memos to
clustering enhance the
model. pace of the
search.

Weng et al. Use the Use the - 0.15 bpp 61.51 dB

[105] k-means improved
clustering. criss-cross
optimization
algorithm.

Wu et al. Same as in Derive the - 0.61 bpp 61.00 dB

[108] Reference [49]. mapping based
on the
reversible
embedding
function
Table 8.2: Overview of the current MHM-based Methods

In this context, the techniques of pixel partitioning and the choice of expansion bins in each
project are briefly presented. The PSNR is computed by evaluating the outcome of
embedding 10,000 bits into the test image of Lena.

The evolutionary method was used to solve the ER allocation model that was [Link] EC of
MHM is improved by using this method, which makes it possible to get a better PSNR for a
certain data payload. In their study [105], Weng et al. developed the k-means clustering
method for categorizing PEs, which made the clustering process more accurate and faster. In
addition to this, they found that there was a connection between the utilization of expansion
bins and the traditional multi-choice knapsack problem. An improved version of the
crisscross optimization algorithm has been used in place of the searching approach. The
generalization of histogram modification-based RDH was investigated by Wu et al. [108],
who also established the reversible embedding function as a mathematical representation for
the generalization. There is a possibility of determining the expansion bins by utilizing the
results that are obtained from the function. This approach is more flexible than the standard
 42

optimization strategy since it takes into account a larger number of potential solutions. Ma et
al. [60] devised an efficient approach for determining expansion bins. The initial
optimization objective is represented as a differentiable function with real values, and
sophisticated analysis tools are employed to solve it. The significant reduction from 41.7
seconds to 0.05 seconds was observed due to this approach in average duration

8.2 Pairwise PEE

In the traditional PEE [90, 94], pixel modification is determined using low-dimensional
statistical analysis of the picture data. The mapping derived from the one-dimensional
Prediction Error Histogram (PEH) lacks the ability to fully exploit the intricate interaction
between neighboring PEs. Pairwise PEE [68] provides a viable solution by taking into
account the alteration in a space with more dimensions. This approach initially groups
adjacent Prediction Error (PEs) in pairs to categorize the pixels and produce a
two-dimensional PEH. Next, the expanding pixels, shifting pixels, and unaltered pixels are
identified by creating modification rules for the 2D PEH bins, specifically by building a 2D
mapping. Next, we will provide a concise overview of the particular structure outlined in
approach [68] and examine its distinctive features.

Case Mapping directions Condition

(a) (x,y)→{(x+1,y),(x,y),(x+1,y x=b,y=b

+1),(x,y+1)}

(b) (x,y)→(x,y+1) x<b,y>b

(c) (x,y)→(x+1,y) x>b,y<b

(d) (x,y)→(x,y) x<b,y<b

(e) (x,y)→(x+1,y+1) x>b,y>b

(f) (x,y)→{(x,y+1),(x,y)} x<b,y=b

(g) (x,y)→{(x+1,y+1),(x+1,y)} x>b,y=b

(h) (x,y)→{(x+1,y+1),(x,y+1)} x=b,y>b

(i) (x,y)→{(x+1,y),(x,y)} x=b,y<b

Table 8.3: The Conventional PEE's Nine Potential PE Pair Mappings
 43

Fig 8.2: Two-dimensional mapping comparison, with the standard PEE expansion bin is
denoted by the symbols b = 0.

In the context of paired PEE, the sequence {(e1, e2), (e3, e4),..., (eN−1, eN)} is created by
merging the two PEs that are next to each other. In a similar manner, the PEH is generated,
and the mapping is made in accordance with its distribution. The distinction lies in the fact
that the PEH is produced inside a 2D coordinate system, specifically expressed as h(x,y) =
#{1 < i ≤ N/2 : e2i−1 = x, e2i = y}, where h represents the frequency count for the PE
pairings of (x,y). Next, the expansion bins are dynamically chosen to create a
two-dimensional mapping. While the embedding process is being carried out, the pixels that
are adjacent to one another will undergo collective changes that will either conceal or shift a
single bit. This method enables greater flexibility in the directional shift of each PEH bin.
The modification approach of pairwise PEE is compatible with 1D mapping-based methods,
serving as an extension of PEE. The alteration of the traditional PEE, as specified in
Equation (4), can also be elucidated in the two-dimensional space. For the standard PEE,
there are nine possible mappings for PE pairings (x,y). The items are enumerated in Table
8.3. It should be noted that only the mapping for the PEs with positive values is shown for
the sake of being concise. The authors in Reference [68] devised an alternative modification
model that relied on an examination of embedding efficiency. The clear enhancement is case
(a) specified in Table 8.3. For the pair (b,b), the pairwise PEE converts the mapping outputs
to {(b,b), (b + 1,b), (b,b + 1)}. Meanwhile, the pair (b + 1,b + 1) is chosen as the expansion
bin, and the self-mapping is included in its outputs. Figure 8.2 depicts two distinct 2D
mappings. The pairwise PEE [68] utilizes a 2D mapping that presents a novel modification
framework.
 44

Fig 8.3: The alteration of pixels using pairwise pixel error estimation (PEE). (a) The process
of embedding in the paired PEE. (b) The entire process of mapping the pairwise PEE.

Fig 8.4: The 2D mapping generated using the adaptive approach for the Lena and Airplane
test image with an EC value of 10,000 bits.

The performance of the recently developed 2D mapping has been exhibited [68]. According
to the established 2D mapping, it is conceivable that the covert message is no more encoded
in binary format. Figure 8.3 displays the data embedding of paired PEE. The paired PEE
offers a novel approach to representing data in the context of RDH (Reversible Data Hiding)
design. The high-dimensional-based modification has gained significant interest, and
numerous 2D RDH approaches have been presented. Subsequent studies have demonstrated
that adaptively modifying the expansion bins based on the provided histogram distribution is
beneficial for enhancing performance. The proposed approach involves producing adaptive
2D mapping based on the given strategy [5, 15, 65, 84, 114, 131]. Ou et al. [65] introduced
two techniques for creating a 2D map of a given image. The initial approach entails
 45

exploring a substantial number of potential solutions. The distortion-capacity ratio of each

candidate can be compared in order to determine which mapping technique is the most
effective. The utilization of the optimal probability matrix is an alternate method that can be
utilized. On the basis of the evaluation of the mapping probability, the final modification
model is validated. The modification method is depending on the image, allowing for
improved performance. Chang et al. [5] and Zhang et al. [131] sequentially developed an
adaptive 2D mapping technique. In the study cited as Reference [5], the ideal
two-dimensional mapping is initially set as a random solution and then determined using an
iterative updating method.

Work Year Complexity PSNR Description

Pairwise PEE 2013 O(210) 59.75 dB 1. The first 2D

[68] mapping-based scheme.
2. The 2D mappings are
constant for various pictures.

M2D-PEHs 2019 O(218) 60.42 dB Create a total of eight

[84] different 2D mappings, and
then choose the mapping
combination that will
produce the best results for
the MHM.

Hybrid PEE 2019 O(210) 60.77 dB 1. The PEs are

[65] adaptively paired.
2. The optimal mapping
is searched in a certain
region.

OPPEE [114] 2019 O(225) 61.00 dB Convert the problem of

optimizing 2D mapping into
the task of determining the
most efficient routine, then
solve it using dynamic
programming.

APPEE [5] 2021 O(210) 61.35 dB Apply the random iteration

approach to modify the 2D
mapping in the region of size
8 x 8.

SMPPEE [15] 2021 O(218) 61.58 dB Create a modified version of

skewed two-dimensional
histogram and implement it
on many skewed PEHs.

AMPPEE [131] 2022 O(218) 60.93 dB 1. The mapping outputs

of the 2D PEH bins are
 46

Work Year Complexity PSNR Description

adjusted in a pre set order.

2. The MHM
framework is utilized.
Table 8.4: Summary of the Adaptive Optimization Algorithms for the Pairwise PEE

In this case, the PSNR is computed on the result when embedding ten thousand bits into
Lena(test image).

Two examples are presented here in Figure 8.4. The two-dimensional mapping can be
modified according to the content of the image. An alternative approach was proposed by
Zhang and Ou [131] as a means of improving the mapping optimization. The method entails
first sorting the acquired PE pairings according to their frequencies and placements, and then
modifying the mapping inputs of each of these pairs on an individual basis. It was suggested
by Fan et al. [15] that the pairwise PEE could be combined with a skewed histogram in order
to improve performance. In addition, they made use of the distorted PEH-based
two-dimensional mapping that was included in the MHM framework, which allowed them to
carry out modifications that were more highly effective. The authors of References [3, 4]
have recently created a method to expand the pairwise modification methodology to color
images and create a related optimization strategy for 3D mapping. The mapping generation
model in Reference [3] is trained using the reinforcement learning approach known as double
deep Q-network. Table 8.4 provides a summary of different adaptive optimization algorithms
for the paired PEE. Furthermore, the utilization of the adaptive pairing strategy [12, 65] is a
highly successful method for improving performance. In the initial pairwise PEE [68], the
pairs of PE are created by simultaneously examining the two neighboring pixels. The
combination is determined solely by the positions, namely the diagonal or anti-diagonal
directions. Citations [12, 65] indicate that this method could result in unsatisfactory
matching, especially for pixels located in certain areas such as the edges. This is because
spatial distance does not always effectively represent pixel similarity. This is the reason for
this. In their study, Dragoi and colleagues [12] suggested that the PEs should be matched not
based on their geographical position but rather on the gray values that they possess.
Amplification of the frequencies of PEH bins that are clustered around zero is achieved
through the utilization of this approach. Hybrid-dimensional histogram modification is the
name of the innovative embedding framework that Ou et al. [65] created for the paired
parametric equation equation. Within the framework of this strategy, the pixels that are
associated with the smooth region and the texture region are processed using different
methods. Only the contiguous pixels with a uniform texture are merged into pairs and
utilized for the process of embedding. The reason for this is that smooth pixels typically
provide a histogram with a more distinct distribution. The texture pixels are gathered to
create the one-dimensional PEH. The objects will undergo a 1D mapping transformation,
without any bit embedding. This method of pairing results in a more accurate categorization
of the pixels. The resulting 2D PEH exhibits decreased entropy and a further reduction in
embedding distortion.
 47

8.3 Pixel value ordering

PVO is a method that makes predictions by taking into account the similarity of pixels inside
a block. In contrast to the approach of producing predictions for each individual pixel, Li et
al. [46] suggested utilizing only the maximum and minimum pixel values within a block for
data embedding. The maximum and minimum pixels are determined by the penultimate
greatest and smallest pixels, respectively. The prediction takes into account both the
coordinates of the pixels and their grayscale values, resulting in a higher level of accuracy. In
addition, just two pixels are altered inside a block, resulting in a lower level of distortion
compared to traditional PEE approaches. Given the similarity of the embedding processes,
we will focus solely on explaining the modification of the maximum for the sake of
simplicity. The pixels in the block (p1, p2, ..., pn) will be arranged in ascending order
depending on their gray values, resulting in the sequence pσ (1) ≤ pσ (2) ≤ ... ≤ pσ (n). Here,
the mapping σ provides a bijective relationship between the new index and the original one.
The new order of two pixels with identical values is determined by their original indexes, so
that σ(i) < σ(j) when i < j. Within the ordered sequence, only the pixel with the highest value
is anticipated and altered. Firstly, it employs the closest element as the forecast, and the
prediction error (PE) is computed as emax = pσ (n)−pσ (n−1). It is evident that the PEs
(Potential Energy) in PVO (Potential Energy of an Object) are always larger than or equal to
0. Furthermore, a PE with a value of 1 is utilized for expansion purposes. The secret bits are
embedded by modifying the biggest pixels in a reversible manner. More precisely, the pixels
with pixel elements labeled as "0" will remain unaltered and will not be included in the
embedding process. The pixels with PEs "1" are enlarged as pσ (n) + m in order to hide the
secret bit. If the associated pixel error exceeds 1, the pixel is displaced and increased by
[Link] that the maximum pixel can only be increased by 1 or remain the same, the
maximum value will always be the largest one within the designated block. The order of
pixel values in the marked block remains unaltered. It ensures the ability to be reversed. The
restoration and extraction of the indicated pixel are performed in a reverse manner. Figure
8.5 illustrates the process of data embedding and extraction using the PVO approach [46].

