Expert Systems With Applications 246 (2024) 123270

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Novel asymmetric CNN-based and adaptive mean predictors for reversible

data hiding in encrypted images
Ping Ping ∗,1 , Junyuan Huo 1 , Bobiao Guo
College of Computer Science and Software Engineering, Hohai University, Nanjing, 210098, China

ARTICLE INFO ABSTRACT

Keywords: With the proliferation of cloud computing and the Internet of Things(IoT), ensuring privacy and robust data
Reversible data hiding protection has become increasingly critical. Reversible data hiding in encrypted images (RDHEI) methods
Encrypted images emerge as a promising solution to address these challenges. In this paper, we propose an asymmetric CNN-
Image prediction
based predictor (ACNNP) that significantly outperforms other predictors in terms of prediction accuracy.
Data embedding
Then, an adaptive mean predictor is proposed to cooperate with ACNNP. According to these two predictors,
Privacy protection
a two-stage prediction and embedding model is designed to further enhance the embedding capacity and
make data embedding more flexible. Experimental results demonstrate that the proposed method exhibits
a high embedding capacity and reversibility when compared to state-of-the-art methods. On the BOWS-2,
BOSSBase, and UCID datasets, it obtains average embedding rates of 3.302, 3.461, and 2.69 bpp, respectively.
Therefore, this method provides an effective solution for applications such as data integrity verification, privacy
protection, and copyright preservation.

1. Introduction 2023), it modifies the value of the pixel located at the peak position
of the image histogram to reduce the impact of the embedded data
In today’s era of multimedia communication and cloud storage, dig- on the visual quality of the image. Then, prediction error expansion
ital images occupy an increasingly important position. How to embed (PEE) methods (Kouhi & Sedaaghi, 2021; Ou et al., 2019, 2013; Thodi
and extract data in images has become an urgent issue in the fields & Rodriguez, 2007) perform data embedding by exploiting the correla-
of digital media management, copyright protection, privacy protection tion between neighbouring pixels. Finally, Pixel value ordering (PVO)
and other fields (Zhang et al., 2016). Traditional data hiding methods methods (Dragoi et al., 2018; Fan et al., 2023; Li, Li, Li et al., 2013; Liu
embed data at the expense of permanent distortion of the original et al., 2023; Melendez-Melendez et al., 2023; Ou et al., 2016) embed
image, and the receiver cannot recover the original image after data data by ordering the pixel values in the division block.
extraction (Jiang et al., 2018). However, distortion of the image is For further privacy protection and facilitating the cloud server to
intolerable (Shi et al., 2016) in some special fields, such as medi- embed additional data into the encrypted image, many researchers
cal, aerospace, and military image processing. Therefore, reversible have explored the reversible data hiding in encrypted images(RDHEI)
data hiding(RDH) methods that are capable of recovering the origi- methods, which was first proposed in Zhang (2011). Based on the
nal image non-destructively after data extraction have been proposed order of image encryption, RDHEI methods can be classified into two
in Honsinger et al. (2001).
schemes: reserving room before encryption(RRBE) and vacating room
Currently, the research focus of many RDH methods is to minimize
after encryption(VARE) (Shi et al., 2016). For the VRAE scheme meth-
the embedding distortion to improve the visual quality of the image
ods, data hiders accomplish data hiding by modifying the pixel values
after data embedding. In the past decades, a large number of RDH
after encryption. Zhang (2011) divided the encrypted original image
methods have been proposed. They can be classified into the following
into several non-overlapping blocks, and then flipped the least sig-
categories according to their methods. The first category is difference
nificant bits(LSBs) in a block according to the embedded bits. Hong
expansion(DE) (Alattar, 2004; Kim et al., 2008; Tian, 2003), which
et al. (2012) improved the method in Zhang (2011) by calculating
embeds the data by enlarging the difference between two pixels. The
the smoothness of each block and considering the correlation between
second category is histogram shift(HS) (Gao et al., 2020; Huang &
Chang, 2013; Jia et al., 2019; Li, Li, Yang et al., 2013; Wang et al., pixels, which reduced the error rate during data extraction. Qin and

∗ Corresponding author.
E-mail addresses: [email protected] (P. Ping), [email protected] (J. Huo), [email protected] (B. Guo).
1
Equal contribution.

https://doi.org/10.1016/j.eswa.2024.123270
Received 26 October 2023; Received in revised form 25 December 2023; Accepted 15 January 2024
Available online 19 January 2024
0957-4174/© 2024 Elsevier Ltd. All rights reserved.
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Fig. 1. Application scenario of expert system in medical field.

Zhang (2015) made a further improvement in the quality of the recov- the physician can obtain the medical image and the clinical report
ered image by reducing the number of flipped LSBs in each block. Wu compiled by the radiologist. Based on the content of the medical image,
and Sun (2014) proposed a joint method and a separable method by the physician gives a reasonable and accurate diagnostic information.
employing prediction error, which improved the correct rate of data Finally, the diagnostic information from physician is embedded into the
extraction. Qian and Zhang (2016) utilized distributed source encoding medical image using the RDHEI methods and uploaded to the cloud
to compress a series of selected bits taken from the encrypted image to data centre for storage.
make room for data embedding. In general, the encrypted image has In this paper, we propose a two-stage RDHEI method using asym-
high information entropy, the embedding capacity of these methods is metric CNN-based and adaptive mean predictors. In the beginning, we
not very high. firstly preprocess the original image into two parts: the ‘Circle’ set
To increase the embedding capacity, Ma et al. (2013) firstly pro- image and the ‘Triangle’ set image as shown in Fig. 2. Then, we can
posed RRBE scheme method, where they preprocess the image to perform the two-stage data embedding procedure. In the first stage, we
reserve room for data embedding before encryption. Zhang et al. (2014) propose an asymmetric CNN-based predictor to predict the ‘Triangle’
hid the data in pixel prediction errors and designed a special encryption set image using the ‘Circle’ set image as input. After that, the steps
scheme to encrypt the prediction errors. Huang et al. (2016) proposed of image encryption, label map generation, and label map embedding
a simple yet effective RRBE scheme method. The original image is are performed sequentially for the content-owner. For the data-hider,
encrypted by block substitution and pixel bit-level stream cipher, and multi-MSB substitution is used to embed the secret data. In the second
the embedding process is the same as the traditional RDH method. Yin stage, we design an adaptive mean predictor to predict the ‘Circle’ set
et al. (2020) Obtained the label map of the original image by using image with itself as the input. Similar to the first stage, we repeat
mean edge detector(MED) predictor, and then hid the data in the the same procedure. The data-hider uses multi-MSB substitution to
multi-MSB of the original image after stream encryption. Wang et al. embed the secret data. Finally, we can get the marked image. The main
(2021) divided the image into non-overlapping blocks, then encrypted contributions of this paper are summarized as follows.
them using block-based encryption, classified and encoded the MSBs
planes in the blocks. After that, the data can be embedded in the MSBs (1) We design an asymmetric CNN-based predictor which can make
planes of each block. Chen and Chang (2022) divided the image into full use of the correlation between neighbouring pixels. Com-
3 × 3 non-overlapping blocks, using a variant of the MED predictor to pared with some classical local linear predictors, our proposed
perform MSB prediction. Then, data can be hid in each of the encrypted asymmetric CNN-based predictor has a better prediction perfor-
blocks. Wang and He (2022) also utilized block-based adaptive MSBs mance.
prediction and encryption, which has lower algorithmic complexity (2) We propose a two-stage prediction and embedding model.
and guaranteed security compared to Chen and Chang (2022). Overall, Through the collaboration of asymmetric CNN-based predictor
these methods greatly improve the embedding capacity compared with and adaptive mean predictor, it not only improves the embed-
the VARE scheme methods. ding capacity, but also makes the data embedding more flexible.
Expert systems have a wide range of applications in many fields. When the length of the secret data is less than the maximum
They not only contain a large amount of knowledge and experience at embedding capacity of the first embedding stage, it is possible
the level of experts in a certain domain, but also are able to simulate to omit the second stage of embedding.
the thought process of human experts to solve difficult problems (Tan (3) The procedures of data extraction and image recovery are in-
et al., 2016). In recent years, with the improved and mature cloud dependent of each other. For the receiver, the image can be
service technology, expert systems have been widely used in various recovered losslessly using the correct key and the embedded data
industries. Fig. 1 shows an RDHEI based application scenario of expert can be accurately extracted using a different key, separately.
system combined with cloud service technology in the medical field.
In the example, a patient’s medical image can be acquired through CT The remainder of the paper is organized as follows. Section 2
scan, X-ray, MRI, etc. The image is then input into the medical imaging describes our proposed method in detail. The results of the experiment
expert system to give an advice about the diagnosis information. The are presented in Section 3. Finally, Section 4 concludes the paper.
radiologist uses the result from the medical imaging expert system
to write a clinical report for the patient. Next, the clinical report is 2. Proposed method
embedded in the medical image using the RDHEI methods, obtaining
the marked image. The marked image is then stored in the cloud data In this section, we provide a detailed description of the proposed
centre. On the physician’s side, the marked image is downloaded. By RDHEI method. Fig. 2 shows the framework diagram of the proposed
performing image recovery and data extraction on the marked image, method. It consists of three parts: (1) prediction, (2) data embedding,

2
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Fig. 2. The framework of the proposed method.

