Computer Science Review 53 (2024) 100662

Contents lists available at ScienceDirect

Computer Science Review

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

Review Article

Digital image watermarking using deep learning: A survey

Khalid M. Hosny * , Amal Magdi, Osama ElKomy, Hanaa M． Hamza
Department of Information Technology, Zagazig University, Egypt

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

Keywords: Lately, a lot of attention has been paid to securing the ownership rights of digital images. The expanding usage of
Image watermarking the Internet causes several problems, including data piracy and data tampering. Image watermarking is a typical
Deep learning method of protecting an image’s copyright. Robust watermarking for digital images is a process of embedding
Neural networks
watermarks on the cover image and extracting them correctly under different attacks. The embedded watermark
might be either visible or invisible. Deep learning extracts image features using neural networks, which are
highly effective in feature extraction. Watermarking techniques that utilize deep learning have gained a lot of
interest due to their remarkable ability to extract features. This article offers an overview of digital image
watermarking and deep learning. This article will discuss several research articles on digital image watermarking
in deep-learning environments.

1. Introduction it is designed to withstand attempts to erase or destroy the mark [16].

Small modifications can easily affect the fragile watermark [17].
Digital image watermarking has received more attention lately Semi-fragile approaches protect concealed data from purposeful attacks
because of the rise in copyright disputes [1]. The process of hiding a while making it vulnerable to malicious attacks [18].
watermark in images to preserve its copyright is known as digital image Watermarking techniques can be blind, non-blind, or semi-blind
watermarking [2]. There are two watermarking processes: spatial based on the embedding and extraction process [19]. Blind water-
domain and frequency domain [3]. Spatial domain techniques include marking extracts a watermark without knowing the original information
patchwork [4], correlation, spread spectrum [5], and least significant bit [20]. Semi-blind watermarking does not need the original data to extract
(LSB) [6]. The cover image’s pixel values are directly altered during the the watermark; it only needs the secret key and the watermark [21]. In
watermarking embedding process in the spatial domain. In contrast, non-blind watermarking, the watermark must be extracted with
image pixels are converted to frequency coefficients in the frequency knowledge of the cover image and the secret key [22]. Fig. 1 summarizes
domain and then altered [7]. Techniques in the transform domain the image watermarking categorization based on human perception,
include singular value decomposition (SVD) [8], discrete cosine trans- transform domain, and procedure type. Traditional watermarking ap-
form (DCT) [9], discrete wavelet transform (DWT) [10], and others. proaches protect picture copyright by embedding an invisible but
Despite being computationally easier to implement, spatial domain detectable signal into the host image. Therefore, watermarked pictures
watermarking provides less protection against geometric assaults. generated using these methods are always distorted [23]. Zero water-
However, watermarking in the transform domain became more robust to marking adds no data to the cover picture itself. Instead, it generates an
attacks [11]. ownership authentication key using a watermark signal and the cover
Watermarks can be categorized as visible or invisible [12]. Visible picture’s selected features [24].
watermarks are clear and embossed on the cover media. However, their The fundamental steps in watermarking are illustrated in Fig. 2.
limited imperceptibility allows for visual identification [13]. Invisible Using a secret key, the secret data is embedded into the cover image
watermarking embeds hidden data into cover media that authorized during the embedding procedure. It is accomplished by applying spatial
users can only identify through specific logical processes. However, it or transform domain approaches [25]. Upon receiving the watermarked
may cause an overhead in cover media [14]. The invisible watermarking image, the receiver proceeds to the extraction step, which is the last
methods are categorized into robust, fragile, and semi-fragile [15]. phase of the process. The extracted watermark is recovered by following
Robust watermarking protects copyright and identifies ownership since the extraction procedure [26].

* Corresponding author.
E-mail address: [email protected] (K.M. Hosny).

https://doi.org/10.1016/j.cosrev.2024.100662
Received 14 May 2024; Received in revised form 20 July 2024; Accepted 27 July 2024
Available online 7 August 2024
1574-0137/© 2024 Elsevier Inc. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

Important characteristics of the watermarking technique are without starting from scratch, for instance, Keras, Torch, and Tensor
measured using a variety of performance metrics, such as bit error rate Flow [37].
(BER), peak signal-to-noise ratio (PSNR), and structural similarity index This paper provides a broad review of deep learning-based digital
(SSIM) [27]. The watermarking approach has several noteworthy uses, image watermarking and outlines some recent research publications.
as seen in Fig. 3. These applications include military, e-voting, copyright The remaining sections of this work are arranged as follows: An over-
protection, tamper detection, etc. The primary objectives of the water- view of the most popular deep learning models and their recent research
marking technique are robustness, security, and imperceptibility. publications is provided in Section 2. A discussion is provided in Section
The robustness of traditional watermarking algorithms is limited, 3. Future directions in deep learning-based image watermarking tech-
and it is frequently challenging to identify the best embedding areas, as niques are provided in Section 4. Section 5 concludes with some
these algorithms adopt the manual method of embedding the watermark findings.
[28]. Neural networks, which are excellent at extracting features, have
been included in digital watermarking to enhance their robustness 2. Popular deep learning models based image watermarking
compared to most conventional techniques [29].
Watermarking with deep learning has recently gained significant 2.1. Generative adversarial network (GAN)
interest due to its numerous widely used applications and significant
impact on the security of image content [30]. Deep learning has several GAN is a sort of machine learning framework. [38]. Goodfellow et al.
benefits for watermarking images since the best embedding positions in proposed the generative model known as GAN [39]. In image process-
the cover image may be found, and the ideal embedding strength can be ing, the deep learning-related GAN has shown impressive results. An
determined while maintaining the robustness and imperceptibility of the artificial intelligence method, generative adversarial networks, was
image. Deep learning can decrease noise in the final watermark and created to address the generative modeling issue [40]. A generative
efficiently simulate attacks to extract watermarks [31]. model looks at a set of training samples and attempts to determine the
Deep learning is a branch of artificial intelligence that uses neural probability distribution that produced each [41]. A generator and a
networks to simulate how the human brain processes information and discriminator are two sub-models found in a GAN that interact and give
makes decisions [32]. Deep learning extracts image features using feedback to one another to create more realistic outputs [42]. The
neural networks and adaptively embeds the watermarking to provide generator is trained to produce samples that resemble the training data
superior outcomes [33]. Deep learning-based watermarking technology set. Inputs are categorized as authentic (from the original training data)
research is still in its early phases. Even if there aren’t many studies, they or fraudulent by the discriminator. The discriminator seeks to detect
appear to have become more numerous recently [34]. Numerous deep fakes, while the generator attempts to replicate reality, as depicted in
learning models exist, such as generative adversarial networks (GAN), Fig. 4 [43].
deep neural networks (DNN), and convolutional neural networks (CNN) There are several variations of the GAN model, including conditional
[35]. Based on the utilized network architecture, deep learning can GAN (CGAN), deep convolutional GAN (DCGAN), laplacian GAN (Lap-
consist of convolutions, fully connected layers, pooling layers, gates, etc. GAN), information maximizing GAN (InfoGAN), etc. [44]. GAN is
[36]. With the help of deep learning platforms, users can easily frequently employed in watermarking for both watermark generation
construct, train, and implement strong neural networks for various uses and verification. Some of the research publications about image

Fig. 1. Digital image watermarking classification.