Fig 8.5: Diagram depicting the process of embedding and extracting data in a PVO-based
PEE system.

Within the PVO framework, the process of identifying the most comparable pixels in a block
involves arranging the pixel values in a sorted manner. It has the ability to acquire precise
forecasts and produce the PEH with reduced entropy, hence enhancing the efficiency of the
 48

embedding process. When implemented with limited memory, PVO generally shows better
performance in terms of peak signal-to-noise ratio (PSNR). Several improvements have been
proposed for PVO, including improved PVO (IPVO), PVO-k, k-pass PVO, pixel-based PVO
(PPVO), pair-based PVO, MHM-based PVO, adaptive block complexity calculation, and
dynamic block size.

8.3.1 IPVO [73]

In IPVO, the calculation of PEs is redefined considering the original location of the target
pixel and its prediction. The PEs are calculated for the sorted block pixels (pσ (1),pσ (2), . . .
, pσ (n)) in the following manner:

Fig 8.6: Comparing the alteration of the PEH for PVO and IPVO during the embedding
process.

1. The prediction can be determined by calculating the difference between the current value
and the previous value, given by the PVO equation: Emax = pσ (n) −pσ (n−1).

2. If the initial index of the maximum is greater than the predicted index, then the PE value is
negative, especially Emax = pσ (n−1) − pσ (n). The PEH obtained using this definition
closely resembles the distribution observed in traditional PEE analysis, especially the
Laplace-like distribution centered at zero. Figure 3.6 shows a comparison of histogram
transformation strategies used for PVO [46]. and IPVO [73]. It is possible that the migration
of PEs can occur in two different directions in IPVO due to the symmetrical distribution of
the histogram. It was decided that expandable PEs would be placed in the top two positions,
specifically PE "0" and PE "1". During the embedding, only those pixels with PEs of "0" or
"1" are selected to carry secret bits. of. These pixels are used to encode a binary value of "0"
or "1" by either preserving the original value or increasing it by 1, respectively. For pixel
errors more than 1 or less than 0, the relevant pixels are moved to the right, meaning they are
increased by 1, in order to guarantee reversibility. It is important to observe that in IPVO, the
direction in which the PEs change is distinct from the direction in which the pixels move.
 49

Regarding the cover pixels, they undergo either an increment of 1 or remain unaltered during
the embedding process. In IPVO, several PEs are employed to transport the secret bits, as
opposed to standard PVO. The classification process is more sophisticated, enabling a greater
payload capacity and improved visual quality.

8.3.2 PVO-k [67]

PVO-k is an extended version of PVO with more features. It considers a greater number of
situations. In traditional PVO (Pixel Value Ordering), only the last pixel in the sequence (pσ
(1), pσ (2),... ,pσ (n)) is used for embedding. However, there is one exceptional scenario
when there are multiple pixels in one block that are the largest, especially when pσ(n-2) =
pσ(n-1) = pσ(n). The standard PVO technique results in the omission of multiple blocks
during the embedding process. In PVO-k, to solve this problem, the ability to change the k
largest pixels simultaneously to embed a single piece was implemented. For a block of pixels
satisfying the condition pσ (1) ≤ ... ≤ pσ (n−k) < pσ (n−k+1) = ... = pσ (n), there are k pixels
that are the largest. The prediction is made by selecting the second largest pixel. Those k
pixels can be assigned to the same PE where Emax is equal to pσ (n) minus pσ (n-k).
Choosing PE "1" as the expansion tank, the immersion and removal procedures are almost
the same as for PVO. It is important to note that the k largest pixels must be updated as a
whole to maintain reversibility. This means that each block still covers only one bit. The
authors of [67] showed that the flexible use of PVO-2 and PVO-1 can lead to better
immersion..

8.3.3 k-pass PVO [23]

The purpose of K-pass PVO is similar to PVO-k, which uses more than one pixel in each
block to add information. The difference is that multiple pixels selected by k-pass PVO may
not be identical. K-pass in the PVO algorithm with a pixel sequence satisfying the condition
pσ (1) ≤ ... ≤ pσ (n−k) ≤ pσ (n−k+1) ≤ ... ≤ pσ (n) . k of the largest pixels are assigned to PE.
The prediction starts with the highest value and continues for the rest of the time. Initially, a
pixel with index n - k is selected in order.

The maximum forecast and PE can be calculated using the formula en = pσ (n) − pσ (n−k).
The PEs of the remaining k − 1 largest pixels are determined based on en. A pixel pσ (i),
where i belongs to the integers n - k + 1 - n - 1, has two possible values ​for its PE. If the
maximum PE (pixel energy) is greater than 1, then the PE is calculated by subtracting the n −
k indexed pixel value from the maximum pixel value, i.e. ei = pσ (i) − pσ (n− k). No
information provided. Alternatively, the largest value becomes the forecast, i.e. PE is
calculated as ei = pσ (i) −(pσ (n) − 2).
 50

Fig 8.7: Embedding comparison of PVO-k and k-pass PVO.

With this prediction technique, each transformed pixel contains a prediction error (PE) and
the original sequence can be recovered after the embedding. During embedding, pixels with a
value of "1" in PEs are identified as extension pixels. Pixels with a PE greater than 1 are
classified as moving pixels, while the remaining pixels are classified as stationary. The
K-pass PVO approach involves inserting a single bit into each pixel instead of processing
them together. The maximum size of each block is k bits. Figure 8.7 illustrates the contrast
between PVO-k [67] and k-pass PVO [23]. In the PVO-k algorithm, the two largest pixels at
coordinates (120, 120) are incremented at the same time (121, 121) to include the hidden bit
"1". In K-pass PVO, the pixels located at coordinates (120, 120) have the binary values ​"1"
and "0", respectively, and are the two largest pixels in the image. In the K-pass PVO, it is
obvious that the arrangement of marked pixels is not the same as that of the mask pixels.
Specifically, a pixel with a value of 120σ (5) increases to 121, becoming the largest in the
labeled sequence. However, the result of the extraction remains unchanged because the
sought bits can still be fine-tuned to their original position before sorting..

8.3.4 PPVO [88]

As part of the PPVO methodology, the PVO prediction model is combined with standard
predictors used in PEE calculations. Specifically, each pixel contained in a block is assigned
to a processing element for processing. At some point during the prediction process, the
target pixel and its neighboring pixels are grouped together and combined into a block of
pixels. If an aligned pixel in a block is the largest element, it is assumed to be the largest
based on the pixel next to it. If the element is the smallest, this prediction results in the
smallest value of the elements next to it. Alternatively, if this pixel is not modified, it will be
set as an unchanged pixel during embedding. An exceptional scenario occurs when all
adjacent pixels have the same values, in which case the target pixel becomes the smallest of
them. In this scenario, the expected target pixel is determined by subtracting one of the
largest values. The type of change selected in the embedding process can be determined by
examining the relationship between the target pixel and its neighboring pixels. If the current
pixel matches the largest or smallest value of nearby pixels, indicated by PE = 0, the pixel is
expanded to transmit information. If the value of the target pixel is greater than the maximum
value of the pixels next to it, the value of the target pixel is increased by one to move it.
Similarly, if the current pixel is smaller than the minimum element, it is decremented by one.
 51

When using PPVO, the hiding of additional bits is improved due to the increase in the
number of embedding pixels.

8.3.5 Pairwise PVO [63]

The diagrams above show that expansion tanks are permanently selected as either PE "1" or
"0". It has no adaptive variants. 2D mapping is a great adaptive approach that offers more
possibilities to change the PVO, which greatly improves performance. Ou et al. [63]
extended the PEE to partial variance optimization (PVO) considering the two largest
elements of each block pair. The third largest pixel of the ordered pixel sequence pσ (1) ≤ pσ
(2) ≤ ... ≤ pσ (n) is used to estimate the values ​of pixels pσ (n) and pσ (n−1). To ensure
reversibility, it is necessary to set limits on the binding queue. The PE pair (e1 max , e2 max
) is formed considering both positions and gray values. The first PE of the pair is calculated
as the difference between the lower index, represented by e1 max = pu − pσ (n−2), where u is
the smallest value between σ(n) and σ(n − 1) . The second PE of the pair is determined by
subtraction with the higher index, specifically e2 max = pv − pσ (n−2), where v = max (σ (n),
σ (n − 1)). Using the generated 2D description, it is possible to change the pixels of the block
in this way. The four mentioned approaches, namely PVO [46], IPVO [73], PVO-k [67] and
k-pass PVO [23], can also be explained in 2D space. Figure 8.8 shows a comparison of these
two-dimensional mappings. In addition to the improvements discussed in Section 3.2, several
studies have been conducted to improve pair PVO by optimizing 2D mapping [21, 130] and
using adaptive pairing [112, 133].

Fig 8.8: Two-dimensional mapping comparison for (a) PVO, (b) IPVO, (c) PVO-k when k =
2, and (d) k-pass PVO when k = 2.
 52

8.3.6 MHM-based PVO [64]

Studies have shown that MHM is an adaptive delivery method that is proving to be quite
successful. In light of this, better performance satisfaction can be achieved by extending the
MHM framework to PVO. For N-mask pixel blocks {B1, B2,..., BN}, it is possible to create
a one-dimensional histogram by collecting the highest value of each block and storing it in an
array. Here Bi = (pσ (1), pσ (2), ... . , pσ (n)). Conventional techniques typically classify a
pixel pσ(n) based on its PE, namely emax. After examining the histogram distribution, the
best embedding values ​are found in an adaptive way. The 1D histogram produced by the
MHM-based PVO method is then divided into many smaller histograms based on the
complexity measure. In reference [64], block complexity ci is defined as the largest
difference of the Nth block. In other words, ci is equal to pσ (n−1) minus pσ (1). Based on
the pair (emax, ci), the largest numbers are placed in different sub-histograms. Then, the
correct set of expansion bins is adaptively selected by looking at the distributions of the
collected sub-histograms.

