Comprehensive survey of image steganography: Techniques, Evaluations, and trends in future research

Comprehensive survey of image steganography: Techniques, Evaluations, and trends in future research

Neurocomputing 335 (2019) 299–326 Contents lists available at ScienceDirect Neurocomputing journal homepage: www.elsevier.com/locate/neucom Compreh...

3MB Sizes 0 Downloads 3 Views

Neurocomputing 335 (2019) 299–326

Contents lists available at ScienceDirect

Neurocomputing journal homepage: www.elsevier.com/locate/neucom

Comprehensive survey of image steganography: Techniques, Evaluations, and trends in future research Inas Jawad Kadhim a,b,∗, Prashan Premaratne a, Peter James Vial a, Brendan Halloran a a b

School of Electrical and Computer and Telecommunications Engineering, University of Wollongong, North Wollongong, NSW, 2522, Australia Electrical Engineering Technical College, Middle Technical University, Baghdad, Iraq

a r t i c l e

i n f o

Article history: Received 4 June 2018 Accepted 28 June 2018 Available online 9 November 2018 Keywords: Data hiding 2D and 3D image steganography Reversible Irreversible Spatial domain Transform domain Adaptive steganography Steganalysis

a b s t r a c t Storing and communicating secret and/or private information has become part of our daily life whether it is for our employment or personal well-being. Therefore, secure storage and transmission of the secret information have received the undivided attention of many researchers. The techniques for hiding confidential data in inconspicuous digital media such as video, audio, and image are collectively termed as Steganography. Among various media types used, the popularity and availability of digital images are high and in this research work and hence, our focus is on implementing digital image steganography. The main challenge in designing a steganographic system is to maintain a fair trade-off between robustness, security, imperceptibility and higher bit embedding rate. This research article provides a thorough review of existing types of image steganography and the recent contributions in each category in multiple modalities. The article also provides a complete overview of image steganography including general operation, requirements, different aspects, different types and their performance evaluations. Different performance analysis measures for evaluating steganographic system are also discussed here. Moreover, we also discuss the strategy to select different cover media for different applications and a few state-ofthe-art steganalysis systems. © 2018 Elsevier B.V. All rights reserved.

1. Introduction to information security In the modern world, advancements in digital communication have a key role in our everyday lives. Innovations in internet technologies along with the digitalization of information boosted the use of data transfer on an exponential scale. Information security has a paramount role in securing information. Even if there are many robust and highly secure methodologies are present, it is still progressing towards making these techniques more secure and robust in terms of performance measures. Without a doubt, data security is the soul of data communication. Generally, information security systems are separated into two major categories, one is encryption and another is information hiding [1]. Both of the categories are responsible for securing the information but their techniques differ. Data encryption and hiding methods were developed by various researchers [2–5]. Fig. 1 represents the classification tree of a general data security system.

∗ Corresponding author at: School of Electrical and Computer and Telecommunications Engineering, University of Wollongong, North Wollongong, NSW, 2522, Australia. E-mail address: [email protected] (I.J. Kadhim).

https://doi.org/10.1016/j.neucom.2018.06.075 0925-2312/© 2018 Elsevier B.V. All rights reserved.

Cryptography is one of the most attractive areas in the data analysis area which deals with different data encryption methodologies. Its objective is to use different encryption techniques and convert secret information into a scribbled encrypted form. The encrypted data is suitable for transmission in terms of data security as it is incomprehensible by any third party. The method is very useful for peer-to-peer communication. As the encrypted data is generated using techniques which involve many permutations and substitutions, any third party is not able to gain access to the original message. To perform cryptographic techniques, a key is required. It is further segregated into two ways. One is Symmetric key cryptography; which requires just a single key for both encryption and decryption while another one is Asymmetric key cryptography where multiple keys are used for data encryption and decoding. In the latter type, a public key is used for the encryption purpose and private keys will be used for decoding by authorised personals/systems [6]. Steganography is another way to secure messages during data communication. The ultimate goal of steganography and cryptography are the same, but their approaches are different. Steganography does not change the format of data or message and keeps the presence of its actual data, whereas cryptography keeps the data secret by converting it into an unreadable form. Cryptographic

300

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Fig. 1. Classification tree of security systems [1].

method’s weakness lies in the presence of original data, even if the original data has been encrypted. Hence steganography techniques are an additive security to cryptographic techniques. With the combination of both, it provides an additional layer of security for the message during data communication. Even though the objective of the cryptography and steganography are the same, which is to provide secure end-to-end data communication, their definitions of robustness to attack are different. A cryptographic system is regarded as broken when the third party gains access to the original data, whereas the steganography system is regarded as broken when the third party gets the presence of secret data [7]. Whenever data security is an issue, Watermarking comes to light. It is one of the popular data security techniques that use authentication and copyright method to serve its purpose. The process of embedding digital data in the multimedia data is called a watermark. Watermarking can be further segregated into two ways based on its visibility on files such as visible watermarking and invisible watermarking. The watermarking process prevents from unethical duplication of the media files and pseudo-ownership claim [8–11]. It is useful for one to many communications. The prime motive behind watermarking is to generate secure, robust and effective watermarks on the media files or documents, which is permanent and unalterable by an unauthorised party [12]. Watermarking is also pertinent to steganography in a broader sense, as both are intending to hide the data in media data. Both of the techniques possess properties such as security, capacity, robustness, and imperceptibility, but they prioritize them differently. Let’s say, imperceptibility is of prime importance for steganography process, whereas watermarking gives the highest priority to robustness [13,14]. Nonetheless, there are some cases where imperceptibility is not of high priority in steganography. One of the situations is where, when the colour tone of a particular object has altered in the image. Still the original message is a secure and third party is not able to retrieve it. The ultimate goal of watermarking and steganography are distinct as steganography aims to provide secure data communication by embedding a secret data in another digital object such as digital image and watermarking aims to provide protection from

violation of copyrights. Fingerprinting and Watermarking are similar in a way that, both require to mark objects. The only distinction is that each fingerprinting requires different fingerprints for proving their ownership and watermarking marks same watermark on the different object [2]. Another major difference in Watermarking and Steganography techniques are, steganography can select an object to hide the secret message in it, but in case of watermarking, particular object is required to cover and cannot be ignored [13,15]. In Table 1, the brief summarizes a thorough comparison among various information security techniques. 2. Background of steganography The word ‘Steganography’ itself is made of two old Greek words, one is “Stegano” and another is “Graphy”. The combined meaning is “Cover Writing”. For thousands of years, it has been used in many forms. During the 5th century, Histaiacus used it for sending a secret message, by tattooing a message on his slave’s skull and the slave moved to send a message with grown hair. Greeks were famous for sending secret messages. One of the ancient Greeks, Demeratus used a wax-layered tablet, in this methodology, the message is written on the wooden made writing tablet by scratching on it. The wax is scrapped off first and then the message was scratched. After that, the wooden tablet layered again with wax, which seems like similar to another blank tablet. It does not create any suspicion and message is sent safely [1,7,16,17]. Almost fifty decades ago, one of the Italian mathematicians, Jerome Carden revived a methodology for writing secret messages, which was used by Chinese in ancient times. The method was to use a paper with grid holes as a mask and write a secret message by putting it on a blank paper and this mask is shared between both the parties; i.e. sender and the receiver. After putting off the grid mask, the blank part of the paper is then filled which appears as the innocuous text. During the first world war, Germans used multiple stages based microdot technology by the use of waste materials of magazines [18]. There were many techniques used to write secret messages in World War II, including writing opencoded messages, Enigma machine, different null ciphers and using invisible ink [7]. One of the kings from Saudi Arabia also started a

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

301

Table 1 Comparison of characteristics of Information Security techniques. Characteristics

Steganography

Watermarking

Cryptography

Goal Cover Selection Challenges Key Output Visibility The system is invalid if Attacks

Preserve the confidential data from the detection Free cover selection Imperceptibility, Security and Capacity Optional Stego-file Certainly not Detected Steganalysis

Preserve the authenticity of the cover media Restriction Robustness Optional Watermarked-file Sometimes Removed or replaced Any image processing

Obfuscate the form or content of data N/A Robustness Compulsory Cipher-text Always De-ciphered Cryptanalysis

project for secret writing at Abdulaziz City of Science and Technology. It was found in a twelve hundred years old manuscript. These manuscripts were collected from Germany and Turkey [1,19]. There are many kinds of literature available, to get more insight about the history of steganography and methodologies used all over the world [7,16,20,21]. With the growth of wireless systems, inter-linked multimedia systems, and electronic digital camera, data digitalization has immensely raised the potential of regenerating and distribution of information. In the past two decades, Steganography techniques had been moved to digital processing, because of the growth in the computer processing power and internet speeds. Recent research and development in the field of signal processing [22,23], encoding techniques and information theory are helping in finding secure and robust steganography techniques. In the digital world, steganography has emerged as the secure data transmission technique and it is attracting various industrial applications, hence there is assurance of the evolution. The latest steganography techniques are not only limited to hiding secret information in the images, but it also aids in embedding data in text [24,25], codes [26], audio [27,28], video [29] and DNA [30,31]. It also includes hiding information in various formats such as Extensible Markup Language(XML), executable (EXE) and Hyper Text Mark-up Language (HTML) [32,33]. The papers [34,35] also presented and reviewed many trending digital techniques in the field of steganography. 3. The general procedure in steganography The basic idea behind a digital steganographic process is to conceal secret or private information inside a cover media in an undetectable manner. The secret data type may vary from binary bits, text data, image and video files. The cover media can also be any popular digital media like image, video or text. The data concealed inside a cover/host media is termed as ‘Secret data’, and the resulting embedded cover media is known as ‘Stego-media’. The stegomedia is intended to share through an open or unsecured communication channel. A block diagram representation of a generic stegano system is depicted in Fig. 2. Systems intended for better security feature may often use a security key and/or an encryption scheme during the embedding process. The key may hold supporting information such as, embedding map, encryption password, the threshold value for selecting embedding coefficients etc. In general, an embedding system can be represented using the equation

C  = Em(C, En(S, k1 ), k2 )

(1)

where, C is the cover media, S is the secret data, C is the stegomedia and k1 and k2 are secret keys used for Encryption function En(.) and embedding function Em(.) respectively. The stego-data is then passed through a channel and a reverse embedding and decryption algorithm is supposed to apply over the stego-media and is represented as

Sr = D(Ex(C ∗ , k2 ), k1 )

(2)

where, Sr is the retrieved secret data, Ex(.) is the extraction function (inverse embedding function), and D(.) is the decryption function. C∗ is the deformed (due to channel noise or intruder’s attacks) stego-media received at the receiver side. 4. Properties of steganography There are mainly three essential properties of any steganographic systems, namely imperceptibility, security and capacity of hiding information [36,37]. While [35,38–40] in their research mentioned that there are four properties, same pervious properties along with robustness. These are the most effective parameters that test the effectiveness of a steganographic system. For some steganographic systems, there are particular requirements for handling it, as per its application. Watermarking and steganography have all the properties of data embedding. There is a trade-off among the properties; when it increases the amount of secret data in the stego-image, the artifacts effect increases and immunity towards modification of stego-file decreases [41,42]. All the properties must be maintained at an optimum level. The highly robust steganographic system in some application is not always a requirement, but high security, capacity and imperceptibility of secret information are required. In the case of digital watermarking, imperceptibility and high capacity are not mandatory. Venkatraman et al. [43] suggested that robustness is highly required in the response of undesirable and malicious attack. Fig. 3 represents the key requirements for any steganography system. The further section clarifies the discussed properties in a more in-depth manner. 4.1. Imperceptibility The highest priority requirement for any data embedding is imperceptibility as the key feature and strength of any steganographic technique is in hiding the secret data in the digital image such that, it could not be comprehended by naked human eye or with the use of statistics [15]. Statistical techniques are also useful tools for attackers to recognize whether any secret data communication is occurring or not. Hence, steganographic techniques must not alter the cover media perception or its statistical means due to secret data embedding. In simple words, if the statistical data for the stego-file and the original data file are similar, then the security is better for data communication. Though embedding secret data in the cover image adds some amount of noise to it, the quality of cover image should not be diminished during sharing via unsecured channels [43]. 4.2. Security In a steganographic system, the word “security” indirectly refers to “un-noticeability” or “undetectability”. Hence any steganography technique is regarded as secure if the secret data is not detectable by statistical means or removal after being detected by the attacker. The key requirement of the steganographic process is the

302

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Fig. 2. Block diagram of a Steganographic system.

Fig. 3. Trade-off between the properties of the data hiding.

secure transmission of secret data. So security is the primary concern in order to avoid data access by unauthorized persons or computer while transmitting through an open channel.

analog format. Whereas for a fingerprint system, robustness is required in case of modifying or manipulating files deliberately. 5. Applications

4.3. Payload capacity An efficient steganographic system always aims at sending maximum information using minimum cover media. This will help to reduce the chance of interception while sending through an unsecure network and therefore usually demands high embedding capacity. Venkatraman et al. [43] defined the embedding rate as the amount of information hidden (in bits) comparative to the size of the cover image. Keeping higher payload capacity without sacrificing imperceptibility and security is a major challenge in steganography. 4.4. Robustness This represents the ability of the embedding and decoding scheme even if the stego-image is corrupted by a third hand using image processing techniques like rotation, scaling, resizing etc. [39,40] . With steganography, active attack scenario is not considered as there is an assumption of sending stego-files via the internet. Hence the stego-file is not affected and the receiver receives the distortion-less stego-file. The steganographic systems are less robust when any modification is done to stego-files such as file format conversion, compression and converting digital files into

Secure and stealth communication is widely used in almost all fields. Main beneficiaries include medical, military, multimedia, and industry where secret communications may be used for internal and external security purposes. In the medical field, crucial private information is hidden in the medical data itself and sequence of DNA and propagated. This will help to avoid the leakage of private details in unauthorized hands. Reversible steganographic systems are more common there since both the cover and secret data should be extracted separately at the receiver side. Security is the primary concern in military and defence communication. Open channels may be compromised and authorized communication is much more important. Such stegano systems use multiple layered encryption techniques before the embedding process. In multimedia applications, steganography is often applied to mark the copy right information. This is termed as watermarking and here, the cover media have more significance than the secret data. In industry and corporate communication, authenticity and security are much important since unsafe communication may result in serious data leakage. Some applications are presented here such as Smartsteg on mobile devices [44], securing multimodal biometric data [45], protection of IP (Intellectual Properties) and embedding individual information in smart identity card are also available [46]. One of the

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

303

advanced steganography techniques is to use it with an advanced data structure; it helps in securing a large amount of information. The end-to-end data transmission could be done with the actual file securely using meta information with it. With the use of advanced data structure, the problem of allocation in the hard-disk memory can be targeted and yield in addressing the big data problems [47]. The well-known Japanese company, Fujitsu is inventing the system to encrypt information in a printed image which is hidden from the naked human eye, but its decryption is possible with the handheld camera. The whole process of decryption takes less than a second as the size of the hidden information is only 12 bytes [1]. The reason for the popularity of steganography is not only because of its ethical uses but also because of some unethical uses, such as embedding viruses, spam and malicious links to the data by the cyber-criminals. Hence in the past decade, the rapid growth in the field of steganography has attracted people from all domains. There are also ready steganography tools available online that lets people with no technical knowledge to delve into unethical cyber-criminal activities. Xiao Steganography is one such tool, which lets any user to leak confidential information from a company with just simple three steps and rest will be handled by the tool [48]. The steps are to select any cover image, write any secret message to embed or select a secret file and hit the Button to embed. In all of these applications, image steganography can perform a vital role since the images are nowadays a common data among digital communication systems. Detecting secret data from millions of images is practically impossible if it is safely communicated.

6.3. Structural similarity index measure (SSIM)

6. Performance evaluation techniques

pixel in the stego-image and Ci is the intensity of ith pixel in the cover image. μs and μc are the mean pixel values of stego-image and cover image respectively. Some other metrics as defined in [35,51–54] can also measure the security of a system and the quality of the image between the cover and stego-image such as :Weighted Peak Signal-to-Noise Ratio (WPSNR), Kullback–Leibler Divergence (K–L divergence), PSNR-HSV, Receiver Operating Characteristic curve (ROC), Universal Image Quality Index (UIQI), Maximum Mean Discrepancy (MMD), Euclidean Distance (ED) and Manhattan Distance (MD).

For evaluating the different aspects of image steganography, different metrics are used [49,50]. Some of the popular metrics include Peak Signal to Noise Ratio, Correlation coefficient, histogram comparison, Structured Similarity Index Measure (SSIM) and Payload Capacity. The different metrics used are defined next. 6.1. Payload capacity It is the measure of the volume of information concealed in the cover image. It is commonly represented as Bits per Pixel (BPP), where

BPP =

Number o f secret bits embedded T otal pixels in the cover image

(3)

Payload capacity is much more crucial in a steganographic system as the communication overhead will be in accordance with the maximum payload capacity. 6.2. Peak signal to noise ratio (PSNR) As the cover image is altered to embed the secret data, there will be changes in the Cover image pixel values. The changes need to be analysed since it directly affect the imperceptibility of the output Stego-image. PSNR is one of the popular and top notch metric used to measure the quality of the Stego-image by analysing the mean squared error value between the Cover and the Stego-image. For 8-bit images, PSNR is calculated as



P SNR(in dB ) = 10 log10 and

MSE =

N

i=1



2552 MSE



(4)

  2

Ci − Ci N

(5)

where MSE is the mean squared error while comparing the stego and cover image, N is the number of cover image pixels, Ci is the intensity of ith pixel in the stego-image and Ci is the intensity of ith pixel in the cover image.