2
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

watermarking-based GAN are illustrated below, and the summary of the max-pooling, dense, and other special layers, make up the DNN hidden
GAN model in these research papers is illustrated in Table 1. Singh et al. layer structure [51]. With these extra layers, the model can better
[45] propose an image watermarking application for healthcare sce- comprehend issues and offer the best solutions for intricate tasks [52].
narios using the GAN model. The medical image is first subjected to Some of the research publications about image watermarking-based
randomized singular value decomposition (RSVD)–based encryption DNN are illustrated below, and the summary of the DNN model in these
and a 3D chaotic map before being watermarked. GAN is used to create a research papers is illustrated in Table 2. Ge et al. [53] proposed a
final mark by embedding the patient’s identifying number and the hash DNN-based document image watermarking method. The watermark is
of the medical picture into the hospital logo image. RSVD-based intended to be embedded and extracted using an encoder and a decoder.
embedding and redundant discrete wavelet transform (RDWT) are A noise layer has been created to simulate the many types of assaults that
used to watermark the encrypted medical image. The suggested might occur. A text-sensitive loss function is applied to limit embedding
approach is undetectable and effectively withstands various assaults, changes to characters. Although the PSNR and SSIM appear to be
according to experimental data. adequate, the watermarked image’s visual impact is unsatisfactory,
Annadurai et al. [46] suggest a convolutional generative adversarial likely because the document image’s backdrop is visible. In light of this,
neural network method for watermarking digital images. DWT quanti- they suggested an embedding strength modification technique to
zation approach based on SVD is used to watermark the image. The improve image quality while maintaining extraction accuracy.
proposed technique uses convolutional generative adversarial neural Jaiswal et al. [54] proposed a DNN method for blind color image
networks to segment and classify watermarked data. The test results watermarking. The input image is transformed using integer wavelet
show how resilient the recommended watermarking method is against transform (IWT), and feature reduction is achieved by principal
re-encoding attacks and how strong it is against image handling attacks. component analysis (PCA). Watermark images are encoded in the blue
Regarding training accuracy, the analysis yielded 98%, validation ac- color channel. Watermark extraction is done using a binary classifica-
curacy of 95%, normalized correlation (NC) of 71%, and SSIM of 56%. tion technique. The DNN is trained on a 512 × 11 training set and a 512
Pang et al. [47] suggested a novel GAN deep learning model-based × 10 testing set. Sinhal et al. [55] suggest a blind and robust medical
blind watermarking open-source dataset protection technique. The image watermarking technology based on DNN. The LZW (Lamp-
blind watermarking technique applies invisible watermarks to some el-Ziv-Welch) compression method is used during watermarking to
images in the data collection. This approach suggests using UNet in maintain the original picture’s region of interest (ROI) data. IWT is used
addition to GAN. When required, the watermark is extracted to fulfill the for embedding the watermark. The hash keys for the ROI and RONI
verification function when there are suspicions that a dubious model (region of non-interest) areas are produced using the SHA-256 tech-
was improperly trained. nique. Robust watermark extraction is done using a deep neural
Huang et al. [48] provide a robust, attention-guided image water- network-based architecture.
marking model based on GAN. Employ the FFM to acquire multi-layered Battarusetty et al. [56] provide a novel method for utilizing DNN
and deep image features for watermark fusion to avoid the disadvantage with a wavelet-based technique to create watermarked images with
of extracting specific image features. An attention module (AM) reduces increased invisibility. The locations where the watermark should be
image distortion brought on by watermark embedding. The discrimi- embedded are selected using the modified DNN. The Squirrel Search–-
nator is utilized to recognize the encoded picture from the original Grey Wolf Optimisation (SS–GWO) optimizes the DNN training method.
image to increase the visibility of watermarks. Adaptive DWT is used for watermark embedding. Wang et al. [57]
provide a reversible image-concealing network-based DNN model
watermarking system to safeguard DNN models’ copyrights. The sug-
2.2. Deep neural network (DNN) gested solution successfully protects DNN model copyrights while
maintaining the key functionality of the host models. The test outcomes
DNN is a neural network with specific intricacy [49]. It is comparable show that the proposed framework can provide accuracy and efficacy.
to stacked neural networks, which have many layers—typically two or The model is applied to five datasets and 11 prominent DNN models.
more—input, output, and at least one hidden layer situated in between,
as depicted in Fig. 5 [50]. Various layers, including convolutional,

Fig. 2. Digital image watermarking procedures.

3
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

Fig. 3. Digital Image Watermarking Applications.

Fig. 4. Illustration of how GAN works.

Table 1
Summary of existing GAN model research papers.
Reference Ref. [45] Ref. [46] Ref. [47] Ref. [48]
number

GAN Role Secure watermark Segmentation and classification enhanced feature sharing UNet with GAN is The encoder-noise-decoder structure is
generation of watermarked data designed used for end-to-end training.
as the extraction network
Objective prevent data leakage in assure the robustness and overcome the drawback of extracting limited protection algorithm suitable for
healthcare scenarios security of watermarking image features multiple paired open-sourced datasets
Working Chaotic map, RSVD SVD-DWT, feature fusion module, attention module UNet with GAN
Strategies
Domain Frequency domain Frequency domain Trained DL network Trained DL network
Robustness NC=0.9999, and BER = 0 BER= 71%, and NCC=71%, bit accuracy (BA) is used to validate BER=1.748%,
Evaluation watermarking robustness under five types of
noises
Invisibility PSNR=48.56 dB, SSIM=56%, PSNR=41.48, SSIM=0.99000 PSNR=28.537, SSIM= 0.856
SSIM=0.9997

4
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

Fig. 5. DNN architecture.

Table 2
Summary of existing DNN model research papers.
Reference Ref. [53] Ref. [54] Ref. [55] Ref. [56] Ref. [57]
number

DNN Role Embed and extract the Watermark extraction Watermark extraction Select the regions for Model watermarking and
watermark. embedding the verification
watermark
Objective Watermarking scheme based on Providing a better trade-off Provide effective security Offer an improved protect the copyrights of DNN
document image between imperceptibility and solutions for medical watermarking approach models
robustness images.
Working Encoder-decoder-noise layer IWT, PCA LZW, IWT, SHA-256 SS-GWO, ADWT DNN model
Strategies
Domain Trained DL network Frequency domain Frequency domain Frequency domain Trained DL network
Robustness bit accuracy is used to validate Average NC= 0.9953, BER=0, NC=1 NC=0.93709 ____
Evaluation watermarking robustness under
six distortions
Invisibility PSNR=41.07 dB, SSIM=0.965 Average PSNR= 35.57 dB, PSNR=40.27, PSNR=19.804 PSNR reaches a maximum of
SSIM=0.9969 SSIM=0.9445 43.10, and SSIM reaches a
maximum of 0.9932.

Fig. 6. CNN architecture.