8.4 Discussion

This section provides a detailed analysis of three commonly used RDH techniques: MHM,
paired PEE, and PVO. The aim is to present their advantages over traditional methods.

Test Image Predictor Var MSE PEH Mean

entropy

Airplane DP [95] 93.95 120.67 4.59 5.31

MED [94] 105.71 134.52 4.67 5.51

GAP [14] 31.87 43.81 4.12 3.53

RP [90] 41.84 56.24 4.19 3.81

PVO [46] 30.46 8.50 1.27 2.84

Lena DP [95] 87.05 127.57 5.05 6.48

MED [94] 72.71 106.16 4.91 5.89

GAP [14] 25.38 40.58 4.38 3.96

RP [90] 32.86 50.86 4.49 4.31

PVO [46] 23.78 7.22 1.39 2.99

Table 8.5. Comparative analysis of prediction accuracy between commonly utilized
predictors and PVO on the Airplane and Lena Test Images, both having a size of 512x512.

In relation to the adaptive delivery strategy, MHM-based techniques can be considered a

logical progression of the strategy. Does not affect the class label applied to certain moving
 53

pixels in previous methods (such as pixel sorting and adaptive selection of extension
platforms). This helps reduce the number of unnecessary exchanges. These procedures can
be considered complex changes. However, these procedures are usually performed using
global statistical properties. This contrasts with a more subtle change. The problem arises
when a significant number of pixels that can be embedded are also classified as immutable.
To solve this problem in MHM, this approach involves the use of clustered local statistical
functions for pixel distribution. As a result, the original histogram is simultaneously
transformed into a large number of sub-histograms. As a result, management occurs at
multiple levels. Finding a more appropriate method to modify a local property is possible.
PEE is an alternative way of making changes on many levels. To maintain translatability, 1D
map adaptations must adhere to certain constraints. This puts some limitations on some of the
more advanced changes. In paired PEE, the use of 2D mapping allows greater freedom to
change the embedding in different directions. This allows fine-tuning of pixels in 2D
mapping, enabling multiple levels of customization. The reason why paired PEE can achieve
higher performance is as follows. PVO uses the inherent properties of local visual
information to make more accurate predictions, rather than relying on made-up formulas.
The two largest pixels in a given pixel block are naturally close to each other. Classical
predictors [14, 90, 94, 95] often select surrounding pixels from a local block for prediction.
In the same block, adjacent pixels with the largest value are all less than or equal to the
second largest value. This improves the accuracy of the PVO forecast by targeting it more
precisely to the specific local context, resulting in a more reliable estimate. The block size in
PVO is defined here as 3 × 3. From the findings in Table 8.5, it is clear that the mean and
variability of the absolute prediction errors obtained by the PVO method are significantly
smaller compared to the other four predictions. The statement suggests that PVO has the
ability to produce smaller PEs that provide more accurate predictions. Similarly, the mean
squared error, commonly known as MSE, between the pixels and the predictions produced by
PVO is the smallest. Regarding the histogram distribution, the entropy of PEH produced by
PVO is significantly lower than that of conventional PEH. Based on this, it seems that its
histogram is clearer. Table 8.6 shows a comparison of the performance of PVO and
conventional predictors in predicting the maxima of each block. The PVO algorithm is able
to make more accurate predictions for a significant part (more than 83 percent) of the eight
test photos. of.
 54

Predictor Lena Airplane Barbara Boat Lake Peppers Elaine Baboon

PVO [46] 91.9 % 91.1 % 93.8 % 93.3 % 91.4 % 92.5 % 91.9 % 88.3 %
vs. DP
[95]

PVO [46] 86.1 % 87.9 % 90.0 % 89.9 % 85.9 % 86.2 % 85.6 % 82.5 %
vs. MED
[94]

PVO [46] 82.9 % 83.0 % 83.8 % 85.9 % 86.1 % 88.2 % 89.0 % 84.2 %
vs. GAP
[14]

PVO [46] 83.8 % 82.5 % 85.7 % 86.0 % 86.1 % 89.0 % 90.7 % 87.5 %
vs. RP
[90]
Table 8.6: Results of the proportion of better predictions achieved by PVO in some test
images from the CVG-UGR dataset.
 55

CHAPTER-9 THE STATES OF THE ART

Dataset Description Image size Link

Kodak The Kodak package contains 768 × 512, [Link]

twenty-four PNG-formatted true colour 512 × 768 /
photos.

RAISE 8,156 high-resolution colour nature 4288 × 2848, [Link]

photos in RAW format are included in 4928 × 3264 E/
this dataset.

BOSSbase1 A sizable PGM-formatted database 512 × 512 [Link]

.0.1 containing 10,000 grayscale pictures. s

BOWS-2 The photos from the second BOWS 512 × 512 [Link]
competition. It is made up of 10,000
PGM-format grayscale pictures.

CVG-UGR A dataset available online for free. It 512 × 512 [Link]

has multiple sets of uncompressed agenes/[Link]
picture files that were gathered from
public databases.

MedPix A collection of freely available medical 512 × 512 [Link]

photos. 59,000 CT scan pictures are in home
there.

USC-SIPI It has sixteen colour photos and 256 × 256, [Link]

twenty-eight grayscale images. They 512 × 512,
are all in the TIFF file format. 1024 × 1024
Table 9.1: Summary of the Datasets

PEE [29, 40, 41, 90], MHM [27, 49, 78, 105], paired PEE [5, 15, 65, 68, 131] and PVO [22,
23, comparative analysis and analysis of sample works, 46, 63, 67, 73, 112, 133] is done in
this section. Before concluding the comparison, a brief explanation of the evaluation criteria
(Table 9.2) and the databases commonly used in the RDH community is provided. Table 9.1
contains commonly used databases. Table 9.1 shows a selection of uncompressed image
databases, as well as their descriptions, image sizes, and Internet URLs. These are usually
nature pictures taken with a camera, recorded either in grayscale or in color. Some of them
contain high contrast black and white medical photos containing information. Typically,
RDH obtains its test images from the USC-SIPI, Kodak, Bossbase, and BOWS-2 databases.
These databases are frequently used. Table 9.1 provides a summary of RDH-related metrics
and their associated application scenarios. For starters, there are four main features which are
Immersion, Stealth, Stamina and Recovery. Common practice is to assess the opacity of
RDH based on the quality of the patient's vision. PSNR is the measure most often chosen.
Reversibility takes into account the quality of the returned image. Generally speaking, RDH
 56

necessitates that the transmission channel be lossless in order to facilitate flawless recovery.
Therefore, standard RDH algorithms don't really handle resilience very well.

Measurement Description Use case Reference

Capacity It assesses how many bits are hidden. Better The primary [49, 90]
(in bits, bpp) transfer efficiency is correlated with bigger performance of RDH.
capacity.

PSNR (in dB) The changes in pixel values brought about by A frequently used [13, 33]
embedding are measured by this score. A measurement
greater PSNR suggests that there is less in the RDH.
difference between the altered and original
images.

SSIM The purpose of Structural Similarity The RDH considering [36, 132]
Measurement (SSIM) is to assess visual human visual system.
differences in terms of texture, brightness,
contrast, etc. It is more in line with how people
see things..

RCE The degree of contrast difference can be The RDH pursing the [109, 134]
determined by using relative contrast error quality enhancement
(RCE) in the terms of the
contrast.

Extracted-bit It estimates how accurately the hidden message RDH in encrypted [75, 100]
error rate was decoded. images and strong
RDH.

NCC The master secret message and the message RDH in encrypted [50, 116]
read by the attacked carrier are compared for images, and Robust
similarity using the Normalized Cross RDH.
Correlation (NCC) technique.. One can gauge
the robustness using it.
Table 9.2: Summary of Evaluation Criteria in RDH

Next, the PSNR data for a given capacity will be used to compare performance.

Method Technique 10000 bits 20000 bits

Sachnev et al. [90] PEE with sorting 58.21 55.04

Xiao et al. [115] Improved predictor having 61.40 57.77

mapping model

Hu et al. [29] CNN predictor with 61.70 57.96

 57

Method Technique 10000 bits 20000 bits

two-dimensional mapping

Abolfazl et al. [41] Asymmetric PEE 60.86 57.26

Kim et al. [40] Asymmetric PEE 59.92 56.67

Li et al. [49] MHM 61.02 57.55

Hou et al. [27] MHM with DNN 61.46 57.65

Qi et al. [78] MHM with Multi-parameter 61.04 57.64

Weng et al. [105] MHM with Clustering 61.51 57.65

Ou et al. [68] Pairwise PEE (two-dimensional 59.75 56.21

mapping)

Zhang et al. [131] Adaptive two-dimensional 60.93 57.38

mapping

Chang et al. [5] Adaptive two-dimensional 61.35 57.83

mapping

Ou et al. [65] Two-dimensional mapping with 60.77 57.25

Adaptive pairing

Fang et al. [15] Two-dimensional mapping with 61.58 58.02

Skewed MHM

Li et al. [46] PVO 60.34 56.21

Ou et al. [67] PVO-k 60.59 56.58

He et al. [23] k-pass PVO 60.64 56.77

Peng et al. [73] IPVO 60.49 56.57

He et al. [22] k-pass PVO with MHM 61.39 57.75

Ou et al. [63] PVO with two-dimensional 60.91 56.69

mapping

Xiang et al. [112] PVO with two-dimensional 61.57 57.69

mapping

Zhang et al. [133] PVO with two-dimensional 61.53 57.74

mapping
Table 9.3: Comparison between the current representative methods of RDH, when 10,000
and 20,000 bits are embedded into the test image Lena, the PSNR (in dB) is obtained.
 58

A comparison of techniques includes ordination-based PEE [90], CNN-based RDH method

[29], pixel-residual PEH transform [115], trend histogram shift [40], dynamic
asymmetry-based PEE [41]. , traditional MHM [49], the improved MHM of Qi et al. [78],
deep neural network (DNN)-based MHM [27], cluster-based MHM [105], conventional
pair-based PEE [68], adaptive. 2D mapping [5, 15, 131] , adaptive pairing [ 65 ], traditional
PVO [ 46 ], IPVO [ 73 ], PVO-k [ 67 ], k-pass PVO [ 22 , 23 ], and paired PVO [ 63 , 112,
133]. Ten thousand bits and twenty thousand bits are encoded in Lena's test image. The
PSNRs and technical characteristics of the two different technologies are compared and can
be found in Table 9.3. All study data were taken from these previously published studies.