SSIM is a mode of comparison metric to check the similarity between two images. It is calculated as

SSIM (x, y ) =

(2μx μy + c1 )(2σxy + c2 )   μx 2 + μy 2 + c 1 σ x 2 + σ y 2 + c 2



(6)

c1 = ( k1 L )2 c2 = ( k2 L )2 where μx and μy are the mean intensity values of images x and y. σ x 2 is the variance of x, σ y 2 is the variance of y and σ xy 2 is the covariance of x and y. c1 and c2 are the two stabilizing parameters, L is the dynamic range of pixel values (2#bits per pixel - 1) and the contents k1 = 0.01 and k2 = 0.03. 6.4. Correlation factor Correlation factor (correlation coefficient) is calculated as

 μs (Ci − μc ) Correlation factor, r =  2 N  N   2 i=1 Ci − μs i=1 (Ci − μc ) N

i=1



Ci −

(7) where N is the cover image pixel count, Ci is the intensity of ith

7. Classification of image steganography techniques The essential feature of Image steganography is to keep communication secure while transmitting Stego-image over any network or communication channel. Different methods for image steganography were proposed based on the application and stages included in the embedding process. Hence, these systems can be classified based on multiple modalities and a detailed description will be presented in the subsequent sections. Here we are addressing the classification of image steganography based on several concerns. The categorization can rely on the type of cover image used (2D or 3D images), target application type or on the retrieval process (reversible or irreversible), nature of embedding process (spatial or transform domain) and adaptive steganography. Based on the research scope, Fig. 4 illustrated a flowchart of classification image steganography techniques. 7.1. Classification based on the cover image dimension Images can be represented in multiple formats. It includes 2 dimensional images such as grayscale or binary images, 3D images such as tri colour RGB images, or multi-slice images like MRI. Based on the dimension of the cover image used, the steganographic system can be classified into 2D image steganography or 3D image steganography [55]. In 2D image steganography, the secret data is embedded over the 2D plane of the cover image. The embedding may be either

304

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Fig. 4. Classification of Image Steganography Techniques.

on the spatial domain pixel values or on transform domain coefficient values. The scheme can be even applied over the 3D images, where each image planes will be used one by one at a time. For example, in a colour image, 2D steganography can be applied by splitting the colour planes and by using red, green and blue planes separately for hiding secret bits. This scheme is more popular than the other one and there are many concepts postulated under this research field. A 2D steganographic approach which makes use of the RGB colour space is explained in the article [56]. The individual colour planes are separated here and the sender and the receiver use an identical shared dictionary key to conceal a mapping between the secret data bits and colour ranges. The shared dictionary key is used to compress and secure the message at the sender side and the same key is required to extract the secret message at the receiver side. The main advantage of 2D steganography is that, popular public images can be used as the cover media and help to share the secret data in a stealth mode and hence more research work is occurring in this field. The 3D image steganography can generally be made in a number of ways: Geometrical domain [57–61], Topological domain [62] and Representation domain [63] steganography. In 3D image steganography, the basic approach is to embed the secret bit streams in the vertices of a 3D cover image. The bits are embedded in the vertices of 3D geometrical models and are preferred rather than other models (topological and representation) when

higher payload capacity is required. Similar to 2D steganography, 3D methods can also be used in both spatial and transform domain. However, the major share of the 3D steganography is plotted in the time domain to avoid unnecessary complications in generating valid 3D geometrical patterns in the frequency domain. With respect to 2D methods, time complexity is a factor here since a lot of points have to be addressed in the 3D space for the embedding purpose. While embedding the secret bits, the feature points such as vertices or edges of specified geometrical shapes are altered in the process in an undetectable pattern. Embedding limit or payload capacity is determined in two different aspects in 2D and 3D steganography. In 2D image steganography, a numberof bits per cover image pixel is taken as the benchmark metric and is known as Bits Per Pixels (BPP). While, in the case of 3D type, the capacity is measured either in terms of bits embedded in voxels (3D pixel groups) or bits per vertex points of the geometrical models in the cover media. Generally speaking, 3D image steganography is complex in nature but have better security and capacity while comparing basic 2D image steganography. Since the embedding is over vertices and faces in 3D steganography, the changes in stego-data due to noise and attacks will be high. Hence, at the receiver side, the stego-image need to be pre-processed using complex recovering techniques before the actual data extraction stage, to reduce the possible distortion occurred while transmitting through an unsecure open channel. Another possible

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

305

Table 2 Summarizes most of 3D image steganography work. Reference

Features

Comments

Cheng et al. [61]

- 3D polygonal and geometric model - Spatial substitute procedure - Embedding data secret based on shifting point by its virtual multilevel

Thiyagarajan et al. [58]

−3D triangle meshes and geometric models - Spatial domain, Pattern based Steganography - Embedding data secret in the vertices of the triangles

Wu and Dugelay [57]

- 3D geometric models - Spatial domain, LBS+ substitution to encode secret - Adjacent two bins method was applied on the 3D VRML models

Li et al. [59]

- Different 3D models are used -Using shifting strategy with a truncated space of data to embed data in the 3D model - Generate suit model for embedding by using pre-processing step for 3D model - Using the library function randint for generation binary data bit stream - Using the PCA in the pre-processing stage to robust against attack - Using a specified threshold to adjust the embedding distortion −3D Polygonal Models - Spatial domain, - Normalized distance between reference vertex is used to generate a histogram - Applying histogram shifting over the peak vertices points - 3D polygonal models ,topological domain - adaptive 3D steganographic to embed varies amount of data - Using vertex decimation procedure to find its referencing neighbours. - Using PCA analysis to robust against attack

- Reversible -Robust against affine transformation - Less distortion - Embedded 3 bits per points - No security key - Not robust against compression - Irreversible - Robust against affine transformation - Less distortion - High capacity, embedded 9 bits per points - Generation secret key -Not robust against compression - Irreversible Not robust against affine transformation - Less distortion - Low payload capacity, Embedded 2 bits per points - Not used security key - Irreversible - Failing to extract secret data since the PCA not able to find the right position - Generation security key - Not Robust against affine transformation - Less distortion - High embedding

Huang and Tsai [60]

Tasi [62]

drawback in 3D steganography is that, the embedding of secret data over the irregular 3D mesh points may create errors while extracting the data from the cover points. 3D methods are also susceptible to attacks such as smoothing, vertices resembling etc. Table 2 summarizes most of the 3D steganography work. To get more insight into the techniques in different application, survey papers in 2D image steganography are available in [1,35,64–66]. Also, survey papers on 3D image steganography and watermark are available in [55,67,68].

- Reversible - Low embedding capacity - Using security key - Good visual quality, -Not Robust against affine transformation - Irreversible - Adaptive - Using security key - Acceptable visual quality - Robustness against similarity transformation

hence, the loss-less retrieval of cover data is not a primary concern. The algorithms in this type are designed to increase the perceptibility, payload capacity, robustness (to save the secret data from stego-image distortions), lossless retrieval of secret data etc. Commonly, in an irreversible technique, the cover image bits get altered in an irreversible way and it is practically impossible to retrieve back the altered cover image bits from the stego-image. Hence the technique is advisable for the applications where no relevant information in the cover image needs to be passed to the receiver.

7.2. Classification based on the retrieval nature In a Steganographic process, usually the secret data bits are embedded in the cover media and at the receiver, the embedded data gets separated from the stego-media. The process can be designed with two motives: reversible and irreversible type [64,69,70] In the former, both the cover\host and secret data gets recovered without any data loss and in the latter type, the error free retrieval is applicable only for the secret data and the cover image may get corrupted. A thorough comparison along with some popular techniques in each type is explained below. 7.2.1. Irreversible type steganography This is one of the popular fields in the image steganography where the importance is given to the secret data and the cover image is just using as a wrapper to hide the secret message. Usually, a widely known public image is used as the cover image and

7.2.2. Reversible type steganography Reversible image steganography is a mode of data hiding system where both cover and secret images are important and both of them need to be separated from the stego-image in a lossless manner. Usually the technique has significance in advanced communication scenarios such as military and medical applications rather than common steganographic communications. In watermarking, reversibility is commonly used than in steganography [71]. Here, the cover data may be more important than that of the secret image and the leakage or loss in these details will be critical and algorithms are designed in such a way that, the retrieval accuracy should be high. Several reversible information hiding systems have been proposed and some of them are detailed here. The popular reversible schemes use error expansion or histogram shifting mechanism or a hybrid/modified version of these two or quantization based [70]. It can be argued that most of the

306

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

reversible techniques are not considering the security and robustness factors and are very critical in certain scenarios. The lossless extraction will be possible only when the stego-image is shared through a secured channel free from noise and geometric attacks. Such systems are hard to maintain and hence the reversible techniques are usually limited to applications where the cover image details are critical. 7.3. Classification based on embedding domain The embedding process used for hiding secret data over the cover image is the backbone of the steganographic process. Since there are possibilities for embedding secret data over the spatial domain and transform domain of the cover image [1,20,72–75], a subtle classification based on the nature of cover domain is used. 7.3.1. Spatial domain image steganography The easiest and the simplest way of data embedding in digital images is to modify the cover image pixel values in the spatial domain itself. These techniques use the cover image pixel intensity value levels directly or indirectly to encode the secret message bits. These methodologies fall into some of the simplest mechanisms in terms of embedding and decoding complexity. Spatial or Image domain techniques uses bit-wise methods that apply bit insertion and noise manipulation using simple mechanisms. The main steganographic schemes come under the spatial domain technique includes Least Significant Bit (LSB), Pixel Value Differencing (PVD), Histogram Shifting, Expansion based, Multiple Bit-planes based, Quantization based, Palette based, Pattern based and Pixel Intensity Modulation based Steganography. Recent works in these areas along with a detailed analysis is explained in the following sections. Also, some examples of the research work are explained in Table 3. Least Significant Bit (LSB) steganography: Least Significant Bit (LSB) technique is one of the easiest and hence popular spatial image steganographic approaches. The idea behind this method is that, in an image, the least significant bits represent only feeble information and small changes in those bits cannot get detected by human eyes. In LSB-based spatial domain techniques, the secret data get embedded directly in the host image by altering the least significant bits of selected pixels without distorting the visual quality of the original cover image. This technique yields when it is used in communication channels with only human attacks, where the intruders cannot find visual quality degradation. But statistically speaking, the embedding process creates a noise in the range of 50% of the average bit embedding rate (embedded bits per pixel). Earlier works in LSB steganography [72,74] merely concentrate to design the system to increase the payload capacity by utilizing most of the cover image pixels. As the time passed by, steganalysis field has become strong enough to break such systems using statistical analysis. Then the research problem became more focused to develop sophisticated robust LSB techniques based cryptography-steganography which can evade such steganalysis attacks [75–77]. In order to progress the efficiency, a lot of improved versions of LSB based image steganography is undertaken in research. The important ones use LSB matching algorithms [78] , Adaptive LSB embedding based on image characteristics like intensity, texture contents or nature of edge pixels [79,80], Optimized LSB substitution based on learning approaches etc [81,82]. Also, to increase embedding capacity, LSB technique can be extended up to 4 LSB planes at the cost of reduced imperceptibility [83]. The main advantage of the LSB technique is its simplicity of embedding and decoding process. Since most of the image format uses an 8-bit representation for representing individual pixels, the least significant bits (usually 6th−8th bit) of some or all of the pixel intensity values of

the cover image are modulated according to the secret data. While using a 24-bit colour image as the cover media, least significant bits of each colour planes (Red, Green and Blue) are changed as per the data. However, LSB methods are susceptible to statistical attacks, with few manipulations in the stego-image. Since the soul of the LSB steganography is the modification of spatial pixel values of the cover image, its decoding performance depends on the effect of noise, compression quantization etc. in the communication channel, even if we disregard the intruder attacks. Pixel value differencing (PVD) steganography: In this method, the secret message is being hidden by comparing the differences between the pixel values of two successive pixels. In the basic PVD method [84], while embedding, a secret message and a cover image are divided into non-overlapping blocks with two adjacent pixels and the difference values of pixels in each block are then lodged into a number of groups. The selection of the range intervals is selected based on the human vision sensitivity to intensity value variations from low fidelity to high frequency. For dense embedding scenario, PVD shows better accuracy over the LSB methods since the embedding is smoother in the PVD techniques [1]. The system design is in such a way that, the alteration will be in the limits of specified range interval. In the field of PVD steganography, many approaches were proposed by analysing the correlation of pixels. Different neighbourhood schemes such as five, six, seven and eight neighbourhoods are used to find the difference to predict the most optimized level of embedding in the cover pixel [85]. Visual distortions seem feeble while comparing with most of the other PVD approaches. The main downside of this PVD method is the lack of security even though it provides much better image perceptibility factor. To improve this security factor of the PVD approach, many additional security features are included in the standard PVD scheme, e.g. Hussain et al. [86] proposed a data concealing method which improves the security by using two modes of the embedding process. The processes are improved rightmost digit replacement (iRMDR) and parity-bit pixel value difference (PBPVD). Another example of enhanced security, histogram analysis based vulnerability PVD [87] is proposed here. In order to combine advantages of different embedding schemes, hybrid embedding schemes were proposed. (i.e. A steganographic methods employing both PVD and LSB substitution [88,89]). In order to advance the efficiency, a lot of improved versions of PVD based image steganography is researched. The important ones use Adaptive PVD block by employing pseudo-random number mechanisms for selecting the blokes [90] and solving falloff boundary issue in PVD by optimizing strategy [91]. Histogram shifting based steganography: The core strategy in this method is by shifting the histogram levels of the cover media. The valley (lowest) and peak points (highest) in the cover image histogram is determined and the embedding process is then by changing these valley and peak points [92,93]. The method preserves imperceptibility and provides higher payload capacity. The main advantage of the histogram-based image hiding is that, the scheme supports reversible data hiding [94]. This technique also prevents the condition of exceeding grey values above 255 and below 0 intensity values. For avoiding overflow and underflow issues and increasing the data hiding capacity, an alternative solution for embedding the maximum and minimum points in the histogram map by using a binary tree structure is proposed [95]. Another advanced approach based on histogram shifting imitated method is explained in [96]. Instead of conventional reference point selection from histogram map, here employs a selection procedure based on image pixel intensity. Prior to the embedding process, the image intensity range is divided into non-overlapping segments. Similar to the other histogram shifting based techniques, the secret data embedding is processed by altering the peak point pixel intensity with others, within the same segment of the peak. Since, the

Table 3 Review of main spatial domain image steganography methods. Reference

Classification Method

Method Description

Reversibility

Advantage

Disadvantage

Results

Sutaone and Khandar [74]

LSB

The embedding of spread out secret data over the cover image is based on random LSB substitution

No

Including some cryptography by using the security key

Rajendran and Doraipandian [75]

LSB

LSB technique which used chaotic map-based image steganography. Chaotic sequence is generated using a 1-D logistic map and the secret data embedding is based on this generated sequence

No

-The use of chaotic sequence helps to increase the security of the hiding process as well as maintaining the simplicity of the process. - Higher visual quality

Grajeda-Marín el at. [91]

PVD

Modifying the Tri-way PVD by finding the best value of each pixel pairs where their difference gives the maximum information without neglecting any. Also proposed Extra Bit insert of optimal Tri-way PVD

No

-Avoiding falling-off boundary issue -Higher payload capacity

- Low payload capacity -Poor Robustness against geometric and compression attacks - Tested only on text -Low payload capacity. -The robustness of the system is questionable in a deeper look as the secret data gets scrambled while undergoing any sort of geometrical attacks or lossy compression. -Poor robustness against geometric and compression attacks -Poor Security -Less visual quality

Swain [88]

LSB + PVD

No

-Higher payload capacity. -Better Imperceptibility -Simple extraction process. -No need to original image.

-Poor security against unauthorized attacks. -Poor Robustness against attacks

Kalita and Tuithung [89]

LSB + PVD

No

-High payload capacity. -Simple extraction process. -Using the edge pixels in all directions.

-Poor security -Poor Robustness against attacks

Capacity bits = 199,211.2 PSNR = 40.606

Tai el at. [95]

Histogram Shifting based

Yes

-Unnecessary communication of histogram points can be avoided -Embedding capacity can be varied by tuning the binary tree structure.