5
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

2.3. Convolution neural network (CNN) combined into a featured picture, and the owner’s watermark sequence
was added with an XOR operation. 2D-LACM ensures security by
One kind of deep learning neural network architecture is CNN. CNN encrypting the watermark and scrambling the feature matrix. Experi-
handles grid-like data structures found in images [58]. CNN, particularly mental findings indicate that the suggested approach is both invisible
well-suited for image analysis, has been used in many contexts as object and resilient. Hsu et al. [72] present a high-capacity QR
recognition and image classification [59]. Three layers are commonly decomposition-based blind watermarking technique for color photos
seen in a CNN, as depicted in Fig. 6: the convolutional layer, the pooling using AI technology. The technique employs particle swarm optimiza-
layer, and the fully connected layer [60]. tion (PSO) and super-resolution convolutional neural networks
The convolution layer is the CNN’s most essential component. It (SRCNN). PSO optimizes parameters for imperceptibility and resilience.
carries the majority of computational strain on the network [61]. The SRCNN refines derived watermark images for better identification. The
Pooling layer decreases the spatial size of the Convolved Feature. This suggested approach is very resilient against many assaults, except JPEG
process reduces the computing resources needed for dimensionality compression.
reduction in the data processing [62]. Pooling comes in two varieties: Mahto et al. [73] create a durable and imperceptible watermarking
max pooling and average pooling. The largest value from the area of the technique using hybrid optimization. The technique uses spatial and
picture that the Kernel covers is returned by max pooling [63]. transform domains to embed hidden markings into cover material. The
Conversely, Average Pooling yields the mean of all the values from the multi-type mark adds legitimacy to the suggested algorithm by using a
area of the picture that the Kernel covers [64]. The representation be- proper scaling factor. The watermarked image is encrypted with an
tween the input and the output is mapped with the aid of the fully enhanced encryption method. A denoising CNN is used to increase the
connected layer [65]. algorithm’s resilience. Barlaskar et al. [74] use a DWT-DCT framework
Different CNN architectures are available, such as [66]: to develop a blind watermarking system with an optimum gain factor for
embedding a binary watermark in any host picture. DCNN extracts
• LeNet features that aid the support vector regression (SVR) in estimating
• AlexNet geometric assaults. A statistical non-geometric correction approach re-
• VGGNet trieves the watermark under different non-geometric attack scenarios.
• GoogLeNet Zhang et al. [75] introduce CNN and DCT to provide a blind resilient
• ResNet image watermarking system. A channel weighting module using a
• ZFNet modified Gaussian distribution is developed to adjust watermarking
information distribution and intensity. A spatial weighting module to
Several research publications about image watermarking-based CNN enhance the visual quality of watermarking is built. A frequency domain
are illustrated below, and the summary of the DNN model in these channel-attention enhancement module improves watermarking resil-
research papers is illustrated in Table 3. ience by sensing and enhancing channel signals. Ingaleshwar et al. [76]
Liu et al. [67] suggest a watermarking method that uses CNN to present a WCFOA-based Deep CNN technique for embedding hidden
create a zero watermark image by superimposing the host image’s color messages into cover images. The proposed WCFOA combines the Water
style over the watermark logo’s information. A styled picture is created Wave Optimization and Chaotic Fruit Fly Optimization techniques. After
by iterating between the host image’s style and the content of a water- extracting features, the Deep CNN classifier trained by the proposed
mark logo with a timestamp. The styled picture is encrypted using WCFOA determines the best place to implant the secret messages. The
Arnold transform. Convolutional neural networks are used to verify the wavelet transform is used for embedding.
copyright of images. Their loss function differs from the original Rai et al. [77] proposed an enhanced chimp optimization (ECO)
watermark logo and the picture produced. Experiments indicate that the technique that uses deep fusion CNN to provide robust watermarking.
suggested zero-watermarking approach is extremely resistant to stan- This approach uses an embedding and extraction network for water-
dard image processing and geometric assaults. marks. The embedding network’s octave convolutional model reduces
Chacko et al. [68] suggest a novel method of embedding technology spatial redundancy while capturing several features. The extraction
that depends on several embedding parameters to watermark images. network’s pyramid feature extraction module extracts local features,
Robust watermarking is achieved by defining a novel approach that uses while dilated convolutions reduce model parameters. Tavakoli et al.
superpixel, DCT, and Harris Hawks Optimisation (HHO). The water- [78] suggested using CNN and wavelet transformations to create a
marking bits have been merged with a deep learning convolution neural watermarking network for embedding and extracting the watermarks.
network (DLCNN) using integration standards. A DLCNN-based pattern The network can take all types of watermarks, is agnostic of host picture
recognition model is used to find and extract watermarks, which are resolution, and has just 11 layers due to wavelet transform. Performance
then recognized using the HHO method. The suggested technique out- is evaluated using the extracted watermark’s similarity to the original
performs previous PSNR, BER, and NCC methods. and the host image’s similarity to the watermarked one.
Wu et al. [69] present a recovery watermarking approach utilizing Guo et al. [79] proposed a blind watermarking system based on deep
convolutional networks and DWT that is fragile. The watermark is based learning. The suggested method utilized a modified CNN to extract
on a downsampled image of the source image. The watermark is inserted watermarks with a block classification approach. The modified CNN
in DWT coefficients. The watermark is utilized for the detection and creates a residual module with fewer batch normalizations, resulting in
recovery. A CNN is proposed for recovering tampered images. The faster watermarking. The suggested approach is effective against typical
experimental findings suggest that the proposed technique detects signal processing assaults, as demonstrated by experimental findings.
tampering more effectively. Geng et al. [80] study the confluence of deep learning with blockchain
Singh et al. [70] Utilise CNN to develop a novel watermarking technologies in copyright protection. A digital picture watermarking
method for digital images. The encoder network extracts latent char- model is designed, and SVD and deep learning approaches are used to
acteristics from cover and secret images, which are then concatenated to create a flexible watermarking system. The blockchain model’s perfor-
create a watermarked image. The recipient uses a denoising autoencoder mance parameters use a practical byzantine fault tolerance consensus
network to eliminate noise from the picture, followed by a CNN to method for consistent and reliable transactions. The algorithm’s per-
extract the secret mark. Darwish et al. [71] suggest a method for formance was examined using normalized cross-correlation (NC), and
zero-watermarking color images with CNN and 2D logistic-adjusted attack tests were conducted on original photos to evaluate its resilience
Chebyshev map (2D-LACM). Deep feature maps are extracted using and practicality.
the pre-trained VGG19 from the original image. The feature maps were Bai et al. [81] present an infrared images steganography method

6
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

Table 3
Summary of existing CNN model research papers.
Reference CNN Role Objective Working Strategies Domain Robustness Evaluation Invisibility
number