(3) data extraction and image recovery, which are described and il- 2.1. Prediction
lustrated in detail in Sections 2.1, 2.2, and 2.3, respectively. In the
prediction part, we propose a new two-stage prediction method. In Convolutional Neural Networks(CNN) have a wide range of applica-
the first stage, an asymmetric convolutional neural network is adopted tions in the field of prediction, such as weather prediction (Park et al.,
to design the predictor. In the second stage, a novel adaptive mean 2022), traffic and travel prediction (Xu et al., 2023), human behaviour
predictor is applied. In the data embedding part, multi-MSB substitu- prediction (Xiang, 2022) and so on. Here we apply it in the field of
tion is adopted to embed the data and the auxiliary information. In pixel prediction. In this section, we first describe the image division
the data extraction and image recovery part, receivers with the data method and the idea of two-stage prediction. Then, we present a new
decryption key and the image decryption key can extract the hidden predictor based on asymmetric convolutional neural network(ACNNP).
data or recover the image losslessly, or both of them. It is described in the order of neural network structure and neural
In our approach, there are three users: content-owner, data-hider, network training. After that, another novel adaptive mean predictor is
and receiver, whose task is different. The pseudo-code of our proposed introduced in detail.
method is shown in the Algorithm 1. We have encapsulated the steps
of image processing according to their roles to facilitate understanding. 2.1.1. Image division and preprocessing
A more comprehensive introduction is as follows. Before image prediction, we adopt the image division method in Hu
Algorithm 1 The whole process of proposed RDHEI method. and Xiang (2021) to divide the pixels in the original image into two
Input: The original image 𝐼 𝑂𝑟𝑖 , encryption key 𝐾𝑇 , encryption key 𝐾𝐶 , parts: the ‘Circle’ set 𝐼 𝐶𝑖𝑟 and the ‘Triangle’ set 𝐼 𝑇 𝑟𝑖 as shown in Fig. 2.
encryption key 𝐾1 , encryption key 𝐾2 , secret data 𝑆𝐷1 , secret data 𝑆𝐷2 The original image can be expressed as 𝐼 𝑂𝑟𝑖 = 𝐼 𝐶𝑖𝑟 + 𝐼 𝑇 𝑟𝑖 . These
Output: The recovered image 𝐼 𝑅𝑒𝑐 , secret data 𝑆𝐷1 , secret data 𝑆𝐷2 two parts are independent of each other but have a certain degree
/* Different users have different operations and their inputs and outputs of correlation. After dividing, the pixel values that have no image
are also different. Here we discuss them case by case. */ information in the ‘Circle’ set and ‘Triangle’ set (the blank squares in
Fig. 2) are all filled with zero.
1: Case 1 : Content-owner
Then, during our two-stage prediction, we first use the ‘Circle’ set
2: (𝐼 𝐶𝑖𝑟 , 𝐼 𝑇 𝑟𝑖 ) = 𝐼𝑚𝑎𝑔𝑒_𝐷𝑖𝑣𝑖𝑠𝑖𝑜𝑛(𝐼 𝑂𝑟𝑖 );
to predict the ‘Triangle’ set with the asymmetric CNN-based predictor.
3: 𝐸 𝑇 𝑟𝑖 = 𝐼𝑚𝑎𝑔𝑒_𝐸𝑛𝑐𝑟𝑦𝑝𝑡𝑖𝑜𝑛(𝐼 𝑇 𝑟𝑖 , 𝐾𝑇 );
Compared with the traditional local linear predictors, it uses more
4: 𝐸 𝐶𝑖𝑟 = 𝐼𝑚𝑎𝑔𝑒_𝐸𝑛𝑐𝑟𝑦𝑝𝑡𝑖𝑜𝑛(𝐼 𝐶𝑖𝑟 , 𝐾𝐶 );
neighbouring pixels as inputs, which can effectively enhance the accu-
5: 𝐼̂𝑇 𝑟𝑖 = 𝐼𝑚𝑎𝑔𝑒_𝑃 𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛_𝐴𝐶𝑁𝑁𝑃 (𝐼 𝐶𝑖𝑟 );
racy of the prediction results. In the second stage, in order to cooperate
6: 𝐼̂𝐶𝑖𝑟 = 𝐼𝑚𝑎𝑔𝑒_𝑃 𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛_𝐴𝑀𝑃 (𝐼 𝐶𝑖𝑟 );
with the asymmetric CNN-based predictor and further improve the
7: 𝑀𝐴𝑃𝑇 𝑟𝑖 = 𝐿𝑎𝑏𝑒𝑙_𝑀𝑎𝑝_𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑜𝑛(𝐼̂𝑇 𝑟𝑖 , 𝐼 𝑇 𝑟𝑖 );
embedding capacity of the original image, we design an adaptive mean
8: 𝑀𝐴𝑃𝐶𝑖𝑟 = 𝐿𝑎𝑏𝑒𝑙_𝑀𝑎𝑝_𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑜𝑛(𝐼̂𝐶𝑖𝑟 , 𝐼 𝐶𝑖𝑟 );
predictor for the ‘Circle’ set image. It is a self-predicting predictor which
9: 𝐿_𝐸 𝑇 𝑟𝑖 = 𝐿𝑎𝑏𝑒𝑙_𝑀𝑎𝑝_𝑒𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔(𝐸 𝑇 𝑟𝑖 , 𝑀𝐴𝑃𝑇 𝑟𝑖 );
takes itself as the input. For the two of our proposed predictors, a
10: 𝐿_𝐸 𝐶𝑖𝑟 = 𝐿𝑎𝑏𝑒𝑙_𝑀𝑎𝑝_𝑒𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔(𝐸 𝐶𝑖𝑟 , 𝑀𝐴𝑃𝐶𝑖𝑟 );
detailed description is given below.
11: 𝐿_𝐸 = 𝐿_𝐸 𝑇 𝑟𝑖 + 𝐿_𝐸 𝐶𝑖𝑟 ;
/* 𝐿_𝐸 is transmitted to the data-hider and will be embedded
with the secret data. */ 2.1.2. Proposed asymmetric CNN-based predictor
12: Case 2 : Data-hider Neural network structure After image division and preprocessing
13: (𝐿_𝐸 𝑇 𝑟𝑖 , 𝐿_𝐸 𝐶𝑖𝑟 ) = 𝐼𝑚𝑎𝑔𝑒_𝐷𝑖𝑣𝑖𝑠𝑖𝑜𝑛(𝐿_𝐸); is completed, we use convolutional neural network to construct our
14: 𝐸 𝑆𝐷1 = 𝐷𝑎𝑡𝑎_𝑒𝑛𝑐𝑟𝑦𝑝𝑡𝑖𝑜𝑛(𝑆𝐷1 , 𝐾1 ); predictor. As shown in Fig. 3, the structure of our designed asymmetric
15: 𝐸 𝑆𝐷2 = 𝐷𝑎𝑡𝑎_𝑒𝑛𝑐𝑟𝑦𝑝𝑡𝑖𝑜𝑛(𝑆𝐷2 , 𝐾2 ); convolutional neural network predictor(ACNNP) consists of three main
16: 𝑀 𝑇 𝑟𝑖 = 𝐷𝑎𝑡𝑎_𝐻𝑖𝑑𝑖𝑛𝑔(𝐿_𝐸 𝑇 𝑟𝑖 , 𝐸 𝑆𝐷1 ); procedures: feature extraction, image prediction, and image reconstruc-
17: 𝑀 𝐶𝑖𝑟 = 𝐷𝑎𝑡𝑎_𝐻𝑖𝑑𝑖𝑛𝑔(𝐿_𝐸 𝐶𝑖𝑟 , 𝐸 𝑆𝐷2 ); tion.
18: 𝑀 = 𝑀 𝑇 𝑟𝑖 + 𝑀 𝐶𝑖𝑟 ; /* The marked image 𝑀 is transmitted to (1) Feature extraction: This procedure consists of three parallel
the receiver, which would be used to extract data and recover the asymmetric convolutional blocks with a parameter 𝐾 related to the size
original image. */ of the convolutional kernel. For a specific asymmetric convolutional
19: Case 3 : Receiver block, it has three parallel convolutional layers with kernel sizes of
20: (𝑀 𝑇 𝑟𝑖 , 𝑀 𝐶𝑖𝑟 ) = 𝐼𝑚𝑎𝑔𝑒_𝐷𝑖𝑣𝑖𝑠𝑖𝑜𝑛(𝑀); 1 × 𝐾, 𝐾 × 𝐾, 𝐾 × 1, followed by two activation unit leakyRELUs and
21: 𝑆𝐷1 = 𝐷𝑎𝑡𝑎_𝐸𝑥𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛(𝑀 𝑇 𝑟𝑖 , 𝐾1 ); two convolutional layers with a kernel size of 3 × 3, as illustrated in
22: 𝑆𝐷2 = 𝐷𝑎𝑡𝑎_𝐸𝑥𝑡𝑟𝑎𝑐𝑡𝑖𝑜𝑛(𝑀 𝐶𝑖𝑟 , 𝐾2 ); Fig. 4. The two activation unit leakyRELUs are located before each of
23: 𝐼 𝐶𝑖𝑟 = 𝐼𝑚𝑎𝑔𝑒_𝑅𝑒𝑐𝑜𝑣𝑒𝑟𝑦(𝑀 𝑇 𝑟𝑖 , 𝐾𝐶 ); the two convolutional layers. As 𝐾 increases, more pixels can be used to
24: 𝐼 𝑇 𝑟𝑖 = 𝐼𝑚𝑎𝑔𝑒_𝑅𝑒𝑐𝑜𝑣𝑒𝑟𝑦(𝑀 𝐶𝑖𝑟 , 𝐾𝑇 ); extract features and predict pixel values. This can reduce the influence
25: 𝐼 𝑅𝑒𝑐 = 𝐼 𝐶𝑖𝑟 + 𝐼 𝑇 𝑟𝑖 ; of local features on prediction, and improve the accuracy of prediction.
26: Return 𝐼 𝑅𝑒𝑐 , 𝑆𝐷1 , 𝑆𝐷2 . The reason for using asymmetric convolution is that it can explicitly
enhance the representation of the convolutional neural network and

3
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Fig. 3. Architecture of proposed asymmetric CNN-based predictor.