-Poor defence against intruder attacks -Low imperceptibility at reasonable payload capacities -Less payload capacity

BPP = 1 PSNR = 30

Nyeem [104]

Bit Plan + Histogram Shifting based

Yes

-Providing higher payload capacity with high imperceptibility among most steganography techniques

-Poor defence against intruder attacks

BPP in the range of 5 at higher PSNR values of above 40

Das et al. [110]

Pixel Intensities Modulation based

Using both LSB substitution and PVD. The image is partitioned into non-overlapping 2 × 2-pixel patches as that in standard PVD and then secret data bits are embedded on the upper-left pixel of every 2 × 2-pixel block using LSB substitution. Then data bits are hidden using the pixel difference in three directions Using both LSB substitution and 8neighbouring PVD for gray scale. .The proposed method partitions the The image is partitioned into non-overlapping 3 × 3-pixel patches as that in standard PVD and then secret data bits are embedded on the upper-left pixel of every 3 × 3-pixel block using LSB substitution. Appling a histogram shifting method which gives an alternative solution for embedding the peak and minimum points in the histogram map by using a binary tree format. In order to avoid overflow and underflow problem, the histogram is tapered at both ends prior to the embedding process. The narrowing range depends on a generated binary tree. An indirect embedding using a pair of bit plane sliced image and histogram shifting. The pixel intensity values are sliced into two based on the bit plane values and histogram shifting based embedding is applied over the histogram bins separately. An intensity adjustment image steganographic is used. The adjacent pixel intensities are adjusted to embed secret bits by keeping minimum changes in the cover media. Here, the cover image is initially divided into 3 × 3-pixel size non-overlapping patches and treated as seed matrices. For embedding secret bits, each seed matrices were considered one by one and the nearby pixel intensities are adjusted as per the secret data bits.

No

-Less distortion -Good imperceptibility

-Lower payload capacity -Poor defence against noise attacks

BPP = 0.88 PSNR up to 40

Maximum text length is 6 characters

BPP = 2 PSNR = 44.53

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

BPP for OTPVD = 2.483 BPP for EOTPVD = 2.733 PSNR for OTPVD = 37.774 PSNR EOTPVD = 37.635 BPP = 3.1 PSNR = 40.4

307

308

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

changes occurred per pixel embedding is very small, multi-layer embedding is possible without compromising the perceptibility of the cover image and error free retrieval of the secret information. Similar to other histogram-based techniques, this is a reversible Stegano system where the cover image can also be extracted along with the secret data. The intensity level segmentation limits the embedding changes in a safe range which yields high image quality of the stego-media. The drawback lies in the payload capacity where, the payload capacity of at most 0.5 bpp is observed while testing with various test images. Also, the scheme cannot be used with compressed stego-image or in communication channel susceptible to geometrical distortions like scaling or with additive noise. Difference expansion steganography: The difference Expansion based steganographic system refers to the secret data embedding over difference pixel pairs. The difference values are expanded using different techniques and secret data bits get embedded over this expanded difference range [97]. Most of the difference expansion techniques came under the reversible stegano systems, where both cover image and secret data can be extracted at the receiver side without errors. Several works in Difference Expansion based image steganographic systems are proposed by various researchers and some of the latest relevant articles are discussed here. In [98], a pre-processing is used to prevent underflow and overflow issues. This helps to avoid the situation where embedded pixel may be of values greater than that of the maximum and minimum ranges in the cover image. The LSB bits of the cover media is saved here to make the system as a reversible one. Huffman coding is applied here to compress the LSB data bit size and is concealed along with the secret data in the embedding phase. For predicting the active embedding portion, GAP [99] method-based prediction scheme is used. The difference image is then generated from the actual cover and predicted one. The secret data is then embedded over the expanded version of the actual difference image and the embedding capacity can be increased or decreased by controlling the expanded range. The more the expanded range, the more will be the embedding rate with a degradation in the image quality. Hence, the expansion range is determined to keep a trade-off between image imperceptibility and payload capacity. A thresholding in the difference image along with the use of flag bits at high difference values helps to limit the errors in an innocuous interval. Another difference expansion based steganographic algorithm is explained by Jung et al. [100]. Here a block level difference expansion based embedding strategy is used along with an interpolation prediction mechanism. The cover image is expanded to multiple scales (> 1) and embedding is attempted on each scale space. At each scale level, the input images are divided into sub blocks and the embedding bits are decided for each expanded sub blocks. The embedding capacity can be adjusted using the key scaling parameter. A payload capacity of ∼4 BPP is observed at a scaling limit of 3 and at a visual quality of ∼30 dB of PSNR. The method suffers all drawbacks of reversible techniques including poor defence against geometrical and statistical attacks. Another high capacity reversible image stegano system is mentioned in the article [101]. The key process here is to find a difference image between the actual cover data and a predicted image based on the reference pixels from the cover media. The secret data gets embedded over this error image and is used as the stego-image for communication. The embedding is limited to the pixel positions where the difference is above a threshold and this control helps to keep the system free from overflow and underflow errors. At the receiver side, based on the pixel values and the pre-determined threshold values, the reference pixels can be identified and the reverse embedding process helps to isolate the secret bits from the stego-image. The main attraction in this scheme is that, the embedding does not need a costly reference peak selection from the histogram map. The embedding rate

can also be reformed as per the needs at a cost of lower visual quality. In total, it is concluded that the difference expanding methods are limited to target applications where cover image is crucial and the communication channel is less susceptible to intruder attacks. Multiple bit-planes based steganography: This technique was first introduced in 2006 as an extension to the basic LSB substitution technique, where bit planes are used to hidden secret data bits [102,103]. Usually, bit plane stegano systems are used along with other methods to boost the performance of the overall system [104]. Hence, most often, it also belongs to other major categories in image steganography. Bit plane segmented steganographic system is plotted in the paper [105]. Here, the complexity of each of the bit planes is assessed initially before the embedding process. Bit planes with higher noise are selected and corresponding Hessenberg Matrix are generated then. Q-R factorization is then applied over this Hessenberg matrix and sub blocks of the secret image get embedded over the Q part of the decomposed matrix. After the embedding process, the decomposed Q and R parts get combined and stego-image is then generated. If the cover image selection is optimum, the bit plane selection will be perfect and secret bits can be embedded without any degradation in the visual quality. The main drawback of this system is that, the image imperceptibility may void if the selection of bit slice is improper. The method also suffers from other demerits associated with average LSB technique. In [106], authors proposed such a method which uses bit planes of pixel intensity values for the secret data wrapping. The first stage is to slice the system into multiple bit planes and the required number of planes is opted using an ANR255 sequence. For maintaining higher security, the secret data is encrypted before the actual embedding process. The method uses a 13-bit plane ANR encoder and the secret data bits are embedded over these expanded bit slices. The cover image after the embedding process is converted back to standard 8-bit representation before using it as the stego-image. The expanded bit plane encoding brings two advantages. One is that, it can host more secret bits than normal 8-bit LSB techniques. Secondly, the degree of randomness of embedding will be high and make the system more robust against intruder’s steganalysis process. Another advantage of the system is that, there is no need for sharing pre-defined dictionary at the sender and receiver side. The main drawback is that, the stego-image is vulnerable against geometrical attacks and even a small change in the pixel alignment may cause the scrambling of embedded data. Palette based steganography: Palette based steganography [107] used palette-based images as the cover media. Image formats such as PNG, GIF and TIFF are suitable for this approach. A secret key is used to generate pseudo random numbers and the selected secret data bit gets engraved on a single cover pixel. Instead of original colour, the colour with the same parity as the secret bit in the palette is used for the embedding process. Generally, palette based stegano systems can be categorized to two: in the first type, the palette colour is altered to make the distortion in a small range. In such systems, the embedding capacity seems less. The other type keeps the colour details of the cover data but modulate the palette entries as per the secret data bits and favours dense embedding. The main attraction of palette-based steganography is that the overall distortion in the stego-image is lesser when compared with other spatial methods. The main disadvantage of this method is the need of images in specific lossless compression formats. The method cannot be applied on common image formats such as JPEG. The approach is less popular due to these drawbacks and used less for real applications. Imaizumi et al. [108] proposed a dense image embedding scheme using palette based steganographic system which keeps the visual quality in a fair level enough for undetectable commu-

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

nication. Multiple bits of the secret data may be embedded here over a single pixel after evaluating the difference using Euclidian distance measures, where most of the palette systems follow one bit per pixel strategy. The scheme uses a parity check for reducing embedding error. One condition for this hiding scheme is to share the embedding location map across the sender and receiver to make sure that, the secret data is retrieved in the correct order. While compared to other palette-based systems, the payload capacity is marginally high and the visual quality is higher around PSNR value of ∼40 dB. In order to improve the security features, random pixel selection is opted for the embedding process [109]. The embedding process starts with the generation of Julia-set fractal images based on the initially given parameters. The colour channels are extracted and then the colour palette is sorted accordingly. The random pixel selection is then employed and the palette index is updated as per the most matching secret bits and stego-image pixel is updated accordingly. The embedding index needs to be shared at the receiver side and the secret data is extracted by comparing the actual index with the embedding index. The method helps to maintain the visual quality in the range of ∼60 dB but all other aspects of steganography are not well considered here. The overall assessment of palette-based steganography is not sufficient while considering advanced LSB steganographic systems. Pixel intensity modulation based steganography: Pixel intensity modulation or adjustment based steganographic systems were introduced as a modification of pixel intensity adjustment embedding as in LSB steganographic type. Here, the secret data bits are embedded in the intensity adjustment between adjacent pixels or of neighbouring blocks which depends on the nature of embedding scheme. These methods help to provide better quality stegoimages while comparing with LSB modification systems due to this indirect embedding process [110]. In order to improve the security, pixel intensity modulation based edge selection is proposed in [111]. In order to increase capacity and perceived quality, another advanced pixel intensity modulation technique is explained [112]. Here, the secret bits are embedded in the host image sub blocks. The sub block averages are updated by smart pixel adjustment based on the secret data bits. Sufficient thresholds are used to keep the adjusted values within the maximum intensity value (overflow). The extraction process is the inverse approach as that is used in the embedding phase. By reviewing these techniques, it can be assessed that the approach is better than an average LSB modulation technique in terms of imperceptibility, security and robustness. But the other drawbacks of LSB systems persist here and research is still needed to tackle these flaws. Quantization based steganography: In this category, the steganographic system uses any sort of compression encoding system to hide secret data bits. The encoding system can be any standard compression codec like JPEG, vector quantization etc. Generally, the secret data is divided into small blocks of data sub samples and these small data fragments are embedded along with the encoded carrier images. At the receiver side, the same coding is applied over the stego-image and the secret data is retrieved back using the inverse embedding process. An image steganographic system where secret data is embedded over jpeg encoder is shown in [113]. JPEG coder uses 8 × 8-pixel blocks for the compression process and the secret data is also divided into small fragments to get hidden over these transformed coefficients. The secret data may need multiple encoder coefficients to host the full fragments. Similarly, the entire secret fragments are embedded on one or more compression coefficients and at the receiver end, the reverse process applied over the stego-image and the secret bits are then combined to generate the original form. The system does not need any information regarding the coefficient location and the embedding location is adaptively selected based on the

309

image statistics. Authors claim that the method is suitable to use with any coders and offers an extra advantage. But the scheme is not sufficient to handle the geometrical attacks and steganalysis. On the part of enhancing the capacity and minimisation of distortion, a lot of improved versions of quantization-based image steganography are discussed in the research field. One of them uses modified DCT quantization table for colour image [114], adaptive hiding steganography based simple optimal quantization [115]. 7.3.2. Transform domain image steganography Transform domain values can also be used for integrating secret bits into a cover image. In transform domain based methods, the secret bits get hidden under the subband frequency coefficients. The process of embedding and decoding is more complicated in the transform domain rather than the techniques that are used in the time domain. This will improve the security of the system. Another advantage is that most of the Frequency domain techniques are less affected by compression, cropping, scaling and rotation attacks. Thus, the transform-based systems are more effective in preserving the stego-image quality and makes less detectable in an unsecured channel. A lot of transform domain methods are used in the field of steganography and the most popular schemes include Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT) and different versions of these basic transforms. Integer Wavelet Transform (IWT), Complex wavelet transforms (CWT), Dual-Tree Complex wavelet transforms (DTCWT), Compressive sensing are examples of such techniques. Some of the relevant research work in this domain are analysed in the following sections. Also, some examples of the research works are explained in Table 4. Discrete fourier transform (DFT) based steganography: DFT is one of the popular transform technique used in signal processing. 2-D Fourier transform has been defined for images and is widely used in various image processing techniques [116]. Using DFT, an image can be decomposed into corresponding frequency components (sines and cosines). In image steganography, these transform coefficients can also be modulated as per the secret information and thus can be used as a significant tool in image data hiding. The modulated transform coefficients are then converted back to image form and can be used as the stego-image. At the receiver side, the stego-image is again decomposed into frequency elements and hidden data can be extracted from there. Various image systems have been proposed in the field of DFT steganography [117–119]. A modified Fourier Transform technique-based steganography; Discrete Fractional Fourier Transform (DFrFT) was proposed in [118]. In DFrFT, both the time and frequency information lay in the transform coefficients and these coefficients house the secret data bits. Here, the embedding bits are hosted over the coefficients using LSB replacement technique. The authors claim that the method yields good visual quality by adjusting the order of DFrFT transform. No steps are taken here to enhance the security or robustness of the system and there is the possibility of higher retrieval error due to quantization of forwarding and inverse transforms. Discrete cosine transform (DCT) based steganography: DCT is one of the most common and effective image transform techniques to convert an image from spatial domain to its frequency domain. In basic DCT based steganography, the DCT coefficients are modified as per the secret data bits. In DCT steganography, the image gets separated into its corresponding high, middle and low frequency components. The most important details lie in the low frequency subbands and high-fidelity details lie in the HF bands. Secret data bits hiding in DCT coefficients follow the JPEG compression model. An image steganographic system where secret data is embedded over jpeg encoder is shown in the paper [35]. The

310

Table 4 Review of main transform domain image steganography methods. Classification Method

Method Description

Reversibility

Advantage

Disadvantage

Results

Khashandarag et al. [117]

DFT

No

-Including some cryptography by using LFSR and DES -High speed -Better capacity

-High complexity -Robustness and imperceptibility are not considered well. -Not test against attacks

Hardware application

Savithri et al. [120]

DCT

No

-High security, Including two cryptography methods -Reducing time of processing because of using parallelized version

Saidi et al. [125]

DCT

No

-Improving the imperceptibility of by incorporating the map

-Lower payload capacity. -Lower image quality -Poor robustness against compression and geometric attacks -Stego-image quality is not satisfied -Poor Robustness against geometric and compression attacks

BPP = 0.5 PSNR = up to 25 The ratios of speed-up of 1.6 and 3.18 Capacity bits = 196,608 PSNR up to 30

Kumar and Kumar [126]

DWT

No

-Enhancing the security by using a secret key -Better visual quality

-The payload capacity is limited in this approach to provide better imperceptibility for the system.

BPP = 2 PSNR = 45.34

Divya and Sasirekha [127]

DWT

DFT based stegano system is proposed. The secret data is encoded first using LZW compression algorithm and these encoded bits are logically XOR with pseudo random numbers to make them encrypted. DES algorithm is used on the encrypted secret data blocks to enhance the security. The cover image planes are transformed to DFT a Coefficients selection and frequency hopping algorithm (CSFH) is used to randomly select the required transform coefficients. The processed secret bits get embedded over these selected coefficients and then inverse DFT is applied to bring the embedded values to the actual image format. Two parallel techniques ( 2D-DCT with Rivest Shamir Adleman) and (2D-DCT with chaotic) are proposed. Before embedded into DCT coefficient, of the larger cover image, Duffing map the secret image is used for encrypting into a set of uniform pixels The approach used both DCT domain subbands for the embedding process and chaotic map. Modifications of the elements in high frequency sub image will not result in considerable changes in the perceptibility. The method applies DCT on the cover image and scans the high frequency coefficients in a zigzag manner and then embedding positions are decided using a chaotic function. A DWT based steganographic is applied. Using three details coefficient (horizontal, vertical and diagonal) for embedding the secret data. Using a secret key computation for minimizing the distortion in the host image. Also, a blocking concept is employed to retain the imperceptibility of the cover image. The embedding procedure using a matching block between the cover and secret image. Modifying steganography based on DWT with satisfactory levels of distortion in the host image along with high level data security. using Haar-wavelet for reducing data loss in the cover image

No

-Better security

-Low imperceptibility -Lower payload capacity -Poor Robustness against geometric attacks

BPP = 0.9 PSNR = 29.85

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

References

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

cover image in steganography based on DCT is divided into nonoverlapping of 8 × 8 pixel blocks. Then, by using Eq. 8 each block is transformed into DCT coefficients. The quantization table is used for quantizing DCT coefficients. The quantization DCT coefficients embed secret data which are afterward coded by utilizing combination coding. Then, the stego-media is intended to share through an unsecured communication channel.