Ref. [67] Watermark extraction to verify construct the zero- Arnold transform Spatial correlation coefficient Correlation coefficients
the image copyright watermark image by under various attacks is for ten watermark logos
addition rather than XOR computed are computed
operation
Ref. [68] Identify and extract the Efficient positioning and DCT, Arnold Frequency NC and BER under PSNR= 46 dB
watermark. parameter optimization of transform, HHO domain different attacks are
image watermarking calculated
Ref. [69] Recovery of the tampered image Image authentication and DWT Frequency ____ PSNR around 41.87 dB
tamper recovery. domain and SSIM around 0.977
Ref. [70] Watermark embedding and Train DL network to reduce Encoder-decoder Trained Dl NC=0.9996 PSNR=44.48 dB,
extraction human intervention network network SSIM=0.9997
Ref. [71] Deep feature extraction Improved zero- 2D-LACM Trained DL BER below 0.0044, NC Average PSNR=33.1537
watermarking resistance network above 0.9929
Ref. [72] Refine the extracted watermark solves the shortcomings of PSO, SRCNN Frequency BER=0.107, NCC=0.89 PSNR=38.02,
image. QRD-based watermarking domain MSSIM=0.915
technology
Ref. [73] The denoising process of the Enhanced watermarking LWT-Schur-T-SVD, Hybrid NC and BER are The average PSNR is
recovered watermark image using hybrid optimization HPSOF, SIE domain approaching 1 and 0, 57.7124 dB,
scrambling respectively,
Ref. [74] Watermarked image deep A genetic algorithm has been DWT-DCT, SVR, Frequency NC~0.9784 The average PSNR is
feature extraction. used for optimizing domain 47.81 dB
watermark embedding
Ref. [75] Watermark embedding and control the distribution and 2D-DCT Hybrid BER is computed under PSNR and SSIM are
extraction intensity of watermarking domain some real-world attacks computed under some
information real-world attacks
Ref. [76] Feature extraction and optimal Robust optimized WCFOA, DWT Frequency NC=1 PSNR=45.2157
region selection watermarking method domain
Ref. [77] Feature extraction Apply optimization ECO, OCM, FFN Hybrid NC=0.98, BER=0.038 PSNR=54.64,
algorithm based on CNN for domain SSIM=0.989
better watermarking
Ref. [78] Watermark embedding and Improve watermarking DWT, Haar wavelet Frequency BER= 0.0003 PSNR=40.1
extraction accuracy and performance domain
Ref. [79] Watermark embedding and Improve watermarking Stochastic Gradient Spatial NC and BER values are PSNR and SSIM values are
extraction performance Descent(SGD) domain computed under some computed under some
attacks attacks
Ref. [80] Feature extraction fusion of DL and blockchain SVD, blockchain Frequency NC values are computed ___
technology domain under some attacks
Ref. [81] Feature extraction- Image pre- Steganography infrared SWT, SRCNN, Hybrid ___ Average PSNR=59.27 dB
processing (SRCNN)- improve images CCNP Domain for 10,000 bits
the accuracy of the pixel
prediction (CNNP)
Ref. [82] Feature extraction Propose a multi- DCT, logistic map, Frequency NC under different PSNR under different
watermarking technique and hash function domain attacks is calculated attacks is calculated
Ref. [83] Feature extraction
Ref. [84] Geometry attack classification balancing the MRFO, DTCWT, Frequency Average BER=0.067, Average PSNR=54.721
and correction procedure imperceptibility and SVD domain NCC=0.841 without
robustness by optimization geometric correction
Ref. [85] Feature extraction robust medical chaotic Henon Trained DL NC under different PSNR under different
watermarking algorithm mapping network attacks is calculated attacks is calculated
Ref. [86] Watermark embedding and watermarking scheme for Embedding Trained DL NC=0.9685, PSNR=38.8790,
extraction high-definition images -revealing network network BER=0.0343 SSIM=0.9581
Ref. [87] Selecting the optimal embedding innovative image HWT, CBEO Frequency Max. NC=0.761, least Max. PSNR=24.989 dB,
regions watermarking approach domain BER=0.047 Max. SSIM=0.969
Ref. [88] Feature extraction Overcome geometric attacks DWT, DCT, chaotic Frequency PSNR is computed PSNR is computed under
scrambling domain under various attacks. various attacks
Most NC larger than
0.50
Ref. [89] Image denoising and super- Directly embeds color pixel DCT Frequency Average ZNCCs=1 PSNR=37.253,
resolution reconstruction values into the designated domain under no attack SSIM=0.961
DCT coefficients.
Ref. [90] Feature extraction- zero- replace the traditional XOR The generator Trained DL NC is computed under PSNR is computed under
watermark verification operation with zero network and the network various attacks various attacks
watermarking detector network
Ref. [91] Feature extraction Improve medical DCT Frequency NC is computed under PSNR is computed under
watermarking robustness domain various attacks various attacks
Ref. [92] Feature extraction Improve medical DCT, chaos map, Frequency NC is computed under PSNR is computed under
watermarking robustness Arnold transform domain various attacks various attacks
Ref. [93] Processing the watermark of Improve medical DWT, DCT, logistic Frequency NC is computed under ___
images watermarking robustness map domain various attacks
Ref. [94] Feature extraction Improve better robustness DCT, Tent mapping Frequency NC is computed under PSNR is computed under
domain various attacks various attacks
Ref. [95] Feature extraction authenticating and securing NSST, SVD, SSFC Frequency NC is computed under PSNR is computed under
healthcare records domain various attacks various attacks