Fig. 5. Illustration on the use of adaptive mean predictor from four direc-
tions: (a) top; (b) left; (c) bottom; (d) right.

Fig. 4. The structure of the asymmetric convolutional block.

(3) Image preprocessing: It should be noted that the images in train

strengthen the ability of feature extraction, which is verified in Ding set and validation set need to be converted to gray scale images with
et al. (2019). In order to ensure the input and output sizes are the the size of 512 × 512.
same, when the convolution kernel size is 1 × 𝐾, 𝐾 × 𝐾, 𝐾 × 1, the (4) Training: After all the above steps have been completed, the
corresponding padding size is 0 × ⌊𝐾∕2⌋, ⌊𝐾∕2⌋ × ⌊𝐾∕2⌋, ⌊𝐾∕2⌋ × 0, back-propagation algorithm (Lecun et al., 1998) are utilized to optimize
the symbol ⌊⌋ represents downward rounding. The reflection padding our proposed asymmetric CNN-based predictor. For gradient descent,
mode is employed while performing the padding. It can resolve the we use Adam optimizer (Kingma & Ba, 2014) with a learning rate 10−3 ,
issue that the pixels on the edge of the image do not have adequate a weight decay 𝜆 10−3 and default hyperparameters. The loss function
pixels for prediction. The stride size of all the convolution operations is given below:
in this procedure is 1 × 1. Each asymmetric convolution block has 1
1 ∑ ( 𝑇 𝑟𝑖 ̂𝑇 𝑟𝑖 )2
𝑞
input channel and 32 output channels. Ultimately, the outputs of the 𝐿𝑂𝑆𝑆 = 𝐼 − 𝐼𝑎 + 𝜆 ‖𝜔‖22 (1)
𝑞 𝑎=1 𝑎
asymmetric convolution blocks with three different parameters 𝐾 are
further fused to obtain the input for the image prediction procedure. where q represents the count of the training images, 𝐼𝑎𝑇 𝑟𝑖 is the a-th
(2) Image prediction: Four convolutional layers and four activation ‘Triangle’ set image and 𝐼̂𝑎𝑇 𝑟𝑖 is the corresponding predicted ‘Triangle’
units leakyRELUs connected in series make up the image prediction set image. 𝜔 is all the weights of this predictor. The model is trained
procedure, and LeakyRELUs are located after each of the convolutional for 200 epochs with batch size 8.
layers. For each convolutional layer, the convolutional kernel size is
3 × 3, the padding is 1 × 1, the stride is 1 × 1, and the number of 2.1.3. Proposed adaptive mean predictor
input and output channels is 32. The output of the feature extraction To interoperate with the asymmetric CNN-based predictor and in-
procedure is fed into the first convolutional layer, passed through a crease the embedding capacity of our method, we design an adaptive
LeakyRELU, a second convolutional layer and a LeakyRELU, then fused mean predictor(AMP) for the ‘Circle’ set image. As shown in Fig. 5, the
with the output from the feature extraction as the input of the third con- centre pixel 𝐼(𝑖+1,𝑗+1) is the predicted pixel, which can be calculated
volutional layer. Finally, the output from the fourth convolutional layer using two pixels among four pixels 𝐼(𝑖,𝑗) , 𝐼(𝑖,𝑗+2) , 𝐼(𝑖+2,𝑗) , 𝐼(𝑖+2,𝑗+2) on the
and the LeakyRELU is used as the input to the image reconstruction diagonal lines. We define four directions: top, left, right and bottom.
procedure. 𝑡𝑜𝑝
For top direction, the predicted pixel value 𝑝𝑐(𝑖+1,𝑗+1) is calculated using
(3) Image reconstruction: the image reconstruction procedure is
made up of a single convolutional layer, its input channel is 32 and the two pixels 𝐼(𝑖,𝑗) and 𝐼(𝑖,𝑗+2) . In this case, the first row, first column,
output channel is 1. The convolutional kernel size, padding size, stride and the last column of the image need to be regarded as reference pixels
are 3 × 3, 1 × 1 and 1 × 1 respectively. The outcome of the image pre- to restore the original image, and their pixel values need to remain
diction procedure is regarded as the input of the image reconstruction unchanged. The specific prediction formula is as follows:
procedure, the final output is the predicted ‘Triangle’ set image 𝐼̂𝑇 𝑟𝑖 . 𝑡𝑜𝑝 ⌊( ) ⌋
𝑝𝑐(𝑖+1,𝑗+1) = 𝐼(𝑖,𝑗) + 𝐼(𝑖,𝑗+2) ∕2 , 1 ≤ 𝑖 < 𝑚, 1 ≤ 𝑗 ≤ 𝑛 − 2 (2)
Neural network training After building the network structure of
the ACNNP using asymmetric convolutional blocks as well as regular where 𝐼(𝑖,𝑗) and 𝐼(𝑖,𝑗+2) are the specific pixel values at the positions
convolutional layers and activation functions, we proceed to train the with coordinates (𝑖, 𝑗) and (𝑖, 𝑗 + 2) in the image. The symbol ⌊⌋ means
network. The detailed description of the train process are given below. downward rounding. The 𝑚 and 𝑛 are the height and width of the
(1) Runtime environment: The proposed asymmetric CNN-based image. If top direction is chosen, it means the prediction is performed
predictor is trained on an Intel(R) Xeon(R) Platinum 8255C CPU from the top to the bottom in the circle set image. It can also predict in
(2.50 GHz) with V100-SXM2-32 GB GPU. the three directions as shown in Fig. 5(b), (c), (d). When predicting
(2) Prepare the dataset: We randomly select 2000 images from from left to right, the first column, first row, and last row remain
the ImageNet as the train set and randomly select 800 images as the unchanged as reference pixels; when predicting from right to left, the
validation set. first row, last column, and last row remain unchanged as reference

4
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

pixels; When predicting from bottom to top, the first column, the last
column, and the last row are kept unchanged. The specific prediction
formulas for these cases are as follows:
𝑙𝑒𝑓 𝑡 ⌊( ) ⌋
𝑝𝑐(𝑖+1,𝑗+1) = 𝐼(𝑖,𝑗) + 𝐼(𝑖+2,𝑗) ∕2 , 1 ≤ 𝑖 ≤ 𝑚 − 2, 1 ≤ 𝑗 < 𝑛 (3)

𝑟𝑖𝑔ℎ𝑡 ⌊( )⌋
𝑝𝑐(𝑖+1,𝑗+1) = 𝐼(𝑖,𝑗+2) + 𝐼(𝑖+2,𝑗+2) ∕2 , 1 ≤ 𝑖 ≤ 𝑚 − 2, 0 ≤ 𝑗 < 𝑛 − 1 (4)
Fig. 6. An instance of pixel labelling.