According to the conclusions presented in Table 9.3, higher performance can be achieved
using MHM-based methods [22, 27, 49, 78, 105] and 2D map-based approaches [5, 15, 29,
63, 65, 112, 131, 133]. CNN-based RDH method [29], cluster-based MHM method [105],
biased MHM-based pair-PEE method [15], location-based PVO [133] and adaptive
complex-based method [112]. ] are five excellent approaches that can be used during the
10,000-bit embedding. None of them have a PSNR below 61.50 dB. One of the latest
achievements in the RDH community achieves a PSNR of 61.70 dB according to reference
[29], which is the highest among them. The higher concentration of PEH is obtained from its
CNN-based prediction. In addition, adaptive two-dimensional mapping is used to enhance
the switching process. One of the recent studies of MHM is the system developed by Weng et
al. [105], which is said to improve the performance of the traditional MHM. This method,
which uses a k-means clustering algorithm, generates multiple histograms. Compared with
the standard MHM [49], its PSNR gain is 0.49 dB below the standard. For adaptive
management, Fang et al. [15] generated twenty possible two-dimensional mappings by
applying an asymmetric predictor to the MHM. This method, which is a mixture of pairwise
PEE and MHM, outperforms many pairwise PEE-based methods [5, 65, 68, 131] and
MHM-based approaches [27, 49, 78, 105]. The methods described in references [133] and
[112] are examples of studies that are considered outstanding in the field of PVO. Compared
to a typical PVO, the PSNR gain of the two techniques is 1.19 and 1.23 dB, respectively
[46]. To achieve a more uniform distribution, Zhang et al. [133] proposed that pixel pairs are
produced in a different way. In order to find a suitable solution, the available options were
extensively searched based on statistical data. This was done to create a 2D mapping strategy
dependent on the image content. Pixel blocks are given an adaptive local context to decide
texture levels, as Xiang et al. [112]. However, a selection of blocks has been added..

Chang et al. developed approaches. [5, He et al. [22] and Xiao et al. [115] also showed good
results (PSNR > 61.30 dB at 10,000 bits). These techniques are in addition to the five
strategies discussed above. Researchers Chang et al. [5] conducted an in-depth study of
pair-PEE capabilities. Additional options are now available for 2D mapping. Reference [22]
investigates the improvement possibilities by combining the MHM framework with the
k-pass PVO technique. He and his colleagues improved the accuracy of the prediction by
considering the absolute positional relationship between the two pixels. Many PEHs are
dynamically created and an adaptive collection strategy is the basis. According to Xiao et al.
[115], a new prediction method was proposed to reduce the loss of accuracy due to rounding
 59

errors. In addition, the MHM and associated PEE mapping model were modified by adding a
larger step size change. This was done to improve the results. After a larger payload (20,000
bits), it is very clear that most of the eight methods discussed earlier still outperform the
others. Six methods including CNN-based RDH method [29], pixel-residual PEH transform
[115], adaptive 2D mapping [5], oblique-MHM-based pair-PEE method [15], MHM-based k
- paso-PVO [ 22] and location. based on PVO [133] can obtain a PSNR value greater than
57.73 dB. of.
 60

CHAPTER-10 FUTURE WORK

Photo RDH has been widely studied. When using photo aggregation, the current framework
has a lot of knowledge. In our opinion, future RDH work can focus mainly on two different
directions: improving the diagram and exploring the empirical theory. There are still
questions about the placement of embedding steps and parameter optimization for currently
used frameworks. Current approaches such as adaptive mapping, histogram construction, and
complexity computation can be further explored to explore their potential. Based on a
combination of multiple strategies, there is also an infinite number of effective variations to
explore. In addition, there is potential for significant improvements in RDH and the potential
for significant advances in schematic design using new techniques such as deep learning
algorithms to address this problem. We then present four different avenues for further
research, described in Figures 10.1–10.4, respectively.

First is high-dimensional mapping. This way of further investigation organizes the mapping
of modifications into higher dimensions, which allows a more extensive inclusion of the data.
As the number of spatial dimensions increases, the mapping shows a greater number of
variants, and the transformation model can be changed according to a specific environment.
Full realization of the possibilities of adaptive immersion is a plausible outcome. Similarly,
there are several studies that need to be investigated in order to develop a suitable
optimization algorithm. The use of advanced computational methods such as Convolutional
Neural Networks (CNN) and reinforcement learning can be useful to achieve a more
favorable balance between power, distortion and time cost. This is shown by the
three-dimensional mapping method presented by Chang et al. [3]..

Fig 10.1: Future work: High dimensional mapping

Fig 10.2: The RDH 3D model

 61

RDH is used for both 2D vector graphics and 3D network modeling. A proposed direction
for future development concerns the use of multimedia data, which is frequently used and
requires new embedding strategies and improved performance. Today, the use of 2D vector
graphics and 3D mesh models in various applications such as design and game modeling has
become more common. Each has strict accuracy standards. Since RDH, 3D mesh models and
2D vector graphics differ from the structures of standard raster images, new versions of these
models should be introduced. According to image content, RDH must consider the unique
location of vertices and the topology of the network. Because floating point values ​are more
difficult to manipulate than integers, the data format is more complex. In addition, it should
be taken into account that 2D vector images and 3D network models used in practical
applications are often subject to encryption. To achieve these goals, it is necessary to study
RDH in the context of cryptography. Lin et al. proposed a two-dimensional design graphics
embedding system. [52]. The watermark is created using their method, which involves
applying a region nesting algorithm to create multiple subregions. Establishing a correlation
between the origin points and the generated subspaces is the first step in performing
reversible embedding. After that, Peng et al. [75] developed a semi-delicate RDH method,
originally adapting this approach to 3D network models. The searched watermark can be
used to learn the tampering history and the type of tampering that occurred.

Fig 10.3: RDH for DNN.

Fig 10.4: Example for reversible adversarial.

Third, RDH for the DNN. The purpose of this course of action for future development is to
address the security flaws that have arisen in the era of artificial intelligence. The amount of
mature models that are available on the internet nowadays makes it impossible to
authenticate the owner of an artificial intelligence model. Additionally, the penalty for
 62

unlawful modification is quite minimal. If owners want to achieve meaningful results, they
may want to ensure that their copyrights are protected while sharing their networks with
other researchers so that more people can access and promote their work. When using RDH,
DNN can be integrated in a reversible way to verify the integrity of the watermark.. It was
still possible for the selected DNN to achieve its goal while maintaining its previous level of
performance. For the purpose of preserving the network's integrity, it is possible to recreate
the network at the appropriate time without incurring any loss. It has been demonstrated by
Guan et al. [17] that the idea of model watermarking is quite interesting.

The fourth case is a reversible adversarial example. Reversible adversarial examples are a
concept that was created not too long ago by Liu and colleagues [53]. Through the process of
reversibly embedding the matching perturbation into the aimed photo, this technique for
future research endeavors to construct adversarial examples. First, the watermarking is
removed for the approved AI model in order to finish the authentication process. After that,
the original image can be remade using the hostile example that was received. Finally, the AI
model is able to proceed with its tasks by making use of the image that has been corrected.
While other artificial intelligences might not function properly because of the hostile impact.
The user's privacy can be protected via the reversible adversarial example, which has the
ability to fool unauthorized artificial intelligence. When this scenario is considered, the target
photo is regarded as protected data, and the optimization switches to confrontation.

Even with many RDH techniques previously proposed for JPEG images, data embedding in
JPEG images is still difficult due to its complexity compared to uncompressed images. JPEG
images have a much lower level of data redundancy than uncompressed images, so changes
to the compressed domain can cause noticeable or severe distortions in the original JPEG
images. In addition, for a JPEG scenario, it is important to consider not only the embedding
capacity and resolution of the tagged image, but also its storage capacity. Although various
data hiding techniques have been discussed in the past, there are still only a small number of
documented RDH techniques for JPEG images. Improving the resolution of the encoded
image and the amount of embedded data is critical to limiting the JPEG file size.

The evaluation of image quality in RDH is a topic worth investigating. Accurately assessing
visual quality allows for better guidance of the RDH process and more satisfying outcomes. It
would be beneficial to use techniques created especially for evaluating contrast-altered
images and no-reference approaches in addition to conventional metrics like PSNR and SSIM
for the assessment of image quality. More comparative tests are needed to determine which of
the several no-reference image quality assessment methodologies suggested in the literature
is best suited for contrast-enhanced images. More advanced techniques capable of
maintaining or even improving image quality through data concealing are expected to be
offered in the future with appropriate image quality assessment.

The field of RDH in the encrypted area is growing rapidly, bringing new challenges and
opportunities. In the future, there are many critical areas that need further growth.
 63

First, new theoretical frameworks specifically designed for RDH-ED are urgently needed.
These frameworks should provide insight into the achievable rate of forgery performance
given the statistical properties of plaintext data and the constraints imposed by cryptographic
keys.
In addition, the universal structure of the RDH-ED is necessary to improve research activities
and support the creation of concrete solutions. This framework would provide a systematic
method to remove various barriers and increase efficiency.
Also, more techniques are needed to fix the current speed distortion in RDH-ED. These
technologies must prioritize improving efficiency, robustness and scalability to meet the
different requirements of different applications.
Finally, there is a growing need for tailored methods to address specific use cases, such as
protecting privacy and managing the security of large volumes of data. For these situations, it
is necessary to create customized solutions to ensure the effectiveness and usability of
RDH-ED in practical situations.
 64

CHAPTER-11 SUMMARY

Due to its better image recovery capabilities, RDH, a specialized technology that ensures
transmission security, stands out for its capacity to obtain extraordinary results in sensitive
uses. The ability allowed RDH to accomplish remarkable results. In this paper, classic
techniques and popular adaptive solutions for Reversible Data Hiding in photos are
investigated. Compression-based methods, DE, IT, HS, PEE, MHM, paired PEE and PVO
are all included in these techniques with their respective abbreviations. A presentation,
analysis, and evaluation of the typical methodologies and approaches utilized by these eight
fields of study are included in this article. Additionally, it discusses the benefits of the three
most frequent techniques and evaluates the performance of the most advanced RDH
algorithms now available. There is no way around the fact that RDH theory and embedding
frameworks will continue to develop in the future, given the growing number of application
scenarios and the increasing needs. Within the context of early motives, the paper proposes
four prospective areas for further research, and it provides insights and comments on each of
these potential avenues.