7
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

using smooth wavelet transform (SWT) and CNN. Half of the input function.
infrared picture is pre-processed using SRCNN and SWT algorithms. The Gong et al. [91] suggested a zero watermarking approach using
remaining half of the infrared picture is then predicted using the con- Residual-DenseNet. This approach uses CNN to enhance the
volutional neural network predictor (CNNP). The suggested CNNP zero-watermarking approach by training Residual-DenseNet to extract
model incorporates an attention strategy to enhance prediction accu- resilient feature vectors. The chaotic matrix is generated using the lo-
racy. This concept does not need costly devices or extensive storage gistic map technique for watermark encryption and decryption. This
space for training. Fan et al. [82] proposed an image watermarking approach performs well against conventional and geometric assaults, as
technique that combines CNN Inception V3 and DCT for image pro- demonstrated by experimental findings. Dong et al. [92] proposed a zero
cessing and feature extraction. This approach uses a logistic map and watermarking technique that utilizes an improved NasNet-Mobile CNN
hash functions to scramble the watermarks. Three watermark forms with DCT. The watermarked data is twice encrypted using the Chen
were used in this method to test the algorithm’s viability and chaotic system and Arnold mapping technique, improving its security.
generalizability. Migration learning is applied to the upgraded NasNet-Mobile network.
Huang et al. [83] suggest a lossless photo watermarking technique The retrieved image features are transformed using DCT to recover
using depthwise overparameterized VGG (DO-VGG) and an attention low-frequency data, which is then processed using the perceptual hash
mechanism. The pre-trained DO-VGG model extracts deep, technique. Finally, medical images are encrypted using
high-dimensional features from medical pictures. The suggested tech- zero-watermarking technology.
nique uses deep neural networks to extract features, making it resistant Nawaz et al. [93] present a zero watermarking approach utilizing
to geometric assaults. Deep-level image features are extracted using DWT, an upgraded MobileNetV2 convolutional neural network, and
CBAM in both channels and spatial dimensions. The suggested approach DCT. An enhanced MobileNetV2 pre-trained network is utilized, fol-
uses enhanced logistic chaotic mapping and encryption to encrypt the lowed by transfer learning to extract features from encrypted medical
watermarking image. Sharma et al. [84] propose an efficient water- images. A chaotic method is used for encryption, and the generated key
marking technique with great imperceptibility and resilience called is maintained in a remote location to enhance the watermark’s security.
MantaRayWmark. The system uses Manta-Ray Foraging Optimization to Li et al. [94] suggested a zero watermarking approach using DCT and an
improve the locally relevant multiple embedding strengths, achieving a enhanced DarkNet53 CNN. DCT is utilized to extract features from
compromise between the imperceptibility and resilience of images. A encrypted medical images. DarkNet53 was selected for migration
geometric correction method based on CNN technology is proposed to learning for feature extraction from the encrypted images. Encryption is
strengthen the watermarking technique against geometric assaults. done using Tent Map and Logistic Map.
Zhang et al. [85] present a reliable multi-watermarking solution using Anand et al. [95] propose a deep learning-based zero-watermarking
GoogLeNet and Henon Map. The GoogLeNet network is trained on a approach to safeguarding healthcare records. The watermark image is
medical training sample set and then used to extract feature vectors from scrambled to increase security using the step space-filling curve
images. Three forms of watermark are inserted and extracted to evaluate approach. Alexnet extracts features from host images that are visibly
the algorithm stability. The feature vectors are combined with the marked. Zero watermarking is implemented using non-subsampled
encrypted watermark to complete the zero watermark embedding and shearlet transform (NSST) and SVD.
blind extraction.
Zhu et al. [86] present a blind image watermarking system using 2.4. Other deep learning models-based image watermarking
multiscale fusion dilated ResNet (MDResNet) and key-point detection.
The watermarking system embeds the watermark in multiple Rai et al. [25] offered a hybrid approach for picture watermarking
non-overlapping areas, each protected by a private key that refers to that combines fuzzy back propagation neural network (BPNN) with
different dominating key points. The watermarks are placed in the shark smell optimization. Fuzzy-BPNN and the Human Visual System
middle section of the region in such a way that they remain inside the (HVS) assist in identifying places with noise that the human eye may
regions even when the picture is geometrically changed. MDResNet miss. The SSO algorithm determines the ideal embedding positions. The
embeds and extracts the watermarks. Suresh et al. [87] proposed a deep analysis shows that the suggested strategy beats other previous strate-
learning-based chronological bald eagle optimization (CBEO) technique gies. Liu et al. [96] introduced a new SVD-based image watermarking
for image watermarking. The host image embeds the watermark by method. The suggested approach improves image quality by embedding
identifying an ideal place using LeNet. The embedding process utilizes the watermark bit in the first singular value of each color block. Blind
the Haar Wavelet Transform (HWT) to enhance resilience. The CBEO watermark extraction uses a generalized regression neural network
method is used to optimize the trainable parameters of the LeNet, which (GRNN) to avoid keeping the original image information and determine
selects the ideal region in the cover image. Nawaz et al. [88] introduced the singular value.
a zero-watermarking system using DWT and ResNet-DCT that can
withstand geometric assaults. Deep features are transformed using DWT 3. Discussion and statistics
and DCT, and a perceptual hash function is used to construct the feature
vector. ResNet is used to extract deep features in medical images. After Watermarking using deep learning has lately acquired popularity
obtaining and decrypting the encrypted watermark, the recovered due to its wide range of applications and considerable influence on
watermark can used to identify the image ownership. image content security. Deep learning can select the optimum embed-
Hu et al. [89] introduced a new approach for blind picture water- ding places in the cover image, reduce noise in the final watermark, and
marking based on DCT. The watermarking approach utilizes a deep simulate assaults to extract watermarks. Deep learning is now used in
learning network to create high-quality, high-resolution watermarks for smart city applications. Smart city researchers have used deep learning
visual assessment. ResNet is used for image denoising and in various applications, including traffic prediction, healthcare, air
super-resolution reconstruction. Xiang et al. [90] introduced a quality prediction, and public safety. Most watermarking systems face
zero-watermark for medical images using style features and residual the challenge of balancing embedding capacity, imperceptibility, and
networks. Deep neural networks obtain high-order statistics from image robustness. Combining encryption with watermarking enhances water-
details and abstract features, making them resistant to geometric as- mark security. However, it also increases system complexity.
saults. This paper suggests using a residual network for zero-watermark The various works on using deep learning in image watermarking
verification instead of the standard XOR procedure. The residual demonstrated that using deep learning will benefit the digital image
network iteratively learns how to extract the watermark from the zero watermarking sector. Deep learning models include recurrent neural
watermark images and improve its quality with the defined loss networks (RNN), BPNN, GAN, DNN, and CNN. This survey contains

8
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

research articles on image watermarking with deep learning. The study any other technology, deep learning offers benefits and drawbacks, as
results revealed that three deep learning models are extensively utilized depicted in Fig. 9. We will discuss the benefits and drawbacks of deep
in image watermarking: GAN, DNN, and CNN. This survey studies the learning.
three deep learning models extensively employed in digital picture Advantages of deep learning-based image watermarking:
watermarking. As depicted in Fig. 7, CNN is the deep learning model
most commonly used in image watermarking. 1. Automatic feature extraction eliminates the need for manual
Many research papers have recently been on deep learning-based intervention.
image watermarking; this indicates a progressive growth in re- 2. Deep learning dynamically extracts high- and low-level features from
searchers’ interest in this topic. Deep learning is utilized in image large datasets.
watermarking for various tasks such as watermark embedding and 3. Deep learning algorithms can process structured and unstructured
extraction, feature selection, and selecting the optimal embedding lo- data like images, audio, and texts.
cations. The study on deep learning-based watermarking technology is 4. The same neural network-based method may be used for various
still in its early stages. Even if there isn’t much research, they have applications and datasets.
increased recently. As depicted in Fig. 8, the number of research papers 5. Handling massive and complex datasets
utilizing CNN in image watermarking increases. 6. GPUs can do massively parallel computations, which are scalable for
Our survey provided many innovative research publications for deep big amounts of data.
learning-based image watermarking. Neural networks are used to 7. Deep learning models are extremely adaptive. The models can be
embed and remove watermark information to maintain invisibility and fine-tuned or tailored to new tasks.
increase robustness against attacks. The most problematic aspect of 8. Deep learning can find non-linear relationships in data.
watermarking models is the trade-off between invisibility and robust- 9. The deep learning architecture is adaptable, allowing it to be used to
ness. A speedier watermarking model with fewer processing resources is solve future issues.
preferred for efficiency, but it must also be resistant to many types of
attacks. CNN is the most often used neural network for image water- While deep learning has surpassed traditional machine learning ap-
marking due to its superior performance. Although CNN models might proaches, it is unsuitable for all applications. Researchers must consider
suffer from several limitations as follows: a variety of restrictions:
Disadvantages of deep learning-based image watermarking:
■ The train and process of a CNN model required a larger dataset.
■ Misconfigured network settings might lead to longer training 1. Deep learning requires a vast amount of data to perform better than
periods. other techniques, which requires a lot of time and resources for data
■ Operations such as max-pool enhance latency in CNN models. collection.
■ Overfitting and underfitting are common issues with CNN networks 2. Training a particular model with large datasets requires more pro-
because of their complexity. cessing resources for extensive computing requirements than other
machine learning models.
Deep learning has emerged as one of artificial intelligence research’s 3. Deep learning models might be challenging to understand. The deep
most popular and fascinating areas. It is a subfield of machine learning learning applications act as a black box that converts inputs to out-
that analyses huge datasets using neural networks. In recent years, deep puts, making it unsuitable for people who want to understand
learning has received a lot of interest in the image processing sector. Like

Fig. 7. Most used deep-learning models for image watermarking.

9
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

Fig. 8. Number of research papers per year in deep-learning-based image watermarking.

Fig. 9. Some deep learning advantages and disadvantages in image watermarking.

application decision-making procedures. As a result, it isn’t easy for 4. Future directions

less trained individuals to implement it.
4. Pre-trained models are commonly employed to reduce training Image watermarking-based deep learning is a new and emerging
complexity, and this can cause model overwriting. research field with numerous opportunities for advancement. The recent
5. The robustness of watermarking is dependent on the number of proliferation of research articles on deep learning-based picture water-
samples used during training and the loss function utilized. marking shows that interest in this field has steadily increased. Here, we
outline our recommendations for potential future research areas:
Therefore, further study is required to fully cover the limitations of
deep learning for image watermarking applications.