𝑏𝑜𝑡𝑡𝑜𝑚
⌊( ) ⌋
𝑝𝑐(𝑖+1,𝑗+1) = 𝐼(𝑖+2,𝑗) + 𝐼(𝑖+2,𝑗+3) ∕2 , 0 ≤ 𝑖 < 𝑚 − 1, 1 ≤ 𝑗 ≤ 𝑛 − 2 (5)
2.2.2. Label map generation
Here, we assume that 𝐼 𝑆 is the predicted image, where
For a current pixel 𝑐 in the image, the predicted pixel 𝑝𝑐 associated
𝑆 ∈ {‘top’, ‘left’, ‘right’, ‘bottom’}. When 𝑆 equals ‘top’, 𝐼 𝑡𝑜𝑝 refers to with it can be obtained by using the asymmetric CNN-based predictor
the image predicted from top to bottom in the original image using or the adaptive mean predictor introduced in Sections 2.1.2 and 2.1.3
the Eq. (2). When 𝑆 takes other values, 𝐼 𝑆 is the image predicted by respectively. We convert both 𝑐 and 𝑝𝑐 into the form of an 8-bit binary
the corresponding prediction method. There are four cases in total but sequence 𝑐 𝑘 , 𝑝𝑐 𝑘 . Then, the maximum value of 𝑙 obtained by Eq. (10)
we only need to apply one case in practice. So we need to determine is the label of this pixel.
the prediction direction in advance. Here we use an exhaustive method
𝑘 𝑘
that gets four predicted images using the Eqs. (2)∼(5) respectively, and 𝑎𝑟𝑔 max 𝑐𝑙−𝑀𝑆𝐵 = 𝑝𝑐𝑙−𝑀𝑆𝐵 , 𝑙 = 0, 1, 2...8 (10)
𝑙
then, the prediction direction 𝑆 is determined by the Eq. (6). 𝑘 𝑘
where 𝑐𝑙−𝑀𝑆𝐵 , 𝑝𝑐𝑙−𝑀𝑆𝐵 are the 𝑙 − 𝑀𝑆𝐵 values of 𝑐 𝑘 and 𝑝𝑐 𝑘 , respec-
𝑆 tively . Since the length of the binary sequence is 8, the label can be
𝑆 = 𝑎𝑟𝑔 𝑚𝑖𝑛 𝑀𝑆𝐸(𝐼, 𝐼 ) (6)
𝑆 any number from 0 to 8.
where 𝐼 is the original image, the MSE() is to calculate the mean For instance, as shown in Fig. 6, we suppose the current pixel
squared error of the pixel prediction error between the original image 𝑐 = 130, and its corresponding predicted pixel 𝑝𝑐 = 135. We first
and the predicted image, the specific calculation formula is shown as use the Eq. (8) to convert them into a binary sequence, where 𝑐 𝑘 =
follows: ‘100000010’, 𝑝𝑐 𝑘 = ‘10000111’. A comparison between 𝑐 𝑘 and 𝑝𝑐 𝑘
shows that the first 5 most significant bits of both are the same,
1 ∑∑
𝑚 𝑛
𝑆 2
𝑀𝑆𝐸(𝐼, 𝐼 𝑆 ) = (𝐼 − 𝐼(𝑖,𝑗) ) (7) so the label of the pixel is 5, which indicates that there are 6 bits
𝑚 × 𝑛 𝑖=1 𝑗=1 (𝑖,𝑗) of information that can be embedded into this pixel. Eventually, by
where 𝑚, 𝑛 are the height and width of the original image. (𝑖, 𝑗) is the traversing all the predicted pixels in the image, all the labels called the
specific coordinates of a pixel in the image, 𝑖 and 𝑗 are two integers label map of the image can be obtained. Note that in our approach, we
divide the original image into the ‘Triangle’ set image and the ‘Circle’
and 𝑖 ∈ [1, 𝑚], 𝑗 ∈ [1, 𝑛].
set image, and the pixels in the blank squares of the ‘Triangle’ and
‘Circle’ set image are not required to generate labels. As a result, for
2.2. The embedding part a 𝑁 × 𝑁 sized image, we can get one label map 𝑀𝐴𝑃𝑡𝑟𝑖 that contains
⌊ 2 ⌋
𝑁 ∕2 labels for the ‘Triangle’ set image and one label map 𝑀𝐴𝑃𝑐𝑖𝑟
⌊ ⌋
In the embedding part, we hide the information through multi- that contains (𝑁 2 − (𝑁 2 )∕2 − 3𝑁 + 2) labels for the ‘Circle’ set image.
MSB substitution. Due to our prediction being divided into two stages,
our embedding is also divided into two parts as shown in Fig. 2. For 2.2.3. Label map embedding
each part, it consists of image encryption, label map generation, label After obtaining the label map of each set image, we need to convert
map embedding and secret data hiding. In the following sections, the them into binary sequences and embed them in the encrypted image.
embedding procedures will be described. The purpose of this operation is to restore the original image losslessly.
Here, we use Huffman coding to compress the label maps to improve
the embedding capacity. Huffman coding is a lossless data compression
2.2.1. Image encryption technique. The idea is to assign variable-length codes to the labels.
Considering the privacy of the image, content-owner needs to en- Based on the label map, we count the frequencies of corresponding
crypt the image before transferring it to the data-hider. He can use two labels and build a Huffman tree whose leaf nodes are labels.
encryption keys 𝐾𝑇 , 𝐾𝐶 and chaotic logistic maps (Pareek et al., 2006) Take Lena image for example, we can get the label map of Lena’s
to generate two random matrices with the same size 𝑚 × 𝑛 of the image. ‘Triangle’ set image through the prediction method presented in Sec-
One matrix is used to encrypt the ‘Triangle’ set image and the other tion 2.1.2. The frequencies of corresponding labels are shown in
matrix is used to encrypt the ‘Circle’ set image. The elements in these Table 1. We take the occurrence frequency of the labels as the weights
two matrices are random integers between 0 and 255. Take ‘Triangle’ of the nodes, then we select the two nodes with the lowest frequency
set image for example, the image and all the elements in the random of occurrence from all the labels as the child nodes to generate a tree,
𝑘
matrix are converted into 8-bit binary sequences, denoted as 𝐵(𝑖,𝑗) and and add their weights as the weights of the parent nodes. This step
𝑅𝑘(𝑖,𝑗) . The conversion process uses the following formula: is then repeated in an analogous manner, and when all the nodes
⌊ ⌋ have been connected together to form a tree, a Huffman tree has
𝑘
𝐼(𝑖,𝑗) been constructed, as shown in Fig. 7. We assign a value of ‘0’ to
𝐵(𝑖,𝑗) = 𝑚𝑜𝑑 2, 𝑘 = 1, 2, 3..., 8 (8)
2𝑘−1 the left branch and a value of ‘1’ to the right branch, and the value
𝑘 assigned to each branch on the path from the root node to each leaf
where 𝐵(𝑖,𝑗) denotes the 𝑘th bit value in the binary form of the pixel
node constitutes a binary string, which is the Huffman codes of the
𝐼(𝑖,𝑗) . Then, the image can be encrypted by
associated leaf node label. Thus, we obtain the Huffman coding rules
𝑘
𝐸(𝑖,𝑗) 𝑘
= 𝐵(𝑖,𝑗) ⊕ 𝑅𝑘(𝑖,𝑗) (9) {111 011,111 010,11 100,1111,110,01,00,100,101} for the label map
of Lena’s ‘Triangle’ set image.
where the ⊕ refers to XOR operation, 𝑘
is the obtained encrypted
𝐸(𝑖,𝑗) After obtaining the Huffman coding rules for the label map 𝑀𝐴𝑃𝑡𝑟𝑖
binary sequence. Finally, the encrypted image can be obtained by and the label map 𝑀𝐴𝑃𝑐𝑖𝑟 , we convert these two label maps into two
converting the binary sequence to pixel matrix. binary sequences with Huffman compression. In order to facilitate the

5
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Table 1
The statistical quantity and Huffman codes of each label for Lena’s ‘Triangle’ set image.
Label 0 1 2 3 4 5 6 7 8
Statistical quantity 3117 2742 5052 12 109 19 092 28 464 25 684 16 458 18 354
Huffman codes 111 011 111 010 11 100 1111 110 01 00 100 101

where 𝑏𝑡 is the information to be embedded. 𝑝_𝐵𝑡 is the t-th bit in 𝑝_𝐵

from right to left. After that, 𝑃 _𝑁𝑒𝑤 is treated as a new pixel in the
current position.
Step 3. Traverse the pixels with labels one by one in the encrypted
image in the order of ‘Z’ scanning, and repeat Step2 to embed the
information until the whole information is embedded or the last pixel
with label has been embedded.
Step 4. Obtain the encrypted image after embedding information.