This report looks at important areas of reversible data hiding (RDH) in relation to audio and
video in digital media. These topics include residual detail headroom (RDH) into digital
images in the spatial domain, JPEG domain, semi-fragile RDH into digitally compressed
images, alternative image quality measurements for RDH that go beyond PSNR, encrypted
digital image RDH, and RDH for audio and video. Every topic is covered in detail, along
with an analysis and presentation of common techniques and approaches. The study looks
forward to more advancements in RDH applications and research in the future. That being
said, given the increasing significance of audio and video in our digital age, it is anticipated
that this will soon alter.
 65

CHAPTER-12 REFERENCES

[1] A. M. Alattar, “Reversible watermark using the difference expansion of a

generalized integer transform,” IEEE Trans. Image Process., vol. 13, no. 8, pp.
1147–1156, 2004.

[2] M. U. Celik, G. Sharma, A. M. Tekalp, and E. Saber, “Lossless generalized-LSB

data embedding,” IEEE Trans. Image Process., vol. 14, no. 2, pp. 253–266,
2005.

[3] J. Chang, G. Zhu, H. Zhang, Y. Zhou, X. Luo, and L. Wu, “Reversible data
hiding for color images based on adaptive 3D prediction-error expansion and
double deep Q-network,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no.
8, pp. 5055–5067, 2022.

[4] Q. Chang, X. Li, and Y. Zhao, “Reversible data hiding for color images based on
adaptive three-dimensional histogram modification,” IEEE Trans. Circuits Syst.
Video Technol., vol. 32, no. 9, pp. 5725–5735, 2022.

[5] Q. Chang, X. Li, Y. Zhao, and R. Ni, “Adaptive pairwise prediction-error

expansion and multiple histograms modification for reversible data hiding,”
IEEE Trans. Circuits Syst. Video Technol., vol. 31, no. 12, pp. 4850–4863, 2021.

[6] K. Chen and C.-C. Chang, “High-capacity reversible data hiding in encrypted
images based on extended run-length coding and block-based MSB plane
rearrangement,” J. Vis. Commun. Image Represent., vol. 58, pp. 334–344, 2019.

[7] X. Chen, X. Sun, H. Sun, Z. Zhou, and J. Zhang, “Reversible watermarking

method based on asymmetric-histogram shifting of prediction errors,” J. Syst.
Softw., vol. 86, no. 10, pp. 2620–2626, 2013.

[8] D. Coltuc, “Improved embedding for prediction-based reversible watermarking,”

IEEE Trans. Inf. Forensics Secur., vol. 6, no. 3, pp. 873–882, 2011.

[9] D. Coltuc and J.-M. Chassery, “Very fast watermarking by reversible contrast
mapping,” IEEE Signal Process. Lett., vol. 14, no. 4, pp. 255–258, 2007.

[10] I.-C. Dragoi and D. Coltuc, “Local-prediction-based difference expansion

reversible watermarking,” IEEE Trans. Image Process., vol. 23, no. 4, pp.
1779–1790, 2014.

[11] I.-C. Dragoi and D. Coltuc, “On local prediction based reversible
watermarking,” IEEE Trans. Image Process., vol. 24, no. 4, pp. 1244–1246,
2015.
 66

[12] I.-C. Dragoi and D. Coltuc, “Adaptive Pairing Reversible Watermarking,” IEEE
Trans. Image Process., vol. 25, no. 5, pp. 2420–2422, 2016.

[13] Y. Du, Z. Yin, and X. Zhang, “High capacity lossless data hiding in JPEG
bitstream based on general VLC mapping,” IEEE Trans. Dependable Secure
Comput., pp. 1–1, 2020.

[14] M. Fallahpour, “Reversible image data hiding based on gradient adjusted

prediction,” IEICE Electron. Express, vol. 5, no. 20, pp. 870–876, 2008.

[15] G. Fan, Z. Pan, Q. Zhou, X. Gao, and X. Zhang, “Multiple histogram based
adaptive pairwise prediction-error modification for efficient reversible image
watermarking,” Inf. Sci. (Ny), vol. 581, pp. 515–535, 2021.

[16] J. Fridrich, M. Goljan, and R. Du, “Invertible authentication,” in SPIE

Proceedings, 2001.

[17] X. Guan, H. Feng, W. Zhang, H. Zhou, J. Zhang, and N. Yu, “Reversible

watermarking in deep convolutional neural networks for integrity
authentication,” in Proceedings of the 28th ACM International Conference on
Multimedia, 2020.

[18] X. Gui, X. Li, and B. Yang, “A high capacity reversible data hiding scheme
based on generalized prediction-error expansion and adaptive embedding,”
Signal Processing, vol. 98, pp. 370–380, 2014.

[19] J. He, J. Chen, W. Luo, S. Tang, and J. Huang, “A novel high-capacity reversible
data hiding scheme for encrypted JPEG bitstreams,” IEEE Trans. Circuits Syst.
Video Technol., vol. 29, no. 12, pp. 3501–3515, 2019.

[20] J. He, J. Chen, and S. Tang, “Reversible data hiding in JPEG images based on
negative influence models,” IEEE Trans. Inf. Forensics Secur., vol. 15, pp.
2121–2133, 2020.

[21] W. He and Z. Cai, “Reversible data hiding based on dual pairwise

prediction-error expansion,” IEEE Trans. Image Process., vol. 30, pp.
5045–5055, 2021.

[22] W. He, G. Xiong, and Y. Wang, “Reversible data hiding based on adaptive
multiple histograms modification,” IEEE Trans. Inf. Forensics Secur., vol. 16,
pp. 3000–3012, 2021.

[23] W. He, K. Zhou, J. Cai, L. Wang, and G. Xiong, “Reversible data hiding using
multi-pass pixel value ordering and prediction-error expansion,” J. Vis. Commun.
Image Represent., vol. 49, pp. 351–360, 2017.
 67

[24] W. Hong, G. Horng, C.-W. Shiu, T.-S. Chen, and Y.-C. Chen, “Reversible
steganographic method using complexity control and human visual system,”
Comput. J., vol. 58, no. 10, pp. 2583–2594, 2015.

[25] D. Hou, H. Wang, W. Zhang, and N. Yu, “Reversible data hiding in JPEG image
based on DCT frequency and block selection,” Signal Processing, vol. 148, pp.
41–47, 2018.

[26] D. Hou, W. Zhang, K. Chen, S.-J. Lin, and N. Yu, “Reversible data hiding in
color image with grayscale invariance,” IEEE Trans. Circuits Syst. Video
Technol., vol. 29, no. 2, pp. 363–374, 2019.

[27] J. Hou, B. Ou, H. Tian, and Z. Qin, “Reversible data hiding based on multiple
histograms modification and deep neural networks,” Signal Process. Image
Commun., vol. 92, no. 116118, p. 116118, 2021.

[28] R. Hu and S. Xiang, “CNN prediction based reversible data hiding,” IEEE Signal
Process. Lett., vol. 28, pp. 464–468, 2021.

[29] R. Hu and S. Xiang, “Reversible data hiding by using CNN prediction and
adaptive embedding,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 12,
pp. 10196–10208, 2022.

[30] Y. Hu, H.-K. Lee, and J. Li, “DE-based reversible data hiding with improved
overflow location map,” IEEE Trans. Circuits Syst. Video Technol., vol. 19, no.
2, pp. 250–260, 2009.

[31] Y. Hu, K. Wang, and Z.-M. Lu, “An improved VLC-based lossless data hiding
scheme for JPEG images,” J. Syst. Softw., vol. 86, no. 8, pp. 2166–2173, 2013.

[32] F. Huang, X. Qu, H. J. Kim, and J. Huang, “Reversible Data Hiding in JPEG
Images,” IEEE Trans. Circuits Syst. Video Technol., vol. 26, no. 9, pp.
1610–1621, 2016.

[33] C.-L. Jhong and H.-L. Wu, “Grayscale-invariant reversible data hiding based on
multiple histograms modification,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 32, no. 9, pp. 5888–5901, 2022.

[34] Y. Jia, Z. Yin, X. Zhang, and Y. Luo, “Reversible data hiding based on reducing
invalid shifting of pixels in histogram shifting,” Signal Processing, vol. 163, pp.
238–246, 2019.

[35] R. Jiang, H. Zhou, W. Zhang, and N. Yu, “Reversible data hiding in encrypted
three-dimensional mesh models,” IEEE Trans. Multimedia, vol. 20, no. 1, pp.
55–67, 2018.
 68

[36] S.-W. Jung, L. T. Ha, and S.-J. Ko, “A new histogram modification based
reversible data hiding algorithm considering the human visual system,” IEEE
Signal Process. Lett., vol. 18, no. 2, pp. 95–98, 2011.

[37] Y. Ke, M. Zhang, X. Zhang, J. Liu, T. Su, and X. Yang, “A reversible data hiding
scheme in encrypted domain for secret image sharing based on Chinese
remainder theorem,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 4,
pp. 2469–2481, 2022.

[38] H. J. Kim, V. Sachnev, Y. Q. Shi, J. Nam, and H.-G. Choo, “A novel difference
expansion transform for reversible data embedding,” IEEE Trans. Inf. Forensics
Secur., vol. 3, no. 3, pp. 456–465, 2008.

[39] S. Kim, R. Lussi, X. Qu, F. Huang, and H. J. Kim, “Reversible data hiding with
automatic brightness preserving contrast enhancement,” IEEE Trans. Circuits
Syst. Video Technol., vol. 29, no. 8, pp. 2271–2284, 2019.

[40] S. Kim, X. Qu, V. Sachnev, and H. J. Kim, “Skewed histogram shifting for
reversible data hiding using a pair of extreme predictions,” IEEE Trans. Circuits
Syst. Video Technol., vol. 29, no. 11, pp. 3236–3246, 2019.

[41] A. Kouhi and M. H. Sedaaghi, “Prediction error distribution with dynamic

asymmetry for reversible data hiding,” Expert Syst. Appl., vol. 184, no. 115475,
p. 115475, 2021.

[42] S.-K. Lee, Y.-H. Suh, and Y.-S. Ho, “Reversible image authentication based on
watermarking,” in 2006 IEEE International Conference on Multimedia and
Expo, 2006.

[43] J. Li and S. Xiang, “Audio-lossless robust watermarking against

desynchronization attacks,” Signal Processing, vol. 198, no. 108561, p. 108561,
2022.

[44] W. Li, X. Li, R. Ni, and Y. Zhao, “PVO-based reversible data hiding using
adaptive multiple histogram generation and modification,” Signal Process.
Image Commun., vol. 99, no. 116405, p. 116405, 2021.

[45] X. Li, B. Li, B. Yang, and T. Zeng, “General framework to

histogram-shifting-based reversible data hiding,” IEEE Trans. Image Process.,
vol. 22, no. 6, pp. 2181–2191, 2013.

[46] X. Li, J. Li, B. Li, and B. Yang, “High-fidelity reversible data hiding scheme
based on pixel-value-ordering and prediction-error expansion,” Signal
Processing, vol. 93, no. 1, pp. 198–205, 2013.
 69

[47] X. Li, B. Yang, and T. Zeng, “Efficient reversible watermarking based on

adaptive prediction-error expansion and pixel selection,” IEEE Trans. Image
Process., vol. 20, no. 12, pp. 3524–3533, 2011.