F (k, l ) = ak al

M−1

N−1

f (x, y ) cos

π (2x + 1)k

x=0 y=0

Where :



ak= √

al =

2M

cos

π ( 2 y + )l 2N

(8)



(1/M ) f or k = 0 (1/M ) f or 1 ≤ k ≤ M − 1





(1/N ) f or l = 0 (1/M ) f or 1 ≤ l ≤ M − 1

Where F(k, l) is the DCT transformed image of position (k, l), M and N are row and colum of image f(x, y), ak and al are scalars. Different steganographic approaches use different styles of embedding schemes depending on the requirements [73,120–124]. The method utilizes the fact that many DCT quantized coefficients are near to zero [124]. In this scheme, two neighbour coefficients whose values near to zero as a group are selected in each block and then data bits get hidden into each group. Here the selected coefficient is modified by just one unit so that the stego-image can retain a high image quality. This approach achieves both image quality and high embedding capacity in the stego-image. But the method may not work well for certain images since the coefficient distribution of some images may be randomly scattered and has bad grouping outcomes which may degrade the embedding capacity. So, a more complicated hiding strategy is required for this scheme by evaluating the coefficient distribution so as to pursue high efficient performances. A modified transform domain irreversible using DCT steganography is plotted in [123], where the authors proposed an image steganography technique which provides high embedding performance while introducing minimal fluctuations in the cover carrier image. The method uses modulus 3 of the difference between two DCT coefficients for embedding two bits of the secret text message in the compressed form. For getting better results, the confidential data is subjected to two phases of compression stages before embedding. In the first stage the message gets compressed by removing some words and replacing some expressions with their commonly used abbreviations. In the second phase the resultant compressed message is again compressed using Huffman lossless compression technique. Finally, the compressed version of the secret message is implanted into the cover image based on the modulus three of the difference between DCT coefficients of the cover image while saving it as JPEG image. The embedding capacity is high but the imperceptibility is low due to dense embedding process. The scheme also did not use any specific strategy to override the geometrical attacks on the stego-image. In order to enhance embedding capacity and perceptibility, Tamer and Ibrahim [71] proposed colour image steganography based on a global adaptive block in DCT. For enhancing the security and imperceptibility, DCT based steganography with the chaotic map is proposed in [125]. As we discussed in the previous section, DCT transform represents the changes of intensity values in the horizontal and vertical direction of each block in the image. The basis function used for this conversion is too long and it may lead to higher quantization error during the forward and backward transformation. These issues can affect the accuracy while using it in steganography. So steganography is further expanded using better frequency transformations like Wavelet transform. While comparing with DCT,

311

wavelet transform possesses many advantages and a detailed review of steganographic methods using different versions of DWT is provided in the next section. Discrete wavelet transform (DWT) based steganography: Wavelet transform of digital images provides both time and frequency localization. DWT involves a series of filtering and down sampling to find the subband images. Based on different wavelet’s nature, DWT can be applied for image steganographic applications. In DWT based image steganography methods, generally the DWT of cover image is taken and the coefficient space is modified as per the secret message bits. The main reason for the primacy of DWT over DCT and other similar transforms is the mode of projection of high and low frequency details. Shorter basis functions are used for projecting high fidelity details and large basis vectors for low frequency details to provide optimized space-frequency localization. Depending on the decomposition levels and nature of wavelet used, the DWT based steganography provides a number of embedding options as per the requirement. Since the coefficients represent both spatial and frequency features of the image efficient embedding schemes can provide better image imperceptibility and non-recoverability of the hidden data. Because of these benefits DWT based embedding schemes are the lion share in the transform domain steganography and some of them are introduced here [126–129]. In basic DWT systems, the lack of intelligent embedding approaches possesses a barrier to achieve maximum possible accuracy. Therefore, several modifications of wavelet transform systems came to the scene to improve the performance of embedding systems. Samer et al. [130] discussed a similar kind of modified DWT scheme for embedding secret bits in between wavelet coefficients. Diamond Encoding (DE) in DWT is proposed here to improve the security and reduce the image distortion in contrast with standard DWT approach. The DE-DWT algorithm is used here to transform the secret message into an array of base-5 digits. Standard Wavelet coefficient subbands of the cover image get divided into blocks and the secret array bits are engraved onto these cover image subbands coefficient pairs. Reduction of Stego-image distortion is the main advantage compared to similar DWT based embedding schemes. The main drawback of the proposed algorithm is the limited embedding capacity. Another important flaw is that the stego-image becomes useless if it is compressed with a factor less than 95%. For reducing some major flaws of standard DWT, Redundant DWT (RDWT) and QR factorization are presented here [131]. RDWT is used by avoiding the decimation step as that in normal wavelet. In order to improve the security, pre-processing step of an encrypting algorithm (Arnold transformation) is used on the secret image. A selection of cover image location is also opted here for better embedding process. The main disadvantages of this method include the inferior defence against intruder attacks and image compression and poor payload capacity. Retrieval error rate is quite high as compared to other existing techniques. Conventional Wavelet transforms are seen with fractional values and need to be pre-processed to make them in the integer format since most of the embedding techniques use integer values for data hiding. This truncation from fractions to an integer will not be the main issue in irreversible type steganography. But such rounding needs to be considered in a reversible type steganography, where both career data and secret data need to be recovered without any errors. This can be accomplished by involving Integer Wavelet Transform (IWT). More details will be provided in the subsequent section. Integer wavelet transform (IWT) based steganography: IWT is an invertible integer-to-integer wavelet analysis algorithm. IWT is a modified representation of wavelet transform, where image details can be represented in multi resolution characteristics. IWT offers a technique to provide a lossless decomposition of image

312

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

pixels at different resolutions and sampling spaces. By employing lifting schemes with floor and ceiling functions, integer coefficients will be obtained without quantization and hence reversible operation is possible. The predicted and updated coefficients in IWT are analogous to the low and high frequency components in discrete wavelet transform. IWT is widely used in decomposing high fidelity large size images with error free compression features. Similar to other transform coefficients, IWT coefficients are also used to hide secret data bits in steganographic systems [132–135]. An IWT based steganography is proposed by Miri and Faez [136]. Here, the cover image is transformed using IWT and the edge coefficients in the subband images are grouped by checking their MSBs. Embedding is in such a way to prevent alterations in MSB so the receiver could extract the information with minimal error. The proposed algorithm has better PSNR as compared to some other wavelet-based algorithms, which can be considered as an advantage of the scheme. Imperceptibility and ability to reduce the retrieval error rate are the main attraction of this scheme. Defensive steps to prevent stego-attacks and performance against image processing operations such as rotation, compression, and scaling are also not addressed here. Two popular methods used in IWT coefficients are Histogram shifting and compression techniques and favours reversible property. In the latter type schemes, the high frequency subbands are replaced by the compressed cover and secret data. Another version of IWT based image hiding scheme is proposed by Shet et al. [137], where the YCbCr colour space of the cover image is used for the embedding process. Since the Cb and Cr components represent the chromatic details and Y represents the intensity features, the idea is to embed the secret bits over the components other than intensity plane. This helps to reduce the detection of embedded data in visual checks. The chroma planes are passed through the lossless IWT transform. The binarized secret bits are then concealed under these Wavelet coefficients using LSB substitution. The embedded coefficients are then converted to the corresponding chroma components of the image and then restore it to the RGB format and used as the stego-image. The reverse process at the receiver side yields the secret data. The LSB method used here for embedding helps to reduce the quantization error of IWT coefficients and hence retrieval error can be limited to a small value. The imperceptibility is good as it helps to provide a visual quality of ∼40 dB for an HD image. But all other aspects of steganography like payload capacity, robustness and security are not properly addressed here. Certain approaches attempted with a combination of multiple transforms during the embedding process and gave rise to higher security and imperceptibility [138]. But, they may be slow processes while compared to normal single transform techniques. Based on the review on DWT and IWT based steganography, we could see the common disadvantage is the poor payload capacity and inferior defence over geometrical attacks. Also, normally, DWT is used in most of the steganography technique, where the real coefficient filters are used which results in the real valued approximation. Thus the DWT eliminate the local phase information. In order to evade these flaws, expanded versions of DWT were proposed and the most popular model among those is the Complex Wavelet Transforms (CWT). A few modalities of CWT based steganographic systems are mentioned below. Complex wavelet transform (CWT) based steganography: CWT is an updated version of basic DWT. Here the coefficients are of the complex in nature. It is another updated version of 2D DWT which derives multi-resolution beneficial features from the image. Since the complex representation of wavelet coefficients provides both real and imaginary coefficients, CWT includes the phase information of the image. CWT is a core research area in image watermarking and steganography. There are many basic CWT based research work has been proposed [139–144]. Most of such work failed to

exploit the benefits of additional coefficient planes of CWT. Further, it is holding another advantage of keeping a high degree of shiftinvariance. Thus, most of the highly cited work uses hybrid algorithms to enhance the performance. Such a method is explained by Singh et al. [143]. They presented a data hiding scheme based on CWT, Chaotic sequence and SVD. The researchers claim the system is robust against geometrical attacks due to the shift invariance property of CWT transform. Firstly, the secret data is converted to its SVD coefficients and the most relevant coefficients are embedded over the LF sub images of the CWT coefficients. For increasing the security, the chaotic sequence is used for scrambling secret data. The main drawback here is that the accuracy of decryption reduces for higher capacity and poor defence against tampering. Dual-Tree complex wavelet transform (DT-CWT) based steganography: DT-CWT became one of the most popular tools in signal and image processing such as image compression [145], restoration and enhancement [146], classification [147] etc. As the hiding data requirements increase, researchers sought for more stable transforms with multiple transform planes. Then DTCWT was proposed. The main difference of DTCWT from the standard CWT is that the former uses dual tree complex wavelet decomposition. The dual tree structure is realized by using two sets of filters and multiple levels of subband images. Here, in DTCWT transformation, the subband matrix size will be reduced into half of the previous level subbands. At the same time, the number of subbands in the level become double than that of the previous level. In total, the number of coefficient elements will be same in each level. So, in a generic manner, it can be stated that, increasing the number of levels will results in the number of transform domain sub-space coefficients and in turn it will result in an advantage for hosting more number of secret bits and can be considered as one of the main highlights in DTCWT based steganography. However, embedding secret data in higher level coefficients may degrade the stego-image quality as the distortion due to higher level embedding is higher than in the lower level. In each level, there are coefficients corresponding to frequency changes in horizontal, vertical and diagonal directions. In each direction, there will be a real part and imaginary part coefficient and give more housing to the secret data. While compared to standard CWT, DTCWT uses two sets of filter coefficients. This results in doubling the total number of subband images and the number of elements. This improves the carrier capacity. Also, these subbands show a nearly shift invariance, good directional selective and perfect reconstruction property of DTCWT which enhances robustness against different geometrical attacks. A basic construction of DT-CWT transform subbands is presented in papers [145,148]. Some examples of watermarking and steganographic [149,150]. Most of the surveyed algorithms used DT-CWT for watermarking as a binary image is embedded in the cover images. These algorithms can be divided into two groups based on the manner in which the logo is embedded; Wavelet Based Fusion and algorithms based on Singular Value Decomposition (SVD). Most CWT and DT-CWT algorithms belong to Wavelet Based Fusion, where the CW T /DT-CW T is applied on both of the watermark/secret and the cover image. Then, the fusion of the coefficients of the watermark/secret and the cover image is performed using one of the following equations in Table 5. In algorithms based on SVD, the CW T or DT-CW T is applied on the cover image, X, and then the SVD is applied on the LL subband, XLL :

XLL = UXLL · SXLL · VXLL

(15)

Then the watermark, w, is imposed in the matrix ‘S’ using the following equation:

SX LL = SXLL + k · w

(16)

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

313

Table 5 Embedding equations for most CW T/DT-CW T based research. No.

Equation

Reference

Notes

(9)

f  (x, y ) = f (x, y ) + α · w (x, y )

[151]

The original DT-CWT coefficients The modified DT-CWT coefficients

(10)

f  (x, y ) = f (x, y ) + g(x, y ) · w (x, y )

f(x, y) f (x, y)

[152,153]

α

(11)

f  (x, y ) = f (x, y ) · (1 + α · w (x, y ))

[154–157]

(12) (13)

f  (x, y ) = f (x, y ) + α · w (x, y ) · M (x, y ) f  (x, y ) = f (x, y ) + α · w (x, y ) · J N D (x, y )

[158] [159]

(14)

f  (x, y ) = Q ( f (x, y ) + w (x, y )) − w (x, y )

[160]

Embedding strength factor. The watermark (a pseudorandom image modulated based on the logo to be embedded). A gain function depends on average squared magnitude of the CWT coefficients in a 3 × 3 neighbourhood centred at location (x, y). Perceptual mask Just Noticeable Difference scaling factor Quantizer

w (x, y) g(x, y)

M(x, y) JND(x, y) Q

where, k is the gain factor which represents the strength of the embedding process. Then modified X LL subband is calculated:  = U · S · V  XLL XLL XLL XLL

(17)

Finally, the processed watermark image, Y, is obtained using the modified X LL . At the receiver side, the watermark, w, is haul out using the following equation:

w=

SYLL − SXLL k

(18)

It means that SXLL should be available at the receiver end [161– 164]. Sending such a large matrix (SXLL has the same size as XLL ) along with the cover image is not practical in some applications. Most of such work failed to exploit the benefits of additional coefficient planes of DT-CWT. Thus, most of the advanced approaches use hybrid algorithms to enhance the performance as explained above. The main drawback of the research work above can be summarized as: DTCWT coefficients are changed slightly during the forward and inverse transforms, the embedded data cannot be extracted exactly [153], the issue of the lost bits can be solved using error checking and correcting codes such as Hadamard [153]. Also, selection of robust coefficients is needed to tackle this issue. Compressive sensing (CS) based steganography: CS is a new, innovative technique of signal and image processing that helps in accurately acquiring and reconstructing the image by using underdetermined linear systems. One of the major benefits of using CS is its “exploitation of signal sparsity”. Previously, the major technique used for signal processing was Shanon’s theorem that states that “the sampling rate must be at least twice the maximum frequency in the signal”. Here the sampling rate is also known as Nyquist rate. So the primary main advantage of using CS is to work upon fewer sampling rate and reconstruct accurately. By the conventional way, the data is initially sampled and then the data is compressed by some way or other so that the data can be represented by few coefficients without much loss so that reconstruction of actual image maybe possible. CS, which is also called compressive sampling or sparse sampling focuses on the principle of obtaining fewer samples compared to the number of measurements required by Shannon’s theory and reconstructing the image by those samples [165,166]. There are two main working aspects of CS. The first one is sparsity and the second is incoherence [167]. Sparsity refers to the signal of interest, here the main focus lies on the “information rate” of signal or image under observation, which maybe smaller than the bandwidth suggested and the incoherence is sensing modality used in the system that works on the relationship of two elements. The theory of compressive sensing demonstrates the relationship that is as under:

y = x

(19)

where y is the observed data or compressed data y ∈ Rm , x consider actual image x ∈ Rn , the  denote measurement or sampling matrix  ∈ Rm × n . Exact recovery is achieved by solving the above linear system, x should be compressible or sparse in some transform domain such as (Wavelet, Discrete Cosine Transform (DCT), ect). It became a surprise that given certain assumptions, it is actually possible to do image reconstruction when m  n, there exists even more efficient reconstruction algorithms to do this. Under this class, the secret data values are first converted to compression coefficients before the actual steganography process. This will help to reduce the net payload required for the embedding process and will result in a comparatively small change in the stego-image. Eventually, the degradation in the visual quality will be less as that of the direct embedding process. Sreedhanya and Soman [168] proposed a secure system by combing encryption data based CS and SVD based steganography. For improving image steganography in term of imperceptibility and capacity, a modified version of SVD based compressive sensing is explained in the paper [169], where both secret data and the cover image is used in the transform domain. In this method, The SVD coefficients are embedded in the lower subbands of the framelet transform of the cover image. For providing high security, compressive sensing based on HVS is presented here [170]. The main advantage of using compressive sensing is not just limited to payload concerns. The apt compression schemes over the secret data help to increase the security as well as improve the robustness against attacks such as noise additions in communication channel [171–174]. For a good CS based system, the compression method will be as according to the nature of secret data; whether it is a text, image audio etc. The main drawback of CS is that sampling matrix always has a large size and it occupies a large space, resulting in a lot of annoyance when sharing and communication. The issue of CS can be solved by combining CS and encryption, researches in this filed are still investigated such measurement matrices [171–173]. 7.4. Adaptive steganography While reviewing the research contributions in image steganography, we cannot skip the importance of adaptive methods. It is a special case of spatial and transforms methods. Adaptive steganography is also called “Model-Based” [175] or “Statistics-aware embedding” or “Masking” [1,20]. Many of the cores work in steganography lies under this section. Adaptive nature can be introduced in the data embedding schemes in a number of ways: in selecting the target pixels in the cover image, nature of modification to be made, number of bits embedded in a pixel etc. Based on the nature and mode of adaptiveness offered by these systems, they can be categorized into a few sections. This includes Region based Steganography, Human Visual System (HVS) based

314

Table 6 Review of main adaptive image steganography methods. Classification Method

Method description

Reversibility

Advantage

Disadvantage

Results

Rabie and Kame [71]

Region based -Transform domain-DCT

No

-Higher payload capacity among most stegano techniques -Acceptable imperceptibility

-Poor security -Did not verified against image processing manipulation

Best result was BPP at 17.25 PSNR at 32.2

Qin et al. [93]

Region basedSpatial domainHistogram Shift

Yes

-The retrieval accuracy is guaranteed here since, the reference points can be correctly identified at the receiver side -Higher imperceptibility and error free secret data removal.