2.2.4. Data hiding

After the data hider receives the encrypted image, he divides it into
‘Triangle’ and ‘Circle’ set image with labels respectively. Then, he needs
to extract two Huffman coding rules first from the reference pixels of
the ‘Circle’ set image. It should be noted that he does not know the
prediction direction of the adaptive mean predictor in advance. That
means he cannot be sure which three edges of the ‘Circle’ set image
are reference pixels. Luckily, however, Huffman encoding is a prefix
Fig. 7. The architecture of the constructed Huffman tree. encoding so he is able to extract Huffman coding from four directions
through an exhaustive method. Only one direction can extract the
correct Huffman coding. After two Huffman coding rules extracted from
the 1st to 32nd reference pixels, the partial label maps 𝑀𝐴𝑃𝑡𝑟𝑖 and
𝑀𝐴𝑃𝑐𝑖𝑟 used in two parts can also be extracted from the 33rd to the
last reference pixels. Subsequently, the remaining label maps 𝑀𝐴𝑃𝑡𝑟𝑖
and 𝑀𝐴𝑃𝑐𝑖𝑟 can be extracted and restored from the pixels with labels
in the encrypted ‘Triangle’ and ‘Circle’ set image.
Fig. 8. The architecture of the constructed Huffman tree. The embedding of secret information also adopts the Eq. (11).
It should be noted that when embedding the secret message in the
encrypted ‘Circle’ set image, the starting position is at the end of the
data hiders and receivers to decode the labels after extracting the local embedded reference pixels; when embedding the secret message in the
map, we embed two Huffman coding rules into the reference pixels encrypted ‘Triangle’ set image, the starting position is at the end of
of the ‘Circle’ set image. Here, we use 4 bits to record the length of the embedded label map 𝑀𝐴𝑃𝑡𝑟𝑖 . The length of the secret information
each Huffman code, so we need 36 bits in total. Then, we use 8 bits embedded in each of the two image sets will be recorded in a space of
to record each Huffman code and the total occupied space is 72 bits. 20 bits reserved earlier. In addition, in order to improve the security of
Also, here we reserve 20 bits for recording other auxiliary information. the secret information, the data hider uses two secret keys 𝐾1 and 𝐾2
In general, the total number of bits of auxiliary information is 128 bits, to encrypt the secret information before embedding.
which is 16 pixels. Since our embedding is divided into two parts, there When both parts of the embedding are done, we can get the marked
will be two Huffman coding rules and two label maps. So a total of ‘Triangle’ set image and the marked ‘Circle’ set image. Then, we can
32 reference pixels are reserved to embed the auxiliary information as fuse them into the marked image which can be transferred to the
shown in Fig. 8. receiver.
After embedding the auxiliary information, in order to extract the In order to understand the specific steps of data hiding more clearly,
label maps more conveniently, the remaining reference pixels are di- we give a simple concrete example here, as shown in Fig. 9. We break
vided into two equal parts. One part is used to embed the partial label it down into five sub-steps for detailed description.
map of the ‘Circle’ set image, and the other part is used to embed (1) Image division: For an encrypted image with labels of size
the partial label map of the ‘Triangle’ set image. The remaining parts 32 × 32, we first divide it into encrypted ‘Triangle’ and ‘Circle’ set
for two label maps are embedded into the pixels with labels in the images according to the image division method introduced in Sec-
encrypted ‘Triangle’ set image and ‘Circle’ set image respectively. Note tion 2.1.1. In encrypted ‘Circle’ image, we mark all reference pixels in
that all reference pixels are also converted to a binary sequence and red and number all reference pixels.
embedded in the ‘Circle’ set image after the label map 𝑀𝐴𝑃𝑐𝑖𝑟 . The
(2) Huffman coding rules extraction: When extracting Huffman
specific embedding process is as follows:
coding rules used in ‘Triangle’ set image, we convert all reference pixels
Step 1. Determine the starting position of the embedding. When
numbered 1 to 16 into 8-bit binary form and splice them into a binary
embedding the remaining label map, the starting position is the first
sequence. The first 36 bits of information are taken out and divided
predicted pixel of the encrypted image where the information can be
by 4 bits. Next, 72 bits of information are taken out and divided by 8
embedded.
bits. Based on the segmented information, Huffman coding rules can
Step 2. For the pixel where information is to be embedded, we
be extracted. For example, for the first 4 bits of information ‘0101’,
convert it to 8-bit binary sequence denoted as 𝑝_𝐵. Then, according to
we convert it to decimal number 5, then we take out the last 5 bits
the label map of the image, we can get the corresponding label of this
of information from the first 8 bits of the binary sequence, which is
pixel denoted as 𝑙. The embedding equation is shown below:
{∑𝑙 used as Huffman code for the label 0 ; for the second 4 bits message
7−𝑡 ) + ∑7−𝑙 (𝑝_𝐵 × 2𝑡−1 ), 0 ≤ 𝑙 < 7
𝑡=0 (𝑏𝑡 × 2 𝑡=1 𝑡 ‘0101’, we convert it to a decimal number of 5, then we take the last 5
𝑃 _𝑁𝑒𝑤 = ∑8 (11)
(𝑏 × 2 8−𝑡 ), 6<𝑙≤8 bits of information from the second 8 bits binary sequence as Huffman
𝑡=1 𝑡

6
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Fig. 9. Example of the data hiding procedure.

code for label 1; for the third 4 bits message ‘0100’, we convert it to rules, we can convert it into the corresponding labels. According to
a decimal number of 4, then we take the last 4 bits of information these extracted labels, we can extract information from the image and
from the third 8 bits binary sequence as Huffman code for label 2. So convert it into labels. By repeating this, we can finally get two complete
on and so forth, we can get the final Huffman coding rules. The same label maps respectively.
as extracting Huffman coding rules used in ‘Circle’ set image which is (5) data embedding: After obtaining the complete label map, we
extracted from reference pixels numbered 17 to 32. can use multi-MSB substitution to embed secret information. In the
encrypted ‘Triangle’ set image, the green triangle represents the pixels
(3) Partial label map extraction: Divide the reference pixels num-
occupied by the embedded partial label map, and the yellow trian-
bered 33 to 47 into two parts according to their quantity, convert them
gle represents the pixels that can embed secret information. In the
into 8-bit binary form and splice them into two binary sequences. Based
encrypted ‘Circle’ set image, the green circle represents the pixels oc-
on the Huffman codes extracted in the previous steps, we can extract cupied by the embedded partial label map, the purple circle represents
two partial label maps used in ‘Triangle’ and ‘Circle’ set image. the pixels occupied by the embedded reference pixels, and the blue
(4) Label map extraction and recovery: Based on the extracted circle represents the pixels that can embed secret information. After
label maps, we can extract part of the information in the ‘Triangle’ embedding secret information, the two images will be fused into the
and ‘Circle’ set images respectively. According to the Huffman coding marked image.

7
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

label map extracted from the reference pixels, the entire label map can
be extracted and recovered from the marked ‘Circle’ set image.
Step 2. Since all reference pixels are embedded after the label map
𝑀𝐴𝑃𝑐𝑖𝑟 and the total number of the reference pixels can be calculated
by the size of the marked image, we are able to extract and restore all
reference pixels with the help of the extracted label map 𝑀𝐴𝑃𝑐𝑖𝑟 . Note
that these reference pixels do not need to be decrypted.
Step 3. Use decryption key 𝐾𝐶 to decrypt the marked ‘Circle’ set
Fig. 10. Illustration of using AMP in ‘top’ direction for a ‘Circle’ set image sized 5 × 8.
image.
Step 4. Recover each pixel except for reference pixels in the de-
crypted marked ‘Circle’ set image and obtain the ‘Circle’ set image.
As shown in Fig. 10, we take ‘top’ prediction direction for example.
2.3. Data extraction and image recovery First, we get the predicted values of the pixel 1, 2, 3 with the predictor
AMP which takes reference pixels as the inputs. Then, each pixel and its
In our proposed RDHEI, the data extraction and image recovery are corresponding predicted value are converted into 8-bit binary sequence
independent from each other. There are three cases for the receiver to respectively, denoted as 𝑝_𝐵 and 𝑝𝑟𝑒_𝐵. According to the label map, we
process the marked image. can get the label of this pixel denoted as 𝑙. The reconstruction process
Case 1: The receiver only has the decryption keys 𝐾1 and 𝐾2 . The can be expressed as follows:
secret data can be obtained by extracting the encrypted secret data from ∑𝑙 ∑7−𝑙
the marked image and directly decrypting the extracted data. However, ⎧ 𝑡=1 𝑝𝑟𝑒_𝐵𝑡 × 28−𝑡 + 𝑡=1 𝑝_𝐵𝑡 × 2𝑡−1
⎪
the original image cannot be recovered since there is no decryption 𝑝_𝑅𝑒𝑐 = ⎨ +𝑏𝑙+1 × 27−𝑙 , 0 ≤ 𝑙 < 8 (13)
keys 𝐾𝑇 and 𝐾𝐶 . ⎪ ∑8 8−𝑡
⎩ 𝑡=1 𝑝𝑟𝑒_𝐵𝑡 × 2 ,𝑙 = 8
Case 2: The receiver only has the decryption keys 𝐾𝑇 and 𝐾𝐶 . The
original image can be recovered losslessly with these two keys. where 𝑝_𝑅𝑒𝑐 is the recovered pixel in the current position. 𝑝𝑟𝑒_𝐵𝑡 is
Case 3: The receiver owns all the keys 𝐾1 , 𝐾2 , 𝐾𝑇 and 𝐾𝐶 . Obvi- the 𝑡th bit of 𝑝𝑟𝑒_𝐵 from left to right. 𝑝_𝐵𝑡 is the t-th bit of 𝑝_𝐵 from
ously, he can reversibly extract the secret data and recover the original right to left. After recovering the pixels 1, 2, 3, we can calculate the
image. predicted values of the pixels 4, 5, 6 and recover these pixels with the
same method. The order of the pixel recovery is from top to bottom in
In the following, we will elaborate the data extraction and image
this example.
recovery.
(2) Recover the ‘Triangle’ set image
After restoring the ‘Circle’ set image, we can restore the ‘Triangle’
2.3.1. Data extraction
set image. Its recovery process is similar to that of ‘Circle’ set image.
When the receiver obtains the marked image, he first divides it
The specific recovery process is as follows:
into two parts: the marked ‘Triangle’ set image and the marked ‘Circle’
Step 1. Extract Huffman coding rules and partial label map of the
set image according to the image division pattern presented in Sec-
‘Triangle’ set image from the reference pixels. Then, the entire label
tion 2.1.1. Then, he can extract two Huffman coding rules and two
map can be extracted and restored from the marked ‘Triangle’ set
local label maps using the same method as the data hiding. The length
image.
of two parts of secret data can also be extracted from the reference
Step 2. Take the ‘Circle’ set image as the input of the predictor
pixels. When extracting the secret data in the marked ‘Circle’ set image,
ACNNP and get the predicted ‘Triangle’ set image.
the starting position is at the end of the embedded reference pixels; Step 3. Use decryption key 𝐾𝑇 to decrypt the marked ‘Triangle’ set
When extracting the secret data in the marked ‘Triangle’ set image, the image.
starting position is at the end of the embedded label map 𝑀𝐴𝑃𝑡𝑟𝑖 . When Step 4. Select a pixel in the decrypted marked ‘Triangle’ set image
performing data extraction, for one pixel containing secret information, and convert it into an 8-bit binary sequence 𝑝_𝐵. In the predicted
we first convert it into a 8-bit binary sequence 𝑝_𝐵. Based on the label ‘Triangle’ set image, the pixel at the corresponding position is also
map, we can get the corresponding label 𝑙 for this pixel. The data taken out and converted into an 8-bit binary sequence denoted as 𝑝𝑟𝑒_𝐵.
extraction formula is as follows: According to the label map, we can get the label of this pixel denoted
{
𝑝_𝐵 (𝑙+1)−𝑀𝑆𝐵 , 0 ≤ 𝑙 < 7 as 𝑙. Then, utilizing the Eq. (13), the pixel in this position can be
𝑃 _𝑆𝐷 = (12)
𝑝_𝐵, 6<𝑙≤8 recovered.
Step 5. Traverse the pixels one by one in the decrypted marked
where 𝑃 _𝑆𝐷 presents the extracted secret data from this pixel.
‘Triangle’ set image in the order of ‘Z’ scanning, and repeat Step4.
𝑃 _𝐵 (𝑙+1)−𝑀𝑆𝐵 is the (𝑙 + 1) most significant bits of 𝑝_𝐵. Next, we scan
Step 6. Obtain the ‘Triangle’ set image.
the pixels in the marked ‘Triangle’ and ‘Circle’ set images except the
After getting the ‘Circle’ set image and ‘Triangle’ set image, we
reference pixels in the order of ‘Z’ scanning respectively, and extract the
can fuse them into the original image. It can be seen from the image
embedded secret data. Finally, the original secret data 𝑆𝐷1 and 𝑆𝐷2
recovery process that only with the two keys for image decryption can
can be obtained by decrypting the extracted data with the decryption the complete original image be recovered without loss.
keys 𝐾1 and 𝐾2 , respectively.
3. Experiment and analysis
2.3.2. Image recovery
The image recovery is independent from the data extraction, and In this section, we first investigate the prediction performance of
they can be done simultaneously. There are two parts for image recov- the asymmetric CNN-based predictor and the adaptive mean predictor,
ery, recovering the ‘Triangle’ set image and recovering the ‘Circle’ set and then show the experimental results and compare them with some
image. When recovering the ‘Triangle’ set image, the ‘Circle’ set image classical predictors. Next, we conduct capacity analysis and security
is needed to predict the ‘Triangle’ set image by the ACNNP. As a result, analysis of our proposed RDHEI method. For these experiments, we
it requires to recover the ‘Circle’ set image firstly. The procedure of use three datasets: BOSSBase (Bas et al., 2011), UCID (Schaefer & Stich,
recovering the ‘Circle’ set image is described below. 2003), BOWS-2 (Bas & Furon, 2017), and six common test images, some
(1) Recover the ‘Circle’ set image of which are downloaded from the dataset USC_SIPI (Weber, 2006), as
Step 1. Extract the Huffman coding rules from the reference pixels shown in Fig. 13(a). In addition, PSNR and SSIM are also used to test
and determine the prediction direction. After that, based on the partial the lossless recovering ability of the method.