[48] X. Li, W. Zhang, X. Gui, and B. Yang, “A novel reversible data hiding scheme
based on two-dimensional difference-histogram modification,” IEEE Trans. Inf.
Forensics Secur., vol. 8, no. 7, pp. 1091–1100, 2013.

[49] X. Li, W. Zhang, X. Gui, and B. Yang, “Efficient reversible data hiding based on
multiple histograms modification,” IEEE Trans. Inf. Forensics Secur., vol. 10,
no. 9, pp. 2016–2027, 2015.

[50] X. Liang and S. Xiang, “Robust reversible audio watermarking based on

high-order difference statistics,” Signal Processing, vol. 173, no. 107584, p.
107584, 2020.

[51] C.-C. Lin, W.-L. Tai, and C.-C. Chang, “Multilevel reversible data hiding based
on histogram modification of difference images,” Pattern Recognit., vol. 41, no.
12, pp. 3582–3591, 2008.

[52] Z.-X. Lin, F. Peng, and M. Long, “A low-distortion reversible watermarking for
2D engineering graphics based on region nesting,” IEEE Trans. Inf. Forensics
Secur., vol. 13, no. 9, pp. 2372–2382, 2018.

[53] J. Liu, W. Zhang, K. Fukuchi, Y. Akimoto, and J. Sakuma, “Unauthorized AI

cannot recognize me: Reversible adversarial example,” Pattern Recognit., vol.
134, no. 109048, p. 109048, 2023.

[54] M. Long, F. Peng, and H.-Y. Li, “Separable reversible data hiding and encryption
for HEVC video,” J. Real Time Image Process., vol. 14, no. 1, pp. 171–182,
2018.

[55] T. Luo, G. Jiang, M. Yu, C. Zhong, H. Xu, and Z. Pan, “Convolutional neural
networks-based stereo image reversible data hiding method,” J. Vis. Commun.
Image Represent., vol. 61, pp. 61–73, 2019.

[56] W.-L. Lyu, L. Cheng, and Z. Yin, “High-capacity reversible data hiding in
encrypted 3D mesh models based on multi-MSB prediction,” Signal Processing,
vol. 201, no. 108686, p. 108686, 2022.

[57] B. Ma and Y. Q. Shi, “A reversible data hiding scheme based on code division
multiplexing,” IEEE Trans. Inf. Forensics Secur., vol. 11, no. 9, pp. 1914–1927,
2016.

[58] B. Ma et al., “Adaptive error prediction method based on multiple linear

regression for reversible data hiding,” J. Real Time Image Process., vol. 16, no.
4, pp. 821–834, 2019.
 70

[59] K. Ma, W. Zhang, X. Zhao, N. Yu, and F. Li, “Reversible data hiding in
encrypted images by reserving room before encryption,” IEEE Trans. Inf.
Forensics Secur., vol. 8, no. 3, pp. 553–562, 2013.

[60] S. Ma, X. Li, M. Xiao, B. Ma, and Y. Zhao, “Fast expansion-bins-determination

for multiple histograms modification based reversible data hiding,” IEEE Signal
Process. Lett., vol. 29, pp. 662–666, 2022.

[61] N. Mao, F. Chen, H. He, and Y. Yang, “Reversible data hiding based on adaptive
IPVO and two-segment pairwise PEE,” Signal Processing, vol. 198, no. 108577,
p. 108577, 2022.

[62] Z. Ni, Y.-Q. Shi, N. Ansari, and W. Su, “Reversible data hiding,” IEEE Trans.
Circuits Syst. Video Technol., vol. 16, no. 3, pp. 354–362, 2006.

[63] B. Ou, X. Li, and J. Wang, “High-fidelity reversible data hiding based on
pixel-value-ordering and pairwise prediction-error expansion,” J. Vis. Commun.
Image Represent., vol. 39, pp. 12–23, 2016.

[64] B. Ou, X. Li, and J. Wang, “Improved PVO-based reversible data hiding: A new
implementation based on multiple histograms modification,” J. Vis. Commun.
Image Represent., vol. 38, pp. 328–339, 2016.

[65] B. Ou, X. Li, W. Zhang, and Y. Zhao, “Improving pairwise PEE via
hybrid-dimensional histogram generation and adaptive mapping selection,” IEEE
Trans. Circuits Syst. Video Technol., vol. 29, no. 7, pp. 2176–2190, 2019.

[66] B. Ou, X. Li, Y. Zhao, and R. Ni, “Reversible data hiding based on PDE
predictor,” J. Syst. Softw., vol. 86, no. 10, pp. 2700–2709, 2013.

[67] B. Ou, X. Li, Y. Zhao, and R. Ni, “Reversible data hiding using invariant
pixel-value-ordering and prediction-error expansion,” Signal Process. Image
Commun., vol. 29, no. 7, pp. 760–772, 2014.

[68] B. Ou, X. Li, Y. Zhao, R. Ni, and Y.-Q. Shi, “Pairwise prediction-error expansion
for efficient reversible data hiding,” IEEE Trans. Image Process., vol. 22, no. 12,
pp. 5010–5021, 2013.

[69] B. Ou and Y. Zhao, “High capacity reversible data hiding based on multiple
histograms modification,” IEEE Trans. Circuits Syst. Video Technol., vol. 30, no.
8, pp. 2329–2342, 2020.

[70] Z. Pan, X. Gao, E. Gao, and G. Fan, “Adaptive complexity for

pixel-value-ordering based reversible data hiding,” IEEE Signal Process. Lett.,
vol. 27, pp. 915–919, 2020.
 71

[71] F. Peng, W.-Y. Jiang, Y. Qi, Z.-X. Lin, and M. Long, “Separable robust reversible
watermarking in encrypted 2D vector graphics,” IEEE Trans. Circuits Syst.
Video Technol., vol. 30, no. 8, pp. 2391–2405, 2020.

[72] F. Peng, X. Li, and B. Yang, “Adaptive reversible data hiding scheme based on
integer transform,” Signal Processing, vol. 92, no. 1, pp. 54–62, 2012.

[73] F. Peng, X. Li, and B. Yang, “Improved PVO-based reversible data hiding,”
Digit. Signal Process., vol. 25, pp. 255–265, 2014.

[74] F. Peng, T. Liao, and M. Long, “A semi-fragile reversible watermarking for

authenticating 3D models in dual domains based on variable direction double
modulation,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 12, pp.
8394–8408, 2022.

[75] F. Peng, B. Long, and M. Long, “A general region nesting-based semi-fragile

reversible watermarking for authenticating 3D mesh models,” IEEE Trans.
Circuits Syst. Video Technol., vol. 31, no. 11, pp. 4538–4553, 2021.

[76] P. Puteaux and W. Puech, “A recursive reversible data hiding in encrypted

images method with a very high payload,” IEEE Trans. Multimedia, vol. 23, pp.
636–650, 2021.

[77] W. Qi, S. Guo, and W. Hu, “Generic reversible visible watermarking via
regularized graph Fourier transform coding,” IEEE Trans. Image Process., vol.
31, pp. 691–705, 2022.

[78] W. Qi, X. Li, T. Zhang, and Z. Guo, “Optimal reversible data hiding scheme
based on multiple histograms modification,” IEEE Trans. Circuits Syst. Video
Technol., vol. 30, no. 8, pp. 2300–2312, 2020.

[79] Z. Qian, H. Xu, X. Luo, and X. Zhang, “New framework of reversible data
hiding in encrypted JPEG bitstreams,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 29, no. 2, pp. 351–362, 2019.

[80] Z. Qian and X. Zhang, “Lossless data hiding in JPEG bitstream,” J. Syst. Softw.,
vol. 85, no. 2, pp. 309–313, 2012.

[81] Z. Qian, X. Zhang, and S. Wang, “Reversible data hiding in encrypted JPEG
bitstream,” IEEE Trans. Multimedia, vol. 16, no. 5, pp. 1486–1491, 2014.

[82] C. Qin, Z. He, H. Yao, F. Cao, and L. Gao, “Visible watermark removal scheme
based on reversible data hiding and image inpainting,” Signal Process. Image
Commun., vol. 60, pp. 160–172, 2018.
 72

[83] C. Qin, X. Qian, W. Hong, and X. Zhang, “An efficient coding scheme for
reversible data hiding in encrypted image with redundancy transfer,” Inf. Sci.
(Ny), vol. 487, pp. 176–192, 2019.

[84] J. Qin and F. Huang, “Reversible data hiding based on multiple two-dimensional
histograms modification,” IEEE Signal Process. Lett., vol. 26, no. 6, pp.
843–847, 2019.

[85] Y. Qiu, Z. Qian, H. He, H. Tian, and X. Zhang, “Optimized lossless data hiding
in JPEG bitstream and relay transfer-based extension,” IEEE Trans. Circuits
Syst. Video Technol., vol. 31, no. 4, pp. 1380–1394, 2021.

[86] Y. Qiu, Z. Qian, and L. Yu, “Adaptive reversible data hiding by extending the
generalized integer transformation,” IEEE Signal Process. Lett., vol. 23, no. 1,
pp. 130–134, 2016.

[87] Y. Qiu, Q. Ying, Y. Yang, H. Zeng, S. Li, and Z. Qian, “High-capacity

framework for reversible data hiding in encrypted image using pixel prediction
and entropy encoding,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 9,
pp. 5874–5887, 2022.

[88] X. Qu and H. J. Kim, “Pixel-based pixel value ordering predictor for

high-fidelity reversible data hiding,” Signal Processing, vol. 111, pp. 249–260,
2015.

[89] R. M. Rad, K. Wong, and J.-M. Guo, “Reversible data hiding by adaptive group
modification on histogram of prediction errors,” Signal Processing, vol. 125, pp.
315–328, 2016.

[90] V. Sachnev, H. J. Kim, J. Nam, S. Suresh, and Y. Q. Shi, “Reversible

watermarking algorithm using sorting and prediction,” IEEE Trans. Circuits
Syst. Video Technol., vol. 19, no. 7, pp. 989–999, 2009.

[91] Y.-Q. Shi, X. Li, X. Zhang, H.-T. Wu, and B. Ma, “Reversible data hiding:
Advances in the past two decades,” IEEE Access, vol. 4, pp. 3210–3237, 2016.

[92] Z. Su, Y. Ye, Q. Zhang, W. Li, and Y. Dai, “Robust 2D engineering CAD
graphics hashing for joint topology and geometry authentication via
covariance-based descriptors,” IEEE Trans. Inf. Forensics Secur., vol. 13, no. 4,
pp. 1018–1030, 2018.