-Lower payload capacity -Poor robustness against geometric and compression attacks

Hamid et al. [176]

Region based-Transform domain-DWT

Colour image steganography based on DCT and a global adaptive-region (GAR) is realized. This helps to obtain high embedding capacities. the similarity of the cover image frequency coefficients and secret data is analysed and data hiding in most correlation block of DCT Data hiding based on Histogram shift. The method used an image inpainting based prediction model. The reference pixels are determined adaptively based on the image characteristics. Less number of reference points gets selected from smooth regions and a high number of reference pixels are chosen from the high-fidelity regions in the image area. The secret information is then hide reversibly over the histogram of the predicted error image. Speeded Up Robust Features (SURF) for detecting the robust regions from the cover image is used. The wavelet coefficients from the selected region are then modulated as per the secret data to complete the stegano process.

No

-The embedding capacity is not satisfied

Hamid et al. [177]

Region based-Transform domain -DWT

No

Luo et al. [189]

HVS based-Spatail domain

Atee et al. [229]

ML- Transform domain -DWT

Miri and Faez [237]

AI-Transform domain -DWT-GA

Two trustworthy attempts are plotted to detect and represent key interest points from an image: Scale Invariant Feature Transform (SIFT) and Speeded Up Robust Features (SURF). The performance of these techniques is evaluated while using these algorithms based on the performance against different types of attacks like different compression schemes, Noise addition etc. Authors claim that both methods provide similar visual quality and retrieval accuracy. The least sensitive area in the cover image is calculated using a predictor function. A ‘Just Noticeable Difference index’ (JND) is defined here corresponding to each pixel which indicates the sensitivity to notice changes while embedding in that pixel. A prediction error Histogram (PEH) method is selected here to embed the secret data bits over the cover image pixels based on the just noticeable difference index (JND). Extreme learning machine (ELM) is proposed. In their method, the cover image gets divided into non-overlapping blocks, and they extracted some texture features from every single block. The extracted data will be used as the input data during training. Then some testing data is embedded (using standard DWT) in those blocks. Adaptive wavelet and genetic algorithm are applied. The similarity of the cover image frequency coefficients and secret data is analysed using a genetic algorithm and the data will be mapped to the closely resembling areas of the frequency space. The genetic algorithm also helps to embed the secret bits over the coefficients in the wavelet coefficients in such a way that the corresponding edge portions in the spatial domain get changed.

-The selection of regions using SURF key points is better since the points can be identified even if the stego-image is manipulated by geometrical attacks. -High Imperceptibility -Low decoding error rate even if compression or noise is added -The SURF based systems have an advantage of less computation time and more robustness against JPEG compression in high DWT level

Embedding rate = 0.065 BPP at PSNR = 65 for medical image Embedding rate = 0.17 BPP at PSNR = 50 for Lena image BPP = 0.37 PSNR = 45

-The embedding capacity is not satisfied

For SURF BPP = 0.42 PSNR = 44 For SITF BPP = 0.32 PSNR = 45

BPP = 0.6 PSNR = 40

Yes

-This technique helps for dense embedding over the least noticeable pixels and minimal bit embedding over sensitive regions without crating noticeable distortion in the stego-image.

- The embedding capacity is not satisfied -Poor robustness against statistical attacks

No

-Increasing robust system by finding robust region for embedding -High imperceptibility -High similarity index

-Low payload capacity

No

-Reduce the distortions in the smooth -Improve the image quality -Using the non-linear embedding mechanism also improves the overall security of the system

-An intruder can hack most of the secret data by just searching over the edge regions of the stego-image -No specific algorithms are used to prevent the data loss due to geometric or compression attacks.

BPP = 2 PSNR = 55

(continued on next page)

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Reference

Table 6 (continued) Classification Method

Method description

Reversibility

Advantage

Disadvantage

Uma Maheswari [233]

AI-Transform domain-GA-PSO

No

Capacity bit 902,136 PSNR = 53

AI- Spatial domain-LSB- GA

-High complexity

BPP = 2 PSNR 53.11 dB

Bandyopadhyay [236]

AI-Spatial domainLSB- GA

Genetic algorithm based on a pixel modification algorithm is projected to reduce the difference error between the cover and stego-image while secret bit embedded randomly based key chaotic scheme using LSB method.

No

-Poor robustness against geometric attacks

BPP = 2 PSNR 48

Dadgostar and Alfrisi [82]

AI- Spatial domainFuzzy -LSB

No

-Did not tested against different attacks

BPP = 2.68 PSNR = 46

Kaur and Juneja [238]

AI-Spatial domain-Fuzzy

Shafi et al. [239]

AI-Transform domain- IWTFuzzy

Alam et al. [243]

AI-Spatial domainLSB-ANN

Zear et al. [246]

AI-Transform domain-DWT-ANN

Islam et al. [245]

AI-Transform domain-LWT-ANN

Fuzzy based LSB modification algorithm is used. Using of the interval valued intuitionistic fuzzy edge detecting with a MLSB substitution for reducing the distortions in the image quality. This helps to hold more bits over the edge regions and less number of bits in the smooth regions of the cover image which intern reduces the degradation in imperceptibility. Dynamic fuzzifier based block edges algorithm is used. Using of the dynamic fuzzifier for the segmentation cover and secret image then evaluating the compatibility of the segmentation in order to robust embedding progress. This helps to hold more bits over the block edge regions of the cover image which intern reduces the degradation in imperceptibility. Fuzzy based on integer wavelet is proposed, the pre-processing of cover image is applied by fuzzification of the cover image and histogram modification. The image is partitioned into non-overlapping 8 × 8-pixel patches and the bits are embedding over the transform coefficients. In order to increase capacity, a bit reduction procedure is used in each secret data bytes and an optimum pixel modification scheme is used in the text data embedded process. LBS steganography based ANN and the chaotic edge is proposed. Firstly, using ANN for finding the edge of the cover image then secret bit embedded randomly based key chaotic scheme. This paper proposed an image watermarking system based on hybrid algorithms (DWT, SVD and NN). Here, the cover image is initially decomposed into a 3-level DWT. SVD is used to transform low frequency subband and then logo data bits are embedded on the S component. For extraction watermark stage , back propagation neural network is used in order to robust against different noise attacks. Here proposed image watermarking based on lifting wavelet transform and NN. The cover image is decomposed into 3-level LWT and then it is randomized in 2 × 2 non-overlapping blocks. Binary data bits are encrypted and then embedded on the LWT coefficient component. For extraction watermark stage , back propagation neural network is used in order to robust against different attacks

-Good payload capacity -Higher imperceptibility -Enhancing the security by scrambling secret data -Highly Imperceptible -Including some cryptography by using the security key -Robustness against histogram attack -Highly -Imperceptible Including some cryptography by using the security key -Robustness against histogram and noise attack -High imperceptibility -High embedding capacity

-High complexity

Shah and Bichkar [234]

Genetic algorithm-based transform domain scheme is proposed. where the transform domain (Fresnelet and Contourlet ) coefficients are selected using both a genetic algorithm and particle swarm optimization. A genetic scheme for detecting the optimum regions from the cover image is used then secret bit embedded randomly based key using steganography based on LSB replacement.

Yes

Results

No

-Higher imperceptibility

-The issue of the capacity has not addressed

PSNR 46.8

No

-Enhancing the security by using a secret key -High visual quality

-Poor robustness against geometric (e.g. rotation and cropping) -Only tested text into an image

Payload capacity rate 20 to 60%. PSNR ranges from 37 to 50 dB

No

- Enhancing the security by including some cryptography -High imperceptibility

BPP = 2 PSNR = 54.5

Yes

-Robustness against noise -Acceptable visual quality

-The payload capacity is limited in the edge region -Did not tested against different attacks -The payload capacity is limited in this approach to provide better imperceptibility for social applications

Yes

-Enhancing the security by using diverse keys for randomizing the blocks and LWT coefficients robustness against noise attacks

-Poor robustness against rotation and compression - less payload capacity

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Reference

maximum PSNR = 36.26 dB

PSNR = 43.8 Capacity = 512 bits

315

316

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Steganography, Machine Learning and Artificial Intelligence Techniques based Steganography. Some of the relevant research work in this classification are analysed in the following sections. Also, some examples of the research works are explained in Table 6. 7.4.1. Region based steganography The real motivation behind steganography is the data sharing in an undetectable manner. This approach uses the most appropriate regions or features in the cover image for the embedding process [176–178]. In the region-based technique, the embedding is concentrated on portions where the texture and edge details are high rather than smooth regions [71,79,80,179–181]. This helps to defend the degradation of image quality since changes in textured and edge regions are not easily detectable as in smooth regions. Robustness and higher imperceptibility are the main advantages of these types of systems while poor embedding rate is the main disadvantage. Researchers tried to enhance the imperceptibility of the steganographic system using different directions. Some examples of those directions are: local complexity measure based on image segmentation that can be used to find “informative” and “noiselike” areas, the embedding is concentrated on the noiselike areas [182], Threshold value that can be used to select the high frequency pixels of cover and then LSBM Revisited algorithm is used to conceal secret data [183], and controlling hiding capacity based on image characteristics [184]. Ant colony optimization is used to discover a complex area of the cover image and then, LSB method is used to conceal secret data in the detected area [185]. 7.4.2. Human vision system (HVS) based steganography Generally speaking, HVS is an optical information processing structure, which can observe, process and comprehend optical signals. HVS based steganography utilizes the way of perceiving an image by a human being. Human vision has many dark spots and difficulties in identifying image details in certain backgrounds. HVS is insensitive to small changes in intensity in a smooth area, and even indistinguishable if dense embedding is applied over complex texture regions and edge areas in the images [186,187]. Hence these techniques usually opt for procedures to detect such target regions from the image. The actual embedding scheme for hiding the secret data may be any of the spatial or transform based techniques as explained in the previous sections. Researchers tried to enhance the imperceptibility of the steganographic system using different directions. Some examples of those directions are: quad tree segmentation method that can be used to find smooth regions suitable for hiding data with low image degradation [188], the predictor function which used to find the least sensitive regions on the image. A just noticeable difference index (JND) is defined here as one corresponding to each pixel which indicates the sensitivity to notice changes while embedding in that pixel [189], a compressive sensing technology along with HVS based region selection in order to improve the security of the embedding process [170]. Image steganographic schemes under this category will help to adjust the embedding density without degrading the stego-image quality. Hence it is very difficult to detect the embedding process by simply seeing the stego-image by a human intruder. Thus, the chance for efficient recovering is much high while compared to many of the standard image steganographic systems. However, the embedding can be detected by using statistical measures and more care need to be given for enhancing the security and robustness of the data hiding system. 7.4.3. Machine learning and artificial intelligence techniques based steganography Machine Learning (ML) based artificial intelligence (AI) became one of the most popular tools in image processing from 20 0 0

onwards [190–196]. In recent years, wide level research in this field results in many advanced machine learning techniques suitable for various kind of applications on data types like images, texts, speech etc. In image processing, machine learning based AI was initially introduced for recognition, optimization and object retrieval purposes [197–212]. Later, it is used for other requirements in image processing applications like enhancement, compression, segmentation, classification etc. [213–227]. For standard image steganographic techniques, the embedding efficiency will be high if the embedding creates less distortion on the stego-image and embedding capacity is high, if the retrieval error is minimum and the secret data is impossible to hack by an intruder. In machine learning based steganographic techniques, advanced machine learning methods are employed to get the abovementioned efficiency enhancement [228–230]. Popular machine learning systems used in the steganographic area include, Support Vector Machine (SVM) [231,232], Decision-tree based analysis, Genetic Algorithm (GA) [233–237], Fuzzy Logic (FL) optimization techniques [82,238–242] and Neural Networks (NN) [243–246]. All of these attempts make use of artificial intelligence techniques for pre-processing and/or for defining embedding nature and/or for limiting or extending the embedding stages or extracting process. Researches under this classification have developed the efficiency of the steganographic system using different directions. Some of the relevant research works in this classification are analysed in the following sections. Support vector machine based steganography scheme: SVM is supervised learning system and can be used as a classifier in multiple aspects in image watermarking in both spatial and transform domain [232,247]. Also, the authors in [248] guided the way in utilizing SVM for optimizing the steganography based techniques. It is suited for finding embedding locations with high security, imperceptibility while providing minimum retrieval error using faster computational stages [249]. This approach is often used for the localizing the less sensitive areas (for human eyes) in the image channels. Usually these method is employed in blue channel in RGB colour image and luminance channel in other colour formats [250]. Several features like average pixel intensity, chroma value shifts, localization of edge details are fed to make the decision. Based on the used features and classification requirements, various kernels are used to expand the feature details to a higher feature space. Radial basis function (RBF), Polynomial function, Quadratic function etc. are some commonly used kernels to achieve this purpose [251–253]. Genetic algorithm based steganography scheme: The basic genetic processes; selection, crossover and mutation are the fundamentals of Genetic Algorithm (GA). These basic operations are employed to choose the best offspring to pass through future generations. By using such ideas in selecting target coefficients in image steganography, the process will be optimized in several ways. The apt coefficient selection favours high payload capacity and it is one of the best solutions for getting apt hidden regions in the cover media. In an approach described by Chang et al. [254]., they tried to optimize the payload capacity by employing a 2-level quantization process where three bits are additionally embedded over image blocks with high fidelity. In a subsequent work [255], they upgraded the methodology by using a common bitmap generation for improving both the payload capacity and the embedding time. The approach in GA is not just a point pixel-based decision making but works with regional pixel populations and no derivative information is required. Thus, GA possesses much robustness against image manipulations. But finding the local optimal solution can be obtained only after a large number of computational stages and results in poor search speed in finding precise embedding locations.

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Fuzzy logic based steganography scheme: Fuzzy logic-based system concentrates more towards the preservation of visual quality to improve the imperceptibility of the stego-media at a higher cost of modelling complexity. Fuzzy approaches have been implemented in various ways. For example, in a work [242], a Fuzzy Inference System (FIS) with HVS is used for the decision-making step from the local statistical, texture and brightness informationbased feature vectors. The method makes use of these features from cover image sub regions and will define the semantic rules required for the embedding process. The concept helps to reduce the stegano image distortions even at higher embedding rates. Another fuzzy based approach is explained in the paper [256]. Here, the cover pixel selection is based on fuzzy pixel classification and secret message is converted to a mode of fuzzy data before the actual embedding process. Another fuzzy based water marking technique is explained by Kiani et al. [257] . This work uses fuzzy-c means clustering algorithm based on the transform domain derivative features in different directions. Fuzzy logic is able to enhance the steganographic schemes in different aspects especially when they are vague and/or ambiguous image textures. It favours the system by recognizing the suitable image patterns in a fast way by reducing irreverent complexities. This will eventually help practical applications using proper imperceptibility. Neural network based steganography scheme: The neural network-based stegano systems focus on the system robustness and imperceptibility of the schemes by analysing the image details in a through manner. Most of such techniques make use of a back-propagation algorithm for getting the appropriate embedding locations or it is incorporated in the secret extraction process [243,244,258]. In certain techniques, the detection of coefficients in the cover media makes use of the human visual system (HVS). In an article [259], Lou et al. used features from an HVS model to study the characteristics of the cover image. Multiple features like frequency content, entropy, luminance and texture details are used here to make the decision. Another two-stage neural network based stegano system is explained in the paper [260]. The former phase is for the secret data compression and artificial neural network is used for this purpose. In the second phase, a LSB embedding scheme is utilized for hiding the compressed secret data bits. In general speaking, NN is superior in its classification capabilities. Their decision making is based on non-linear and adaptive stages and hence useful in different scenarios. However, the ‘Black Box’ nature is a major concern while proceeding with a NN approach [261–267]. The relationship with weight changes in training stage is the main reason behind this effect and hence the main hurdle in neural network-based systems is to assign the parameters like number of layers, number of neurons etc. Time cost is another drawback in these techniques. In conclusion, similar to other fields of machine learning schemes, here the efficiency is also high since the method is to adaptively select the most accurate solutions for embedding secret bits over the carrier media. Intelligent selection of coefficients by using the prediction results from such machine learning systems will help to embed the data with minimal distortion to the cover image at high payload capacity, data loss at the receiver side due to quantization errors etc. This helps to enhance all aspects of image steganography from imperceptibility to carrier capacity while comparing to conventional sequential embedding schemes. Also, the embedding nature can be changed by tuning parameters. This will help to give additional weightage for any particular aspects of steganography. The possible drawback includes the requirement of training data. Also, the efficiency very much relies on the trained parameters and properly supervised learning hence became a critical factor.