8
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Table 2 Table 3
Mean, VAR and MSE on UCID dataset for different predictors. Mean, VAR and MSE on UCID dataset for different non-linear predictors.
Method Mean VAR MSE Prediction patten Method Mean VAR MSE
Coltuc (2011) 6.772 138.237 201.394 Hu and Xiang (2021) 4.466 53.699 81.067
1:1
Thodi and Rodriguez (2007) 7.222 146.475 204.102 Proposed ACNNP 4.107 52.692 78.821
Sachnev et al. (2009) 5.758 81.144 128.244
Hu and Xiang (2022) 3.932 40.546 61.564
Chen and Chang (2022) 8.249 201.159 297.857 3:1
Proposed ACNNP 3.838 38.810 58.623
Qin et al. (2022) 6.213 113.528 167.924
Ni and Bi (2023) 6.428 138.568 196.357
Proposed AMP 10.623 285.460 429.731
Proposed ACNNP 4.107 52.692 78.821
are 3.838, 38.810, 58.623, which are the smallest under comparison
with the predictor in Hu and Xiang (2022). The experiment results
show that our proposed ACNNP has better prediction performance
3.1. Predictor performance than state-of-the-art predictors. This further verifies the importance and
criticality of asymmetric convolutional neural networks in improving
To measure the performance of our proposed predictors, we com- pixel prediction performance.
pare our predictors with six local linear predictors (Chen & Chang,
2022; Coltuc, 2011; Ni & Bi, 2023; Qin et al., 2022; Sachnev et al., 3.2. Capacity analysis
2009; Thodi & Rodriguez, 2007) and two non-linear CNN-based pre-
dictors (Hu & Xiang, 2021, 2022). In the experiments, 1338 images In our method, since two label maps and the auxiliary information
from the UCID dataset are used as the test dataset. For each image, we need to be embedded in the image, the net embedding capacity(EC)
calculate the prediction errors between the original pixels and the cor- of our method equals the total embedding capacity minus the space
responding predicted pixels. Three evaluation indicators named Mean occupied by the auxiliary information and label maps. Table 4 records
of absolute value (Mean), Variance (Var), and Mean square error (MSE) the total EC, the length of label maps, the length of auxiliary infor-
are utilized to reveal the performance of the predictors. The smaller mation and the net EC of the six test images. The embedding rate(ER)
the values of Mean, VAR, and MSE, the higher the accuracy of the calculated by net EC is also recorded in Table 4. It can be seen that the
predictor’s prediction results and the better the prediction performance. Jetplane image has the highest EC of 834590 bits and the ER is 3.183
Table 2 shows the average values of Mean, VAR and MSE for our bpp. The Baboon has the EC of 314330 bits and its ER is 1.199 bpp,
proposed predictors and six local linear predictors on UCID dataset. which is the lowest in the six test images.
We can see that our proposed ACNNP has the lowest values of Mean, To better demonstrate the superiority of our method, we compare
VAR and MSE, which are 4.107, 52.692, and 78.821 respectively. It our embedding capacity with eight recent works (Chen, 2020; Chen &
indicates that the ACNNP has the best prediction performance among Chang, 2022; Fu et al., 2019; Wang et al., 2021; Wang & He, 2022; Wu
these predictors. The reason is that, compared with traditional linear et al., 2020; Yin et al., 2020; Yu et al., 2022). We conduct experiments
predictors, ACNNP constructed by convolutional neural networks has on six test images, and the final results are shown in Fig. 11. Among
multiple receptive fields, which can better represent the relationship the six test images, the six methods all achieved higher embedding rates
features between pixels. Besides, it has global optimization capability on the ‘Jetplane’ image, which were 1.696 bpp, 2.503 bpp, 3.03 bpp,
which greatly improves the prediction performance. However, our 2.672 2.772 bpp, 2.022 bpp. The embedding rate of our method on
proposed AMP has a higher Mean, VAR, and MSE, which are 10.623, ‘Jetplane’ image was 3.183 bpp, it is the highest among all methods.
285.460, 429.731. The reason is that AMP is the second stage of our The embedding rate of all methods on the ‘baboon’ image is low. only
prediction and it uses only two pixels on the diagonal as the context to the method in Yin et al. (2020) and our method have an embedding
predict the current pixel. Although AMP is not as good as the other rate greater than 1 bpp on this image. The reason for this phenomenon
predictors, it can coordinate with ACNNP and further increase the is that the embedding capacity of an image is related to the pixel
embedding capacity of our method. value prediction performance, which is closely related to the texture
In addition to the traditional local linear predictors, we also com- complexity of the image. When the texture complexity of an image
pare our ACNNP with two non-linear CNN-based predictors which are is higher, then the pixel prediction of that image is less effective,
presented in Hu and Xiang (2021, 2022). For CNN-based predictors, which in turn affects the embedding capacity of that image. The image
there are two prediction patterns. One prediction pattern is using half ‘Jetplane’ has a lower texture complexity and the image ‘baboon’ has
of the pixels in the image to predict the other half which is adopted a higher texture complexity, which is why all the methods achieve
in Hu and Xiang (2021). The other prediction pattern is using three higher and lower capacities on these two images. As can be seen,
quarters of the pixels in the image to predict the remaining pixels which our method achieves the highest embedding rate in all test images
is adopted in Hu and Xiang (2022). Since there is a certain correlation compared with other methods. The high embedding capacity of our
between the pixels in the image, the more pixels that are used as the method benefits from two aspects. First, the collaboration between
input of the predictor, the better the prediction will be theoretically. ACNNP and AMP increases prediction performance. Then, compressing
In order to make a fair comparison, we compare our proposed AC- the auxiliary information using the Huffman codes vacates more room
NNP with these predictors separately by different prediction patterns. to embed.
All the two predictors and our proposed ACNNP are trained using the In order to verify the generality of our method, we test it on three
same training set, validation set and training environment introduced datasets: BOSSBase, BOWS-2 and UCID. Among them, BOSSBase and
in Section 2.1.2. The number of training epoch is set to 140. For the two BOWS-2 have 10 000 gray scale images with the size of 512 × 512
predictors compared, they adopt the image division pattern in their re- and UCID has 1338 gray scale images with the size of 384 × 512
spective papers. For our proposed ACNNP, we adopt the image division or 512 × 384. The experimental results are shown in Table 5. The
pattern introduced in Section 2.1.1 when the prediction pattern is 1:1 embedding rate in the best and worst case are 6.869 bpp, 0.786 bpp
and also adopt the image division pattern introduced in Hu and Xiang and the average embedding rate is 3.461 bpp for the BOSSBase. For
(2022) when the prediction pattern is 3:1. Table 3 shows the prediction the BOWS-2 dataset, the embedding rate of its best case, worst case,
results under different patterns. When the prediction pattern is 1:1, and average case are 6.013 bpp, 0.755 bpp, and 3.302 bpp respectively,
the mean, VAR, and MSE of our ACNNP are 4.107, 52.692, 78.821, while it is 4.879 bpp, 0.568 bpp, and 2.690 bpp for the UCID dataset.
which are lower than the predictor in Hu and Xiang (2021). When the In addition, the PSNR and SSIM of the recovered images are also used
prediction pattern is 3:1, the mean, VAR, and MSE of our predictor to measure the performance of our method. From the data in the table,

9
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Fig. 11. The ER(bpp) of different methods on six test images.