[93] X. Tang, H. Wang, and Y. Chen, “Reversible data hiding based on a modified
difference expansion for H.264/AVC video streams,” Multimed. Tools Appl., vol.
79, no. 39–40, pp. 28661–28674, 2020.
 73

[94] D. M. Thodi and J. J. Rodriguez, “Expansion embedding techniques for

reversible watermarking,” IEEE Trans. Image Process., vol. 16, no. 3, pp.
721–730, 2007.

[95] J. Tian, “Reversible data embedding using a difference expansion,” IEEE Trans.
Circuits Syst. Video Technol., vol. 13, no. 8, pp. 890–896, 2003.

[96] J. Wang, X. Chen, J. Ni, N. Mao, and Y. Shi, “Multiple histograms-based

reversible data hiding: Framework and realization,” IEEE Trans. Circuits Syst.
Video Technol., vol. 30, no. 8, pp. 2313–2328, 2020.

[97] J. Wang, N. Mao, X. Chen, J. Ni, C. Wang, and Y. Shi, “Multiple histograms
based reversible data hiding by using FCM clustering,” Signal Processing, vol.
159, pp. 193–203, 2019.

[98] J. Wang, J. Ni, X. Zhang, and Y.-Q. Shi, “Rate and distortion optimization for
reversible data hiding using multiple histogram shifting,” IEEE Trans. Cybern.,
pp. 1–12, 2016.

[99] X. Wang, J. Ding, and Q. Pei, “A novel reversible image data hiding scheme
based on pixel value ordering and dynamic pixel block partition,” Inf. Sci. (Ny),
vol. 310, pp. 16–35, 2015.

[100] X. Wang, X. Li, and Q. Pei, “Independent embedding domain based two-stage
robust reversible watermarking,” IEEE Trans. Circuits Syst. Video Technol., vol.
30, no. 8, pp. 2406–2417, 2020.

[101] X. Wang, X. Li, B. Yang, and Z. Guo, “Efficient generalized integer transform
for reversible watermarking,” IEEE Signal Process. Lett., vol. 17, no. 6, pp.
567–570, 2010.

[102] X. Wang, X. Wang, B. Ma, Q. Li, and Y.-Q. Shi, “High precision error prediction
algorithm based on ridge regression predictor for reversible data hiding,” IEEE
Signal Process. Lett., vol. 28, pp. 1125–1129, 2021.

[103] S. Weng, T. Hou, T. Zhang, and J.-S. Pan, “Adaptive smoothness evaluation and
multiple asymmetric histogram modification for reversible data hiding,” J. Vis.
Commun. Image Represent., vol. 90, no. 103732, p. 103732, 2023.

[104] S. Weng, Y. Shi, W. Hong, and Y. Yao, “Dynamic improved pixel value ordering
reversible data hiding,” Inf. Sci. (Ny), vol. 489, pp. 136–154, 2019.

[105] S. Weng, W. Tan, B. Ou, and J.-S. Pan, “Reversible data hiding method for
multi-histogram point selection based on improved crisscross optimization
algorithm,” Inf. Sci. (Ny), vol. 549, pp. 13–33, 2021.
 74

[106] S. Weng, G. Zhang, J.-S. Pan, and Z. Zhou, “Optimal PPVO-based reversible
data hiding,” J. Vis. Commun. Image Represent., vol. 48, pp. 317–328, 2017.

[107] S. Weng, Y. Zhou, and T. Zhang, “Adaptive reversible data hiding for JPEG
images with multiple two-dimensional histograms,” J. Vis. Commun. Image
Represent., vol. 85, no. 103487, p. 103487, 2022.

[108] H. Wu, X. Li, X. Luo, X. Zhang, and Y. Zhao, “General expansion-shifting

model for reversible data hiding: Theoretical investigation and practical
algorithm design,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 9, pp.
5989–6001, 2022.

[109] H.-T. Wu, X. Cao, R. Jia, and Y.-M. Cheung, “Reversible data hiding with
brightness preserving contrast enhancement by two-dimensional histogram
modification,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 11, pp.
7605–7617, 2022.

[110] H.-T. Wu, J.-L. Dugelay, and Y.-Q. Shi, “Reversible image data hiding with
contrast enhancement,” IEEE Signal Process. Lett., vol. 22, no. 1, pp. 81–85,
2015.

[111] Y. Wu, Y. Xiang, Y. Guo, J. Tang, and Z. Yin, “An improved reversible data
hiding in encrypted images using parametric binary tree labeling,” IEEE Trans.
Multimedia, vol. 22, no. 8, pp. 1929–1938, 2020.

[112] S. Xiang and G. Ruan, “Efficient PVO-based reversible data hiding by selecting
blocks with full-enclosing context,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 32, no. 5, pp. 2868–2880, 2022.

[113] M. Xiao, X. Li, B. Ma, X. Zhang, and Y. Zhao, “Efficient reversible data hiding
for JPEG images with multiple histograms modification,” IEEE Trans. Circuits
Syst. Video Technol., vol. 31, no. 7, pp. 2535–2546, 2021.

[114] M. Xiao, X. Li, Y. Wang, Y. Zhao, and R. Ni, “Reversible data hiding based on
pairwise embedding and optimal expansion path,” Signal Processing, vol. 158,
pp. 210–218, 2019.

[115] M. Xiao, X. Li, Y. Zhao, B. Ma, and G. Guo, “A novel reversible data hiding
scheme based on pixel-residual histogram,” ACM Trans. Multimed. Comput.
Commun. Appl., vol. 19, no. 1s, pp. 1–19, 2023.

[116] L. Xiong, X. Han, C.-N. Yang, and Y.-Q. Shi, “Robust reversible watermarking
in encrypted image with secure multi-party based on lightweight cryptography,”
IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 1, pp. 75–91, 2022.
 75

[117] L. Xiong, X. Han, C.-N. Yang, and Z. Xia, “RDH-DES: Reversible Data Hiding
over Distributed Encrypted-Image Servers based on secret sharing,” ACM Trans.
Multimed. Comput. Commun. Appl., vol. 19, no. 1, pp. 1–19, 2023.

[118] D. Xu and Y. Liu, “Reversible data hiding in H.264/AVC videos based on

hybrid-dimensional histogram modification,” Multimed. Tools Appl., vol. 81, no.
20, pp. 29305–29319, 2022.

[119] D. Xu and R. Wang, “Two-dimensional reversible data hiding-based approach

for intra-frame error concealment in H.264/AVC,” Signal Process. Image
Commun., vol. 47, pp. 369–379, 2016.

[120] X. Yang and F. Huang, “New CNN-based predictor for reversible data hiding,”
IEEE Signal Process. Lett., vol. 29, pp. 2627–2631, 2022.

[121] Y. Yang, W. Zhang, D. Liang, and N. Yu, “Reversible data hiding in medical
images with enhanced contrast in texture area,” Digit. Signal Process., vol. 52,
pp. 13–24, 2016.

[122] Y. Yang, T. Zou, G. Huang, and W. Zhang, “A high visual quality color image
reversible data hiding scheme based on B-R-G embedding principle and
CIEDE2000 assessment metric,” IEEE Trans. Circuits Syst. Video Technol., vol.
32, no. 4, pp. 1860–1874, 2022.

[123] Y. Yao, W. Zhang, H. Wang, H. Zhou, and N. Yu, “Content-adaptive reversible

visible watermarking in encrypted images,” Signal Processing, vol. 164, pp.
386–401, 2019.

[124] Y. Yao, W. Zhang, and N. Yu, “Inter-frame distortion drift analysis for reversible
data hiding in encrypted H.264/AVC video bitstreams,” Signal Processing, vol.
128, pp. 531–545, 2016.

[125] S. Yi and Y. Zhou, “Separable and reversible data hiding in encrypted images
using parametric binary tree labeling,” IEEE Trans. Multimedia, vol. 21, no. 1,
pp. 51–64, 2019.

[126] Z. Yin, Y. Ji, and B. Luo, “Reversible data hiding in JPEG images with
multi-objective optimization,” IEEE Trans. Circuits Syst. Video Technol., vol. 30,
no. 8, pp. 2343–2352, 2020.

[127] Z. Yin, Y. Peng, and Y. Xiang, “Reversible data hiding in encrypted images
based on pixel prediction and bit-plane compression,” IEEE Trans. Dependable
Secure Comput., pp. 1–1, 2020.
 76

[128] Z. Yin, Y. Xiang, and X. Zhang, “Reversible data hiding in encrypted images
based on multi-MSB prediction and Huffman coding,” IEEE Trans. Multimedia,
vol. 22, no. 4, pp. 874–884, 2020.

[129] C. Yu, X. Zhang, G. Li, S. Zhan, and Z. Tang, “Reversible data hiding with
adaptive difference recovery for encrypted images,” Inf. Sci. (Ny), vol. 584, pp.
89–110, 2022.

[130] C. Yu, X. Zhang, D. Wang, and Z. Tang, “Reversible data hiding with pairwise
PEE and 2D-PEH decomposition,” Signal Processing, vol. 196, no. 108527, p.
108527, 2022.

[131] C. Zhang and B. Ou, “Reversible data hiding based on multiple adaptive
two-dimensional prediction-error histograms modification,” IEEE Trans.
Circuits Syst. Video Technol., vol. 32, no. 7, pp. 4174–4187, 2022.

[132] C. Zhang, B. Ou, X. Li, and J. Xiong, “Human visual system guided reversible
data hiding based on multiple histograms modification,” Comput. J., vol. 66, no.
4, pp. 888–906, 2023.

[133] T. Zhang, X. Li, W. Qi, and Z. Guo, “Location-based PVO and adaptive pairwise
modification for efficient reversible data hiding,” IEEE Trans. Inf. Forensics
Secur., vol. 15, pp. 2306–2319, 2020.

[134] T. Zhang, T. Hou, S. Weng, F. Zou, H. Zhang, and C.-C. Chang, “Adaptive
reversible data hiding with contrast enhancement based on multi-histogram
modification,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 8, pp.
5041–5054, 2022.

[135] W. Zhang, X. Hu, X. Li, and N. Yu, “Recursive histogram modification:

Establishing equivalency between reversible data hiding and lossless data
compression,” IEEE Trans. Image Process., vol. 22, no. 7, pp. 2775–2785, 2013.

[136] X. Zhang, “Reversible data hiding in encrypted image,” IEEE Signal Process.
Lett., vol. 18, no. 4, pp. 255–258, 2011.

[137] X. Zhang, “Separable reversible data hiding in encrypted image,” IEEE Trans.
Inf. Forensics Secur., vol. 7, no. 2, pp. 826–832, 2012.

[138] X. Zhang, “Reversible data hiding with optimal value transfer,” IEEE Trans.
Multimedia, vol. 15, no. 2, pp. 316–325, 2013.

[139] X. Zhang, Y. Yao and N. Yu, "Convolutional Neural Network-driven Optimal

Prediction for Image Reversible Data Hiding," 2021 IEEE 23rd International
 77

Workshop on Multimedia Signal Processing (MMSP), Tampere, Finland, 2021,

pp. 1-6.