317

8. Cover image selection In a steganographic process, the main concern is to embed a secret data in the cover media without leaving any suspicious noise in the embedded image. Thus, the embedding efficiency mainly depends on the nature of embedding processes and are explained in the previous sections. However, the efficiency also depends on the selected cover image. In this section, the discussion is about the significance of cover image in a steganographic process. As we detailed before, the embedding will be more detectable in the smooth regions in the image while it is difficult to find over the textured or edge regions. Hence, it can be stated that the cover image will possess high fidelity regions to make the embedding in an imperceptible way. This will also help to conceal more secret data bits without creating any visible noise in the stego-image. For selecting the appropriate cover image, a lot of methods were proposed and most of them select the ones with more frequency components rather than smooth regions [268,269]. Some of the earlier work uses correlation-based cover selection procedure, the similarity of image blocks [270] and over statistical measures features. Then, more image information is taken for the selection of a suitable cover image from an available dataset. In the article [268], a correlation-based cover image selection process is employed. The LSB bits of wavelet coefficients of the cover images are correlated with the secret data and the most matching cover image is selected from the dataset. Another approach of cover image selection is demonstrated in the paper [271]. A fuzzy based image complexity model is used to find a least detectable stego-image by checking the image dataset. Another cover image selection method is explained in the paper [272]. This scheme uses Fisher Information Matrix and Gaussian Mixture Model for measuring the embedding ability for comparison. The main drawback of previous works is time consumption. An improved version of cover selection was proposed by Hajduk and Levický [273] to solve the above flaw. Here, accelerated optional cover selection based on shortening vector of a secret message is used. In all of the cover image selection methods, the ultimate aim is to use a suitable cover image for the intended application. This will help to improve all required properties of a steganographic process especially when the stego-image is possibly analysed using a steganalysis tool in the communication channel. 9. Steganalysis: an overview Steganalysis is the study of extracting the secret data from a stego-images as in an intruder’s point of view [274,275]. Steganalysis is complex process since the embedded cover image is often secured using encryption standards and can be corrupted with noise or any attacks. Also, there are multiple type encryptions, noise and attacks are possible and the intruder needs to find which one in which order to make the data extraction intact. There are two main reasons for developing steganalysis techniques. The first one is for the ethical hacking purpose where one needs to retrieve data while communicating with illegal agencies or suspicious subjects. The latter use is for improving the robustness of existing steganographic techniques by detecting the flaws in the embedding schemes. Since different types of steganographic approaches are used for hiding secret data in cover images, steganalysis techniques will also be different based on the embedding scheme that has used there. A broad classification of Steganalysis is given in Fig. 5. The main classification is based on the basic detection procedure used in the steganalysis. It can be of two types: signature steganalysis and statistical steganalysis [274,276,277]. In the former type, the system will search for any particular patterns or signatures of any steganographic tool. The key idea behind this concept

318

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Fig. 5. Steganalysis techniques [274].

is that, the stego-image may have some repeated patterns due to the embedded secret bits. The latter one utilizes statistical parameters to check whether any secret bits are hidden there. Obviously, this will yield better as it uses mathematical models to evaluate the stego-image characteristics rather than guess analysis as used in the signature steganalysis. 9.1. Signature steganalysis Under the signature steganalysis class, the main sub divisions are ‘specific’ and ‘universal’ types. In the specific type, the method search for unique patterns or signatures for known steganographic systems. Steganographic schemes create some special patterns such as histogram arrangement, minimum and maximum intensity range etc. These specific signatures are identified here and then detect the counter method for that particular steganographic embedding scheme. This works well if the stego-image is created using popular steganographic tools. The other one is the universal type where a change in expected jpeg compression quantization nature is used to detect whether an image is a stego-image or not. 9.2. Statistical steganalysis

Fig. 6. Lena image at different embedding levels (a) Original image, (b) 2 BPP, (c) 3 BPP, (e) 4 BPP, (f) 5 BPP, (g) 6 BPP.

In this category, the mathematical statistical models are used for detecting hidden data and similar to a signature class, this can also be classified as ‘specific’ and ‘universal’ types. The specific type uses exclusive stages of statistical decomposition to check for any specific steganographic systems. This method can be further categorized on the basis of the target embedding schemes and some of the popular methods includes LSB embedding steganalysis, LSB matching steganalysis, Spread-spectrum steganography steganalysis, BPCS-steganography steganalysis, JPEG-compression steganography steganalysis, Transform domain steganography steganalysis and Additive noise steganography steganalysis. On the other hand, Universal statistical steganalysis is a meta-detection process not intended for any special category. The detection sensing is in multiple stages and after each stage the detec-

tion parameters are tuned to beat the possible steganographic system. Here, some steganalysis techniques, visual inspection, content and histogram analysis, Chi-squared test and RS Steganalysis are considered the most basic techniques to understand whether a cover image does have embedded bits or not. A sample of visual inspection is shown in Fig. 6 which represents the stego-image at different embedding levels. A sample of histogram analysis is shown in Fig. 7 which represents the difference in the histogram of cover and a stego image. While comparing different aspects like modelling complexity and detection accuracy, statistical based steganography seems superior over the signature type and hence it is preferred in cru-

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

319

Fig. 7. (a) Histogram of original Lena cover image, (b) Histogram of stego-image.

cial systems. Among the statistical taxonomy, the universal type is much more useful since its detection procedure is adaptive to the given stego-image and its operation is not limited to any specific steganographic system. However, the computational field becomes more and more advanced in day by day and it is not practical to retrieve the secret data while using robust secret keys and encryption algorithms. Even though, the stego-image can tamper and the data transfer can be prevented if the right choice of steganalysis is used. 10. Future directions The main challenges in image steganography are to hide secret data inside a cover image with minimum detectability, high security, good defence against intruder attacks and high payload capacity. Even though researchers have been working on this field for more than a decade, the fulfilment of these above said requirements is not yet completely achieved. Since the steganographic properties are mutually related, enhancing some properties may lower the efficiency in other aspects. The issue is that a practical solution is not yet achieved to resolve all these requirements at once. Some of the steganographic concepts are given here for improving the efficiency of current image steganographic techniques. • The emphasis over Transform domain techniques Time domain systems are limited in embedding security since the direct embedding may often void the original visual quality. Thus, the better option is to choose transform domain techniques for advanced techniques. Nowadays, computation cost is not a serious problem as processors are too powerful to use in case of such requirements. Transform domain techniques offer multiple host planes from the cover image and it boosts the choice for selecting the more efficient embedding locations. It also helps to embed with more number of bits per cover media. Researches in transform domain planes are still unsaturated and possibilities are there to improve the conventional steganographic systems. • The emphasis over Adaptive embedding approaches Adaptiveness is the primary idea to get an optimized algorithm. It not only improves the embedding efficiency but also defends the steganalysis attempts with suitable counter measures. Machine learning techniques are the best option to get an adaptive

nature to the system. Initially, the system can be trained using possible known parameters and a new attempt can be made without creating previously occurred flaws in the system. Intelligent selection of coefficients by using the prediction results from such machine learning systems will help to embed the data with minimal distortion to the cover image at high payload capacity, data loss at the receiver side due to quantization errors etc. This helps to enhance all aspects of image steganography from imperceptibility to carrier capacity while comparing to conventional sequential embedding schemes. • Improving security concerns Steganographic systems can be designed with robust encryption algorithms with secret keys and such systems possess better authenticity and security over standard schemes. Hence the research and development of these suitable encryption systems is a need of the hour. Encryption based on existing popular image formats such as jpeg is another unsaturated field where the encryption and the steganographic process can be combined in the image format itself. Such systems help to keep the undetectability of the hidden secret bits up to a great extent. • Improving robustness against attacks While investigating the image steganographic area, stego-image robustness is one of the least considered parameters in most of the research work. But this is very critical since the stego-image is susceptible to various attacks from a communication channel and from intruders. Most of the times, the intruder can easily alter the properties of stego-images like intensity, smoothness etc. This will create serious errors while retrieving the secret data at the receiver side. Hence, supportive counter measures are required to nullify the effects of attacks taken place over the stego-media. • Concentrating over 2D image steganography As we discussed in the previous discussions, we found that, 2D image steganography provides better efficient throughput while comparing with the 3D version. The embedding over 3D vertices is a messy process with complex mathematical models. 3D images are not much popular as 2D images and while sending such images through open channels, the stego process can be easily detected. Hence, it is better to concentrate over the 2D space rather than using 3D carrier images for advanced steganographic processes.

320

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

Table 7 The advantage and disadvantage of all image steganography techniques.

Spatial Domain

Technique

Advantage

Disadvantage

LSB

- Simplicity of embedding and decoding process - Good payload capacity - Good visual quality - High embedding capacity - Good Imperceptibility - Supports reversible data hiding - Preserve good visual quality - Better embedding capacity - Better visual quality compare to previous techniques - Higher embedding capacity - More robust against intruder’s steganalysis process - Distortion in the stego-image is lesser while comparing with other spatial methods - Dense embedding capacity - Better quality stego-images than LSB - Less distortion - Good imperceptibility Simple transform domain that used in data hiding technique

- Poor defence against geometric, compression and statistical attacks - Lack of security

PVD Histogram Shifting Difference Expansion Multiple Bit-Planes Palette based Steganography Pixel Intensity Modulation Transform Domain

DFT

DCT

Better visual quality than DFT

DWT

- Higher secure than DCT - Robust data hiding technique

IWT

- Integer coefficients, supports reversible data hiding - Better security - There will be a real and imaginary part coefficient and give more housing to the secret data - Robustness against attacks - It is provided more housing to the secret data than CWT - Robustness against attacks - It is provided high imperceptibility - Less degradation in the visual quality - Optimal compression schemes over the secret data help to increase the security - Robustness against attacks - It is defend the degradation of image quality since changes in textured and edge or robust regions are not much easily detectable as in smooth regions. - Very difficult to detect the embedding process by simply seeing the stego-image by a human intruder - Good visual quality - It is suited for finding embedding locations with high security, imperceptibility and where provides minimum retrieval error using faster computational or it is incorporated in the secret extraction process The apt coefficient selection favours high payload capacity and it is one of the best solutions for getting apt hidden regions in the cover media - Possesses much robustness against image manipulations - It is used as pre-processing embedding nature and/or for limiting or extending the embedding stages or extracting process - It favours the system by recognizing the suitable image patterns in a fast way by reducing irreverent complexities this will eventually help to use scheme in real world cases with proper imperceptibility - It is used as pre-processing embedding nature and/or for limiting or extending the embedding stages or extracting process - It is superior in its classification capabilities It is helped to provide high robustness and imperceptibility system

CWT

DT-CWT

Compressive Sensing

Adaptive Steganography based on

Region Selection

HVS

SVM

Genetic Algorithm

Fuzzy

Neural Networks

Poor defence against geometric, compression and statistical attacks - Limit payload capacity - Poor defence against attacks - Need large location data for extracting secret data - poor control of capacity - Vulnerable against geometrical attacks - Need of images in specific lossless compression formats. Not favour in real applications. - lack of security - Low payload capacity - Poor defence against noise attacks -

Poor embedding capacity Less visual quality Lack of security Poor embedding capacity Lake of security Poor robustness against attacks Moderate embedding payload capacity Need high supplementary data to achieve Reversibility - Low embedding capacity - Accurate coefficient should be selection to avoid loss data during embedding and extraction data - Accurate coefficient should be selection to avoid loss data during embedding and extraction data

- Sampling matrix always has a large size and it occupies a large space

Poor embedding rate

- Poor embedding rate - Poor robustness against statistical measures

- Finding the local optimal solution can be obtained only after a large number of computational stages and results in poor search speed in finding precise embedding locations - Higher cost of modelling complexity

- The requirement of training data - The efficiency much relies on the trained parameters and properly supervised learning is hence became a critical factor - Time cost

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

11. Conclusion In this article we explored the basic concepts, taxonomy and performance evaluation methods for image steganography. Different types of steganographic systems are illustrated with recent and relevant papers. Different aspects and requirements of steganography are detailed and highly-cited research works are explained to enlighten the concepts. The significance of embedding location selection, cover image selection, the relation of embedding capacity with imperceptibility are explained and compared with various survey papers. Different embedding systems are explained with its advantages and disadvantages. The major challenges of previous research in the image steganography can be listed as: (i) Keeping the imperceptibility in a higher level, (ii) Providing high security for hidden data, (iii) Robustness against intruder attacks and (iv) High payload capacity. As a whole, most of the spatial domain methods are suitable if the requirement is high payload capacity and less security, the common flaw found is that demonstrate poor defence against geometric attacks, also only lossless image compression formats such as TIFF and PNG etc. will guarantee high retrieval quality. While using JPEG, some of the message information may be lost due to compression. Also, spatial domain methods are not preferred if the stego-image is prone to any sort of corruption in the channel. On the other hand, the transform domain methods lead to maintaining the security of the secret data with only providing less payload capacity with reasonable robustness while maintaining good imperceptibility. Based on the survey, it can be seen that the systems with adaptive learning schemes possess efficient steganographic systems and hence research may be directed towards machine learning based schemes for high quality steganographic systems. Based on the survey and analysis, a comparison table is created to illustrate the advantages and disadvantages of all the techniques and is shown in Table 7. Acknowledgments The first author would like to acknowledge higher committee of education development in Iraq (HCED) for the scholarship funding. References [1] A. Cheddad, J. Condell, K. Curran, P. Mc Kevitt, Digital image steganography: survey and analysis of current methods, Signal Processing 90 (2010) 727–752, doi:10.1016/j.sigpro.2009.08.010. [2] R.J. Anderson, F.A.P. Peticolas, On the Limits of Steganography, IEEE J. Sel. Areas Comm (1998) 16. [3] D. Artz, Digital steganography: hiding data within data, IEEE Int. Comput 5 (2001) 75–80. [4] M. Juneja, P.S. Sandhu, Designing of robust image steganography technique based on LSB insertion and encryption, in: Proceedings of the 2009 International Conference on Advances in Recent Technologies in Communication and Computing, 2009, pp. 302–305, doi:10.1109/ARTCom.2009.228. [5] R. Srikumar, C.S. Malarvizhi, Strong encryption using steganography and digital watermarking, in: Proceedings of the 22nd Picture Coding Symposium, 2001, pp. 425–428. https://www.scopus.com/inward/record.uri?eid=2-s2. 0-0035786695&partnerID=40&md5=4db801c2a93ff4e18a0e4fc90ead5813. [6] S.B. Sasi, N. Sivanandam, A survey on cryptography using optimization algorithms in WSNs, Indian J. Sci. Technol 8 (2015) 216–221. [7] N.F. Johnson, S. Jajodia, Exploring steganography: Seeing the unseen, Computer (Long. Beach. Calif) (1998) 31. [8] P. Premaratne, F. Safaei, 2D barcodes as watermarks in image authentication, in: Proceedings of the 6th IEEE/ACIS International Conference on Computer and Information Science (ICIS 2007), IEEE, 2007, pp. 432–437. [9] P. Premaratne, L.C. DeSilva, F. Safaei, Copyright protection scheme for digital television content, Int. J. Inf. Technol. 11 (2005) 101–108. [10] P. Premaratne, M. Premaratne, Key-based scrambling for secure image communication, in: International Conference on Intelligent Computing, Springer, 2012, pp. 259–263. [11] P. Premaratne, L.C. De Silva, I. Burnett, Low frequency component-based watermarking scheme using 2D data matrix, Int. J. Inf. Technol. 12 (2006) 1–12. [12] T.H.N. Le, K.H. Nguyen, H.B. Le, Literature survey on image watermarking tools, watermark attacks, and benchmarking tools, in: Proceedings of the 2nd International Conference on Advance Multimedia, IEEE, 2010, pp. 67–73, doi:10.1109/MMEDIA.2010.37.