Table 4
The embedding capacity and payload of six test images.
Test images Total EC (bits) Label maps (bits) Auxiliary information (bits) Net EC (bits) Payload (bpp)
Lena 1 515 487 775 585 256 739 646 2.821
Baboon 1 103 414 788 828 256 314 330 1.199
Jetplane 1 600 951 766 105 256 834 590 3.183
Lakeboat 1 342 527 795 476 256 546 804 2.085
Peppers 1 434 249 780 040 256 653 953 2.494
Barbara 1 375 110 803 521 256 571 333 2.179

Table 5
Experimental results of ER, PSNR and SSIM on three datasets.
Datasets Indicators Best case Worst case Average
ER(bpp) 6.869 0.786 3.461
BOSSbase PSNR +∞ +∞ +∞
SSIM 1 1 1
ER(bpp) 6.013 0.755 3.302
BOWS-2 PSNR +∞ +∞ +∞
SSIM 1 1 1
ER(bpp) 4.879 0.568 2.690
UCID PSNR +∞ +∞ +∞
SSIM 1 1 1

it can be seen that the PSNR of the recovered image tends to infinity
and the values of SSIM are all 1, which indicates that our method is
universal and is able to restore the original image losslessly.
Also, to further prove the preponderance of our method, we com-
pare it with six methods (Chen & Chang, 2022; Fu et al., 2019; Wang
Fig. 12. The average ER(bpp) of different methods on three datasets.
& He, 2022; Wu et al., 2020; Yin et al., 2020; Yu et al., 2022) in terms
of average embedding rate on three datasets. The results are shown in
Fig. 12. As can be seen from the figure, the average embedding rate of
our method on the three datasets is 3.302 bpp, 3.461 bpp, and 2.690 recovered images, and Fig. 13(e) is the difference images between the
bpp which are the highest among all the methods. This further validates original images and the recovered images. The difference images are
the superiority of our method. all-black, we cannot obtain any information from them. The fact that
the PSNR of all recovered images in Table 5 is positive infinity and
3.3. Security analysis SSIM is 1 verifies the complete reversibility of our method.
Next, we analysed the security of our method from the view of
Before performing data embedding, we will encrypt the ‘Triangle’ statistical analysis, which includes histogram analysis, correlation anal-
and ‘Circle’ set image. Fig. 13(c) shows the marked images obtained ysis, and information entropy analysis. The histogram of an image can
after embedding their respective maximum capacity of information intuitively reflect the distribution of the pixel values of the image. For
for the six test images while the ‘Triangle’ and ‘Circle’ set image are the encrypted image, a uniform histogram means the image displays
encrypted using the same encrypted key. They are noise-like images, less statistical information, and the information leakage is less prob-
we cannot get any information from these images. Fig. 13(e) is the able. As Fig. 13(b) and (d) shown, the distribution of pixel values in

10
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

Table 7
The information entropy of six test images in different stages.
Images Original image Encrypted image Marked image
Lena 7.44506 7.99945 7.98744
Baboon 7.35833 7.99933 7.98003
Jetplane 6.70246 7.99927 7.98517
Lakeboat 7.48421 7.99923 7.98067
Peppers 7.59365 7.99922 7.98473
Barbara 7.63211 7.99925 7.98557

Table 7 shows the information entropy of the original image, the

encrypted image, and the marked image of the six test images. Accord-
ing to Table 7, the information entropy of the marked image is very
close to 8 and is larger than the original image, which indicates that
the security of our proposed method is guaranteed.

4. Conclusion

In this paper, a high-capacity RDHEI method is proposed. There are

two stages in the data embedding part. in the first stage, we design an
asymmetric CNN-based predictor by exploiting the multiple sense fields
Fig. 13. Experiment results: (a) the original images; (b) histograms of original and global optimization capabilities of neural networks. In the second
images; (c) the marked images; (d) histograms of the marked images; (e) recovered stage, we propose an adaptive mean predictor that can be used for
images; (f) the difference between (a) and (e). multi-MSB prediction in conjunction with the asymmetric CNN-based
predictor. After multi-MSB prediction, the steps of image encryption,
Table 6 label map generation, label map embedding and data hiding will be
The correlation coefficients of the six marked images in three directions. performed in turn for the two stages. Then, a marked image can be
Directions Lena Lakeboat Barbara Jetplane Baboon Peppers obtained and transmit to the receivers. When the receiver receives the
Horizontal 0.01179 0.00793 0.00364 0.00596 0.00301 0.01043 marked image, he can extract the data and recover the image separately
Vertical 0.00823 0.01212 0.00239 0.00962 0.00419 0.01017
according to the key he has. The experimental results and analysis
Diagonal 0.01262 0.01242 0.00699 0.01001 0.00919 0.01412
show that our method not only has high embedding capacity, but also
guarantees security.
In the future research, we will improve the performance of the
the original images is different and has an obvious statistical regularity algorithm from two aspects. On the one hand, we will continue to work
while the histograms of the marked images are significantly uniform. on the CNN-based predictor and try to design a predictor with higher
A closer look shows that these histograms still have local fluctuations, prediction accuracy that can work with the CNN-based predictor. On
but are overall homogeneous, which shows that our proposed method the other hand, we will try to improve our algorithm to increase the
can resist histogram analyses. robustness of it while increasing the capacity.
Due to the high correlation between adjacent pixels in an image, one
pixel often leaks information about its surrounding pixels, and attackers CRediT authorship contribution statement
may exploit this feature to obtain secret information. The correlation
between pixels can be quantified, and the specific calculation formula Ping Ping: Conceptualization, Methodology, Supervision, Writing
is as follows: – review & editing, Funding acquisition. Junyuan Huo: Methodol-
∑𝑁 1 ∑𝑁 1 ∑𝑁 ogy, Conceptualization, Investigation, Software, Validation, Writing –
𝑖=1 (𝑥𝑖 − 𝑁 1 𝑥𝑖 )(𝑦𝑖 − 𝑁 1 𝑦𝑖 ) original draft. Bobiao Guo: Validation, Reviewing and editing.
𝐶(𝑋, 𝑌 ) = √ (14)
∑𝑁 1 ∑𝑁 2×
∑ 𝑁 1 ∑𝑁 2
(𝑥
𝑖=1 𝑖 − 𝑁
𝑥
1 𝑖 ) (𝑦
𝑖=1 𝑖 − 𝑁
𝑦
1 𝑖 )
Declaration of competing interest
where 𝑋, 𝑌 here denote two sets of pixel values, 𝑁 is the number of
pixel values in the corresponding set. 𝑥𝑖 and 𝑦𝑖 denote the specific two The authors declare the following financial interests/personal rela-
pixel values. 𝐶(𝑋, 𝑌 ) is the calculated correlation coefficient between tionships which may be considered as potential competing interests:
𝑋 and 𝑌 . we analysed the adjacent pixel correlation of our proposed Ping Ping reports financial support was provided by National Natural
method in three directions: horizontal, vertical and diagonal. Table 6 Science Foundation of China. If there are other authors, they declare
records the correlation coefficients of the six marked images in three that they have no known competing financial interests or personal
directions. The correlation coefficients of all the images in three direc- relationships that could have appeared to influence the work reported
tions are close to zero, which shows that the attacker cannot get any in this paper.
information by using neighbouring pixels.
Information entropy can be used to measure the randomness of an Data availability
image, its ideal value is 8, the higher the value, it means that the image
pixel distribution the more uniform, then the image is more resistant Data will be made available on request.
to statistical attacks. Its specific formula is as follows:

∑
255 Acknowledgement
𝐻(𝐼) = − 𝑃 (𝑉𝑖 )𝑙𝑜𝑔2 (𝑃 (𝑉𝑖 )) (15)
𝑖=0 This work is supported by the National Natural Science Foundation
𝑃 (𝑉𝑖 ) is the probability of a pixel value 𝑉𝑖 and 𝑉𝑖 ∈ [0, 255]. of China (Grant No. 61902110)