10
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

• The deep watermarking models addressed in this review are resilient [11] S. Sharma, J.J. Zou, G. Fang, P. Shukla, W. Cai, A review of image watermarking
for identity protection and verification, Multimed. Tools Appl. (2023) 1–63.
and effective for image watermarking and may be applied to various
[12] Z. Zainol, J.S. Teh, M. Alawida, A. Alabdulatif, Hybrid SVD-based image
media types because this is a new research topic. watermarking schemes: a review, IEEE Access 9 (2021) 32931–32968.
• Safeguarding machine learning models as intellectual property is [13] A.A. Embaby, M.A.W. Shalaby, K.M. Elsayed, Digital watermarking properties,
also a significant research topic for digital watermarking. classification, and techniques, Int. J. Eng. Adv. Technol. (IJEAT) 9 (3) (2020)
2742–2750.
• Compared to spatial domain-based methods, frequency domain- [14] R. Wazirali, R. Ahmad, A. Al-Amayreh, M. Al-Madi, A. Khalifeh, Secure
based watermarking techniques demonstrate higher durability. We watermarking schemes and their approaches in the IoT technology: an overview,
can incorporate multiple frequency domains into the watermarking Electronics 10 (14) (2021) 1744.
[15] B. Bolourian Haghighi, A.H. Taherinia, R. Monsefi, An effective semi-fragile
technique to improve security. watermarking method for image authentication based on lifting wavelet transform
and feed-forward neural network, Cognit. Comput. 12 (2020) 863–890.
5. Conclusion [16] A.A. Elrowayati, M.A. Alrshah, M.F.L. Abdullah, R. Latip, HEVC watermarking
techniques for authentication and copyright applications: challenges and
opportunities, IEEE Access 8 (2020) 114172–114189.
Image watermarking has lately acquired popularity due to the rise of [17] N.N. Raj, R. Shreelekshmi, A survey on fragile watermarking-based image
copyright concerns. Several watermarking approaches have been authentication schemes, Multimed. Tools Appl. 80 (2021) 19307–19333.
[18] F. Peng, Z.X. Lin, X. Zhang, M. Long, A semi-fragile reversible watermarking for
developed to improve image quality and resilience against various as- authenticating 2D engineering graphics based on improved region nesting, IEEE
saults. Deep learning improves computer vision by fitting and general- Trans. Circuits Syst. Video Technol. 31 (1) (2020) 411–424.
izing complicated features. Watermarking using deep learning has lately [19] A.M. Cheema, S.M. Adnan, Z. Mehmood, A novel optimized semi-blind scheme for
color image watermarking, IEEE Access 8 (2020) 169525–169547.
attracted substantial interest due to its various widely used applications [20] S.B.B. Ahmadi, G. Zhang, M. Rabbani, L. Boukela, H. Jelodar, An intelligent and
and a major influence on image security. Deep learning enables adaptive blind dual-color image watermarking for authentication and copyright protection,
image watermarking by dynamically extracting high- and low-level Appl. Intell. 51 (2021) 1701–1732.
[21] A.A. Mohammed, D.A. Salih, A.M. Saeed, M.Q. Kheder, An imperceptible semi-
features from large datasets. This survey thoroughly examines various
blind image watermarking scheme in DWT-SVD domain using a zigzag embedding
research publications on digital image watermarking in deep-learning technique, Multimed. Tools Appl. 79 (43–44) (2020) 32095–32118.
environments. We reviewed the watermarking idea, steps, perfor- [22] B. Dappuri, M.P. Rao, M.B. Sikha, Non-blind RGB watermarking approach using
SVD in translation invariant wavelet space with enhanced Grey-wolf optimizer,
mance measurements, and applications and illustrated the deep learning
Multimed. Tools Appl. 79 (2020) 31103–31124.
definition and the different types of neural networks. Finally, we high- [23] O.P. Singh, A.K. Singh, G. Srivastava, N. Kumar, Image watermarking using soft
lighted outstanding issues and future research paths for image water- computing techniques: a comprehensive survey, Multimed. Tools Appl. 80 (2021)
marking with deep learning. 30367–30398.
[24] H.Y. Yang, S.R. Qi, P.P. Niu, X.Y. Wang, Color image zero-watermarking based on
fast quaternion generic polar complex exponential transform, Signal Process.
Image Commun. 82 (2020) 115747.
Declaration of competing interest
[25] M. Rai, S. Goyal, A hybrid digital image watermarking technique based on fuzzy-
BPNN and shark smell optimization, Multimed. Tools Appl. 81 (27) (2022)
No conflict of interest exists. 39471–39489.
The second author received a fund, which is acknowledged in the [26] X. Liao, J. Peng, Y. Cao, GIFMarking: the robust watermarking for animated GIF-
based deep learning, J. Vis. Commun. Image Represent. 79 (2021) 103244.
manuscript. [27] S.S. Rajput, B. Mondal, F.Q. Warsi, A robust watermarking scheme via
The authors understand that the Corresponding Author is the sole optimization-based image reconstruction technique, Multimed. Tools Appl. 82
contact for the Editorial process. He is responsible for communicating (2023) 25039–25060.
[28] X. Zhang, R. Jiang, W. Sun, A. Song, X. Wei, R. Meng, LKAW: a robust
with the other authors about progress, submissions of revisions, and watermarking method based on large kernel convolution and adaptive weight
final approval of proofs. assignment, Comput. Mater. Continua 75 (1) (2023) 1–17.
[29] W. Zhang, R. Chen, B. Wang, A robust watermarking algorithm against JPEG
compression based on multiscale autoencoder, IET Image Process. 00 (2023) 1–15.
Data availability [30] J. Dong, H. Wang, Z. He, J. Niu, X. Zhu, G. Wu, TADW: traceable and anti-detection
dynamic watermarking of deep neural networks, Secur. Commun. Netw. 2022
Data will be made available on request. (2022) 1–11.
[31] H.K. Singh, A.K. Singh, Digital image watermarking using deep learning,
Multimed. Tools Appl. 83 (2023) 2979–2994.
[32] S. Dargan, M. Kumar, M.R. Ayyagari, G. Kumar, A survey of deep learning and its
References applications: a new paradigm to machine learning, Arch. Comput. Methods Eng. 27
(2020) 1071–1092.
[33] S. Ge, Z. Xia, J. Fei, Y. Tong, J. Weng, M. Li, A robust document image
[1] K.M. Hosny, A. Magdi, N.A. Lashin, O. El-Komy, A. Salah, Robust color image
watermarking scheme using deep neural network, Multimed. Tools Appl. 82 (2023)
watermarking using multi-core Raspberry Pi cluster, Multimed. Tools Appl. 81 (12)
38589–38612.
(2022) 17185–17204.
[34] S. Dong, P. Wang, K. Abbas, A survey on deep learning and its applications,
[2] D.S. Khafaga, S.M. Alhammad, A. Magdi, O. ElKomy, N.A. Lashin, K.M. Hosny,
Comput. Sci. Rev. 40 (2021) 100379.
Securing transmitted color images using zero watermarking and advanced
[35] M. Kaczyński, Z. Piotrowski, D. Pietrow, High-quality video watermarking based
encryption standard on raspberry Pi, Comput. Syst. Sci. Eng. 47 (2) (2023)
on deep neural networks for video with HEVC compression, Sensors 22 (19) (2022)
1967–1986.
7552.
[3] S. Gaur, V. Barthwal, An extensive analysis of digital image watermarking
[36] A. Shrestha, A. Mahmood, Review of deep learning algorithms and architectures,
techniques, Int. J. Intell. Syst. Appl. Eng. 12 (1) (2024) 121–145.
IEEE Access 7 (2019) 53040–53065.
[4] Y. Li, J. Li, C. Shao, U.A. Bhatti, J. Ma, Robust multi-watermarking algorithm for
[37] F. Deeba, S. Kun, F.A. Dharejo, H. Langah, H. Memon, Digital watermarking using
medical images using patchwork-DCT, in: International Conference on Artificial
deep neural network, Int. J. Mach. Learn. Comput. 10 (2) (2020) 277–282.
Intelligence and Security, Cham, Springer International Publishing, 2022,
[38] A. Aggarwal, M. Mittal, G. Battineni, Generative adversarial network: an overview
pp. 386–399.
of theory and applications, Int. J. Inf. Manage. Data Insights 1 (1) (2021) 100004.
[5] S. Thakur, A.K. Singh, S.P. Ghrera, M. Dave, Watermarking techniques and its
[39] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair,
applications in telehealth: a technical survey. Cryptographic and Information
Y. Bengio, Generative adversarial nets, Adv. Neural Inf. Process. Syst. 27 (2014)
Security Approaches for Images and Videos, CRC Press, 2018, pp. 467–508.
139–144.
[6] M.K. Venkatesh, G. Sailaja, C. Tapaswi, G. Chandramani, C. Rohith, Digital
[40] D. Saxena, J. Cao, Generative adversarial networks (GANs) challenges, solutions,
watermarking techniques using LSB, J. Eng. Sci. 14 (07) (2023).
and future directions, ACM Comput. Surv. (CSUR) 54 (3) (2021) 1–42.
[7] K. Ali, B. Ali, R. Abdelhalim, Medical image watermarking techniques in the
[41] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair,
transform domain: a review, AS-Proceedings 1 (7) (2023) 120–129.
Y. Bengio, Generative adversarial networks, Commun. ACM 63 (11) (2020)
[8] M. Ali, Robust image watermarking in spatial domain utilizing features equivalent
139–144.
to SVD transform, Appl. Sci. 13 (10) (2023) 6105.
[42] J. Gui, Z. Sun, Y. Wen, D. Tao, J. Ye, A review on generative adversarial networks:
[9] S.A. Hosseini, P. Farahmand, An attack-resistant hybrid blind image watermarking
algorithms, theory, and applications, IEEE Trans. Knowl. Data Eng. 35 (4) (2021)
scheme based on combination of DWT, DCT, and PCA, Multimed. Tools Appl. 83
3313–3332.
(7) (2023) 18829–18852.
[43] A. Borji, Pros and cons of GAN evaluation measures: new developments, Comput.
[10] N. Alaoui, DWT-based digital watermarking for various attacks, Acadlore Trans.
Vis. Image Understanding 215 (2022) 103329.
Mach. Learn 2 (4) (2023) 226–234.