321

[13] J.C. Ingemar, M.L. Miller, A.B. Jeffrey, J. Fridrich, T. Kalker, Digital watermarking and steganography, 2008. doi:10.1016/B978- 0- 12- 372585- 1.X5001- 3. [14] F.Y. Shih, Digital Watermarking and steganography: Fundamentals and Techniques, CRC Press, 2017. [15] I. Cox, M. Miller, J. Bloom, J. Fridrich, T. Kalker, Digital Watermarking and Steganography, Morgan Kaufmann, 2007. [16] J.C. Judge, Steganography: Past, Present, Future, Lawrence Livermore National Lab., CA (US), 2001. [17] D. Kahn, The history of steganography, Information Hiding, Springer, 1996, pp. 1–5. [18] F.A.P. Petitcolas, S. Katzenbeisser, Information Hiding Techniques for Steganography and Digital Watermarking (Artech House Computer Security Series), Artech House, 20 0 0. [19] M. Mishra, P. Mishra, M.C. Adhikary, Digital image data hiding techniques: A comparative study, ArXiv: 1408.3564. (2014). [20] N. Provos, P. Honeyman, Hide and seek: An introduction to steganography, IEEE Secur. Priv. 99 (2003) 32–44. [21] F.A.P. Petitcolas, R.J. Anderson, M.G. Kuhn, Information hiding-a survey, Proc. IEEE. 87 (1999) 1062–1078. [22] J.-X. Du, D.-S. Huang, X.-F. Wang, X. Gu, Computer-aided plant species identification (CAPSI) based on leaf shape matching technique, Trans. Inst. Meas. Control. 28 (2006) 275–285. [23] C.-H. Zheng, D.-S. Huang, Z.-L. Sun, M.R. Lyu, T.-M. Lok, Nonnegative independent component analysis based on minimizing mutual information technique, Neurocomputing 69 (2006) 878–883. [24] A.K. Bhattacharjya, H. Ancin, Data embedding in text for a copier system, in: Proceedings of the 2018 IEEE International Conference on Image Processing, 1999, pp. 245–249. https://www.scopus.com/inward/record.uri?eid=2-s2. 0-0033346133&partnerID=40&md5=943015776382f004d054569a623b53a0. [25] S.S. Baawi, M.R. Mokhtar, R. Sulaiman, A comparative study on the advancement of text steganography techniques in digital media, ARPN J. Eng. Appl. Sci. 13 (2018) 1854–1863. https://www.scopus. com/inward/record.uri?eid=2- s2.0- 85044143063&partnerID=40&md5= 677ec16a7b599b5fcbd51ff975a86cf6. [26] M.S.A. Karim, K. Wong, K. Tanaka, Reversible data embedding in GolombRice code, in: Proceedings of the 2011 IEEE International Conference on Signal and Image Processing Applications, 2011, pp. 515–519, doi:10.1109/ICSIPA. 2011.6144149. [27] I.J. Kadhim, A new audio steganography system based on auto-key generator, AL-Khwarizmi Eng. J. (2012) 8. [28] B. Santhi, G. Radhika, S. Ruthra Reka, Information security using audio steganography - a survey, Res. J. Appl. Sci. Eng. Technol. 4 (2012) 2255–2258. https://www.scopus.com/inward/record.uri?eid=2-s2. 0-84862743361&partnerID=40&md5=31f19d1c4b98e21fb2527c64bdc9b74c. [29] S. Limkar, A. Nemade, A. Badgujar, R. Kate, Improved data hiding technique based on audio and video steganography, Smart Computer Informatics, 2018, pp. 581–588. Springer. [30] R.E. Vinodhini, P. Malathi, T.G. Kumar, A survey on DNA and image steganography, in: Proceedings of the 2017 4th International Conference on Advanced Computing and Communication Systems (ICACCS), 2017, doi:10.1109/ICACCS. 2017.8014670. [31] D. Haughton, F. Balado, A modified watermark synchronisation code for robust embedding of data in DNA, in: Proceedings of the 2017 IEEE International Conference on Acoustics, Speech and Signal Processing, 2013, pp. 1148– 1152, doi:10.1109/ICASSP.2013.6637830. [32] A. Odeh, K. Elleithy, M. Faezipour, E. Abdelfattah, Novel steganography over HTML Code, in: Innovations and Advances in Computing, Informatics, Systems Sciences, Networking and Engineering, Springer, 2015, pp. 607–611. [33] A.G. Memon, S. Khawaja, A. Shah, Steganography: a new horizon for safe communication through xml., J. Theor. Appl. Inf. Technol. (2008) 4. ´ [34] E. Zielinska , W. Mazurczyk, K. Szczypiorski, Trends in steganography, Commun. ACM. 57 (2014) 86–95. [35] M.S. Subhedar, V.H. Mankar, Current status and key issues in image steganography: a survey, Comput. Sci. Rev. 13 (2014) 95–113. [36] H. Wang, S. Wang, Cyber warfare: steganography vs. steganalysis, Commun. ACM. 47 (2004) 76–82. [37] B. Li, J. He, J. Huang, Y.Q. Shi, A survey on image steganography and steganalysis, J. Inf. Hiding Multimed. Signal Process 2 (2011) 142–172. [38] L.M. Marvel, C.T. Retter, C.G. Boncelet, A methodology for data hiding using images, in: Proceedings of the IEEE Military Communications Conference, IEEE, 1998, pp. 1044–1047. [39] H. Mathkour, B. Al-Sadoon, A. Touir, A new image steganography technique, in: Proceedings of the 2008 4th International Conference on Wireless Communications, Networking and Mobile Computing, IEEE, 2008, pp. 1–4. [40] A.A.J. Altaay, S. Bin Sahib, M. Zamani, An introduction to image steganography techniques, in: Proceedings of the 2012 International Conference on Advanced Computer Science Applications and Technologies (ACSAT), IEEE, 2012, pp. 122–126. [41] L. Li, J. Qian, J.-S. Pan, Characteristic region based watermark embedding with RST invariance and high capacity, AEU-Int. J. Electron. Commun. 65 (2011) 435–442. [42] S. Edward Jero, P. Ramu, R. Swaminathan, Imperceptibility - robustness tradeoff studies for ECG steganography using continuous ant colony optimization, Expert Syst. Appl. 49 (2016) 123–135, doi:10.1016/j.eswa.2015.12.010.

322

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

[43] A. Venkatraman, S. Abraham, M. Paprzycki, Significance of steganography on data security, in: Proceedings of the ITCC 2004 International Conference on Information Technology: Coding and Computing, IEEE, 2004, pp. 347–351. [44] D. Bucerzan, C. Rat¸ iu, Testing methods for the efficiency of modern steganography solutions for mobile platforms, in: Proceedings of the 2016 6th International Conference on Computers Communications and Control, IEEE, 2016, pp. 30–36. [45] C. Whitelam, N. Osia, T. Bourlai, Securing multimodal biometric data through watermarking and steganography, in: Proceedings of the 2013 IEEE International Conference on Technologies for Homeland Security, IEEE, 2013, pp. 61–66. [46] N. Lofgren, S.K. Decker, H.L. Brunk, J.S. Carr, Digitally watermaking holograms for use with smart cards, (2003). [47] R.A. Wazirali, Optimization of perceptual steganography capacity using the human visual system and evolutionary computation. (2016). [48] A.M. Zeki, A.A. Ibrahim, A.A. Manaf, Steganographic software: analysis and implementation, Int. J. Comput. Commun. 6 (2012) 35–42. [49] A. Pradhan, A.K. Sahu, G. Swain, K.R. Sekhar, Performance evaluation parameters of image steganography techniques, in: Proceedings of the 2016 International Conference on Research Advances in Integrated Navigation Systems, 2016, doi:10.1109/RAINS.2016.7764399. [50] S. Wang, D. Zheng, J. Zhao, W.J. Tam, F. Speranza, An image quality evaluation method based on digital watermarking, IEEE Trans. Circuits Syst. Video Technol. 17 (2007) 98–105, doi:10.1109/TCSVT.2006.887086. [51] S. Sidhik, S.K. Sudheer, V.P. Mahadhevan Pillai, Performance and analysis of high capacity Steganography of color images involving Wavelet Transform, Optik (Stuttg) 126 (2015) 3755–3760, doi:10.1016/j.ijleo.2015.08.208. [52] F. Zhang, Z. Pan, K. Cao, F. Zheng, F. Wu, The upper and lower bounds of the information-hiding capacity of digital images, Inf. Sci. (Ny). 178 (2008) 2950–2959. [53] P. Premaratne, M. Premaratne, Image similarity index based on moment invariants of approximation level of discrete wavelet transform, Electron. Lett. 48 (2012) 1465–1467. [54] P. Premaratne, M. Premaratne, New structural similarity measure for image comparison, in: Proceedings of the International Conference on Intelligent Computing, Springer, 2012, pp. 292–297. [55] Erratum, Comprehensive survey of 3D image steganography techniques” [IET Image Processing, 12, 1, (1-10)] doi: 10.1049/iet-ipr.2017.0162, IET Image Process 12 (2018) 619, doi:10.1049/iet-ipr.2018.0088. [56] Z.Y. Al-Omari, A.T. Al-Taani, Secure LSB steganography for colored images using character-color mapping, in: Proceedings of the 8th International Conference on Communication System, IEEE, 2017, pp. 104–110, doi:10.1109/IACS. 2017.7921954. [57] H.T. Wu, J.L. Dugelay, Steganography in 3D geometries and images by adjacent bin mapping, Eurasip J. Inf. Secur. (2009), doi:10.1155/2009/317165. [58] P. Thiyagarajan, V. Natarajan, G. Aghila, V. Prasanna Venkatesan, R. Anitha, Pattern based 3D image Steganography, 3D Res. 4 (2013) 1–8, doi:10.1007/ 3DRes.01(2013)1. [59] N. Li, J. Hu, R. Sun, S. Wang, Z. Luo, A high-capacity 3d steganography algorithm with adjustable distortion, IEEE Access 5 (2017) 24457–24466. [60] Y.-H. Huang, Y.-Y. Tsai, A reversible data hiding scheme for 3d polygonal models based on histogram shifting with high embedding capacity, 3D Res. (2015) 6, doi:10.1007/s13319-015-0051-x. [61] Y.-M. Cheng, C.-M. Wang, Y.-Y. Tsai, C.-H. Chang, P.-C. Wang, Steganography for three-dimensional models, Adv. Comput. Graph. (2006) 510–517. [62] Y.Y. Tsai, An adaptive steganographic algorithm for 3D polygonal models using vertex decimation, Multimed. Tools Appl. 69 (2014) 859–876, doi:10.1007/ s11042- 012- 1135- 8. [63] Y.M. Cheng, C.M. Wang, A high-capacity steganographic approach for 3D polygonal meshes, Vis. Comput. 22 (2006) 845–855, doi:10.1007/ s0 0371-0 06-0 069-4. [64] V. Thanikaiselvan, S. Shastri, S. Ahmad, Information Hiding: Steganography, Intell. Tech. Signal Process. Multimed. Secur. (2017) 65–91. [65] S. Ashwin, J. Ramesh, S.A. Kumar, K. Gunavathi, Novel and secure encoding and hiding techniques using image steganography: a survey, in: Proceedings of the 2012 International Conference on Emerging Trends in Electrical Engineering and Energy Management, 2012, pp. 171–177, doi:10.1109/ICETEEEM. 2012.6494463. [66] A. Habibi Lashkari, A. Abdul Manaf, M. Masrom, S. Mohd Daud, A survey on image steganography algorithms and evaluation, Commun. Comput. Inf. Sci. (2011) 406–418, doi:10.1007/978- 3- 642- 22389- 1_36. [67] K. Wang, G. Lavoué, F. Denis, A. Baskurt, A comprehensive survey on three-dimensional mesh watermarking, IEEE Trans. Multimed. 10 (2008) 1513–1527. [68] P.R. Alface, B. Macq, From 3D mesh data hiding to 3D shape blind and robust watermarking: a survey, Trans. Data Hiding Multimed. Secur. (2007) 91–115. [69] M.M. Krishna, N. Roberts, Enhancement of embedding capacity and security in reversible steganography, in: Proc. 2016 IEEE Int. Conf. Wirel. Commun. Signal Process. Networking, WiSPNET 2016, 2016, pp. 803–807, doi:10.1109/ WiSPNET.2016.7566244. [70] K.-H. Jung, A survey of reversible data hiding methods in dual images, IETE Tech. Rev. (Institution Electron. Telecommun. Eng. India). 33 (2016) 441–452, doi:10.1080/02564602.2015.1102099. [71] T. Rabie, I. Kamel, High-capacity steganography: a global-adaptive-region discrete cosine transform approach, Multimed. Tools Appl. 76 (2017) 6473–6493. [72] R. Chandramouli, N. Memon, Analysis of LSB based image steganography techniques, IEEE Int. Conf. Image Process (2001) 1019–1022.

[73]

[74]

[75]

[76]

[77]

[78]

[79] [80]

[81]

[82]

[83] [84]

[85]

[86]

[87]

[88] [89]

[90]

[91]

[92]

[93]

[94] [95]

[96] [97]

[98]

https://www.scopus.com/inward/record.uri?eid=2-s2.0-0035167458& partnerID=40&md5=6726e829c7b970d5d42d12ddb00214dd. W.S. Sari, E.H. Rachmawanto, D.R.I.M. Setiadi, C.A. Sari, A Good Performance OTP encryption image based on DCT-DWT steganography, Telkomnika (Telecommunication Comput, Electron. Control. 15 (2017) 1987–1995, doi:10.12928/TELKOMNIKA.v15i4.5883. M.S. Sutaone, M.V Khandare, Image based steganography using LSB insertion, in: Proceedings of the 2008 IET International Conference on Wireless, Mobile and Multimedia Networks, 2008, pp. 146–151, doi:10.1049/cp:20080166. S. Rajendran, M. Doraipandian, Chaotic map based random image steganography using LSB technique, Int. J. Netw. Secur. 19 (2017) 593–598, doi:10.6633/ IJNS.201707.19(4).12. N. Patel, S. Meena, LSB based image steganography using dynamic key cryptography, in: Proceedings of the 2016 International Conference on Emerging Trends in Communication Technologies, 2017, doi:10.1109/ETCT.2016.7882955. X. Zhou, W. Gong, W. Fu, L. Jin, An improved method for LSB based color image steganography combined with cryptography, in: Proceedings of the 2016 IEEE/ACIS 15th International Conference on Computer and Information Science, 2016, doi:10.1109/ICIS.2016.7550955. V. Ajith, S. Kanagaraj, P. Malathi, Image steganography based on LSB matching revisitedusing secret sharing application, Int. J. Appl. Eng. Res. 10 (2015) 2931–2938. https://www.scopus.com/inward/record.uri?eid=2-s2. 0-84942312316&partnerID=40&md5=f7c3c3f93db70137f339039964903ef9. W. Luo, F. Huang, J. Huang, Edge adaptive image steganography based on LSB matching revisited, IEEE Trans. Inf. Forensics Secur. 5 (2010) 201–214. S. Chakraborty, A.S. Jalal, C. Bhatnagar, LSB based non blind predictive edge adaptive image steganography, Multimed. Tools Appl. 76 (2017) 7973–7987, doi:10.1007/s11042- 016- 3449- 4. S. Bhatt, A. Ray, A. Ghosh, A. Ray, Image steganography and visible watermarking using LSB extraction technique, in: Proceedings of the 2013 7th International Conference on Intelligent Systems and Control, 2015, doi:10.1109/ ISCO.2015.7282315. H. Dadgostar, F. Afsari, Image steganography based on interval-valued intuitionistic fuzzy edge detection and modified LSB, J. Inf. Secur. Appl. 30 (2016) 94–104, doi:10.1016/j.jisa.2016.07.001. T.-C. Lu, C.-Y. Tseng, J.-H. Wu, Dual imaging-based reversible hiding technique using LSB matching, Signal Processing 108 (2015) 77–89. F. Pan, J. Li, X. Yang, Image steganography method based on PVD and modulus function, in: Proceedings of the 2011 International Conference on Electronics, Communications and Control (ICECC), 2011, pp. 282–284, doi:10.1109/ICECC. 2011.6067590. G. Swain, S.K. Lenka, Pixel value differencing steganography using correlation of target pixel with neighboring pixels, in: Proceedings of 2015 IEEE International Conference on Electrical, Computer and Communication Technologies, ICECCT 2015, 2015, doi:10.1109/ICECCT.2015.7226029. M. Hussain, A.W. Abdul Wahab, A.T.S. Ho, N. Javed, K.H. Jung, A data hiding scheme using parity-bit pixel value differencing and improved rightmost digit replacement, Signal Process. Image Commun. 50 (2017) 44–57, doi:10.1016/j. image.2016.10.005. X. Zhang, S. Wang, Vulnerability of pixel-value differencing steganography to histogram analysis and modification for enhanced security, Pattern Recognit. Lett. 25 (2004) 331–339. G. Swain, A Steganographic Method Combining LSB Substitution and PVD in a Block, Procedia Comput. Sci. (2016) 39–44, doi:10.1016/j.procs.2016.05.174. M. Kalita, T. Tuithung, A novel steganographic method using 8-neighboring PVD (8nPVD) and LSB substitution, in: Proceedings of the 25th International Conference on Systems, Signals and Image Processing, 2016, doi:10. 1109/IWSSIP.2016.7502756. O. Hosam, N. Ben Halima, Adaptive block-based pixel value differencing steganography, Secur. Commun. Networks 9 (2016) 5036–5050, doi:10.1002/ sec.1676. I.R. Grajeda-Marín, H.A. Montes-Venegas, J.R. Marcial-Romero, J.A. HernándezServín, V. Muñoz-Jiménez, G.D.I. Luna, A new optimization strategy for solving the fall-off boundary value problem in pixel-value differencing steganography, Int. J. Pattern Recognit. Artif. Intell. (2018) 32, doi:10.1142/ S0218001418600108. C.-L. Liu, H.-H. Liu, Reliable detection of histogram shift-based steganography using payload invariant features, Appl. Mech. Mater. 284–287 (2013) 3517– 3521, doi:10.4028/www.scientific.net/AMM.284-287.3517. C. Qin, C.-C. Chang, Y.-H. Huang, L.-T. Liao, An inpainting-assisted reversible steganographic scheme using a histogram shifting mechanism, IEEE Trans. Circuits Syst. Video Technol. 23 (2013) 1109–1118, doi:10.1109/TCSVT.2012. 2224052. Z. Ni, Y.-Q. Shi, N. Ansari, W. Su, Reversible data hiding, IEEE Trans. Circuits Syst. Video Technol. 16 (2006) 354–361, doi:10.1109/TCSVT.2006.869964. W.-L. Tai, C.-M. Yeh, C.-C. Chang, Reversible data hiding based on histogram modification of pixel differences, IEEE Trans. Circuits Syst. Video Technol. 19 (2009) 906–910, doi:10.1109/TCSVT.2009.2017409. Z.-H. Wang, C.-F. Lee, C.-Y. Chang, Histogram-shifting-imitated reversible data hiding, J. Syst. Softw. 86 (2013) 315–323, doi:10.1016/j.jss.2012.08.029. C.-F. Lee, H.-L. Chen, H.-K. Tso, Embedding capacity raising in reversible data hiding based on prediction of difference expansion, J. Syst. Softw. 83 (2010) 1864–1872, doi:10.1016/j.jss.2010.05.078. C.-C. Chang, Y.-H. Huang, T.-C. Lu, A difference expansion based reversible information hiding scheme with high stego image visual quality, Multimed. Tools Appl. 76 (2017) 12659–12681, doi:10.1007/s11042- 016- 3689- 3.