[140] H. Zheng, C. Wang, J. Wang, and S. Xiang, “A new reversible watermarking

scheme using the content-adaptive block size for prediction,” Signal Processing,
vol. 164, pp. 74–83, 2019.

[141] C. D. Vleeschouwer, J. F. Delaigle, and B. Macq, “Circular interpretation of

histogram for reversible watermarking,” in Proc. IEEE Workshop on Multimedia
Signal Processing, 2001, pp. 345–350

[142] C. D. Vleeschouwer, J. F. Delaigle, and B. Macq, “Circular interpretation of

bijective transformations in lossless watermarking for media asset management,”
IEEE Trans. Multimedia, vol. 5, no. 1, pp. 97–105, Mar. 2003.

[143] Z. Ni, Y. Q. Shi, N. Ansari, W. Su, Q. Sun, and X. Lin, “Robust lossless image
data hiding,” in Proc. IEEE International Conference on Multimedia and Expo,
2004, pp. 2199–2202.

[144] Z. Ni, Y. Q. Shi, N. Ansari, W. Su, Q. Sun, and X. Lin, , “Robust lossless image
data hiding designed for semi-fragile image authentication,” IEEE Trans.
Circuits Syst. Video Techn., vol. 18, no. 4, pp. 497–509, 2008.

[145] H.-H. Tsai, H.-C. Tseng, and Y.-S. Lai, “Robust lossless watermarking using
alpha-trimmed mean and SVM,” in Proc. International Conference on Machine
Learning and Cybernetics, 2008, pp. 3347–3353.

[146] H.-H. Tsai, H.-C. Tseng, and Y.-S. Lai, “Robust lossless image watermarking
based on alpha-trimmed mean algorithm and support vector machine,” Journal of
Systems and Software, vol. 83, no. 6, pp. 1015–1028, 2010.

[147] H. T. Wu, J. Huang, and Y. Q. Shi, “A reversible data hiding method with
contrast enhancement for medical images,” Journal of Visual Communication
and Image Representation, vol. 31, pp. 146–153, 2015.

[148] G. Gao and Y. Q. Shi, “Reversible data hiding using controlled contrast
enhancement and integer wavelet transform,” IEEE Signal Process. Lett., vol.
22, no. 11, pp. 2078–2082, 2015.

[149] S. Kim, R. Lussi, X. Qu, and H. J. Kim, “Automatic contrast enhancement using
reversible data hiding,” in Proc. IEEE International Workshop on Information
Forensics and Security, 2015, pp. 1–5.
 78

[150] A. K. Moorthy and A. C. Bovik, “Blind image quality assessment: From natural
scene statistics to perceptual quality,” IEEE Trans. Image Process., vol. 20, no.
12, pp. 3350–3364, 2011.

[151] M. A. Saad, A. C. Bovik, and C. Charrier, “Blind image quality assessment: A

natural scene statistics approach in the DCT domain,” IEEE Trans. Image
Process., vol. 21, no. 8, pp. 3339–3352, 2012.

[152] A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality

assessment in the spatial domain,” IEEE Trans. Image Process., vol. 21, no. 12,
pp. 4695–4708, 2012.

[153] Y. Zhang, A. K. Moorthy, D. M. Chandler, and A. C. Bovik, “CDIIVINE:

No-reference image quality assessment based on local magnitude and phase
statistics of natural scenes,” Signal Processing: Image Communication, vol. 29,
no. 7, pp. 725–747, 2014.

[154] L. Liu, B. Liu, H. Huang, and A. C. Bovik, “No-reference image quality

assessment based on spatial and spectral entropies,” Signal Processing: Image
Communication, vol. 29, no. 8, pp. 856–863, 2014.

[155] Y. Fang, K. Ma, Z. Wang, W. Lin, Z. Fang, and G. Zhai, “No-reference quality
assessment of contrast-distorted images based on natural scene statistics,” IEEE
Signal Process. Lett., vol. 22, no. 7, pp. 838–842, 2015.

[156] K. Ma, W. Zhang, X. Zhao, N. Yu, and F. Li, “Reversible data hiding in
encrypted images by reserving room before encryption,” IEEE Trans. Inf.
Forens. Security, vol. 8, no. 3, pp. 553–562, 2013.

[157] W. Zhang, K. Ma, and N. Yu, “Reversibility improved data hiding in encrypted
images,” Signal Processing, vol. 94, no. 1, pp. 118–127, 2014.

[158] X. Cao, L. Du, X. Wei, D. Meng, and X. Guo, “High capacity reversible data
hiding in encrypted images by patch-level sparse representation,” IEEE
Transactions on Cybernetics, 2016, to appear.

[159] C. W. Shiu, Y. C. Chen, and W. Hong, “Encrypted image-based reversible data

hiding with public key cryptography from difference expansion,” Signal
Processing: Image Communication, vol. 39, pp. 226–233, 2015

[160] Y. Liu, L. Ju, M. Hu, X. Ma, and H. Zhao, “A robust reversible data hiding
scheme for H.264 without distortion drift,” Neurocomputing, vol. 151, pp. 1053
– 1062, 2015.
 79

[161] G. Song, Z. Li, J. Zhao, J. Hu, and H. Tu, “A reversible video steganography
algorithm for MVC based on motion vector,” Multimedia Tools and
Applications, vol. 74, no. 11, pp. 3759–3782, 2015.

[162] J. Zhao, Z.-T. Li, and B. Feng, “A novel two-dimensional histogram

modification for reversible data embedding into stereo H.264 video,”
Multimedia Tools and Applications, 2016, to appear

[163] C. Vural and B. Barakli, “Reversible video watermarking using

motioncompensated frame interpolation error expansion,” Signal, Image and
Video Processing, vol. 9, no. 7, pp. 1613–1623, 2015.

[164] D. Xu, R. Wang, and Y. Q. Shi, “An improved reversible data hidingbased
approach for intra-frame error concealment in H.264/AVC,” Journal of Visual
Communication and Image Representation, vol. 25, no. 2, pp. 410–422, 2014.

[165] D. Xu, R. Wang, and Y.-Q. Shi, “Reversible data hiding in encrypted H.264/AVC
video streams,” in Proc. International Workshop on Digital-forensics and
Watermarking, 2013, pp. 141–152

[166] W. Puech, M. Chaumont, and O. Strauss, “A reversible data hiding method for
encrypted images,” in Proc. SPIE 6819, Security, Forensics, Steganography, and
Watermarking of Multimedia Contents X, 2008, pp. 68 191E–68 191E–9.

[167] Z. Qian, X. Han, and X. Zhang, “Separable reversible data hiding in encrypted
images by n-nary histogram modification,” in Proc. International Conference on
Multimedia Technology, 2013, pp. 869– 876.

[168] Z. Yin, B. Luo, and W. Hong, “Separable and error-free reversible data hiding in
encrypted image with high payload,” The Scientific World Journal, vol. 2014,
2014, article ID 604876.

[169] X. Zhang, Z. Qian, G. Feng, and Y. Ren, “Efficient reversible data hiding in
encrypted images,” Journal of Visual Communication and Image Representation,
vol. 25, no. 2, pp. 322–328, 2014.

[170] M. S. A. Karim and K. Wong, “Universal data embedding in encrypted domain,”

Signal Processing, vol. 94, pp. 174–182, 2014.

[171] Z. Qian, X. Zhang, Y. Ren, and G. Feng, “Block cipher based separable
reversible data hiding in encrypted images,” Multimedia Tools and Applications,
2016, to appear.
 80

[172] S. Zheng, D. Li, D. Hu, D. Ye, L. Wang, and J. Wang, “Lossless data hiding
algorithm for encrypted images with high capacity,” Multimedia Tools and
Applications, 2016, to appear.

[173] Z. Qian and X. Zhang, “Reversible data hiding in encrypted image with
distributed source encoding,” IEEE Trans. Circuits Syst. Video Technol., 2016,
to appear.

[174] X. Zhang, Z. Wang, J. Yu, and Z. Qian, “Reversible visible watermark embedded
in encrypted domain,” in Proc. IEEE China Summit and International
Conference on Signal and Information Processing, 2015, pp. 826–830.

[175] X. Zhang, C. Qin, and G. Sun, “Reversible data hiding in encrypted images
using pseudorandom sequence modulation,” in Proc. International Workshop on
Digital-forensics and Watermarking, 2012, pp. 358–367.

[176] X. Zhang, “Commutative reversible data hiding and encryption,” Security and
Communication Networks, vol. 6, no. 11, pp. 1396–1403, 2013.

[177] Z. Yin, H. Wang, H. Zhao, B. Luo, and X. Zhang, “Complete separable

reversible data hiding in encrypted image,” in Proc. First International
Conference on Cloud Computing and Security, 2015, pp. 101–110.

[178] M. Li, D. Xiao, Y. Zhang, and H. Nan, “Reversible data hiding in encrypted
images using cross division and additive homomorphism,” Signal Processing:
Image Communication, vol. 39, pp. 234–248, 2015.

[179] B. Ou, X. Li, and W. Zhang, “PVO-based reversible data hiding for encrypted
images,” in Proc. IEEE China Summit and International Conference on Signal
and Information Processing, 2015, pp. 831–835.

[180] D. Xu and R. Wang, “Separable and error-free reversible data hiding in

encrypted images,” Signal Processing, vol. 123, pp. 9–21, 2016.

[181] X. Zhang, J. Long, Z. Wang, and H. Cheng, “Lossless and reversible data hiding
in encrypted images with public key cryptography,” IEEE Trans. Circuits Syst.
Video Technol., 2016, to appear.

[182] D. Yan and R. Wang, “Reversible data hiding for audio based on prediction error
expansion,” in Proc. International Conference on Intelligent Information Hiding
and Multimedia Signal Processing, 2008, pp. 249– 252.
 81

[183] A. Nishimura, “Reversible audio data hiding using linear prediction and error
expansion,” in Proc. International Conference on Intelligent Information Hiding
and Multimedia Signal Processing, 2011, pp. 318– 321.

[184] ——, “Reversible audio data hiding based on variable error-expansion of linear
prediction for segmental audio and G.711 speech,” IEICE Transactions on
Information Systems, vol. 99-D, no. 1, pp. 83–91, 2016.

[185] F. Wang, Z. Xie, and Z. Chen, “High capacity reversible watermarking for audio
by histogram shifting and predicted error expansion,” The Scientific World
Journal, vol. 2014, 2014, article ID 656251.

[186] A. Nishimura, “Reversible and robust audio watermarking based on quantization

index modulation and amplitude expansion,” in Proc. International Workshop on
Digital-forensics and Watermarking, 2013, pp. 275–287.

[187] ——, “Reversible and robust audio watermarking based on spread spectrum and
amplitude expansion,” in Proc. International Workshop on Digital-forensics and
Watermarking, 2014, pp. 215–229.