11
K.M. Hosny et al. Computer Science Review 53 (2024) 100662

[44] A. Jabbar, X. Li, B. Omar, A survey on generative adversarial networks: variants, [71] M.M. Darwish, A.A. Farhat, T.M. El-Gindy, Convolutional neural network and 2D
applications, and training, ACM Comput. Surv. (CSUR) 54 (8) (2021) 1–49. logistic-adjusted-Chebyshev-based zero-watermarking of color images, Multimed.
[45] H.K. Singh, N. Baranwal, K.N. Singh, A.K. Singh, H. Zhou, GAN-based Tools Appl. (2023) 1–27.
watermarking for encrypted images in healthcare scenarios, Neurocomputing 560 [72] L.Y. Hsu, H.T. Hu, H.H. Chou, A high-capacity QRD-based blind color image
(2023) 126853. watermarking algorithm incorporated with AI technologies, Expert Syst. Appl. 199
[46] C. Annadurai, I. Nelson, K.N. Devi, R. Manikandan, A.H. Gandomi, Image (2022) 117134.
watermarking based data hiding by discrete wavelet transform quantization model [73] D.K. Mahto, A. Anand, A.K. Singh, Hybrid optimisation-based robust watermarking
with convolutional generative adversarial architectures, Appl. Sci. 13 (2) (2023) using denoising convolutional neural network, Soft Comput. 26 (16) (2022)
804. 8105–8116.
[47] Z. Pang, M. Wang, L. Cao, X. Chai, Z. Gan, Pairwise open-sourced dataset [74] S.A. Barlaskar, S.V. Singh, K. Anish Monsley, R.H. Laskar, Genetic algorithm-based
protection based on adaptive blind watermarking, Appl. Intell. 53 (14) (2023) optimized watermarking technique using hybrid DCNN-SVR and statistical
17391–17410. approach for watermark extraction, Multimed. Tools Appl. 81 (5) (2022)
[48] J. Huang, T. Luo, L. Li, G. Yang, H. Xu, C.C. Chang, ARWGAN: attention-guided 7461–7500.
Robust Image Watermarking Model Based on GAN, IEEE Trans. Instrum. Meas. 72 [75] Z. Zhang, H. Wang, H. Fu, A convolutional neural network-based blind robust
(2023) 1–17. image watermarking approach exploiting the frequency domain, Vis. Comput. 39
[49] S. Shukla, M. Jain, A novel stochastic deep resilient network for effective speech (8) (2023) 3533–3544.
recognition, Int. J. Speech Technol. 24 (3) (2021) 797–806. [76] S. Ingaleshwar, N.V. Dharwadkar, Water chaotic fruit fly optimization-based deep
[50] R. Ravindran, M.J. Santora, M.M. Jamali, Multi-object detection and tracking, convolutional neural network for image watermarking using wavelet transform,
based on DNN, for autonomous vehicles: a review, IEEE Sens. J. 21 (5) (2020) Multimed. Tools Appl. 82 (14) (2023) 21957–21981.
5668–5677. [77] M. Rai, S. Goyal, M. Pawar, An optimized deep fusion convolutional neural
[51] X. Shu, L. Zhang, Z. Wang, Q. Lv, Z. Yi, Deep neural networks with region-based network-based digital color image watermarking scheme for copyright protection,
pooling structures for mammographic image classification, IEEE Trans. Med. Circuits Syst. Signal Process. 42 (7) (2023) 4019–4050.
Imaging 39 (6) (2020) 2246–2255. [78] A. Tavakoli, Z. Honjani, H. Sajedi, Convolutional neural network-based image
[52] M. Xue, J. Wang, W. Liu, DNN intellectual property protection: taxonomy, attacks, watermarking using discrete wavelet transform, Int. J. Inf. Technol. 15 (4) (2023)
and evaluations, in: Proceedings of the 2021 Great Lakes Symposium on VLSI, 2021–2029.
2021, pp. 455–460. [79] X. Guo, W. Yang, L. Zhang, Y. Shi, J. Li, J. Sun, W. Wan, Deep image watermarking
[53] S. Ge, Z. Xia, J. Fei, Y. Tong, J. Weng, M. Li, A robust document image with loss-driven modification, Multimed. Tools Appl. 83 (13) (2023)
watermarking scheme using deep neural network, Multimed, Tools Appl 82 (25) 37665–37685.
(2023) 38589–38612. [80] H. Geng, M. Zhou, Novel post-photographic technique based on deep convolutional
[54] S. Jaiswal, M.K. Pandey, Color watermark extraction using deep neural network in neural network and blockchain technology, J. Supercomput. 80 (5) (2023)
IWT domain with PCA-based statistical feature reduction, SN Comput. Sci. 4 (5) 6119–6139.
(2023) 669. [81] Y. Bai, L. Li, J. Lu, S. Zhang, N. Chu, A novel steganography method for infrared
[55] R. Sinhal, I.A. Ansari, Machine learning-based multipurpose medical image image based on smooth wavelet transform and convolutional neural network,
watermarking, Neural Comput. Appl. 35 (31) (2023) 23041–23062. Sensors 23 (12) (2023) 5360.
[56] L.R. Battarusetty, G.R.N. Kumari, R. Tamilkodi, B.S. Kumar, Modeling automated [82] Y. Fan, J. Li, U.A. Bhatti, C. Shao, C. Gong, J. Cheng, Y. Chen, A multi-
image watermarking using meta-heuristic-based deep learning with wavelet watermarking algorithm for medical images using inception V3 and DCT, Comput.
approach, Sens. Imaging 24 (1) (2023) 24. Mater. Continua 74 (1) (2023) 1279–1302.
[57] L. Wang, Y. Song, D. Xia, Deep neural network watermarking based on a reversible [83] T. Huang, J. Xu, S. Tu, B. Han, Robust zero-watermarking scheme based on a
image hiding network, Pattern Anal. Appl. 26 (3) (2023) 861–874. depthwise overparameterized VGG network in healthcare information security,
[58] A. Verma, P. Pandey, M. Khari, ELA-Conv: forgery Detection in Digital Images Biomed. Signal Process. Control 81 (2023) 104478.