11
P. Ping et al. Expert Systems With Applications 246 (2024) 123270

References Ni, B., & Bi, W. (2023). New predictor-based schemes for reversible data hiding.
Multimedia Tools and Applications, 82(4), 5923–5948.
Alattar, A. (2004). Reversible watermark using the difference expansion of a generalized Ou, B., Li, X., & Wang, J. (2016). High-fidelity reversible data hiding based on
integer transform. IEEE Transactions on Image Processing, 13(8), 1147–1156. pixel-value-ordering and pairwise prediction-error expansion. Journal of Visual
Bas, P., Filler, T., & Pevnỳ, T. (2011). ’’ Break our steganographic system’’: the ins Communication and Image Representation, 39, 12–23.
and outs of organizing BOSS. In International workshop on information hiding (pp. Ou, B., Li, X., Zhang, W., & Zhao, Y. (2019). Improving pairwise PEE via
59–70). Springer. hybrid-dimensional histogram generation and adaptive mapping selection. IEEE
Bas, P., & Furon, T. (2017). Image database of BOWS-2. Accessed: Jun, 20, 2016–2017. Transactions on Circuits and Systems for Video Technology, 29(7), 2176–2190.
Chen, K.-M. (2020). High capacity reversible data hiding based on the compression of Ou, B., Li, X., Zhao, Y., Ni, R., & Shi, Y.-Q. (2013). Pairwise prediction-error expansion
pixel differences. Mathematics, 8(9), Article 1435. for efficient reversible data hiding. IEEE Transactions on Image Processing, 22(12),
Chen, S., & Chang, C. C. (2022). Reversible data hiding in encrypted images using block- 5010–5021.
based adaptive MSBs prediction. Journal of Information Security and Applications, 69, Pareek, N. K., Patidar, V., & Sud, K. K. (2006). Image encryption using chaotic logistic
Article 103297. map. Image and Vision Computing, 24(9), 926–934.
Coltuc, D. (2011). Improved embedding for prediction-based reversible watermarking. Park, H. S., Hong, T., Lee, D.-E., Oh, B. K., & Glisic, B. (2022). Long-term structural
IEEE Transactions on Information Forensics and security, 6(3), 873–882. response prediction models for concrete structures using weather data, fiber-optic
Ding, X., Guo, Y., Ding, G., & Han, J. (2019). ACNet: Strengthening the kernel sensing, and convolutional neural network. Expert Systems with Applications, 201.
skeletons for powerful CNN via asymmetric convolution blocks. In 2019 IEEE/CVF Qian, Z., & Zhang, X. (2016). Reversible data hiding in encrypted images with
international conference on computer vision (pp. 1911–1920). distributed source encoding. IEEE Transactions on Circuits and Systems for Video
Dragoi, I. C., Caciula, I., & Coltuc, D. (2018). Improved pairwise pixel-value-ordering Technology, 26(4), 636–646.
for high-fidelity reversible data hiding. In 2018 IEEE international conference on Qin, J., He, Z., Xiang, X., & Tan, Y. (2022). Reversible data hiding in encrypted images
image processing (pp. 1668–1672). based on adaptive gradient prediction. Security and Commuication Networks, 2022.
Fan, G., Pan, Z., Zhou, Q., & Zhang, X. (2023). Flexible patch moving modes for Qin, C., & Zhang, X. (2015). Effective reversible data hiding in encrypted image with
pixel-value-ordering based reversible data hiding methods. Expert Systems with privacy protection for image content. Journal of Visual Communication and Image
Applications, 214, Article 119154. Representation, 31, 154–164.
Fu, Y., Kong, P., Yao, H., Tang, Z., & Qin, C. (2019). Effective reversible data hiding in Sachnev, V., Kim, H. J., Nam, J., Suresh, S., & Shi, Y. Q. (2009). Reversible
encrypted image with adaptive encoding strategy. Information Sciences, 494, 21–36. watermarking algorithm using sorting and prediction. IEEE Transactions on Circuits
Gao, X., Pan, Z., Gao, E., & Fan, G. (2020). Reversible data hiding for high dynamic and Systems for Video Technology, 19(7), 989–999.
range images using two-dimensional prediction-error histogram of the second time Schaefer, G., & Stich, M. (2003). UCID: An uncompressed color image database. In
prediction. Signal Processing, 173, Article 107579. Storage and retrieval methods and applications for multimedia 2004: Vol. 5307, (pp.
Hong, W., Chen, T.-S., & Wu, H.-Y. (2012). An improved reversible data hiding in 472–480). SPIE.
encrypted images using side match. IEEE Signal Processing Letters, 19(4), 199–202. Shi, Y.-Q., Li, X., Zhang, X., Wu, H.-T., & Ma, B. (2016). Reversible data hiding:
Honsinger, C. W., Jones, P. W., Rabbani, M., & Stoffel, J. C. (2001). Lossless recovery Advances in the past two decades. IEEE Access, 4, 3210–3237.
of an original image containing embedded data. US Patent 6, 278, 791. Tan, C., Wahidin, L., Khalil, S., Tamaldin, N., Hu, J., & Rauterberg, G. (2016). The
Hu, R., & Xiang, S. (2021). CNN prediction based reversible data hiding. IEEE Signal application of expert system: a review of research and applications. ARPN Journal
Processing Letters, 28, 464–468. of Engineering and Applied Sciences, 11(4), 2448–2453.
Hu, R., & Xiang, S. (2022). Reversible data hiding by using CNN prediction and adaptive Thodi, D. M., & Rodriguez, J. J. (2007). Expansion embedding techniques for reversible
embedding. IEEE Transactions on Pattern Aalysis and Machine Intelligence, 44(12), watermarking. IEEE Transactions on Image Processing, 16(3), 721–730.
10196–10208. Tian, J. (2003). Reversible data embedding using a difference expansion. IEEE
Huang, H.-C., & Chang, F.-C. (2013). Hierarchy-based reversible data hiding. Expert Transactions on Circuits and Systems for Video Technology, 13(8), 890–896.
Systems with Applications, 40(1), 34–43. Wang, X., Chang, C.-C., & Lin, C.-C. (2021). Reversible data hiding in encrypted images
Huang, F., Huang, J., & Shi, Y.-Q. (2016). New framework for reversible data hiding in with block-based adaptive MSB encoding. Information Sciences, 567, 375–394.
encrypted domain. IEEE Transactions on Information Forenscis and Security, 11(12), Wang, Y., & He, W. (2022). High capacity reversible data hiding in encrypted image
2777–2789. based on adaptive MSB prediction. IEEE Transactions on Multimedia, 24, 1288–1298.
Jia, Y., Yin, Z., Zhang, X., & Luo, Y. (2019). Reversible data hiding based on reducing Wang, Y., Xiong, G., & He, W. (2023). High-capacity reversible data hiding in encrypted
invalid shifting of pixels in histogram shifting. Signal Processing, 163, 238–246. images based on pixel-value-ordering and histogram shifting. Expert Systems with
Jiang, R., Zhou, H., Zhang, W., & Yu, N. (2018). Reversible data hiding in encrypted Applications, 211, Article 118600.
three-dimensional mesh models. IEEE Transactions on Multimedia, 20(1), 55–67. Weber, A. G. (2006). The USC-SIPI image database: Version 5. http://sipi.usc.edu/
Kim, H. J., Sachnev, V., Shi, Y. Q., Nam, J., & Choo, H.-G. (2008). A novel difference database/.
expansion transform for reversible data embedding. IEEE Transactions on Information Wu, X., & Sun, W. (2014). High-capacity reversible data hiding in encrypted images
Forenscis and Security, 3(3), 456–465. by prediction error. Signal Processing, 104, 387–400.
Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv Wu, Y., Xiang, Y., Guo, Y., Tang, J., & Yin, Z. (2020). An improved reversible data
preprint arXiv:1412.6980. hiding in encrypted images using parametric binary tree labeling. IEEE Transactions
Kouhi, A., & Sedaaghi, M. H. (2021). Prediction error distribution with dynamic on Multimedia, 22(8), 1929–1938.
asymmetry for reversible data hiding. Expert Systems with Applications, 184, Article Xiang, N. (2022). Individual online learning behavior analysis based on hadoop.
115475. Computational Intelligence and Neuroscience, 2022.
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied Xu, Z., Lv, Z., Li, J., Sun, H., & Sheng, Z. (2023). A novel perspective on travel
to document recognition. Proceedings of the IEEE, 86(11), 2278–2324. demand prediction considering natural environmental and socioeconomic factors.
Li, X., Li, J., Li, B., & Yang, B. (2013). High-fidelity reversible data hiding scheme based IEEE Intelligent Transportation Systems Magazine, 15(1), 136–159.
on pixel-value-ordering and prediction-error expansion. Signal Processing, 93(1), Yin, Z., Xiang, Y., & Zhang, X. (2020). Reversible data hiding in encrypted images based
198–205. on multi-MSB prediction and huffman coding. IEEE Transactions on Multimedia,
Li, X., Li, B., Yang, B., & Zeng, T. (2013). General framework to histogram- 22(4), 874–884.
shifting-based reversible data hiding. IEEE Transactions on Image Processing, 22(6), Yu, C., Zhang, X., Li, G., Zhan, S., & Tang, Z. (2022). Reversible data hiding
2181–2191. with adaptive difference recovery for encrypted images. Information Sciences, 584,
Liu, W., Meng, L., Wang, H., Sun, J., & Wan, W. (2023). Multi-dimensional constraints- 89–110.
based PPVO for high fidelity reversible data hiding. Expert Systems with Applications, Zhang, X. (2011). Reversible data hiding in encrypted image. IEEE Signal Processing
213(C), Article 119190. Letters, 18(4), 255–258.
Ma, K., Zhang, W., Zhao, X., Yu, N., & Li, F. (2013). Reversible data hiding in encrypted Zhang, W., Ma, K., & Yu, N. (2014). Reversibility improved data hiding in encrypted
images by reserving room before encryption. IEEE Transactions on Information images. Signal Processing, 94, 118–127.
Forenscis and Security, 8(3), 553–562. Zhang, W., Wang, H., Hou, D., & Yu, N. (2016). Reversible data hiding in encrypted
Melendez-Melendez, G., Morales-Reyes, A., & Cumplido, R. (2023). An adaptive pixel images by reversible image transformation. IEEE Transactions on Multimedia, 18(8),
value ordering based reversible data hiding scheme for images. Expert Systems with 1469–1479.
Applications, 232, Article 120809.

12