Based on ELA and CNN, in: International Conference on Recent Trends in Image [84] N.K. Sharma, S. Kumar, A. Rajpal, N. Kumar, MantaRayWmark: an image adaptive
Processing and Pattern Recognition, Cham, Springer Nature Switzerland, 2023, multiple embedding strength optimization-based watermarking using Manta-Ray
pp. 213–226. Foraging and bi-directional ELM, Expert Syst. Appl. 200 (2022) 116860.
[59] L. Chen, S. Li, Q. Bai, J. Yang, S. Jiang, Y. Miao, Review of image classification [85] W. Zhang, J. Li, U.A. Bhatti, J. Liu, J. Zheng, Y.W. Chen, Robust multi-
algorithms based on convolutional neural networks, Remote Sens. 13 (22) (2021) watermarking algorithm for medical images based on GoogLeNet and Henon map,
4712. Comput. Mater. Contin 75 (2023) 565–586.
[60] S.S. Basha, S.R. Dubey, V. Pulabaigari, S. Mukherjee, Impact of fully connected [86] L. Zhu, X. Wen, L. Mo, J. Ma, D. Wang, Robust location-secured high-definition
layers on performance of convolutional neural networks for image classification, image watermarking based on key-point detection and deep learning, Optik 248
Neurocomputing 378 (2020) 112–119. (2021) 168194.
[61] R. Ayachi, M. Afif, Y. Said, M. Atri, Strided convolution instead of max pooling for [87] G. Suresh, G. Bhuvaneswari, G. Manikandan, P. Shanthakumar, Chronological bald
memory efficiency of convolutional neural networks, in: Proceedings of the 8th eagle optimization-based deep learning for image watermarking, Expert Syst. Appl.
International Conference on Sciences of Electronics, Technologies of Information 238 (2024) 121545.
and Telecommunications (SETIT’18) 1, Springer International Publishing, 2020, [88] S.A. Nawaz, J. Li, M.U. Shoukat, U.A. Bhatti, M.A. Raza, Hybrid medical image
pp. 234–243. zero watermarking via discrete wavelet transform-ResNet101 and discrete cosine
[62] Y. Zhang, L. He, G. Cheng, MLPC-CNN: a multi-sensor vibration signal fault transform, Comput. Electr. Eng. 112 (2023) 108985.
diagnosis method under less computing resources, Measurement 188 (2022) [89] H.T. Hu, L.Y. Hsu, Blind color image watermarking incorporating a residual
110407. network for watermark denoising and super-resolution reconstruction, Soft
[63] N. Akhtar, U. Ragavendran, Interpretation of intelligence in CNN-pooling Comput. 27 (2) (2023) 917–934.
processes: a methodological survey, Neural Comput. Appl. 32 (3) (2020) 879–898. [90] R. Xiang, G. Liu, K. Li, J. Liu, Z. Zhang, M. Dang, Zero-watermark scheme for
[64] M. Momeny, A. Jahanbakhshi, K. Jafarnezhad, Y.D. Zhang, Accurate classification medical image protection based on style feature and ResNet, Biomed. Signal
of cherry fruit using deep CNN based on hybrid pooling approach, Postharvest Biol. Process. Control 86 (2023) 105127.
Technol. 166 (2020) 111204. [91] C. Gong, J. Liu, M. Gong, J. Li, U.A. Bhatti, J. Ma, Robust medical zero-
[65] P. Wang, E. Fan, P. Wang, Comparative analysis of image classification algorithms watermarking algorithm based on Residual-DenseNet, IET Biom. 11 (6) (2022)
based on traditional machine learning and deep learning, Pattern Recognit. Lett. 547–556.
141 (2021) 61–67. [92] F. Dong, J. Li, U.A. Bhatti, J. Liu, Y.W. Chen, D. Li, Robust zero watermarking
[66] L. Alzubaidi, J. Zhang, A.J. Humaidi, A. Al-Dujaili, Y. Duan, O. Al-Shamma, algorithm for medical images based on improved NasNet-mobile and DCT,
L. Farhan, Review of deep learning: concepts, CNN architectures, challenges, Electronics 12 (16) (2023) 3444.
applications, future directions, J .Big Data 8 (2021) 1–74. [93] S.A. Nawaz, J. Li, U.A. Bhatti, M.U. Shoukat, D. Li, M.A. Raza, Hybrid
[67] G. Liu, R. Xiang, J. Liu, R. Pan, Z. Zhang, An invisible and robust watermarking watermarking algorithm for medical images based on digital transformation and
scheme using convolutional neural networks, Expert Syst. Appl. 210 (2022) MobileNetV2, Inf. Sci. 653 (2024) 119810.
118529. [94] D. Li, J. Li, U.A. Bhatti, S.A. Nawaz, J. Liu, Y.W. Chen, L. Cao, Hybrid encrypted
[68] A. Chacko, S. Chacko, Deep learning-based robust medical image watermarking watermarking algorithm for medical images based on DCT and improved
exploiting DCT and Harris Hawks optimization, Int. J. Intell. Syst. 37 (8) (2022) DarkNet53, Electronics 12 (7) (2023) 1554.
4810–4844. [95] A. Anand, J. Bedi, A. Aggarwal, M.A. Khan, I. Rida, Authenticating and securing
[69] H.C. Wu, W.L. Fan, C.S. Tsai, J.J.C. Ying, An image authentication and recovery healthcare records: a deep learning-based zero watermarking approach, Image Vis.
system based on discrete wavelet transform and convolutional neural networks, Comput. 145 (2024) 104975.
Multimed. Tools Appl. 81 (14) (2022) 19351–19375. [96] X. Liu, Y. Wu, P. Gao, J. Ouyang, Z. Shao, Color image watermarking based on
[70] H.K. Singh, A.K. Singh, Digital image watermarking using deep learning, singular value decomposition and generalized regression neural network,
Multimed. Tools Appl. 83 (1) (2024) 2979–2994. Multimed. Tools Appl. 81 (22) (2022) 32073–32091.

12