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326 [99] X. Wu, N. Memon, Context-based, adaptive, lossless image coding, IEEE Trans. Commun. 45 (1997) 437–444, doi:10.1109/26.585919. [100] K.-H. Jung, High-capacity reversible data hiding method using block expansion in digital images, J. Real-Time Image Process 14 (2018) 159–170, doi:10. 1007/s11554- 016- 0618- 7. [101] T.-C. Lu, C.-C. Chang, Y.-H. Huang, High capacity reversible hiding scheme based on interpolation, difference expansion, and histogram shifting, Multimed. Tools Appl. 72 (2014) 417–435, doi:10.1007/s11042- 013- 1369- 0. [102] B.C. Nguyen, S.M. Yoon, H.-K. Lee, Multi bit plane image steganography, Lect. Notes Comput. Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 4283 (2006) 61–70. LNCS https: //www.scopus.com/inward/record.uri?eid=2- s2.0- 33845456177&partnerID= 40&md5=3ab23ddb8ccbfd007b3774782991ab15 . [103] U. Ghosh, D. Burman, S. Maity, I. Mukherjee, Adaptive Multi-bit Image Steganography Using Pixel-Pair Differential Approach, Adv. Intell. Syst. Comput. 563 (2018) 47–56, doi:10.1007/978- 981- 10- 6872- 0_5. [104] H. Nyeem, Reversible data hiding with image bit-plane slicing, in: Proceedings of the 20th International Conference of Computer and Information Technology, IEEE, 2017, pp. 1–6. [105] B. Gupta Banik, S.K. Bandyopadhyay, Image steganography using BitPlane complexity segmentation and hessenberg QR method, Adv. Intell. Syst. Comput. 458 (2017) 623–633, doi:10.1007/978- 981- 10- 2035- 3_63. [106] J. Collins, S. Agaian, High Capacity Image Steganography Using Adjunctive Numerical Representations With Multiple Bit-Plane Decomposition Methods, ArXiv:1606.02312. (2016). [107] M. Niimi, H. Noda, E. Kawaguchi, R.O. Eason, High capacity and secure digital steganography to palette-based images, in: Proceedings of the IEEE International Conference on Image Processing, 2002 p. II/917-II/920. [108] S. Imaizumi, K. Ozawa, Multibit embedding algorithm for steganography of palette-based images, Lect. Notes Comput. Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), 8333, 2014, pp. 99–110. LNCS, doi:10. 1007/978- 3- 642- 53842- 1_9. [109] H.N. Patel, D.R. Khant, D. Prajapati, Design of a color palette based image steganography algorithm for fractal image, in: Proceedings of the 2017 International Conference on Wireless Communications, Signal Processing and Networking, IEEE, 2017, pp. 2584–2589. [110] S. Das, S. Sharma, S. Bakshi, I. Mukherjee, A framework for pixel intensity modulation based image steganography, Adv. Intell. Syst. Comput. 563 (2018) 3–14, doi:10.1007/978- 981- 10- 6872- 0_1. [111] S. Islam, P. Gupta, Robust edge based image steganography through pixel intensity adjustment, in: Proceedings of the 2014 IEEE International Confernce on High Performance Computing and Communications, 2014 IEEE 6th International Symposium on Cyberspace Safety and Security, 2014 IEEE 11th International Conference on Embedded Software and System (HPCC,CSS,ICESS), 2014, pp. 771–777, doi:10.1109/HPCC.2014.129. [112] C.-Y. Yang, W.-F. Wang, Block-based colour image steganography using smart pixel-adjustment, Adv. Intell. Syst. Comput. 329 (2015) 145–154, doi:10.1007/ 978- 3- 319- 12286- 1_15. [113] Y. Seki, H. Kobayashi, M. Fujiyoshi, H. Kiya, Quantization-based image steganography without data hiding position memorization, in: Proceedings of the IEEE International Symposium on Circuits and Systems, 2005, pp. 4987– 4990, doi:10.1109/ISCAS.2005.1465753. [114] S. Sachdeva, A. Kumar, Colour image steganography based on modified quantization table, in: Proceedings of the 2nd International Conference on Advanced Computing and Communication Technologies, 2011, pp. 309–313, doi:10.1109/ACCT.2012.37. [115] G. Liu, Y. Dai, Z. Wang, J. Yang, Optimal quantization-based steganography for adaptive hiding, in: Proceedings of the Third International Conference on Communications and Communications and Networking in China, 2008, pp. 1413–1417, doi:10.1109/CHINACOM.2008.4685268. [116] P. Premaratne, C.C. Ko, Image blur recognition using under-sampled discrete Fourier transform, Electron. Lett. 35 (1999) 889–890. [117] A.S. Khashandarag, A.H. Navin, M.K. Mirnia, H.H. Agha Mohammadi, An optimized color image steganography using LFSR and DFT techniques, Commun. Comput. Inf. Sci. 176 (2011) 247–253 CCIS, doi:10.1007/978- 3- 642- 21802- 6_ 40. [118] A. Soni, J. Jain, R. Roshan, Image steganography using discrete fractional Fourier transform, in: Proceedings of the 2013 International Conference on Intelligent Systems and Signal Processing, 2013, pp. 97–100, doi:10.1109/ISSP. 2013.6526882. [119] J. Sang, H. Xiang, H. Hu, Discrete Fourier transform-based information steganography, Huazhong Keji Daxue Xuebao (Ziran Kexue Ban)/Journal Huazhong Univ. Sci. Technol. 36 (2008) 5–8. (Natural Sci. Ed. https: //www.scopus.com/inward/record.uri?eid=2- s2.0- 52249117060&partnerID= 40&md5=9b9f24068d76af8b75ec6aee531e92e9 . [120] G. Savithri, S. Mane, J.S. Banu, Parallel Implementation of RSA 2D-DCT Steganography and Chaotic 2D-DCT Steganography, in: Proceedings of the IEEE International Conference on Computer Vision and Image Processing, Springer, 2017, pp. 593–605. [121] Y. Li, C. Xiong, X. Han, R. Xiang, F. He, H. Du, Image steganography using cosine transform with large-scale multimedia applications, Multimed. Tools Appl. (2018) 1–11. [122] P. Patel, Y. Patel, Secure and authentic DCT image steganography through DWT - SVD based digital watermarking with RSA encryption, in: Proceedings of the Conference On Communication System and Network Technologies, 2015, pp. 736–739, doi:10.1109/CSNT.2015.193.

323

[123] A.A. Attaby, M.F.M. Mursi Ahmed, A.K. Alsammak, Data hiding inside JPEG images with high resistance to steganalysis using a novel technique: DCT-M3, AIN SHAMS Eng. J. (2016), doi:10.1016/j.asej.2017.02.003. [124] C.Y. Weng, C.T. Huang, H.W. Kao, DCT-based compressed image with reversibility using modified quantization, in: Proceedings of the International Conference on Intelligent Information Hiding and Multimedia Signal Processing, Springer, 2018, pp. 214–221, doi:10.1007/978- 3- 319- 63856- 0_27. [125] M. Saidi, H. Hermassi, R. Rhouma, S. Belghith, A new adaptive image steganography scheme based on DCT and chaotic map, Multimed. Tools Appl. 76 (2017) 13493–13510, doi:10.1007/s11042- 016- 3722- 6. [126] V. Kumar, D. Kumar, A modified DWT-based image steganography technique, Multimed. Tools Appl. (2017) 1–30, doi:10.1007/s11042- 017- 4947- 8. [127] V. Divya, N. Sasirekha, High capacity steganography technique based on wavelet transform, in: Proceedings of the 2016 Online International Conference on Green Engineering Technologies IC-GET 2016, 2017, doi:10.1109/GET. 2016.7916612. [128] K. Patel, L. Ragha, Binary image Steganography in wavelet domain, in: Proceedings of the 2015 International Conference on Industrial Instrumentation and Control, 2015, pp. 1635–1640, doi:10.1109/IIC.2015.7151012. [129] M.R.D. Farahani, A. Pourmohammad, A DWT based perfect secure and high capacity image steganography method, in: Proceedings of the Parallel and Distributed Computing, Applications and Technologies, 2014, pp. 314–317, doi:10.1109/PDCAT.2013.56. [130] S. Atawneh, A. Almomani, H. Al Bazar, P. Sumari, B. Gupta, Secure and imperceptible digital image steganographic algorithm based on diamond encoding in DWT domain, Multimed. Tools Appl. 76 (2017) 18451–18472, doi:10.1007/ s11042- 016- 3930- 0. [131] M.S. Subhedar, V.H. Mankar, Image steganography using redundant discrete wavelet transform and QR factorization, Comput. Electr. Eng. 54 (2016) 406– 422, doi:10.1016/j.compeleceng.2016.04.017. [132] A.H. Yajurvedi, S. Kulkarni, An adaptive image steganography algorithm using successive pixel difference embedding and integer wavelet transform, Lect. Notes Electr. Eng. 221 LNEE (2013) 203–212, doi:10.1007/ 978- 81- 322- 0997- 3_19. [133] N. Raftari, A.M.E. Moghadam, Digital image steganography based on Integer Wavelet Transform and assignment algorithm, in: Proceedings of the 6th Asia International Conference on Mathematical Modelling and Computer Simulation, 2012, pp. 87–92, doi:10.1109/AMS.2012.15. [134] P. Liu, C. Chen, L. Ge, Y. Luo, Efficient self-adaptive image steganography scheme based on iterative blending and integer wavelet transform, Lecture Notes on Electrical Engineering, Springer, 2014, pp. 1159–1167, doi:10.1007/ 978- 1- 4614- 4981- 2_126. [135] A. Delforouzi, M. Pooyan, Adaptive digital audio steganography based on integer wavelet transform, Circuits, Syst. Signal Process. 27 (2008) 247–259. [136] A. Miri, K. Faez, An image steganography method based on integer wavelet transform, Multimed. Tools Appl. (2017) 1–12, doi:10.1007/ s11042- 017- 4935- z. [137] K. Sathish Shet, A.R.Aswath Nagaveni, Image steganography using integer wavelet transform based on color space approach, Adv. Intell. Syst. Comput. 327 (2014) 839–848, doi:10.1007/978- 3- 319- 11933- 5_95. [138] N. Dey, D. Biswas, A.B. Roy, A. Das, S.S. Chaudhuri, DWT-DCT-SVD based blind watermarking technique of gray image in electrooculogram signal, in: Proceedings of the International Conference on Intelligent Systems Design and Applications (ISDA), IEEE, 2012, pp. 680–685. [139] T. Narasimmalou, R. Allen Joseph, Discrete Wavelet Transform based steganography for transmitting images, in: Proceedings of the IEEE-International Conference on Advances in Engineering, Science and Management, 2012, pp. 370–375. https://www.scopus. com/inward/record.uri?eid=2- s2.0- 84863922094&partnerID=40&md5= 4d893c85183cfe02e102d89ad49e6907. [140] N. Terzija, W. Geisselhardt, Digital image watermarking using complex wavelet transform, in: Proceedings on 2004 Workshop on Multimedia Security, ACM, 2004, pp. 193–198. [141] J. Fan, Y. Wu, Watermarking algorithm based on kernel fuzzy clustering and singular value decomposition in the complex wavelet transform domain, in: Proceedings of the International Conference on Information Technology and Management Science, IEEE, 2011, pp. 42–46. [142] P. Loo, N. Kingsbury, Watermarking using complex wavelets with resistance to geometric distortion, in: Proceedings of the European Signal Processing Conference, 20 0 0. [143] S. Singh, T.J. Siddiqui, Robust image steganography using complex wavelet transform, in: Proceedings of the 2017 International Conference on Multimedia, Signal Processing and Communication Technologies (IMPACT), 2013, pp. 56–60, doi:10.1109/MSPCT.2013.6782087. [144] S. Kumar, S.K. Muttoo, Data hiding techniques based on wavelet-like transform and complex wavelet transforms, in: Proceedings of the Intelligence Information Processing and Trusted Computing, 2010, pp. 1–4, doi:10.1109/IPTC. 2010.46. [145] I.J. Kadhim, P. Premaratne, P.J. Vial, B. Halloran, A comparative analysis among dual tree complex wavelet and other wavelet transforms based on image compression, in: Proceedings of the International Conference on Intelligent Computing, Springer, 2017, pp. 569–580. [146] N. Kingsbury, The dual-tree complex wavelet transform: a new efficient tool for image restoration and enhancement, in: Proceedings of the 9th European Signal Processing Conference, IEEE, 1998, pp. 1–4.

324

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

[147] S. Hatipoglu, S.K. Mitra, N. Kingsbury, Texture classification using dual-tree complex wavelet transform, (1999). [148] I.W. Selesnick, R.G. Baraniuk, N.G. Kingsbury, The dual-tree complex wavelet transform, IEEE Signal Process. Mag. 22 (2005) 123–151, doi:10.1109/MSP. 2005.1550194. [149] N. Sathisha, R. Priya, K.S. Babu, K.B. Raja, K.R. Venugopal, L.M. Patnaik, DTCWT based high capacity steganography using coefficient replacement and adaptive scaling, in: Proceedings of the SPIE - The International Society for Optical Engineering, 2013, doi:10.1117/12.2051889. [150] T. Krishnamurthy, performing image steganography using dual tree complex wavelet transform (DTCWT) and Haar wavelet, Int. J. Adv. Inf. Sci. Technol. 34 (2015) 280–284. [151] A. Bouridane, F. Kurugollu, R. Beggs, S. Boussakta, Colour image watermarking in the complex wavelet domain, in: Proceedings of the 10th IEEE International Conference on Electronics, Circuits and Systems„ IEEE, 2003, pp. 1196–1199. [152] A.I. Thompson, A. Bouridane, C. Tanougast, U.P. Verlaire, Watermarking for multimedia security using complex wavelets, J. Multimed. 5 (2010) 443–457, doi:10.4304/jmm.5.5.443-457. [153] P. Loo, N. Kingsbury, Digital watermarking using complex wavelets, in: Proceedings of the 2018 IEEE International Conference on Image Processing, 20 0 0, pp. 29–32. [154] T. Minamoto, R. Ohura, A non-blind digital image watermarking method based on the dual-tree complex discrete wavelet transform and interval arithmetic, in: Proceedings of the 2012 Ninth International Conference on Information Technology - New Generations, 2012, pp. 623–628, doi:10.1109/ITNG. 2012.129. [155] J. Panda, S. Maurya, R. Dang, B.L. Narayanapuram, Analysis of robustness of an image watermarking algorithm using the dual tree complex wavelet transform and just noticeable difference, in: Proceedings of the 2018 Conference on Signal Processing And Communication Engineering Systems, 2016, pp. 255–260. [156] S. Mabtoul, D. Aboutajdine, A robust digital image watermarking method using dual tree complex wavelet transform, in: Proceedings of the IEEE Symposium on Computing Communication, 2008, pp. 1000–1004. [157] J. Adabala, K.N. Prakash, Dual tree complex wavelet T ransform for, Int. J. Adv. Eng. Technol. 4 (2012) 482–492. [158] M. Asikuzzaman, M.J. Alam, A.J. Lambert, M.R. Pickering, Imperceptible and robust blind video watermarking using chrominance embedding : a set of approaches in the DT CWT domain, 1502 IEEE Trans. Inf. FORENSICS Secur 9 (2014) 1502–1517. [159] A.I. Thompson, A. Bouridane, F. Kurugollu, Colour spread transform data hiding capacity using complex wavelets, in: Proceedings of the 26th European Signal Processing Conference, 2007, pp. 2534–2538. https: //www.scopus.com/inward/record.uri?eid=2- s2.0- 84863758356&partnerID= 40&md5=e48f9509257c9494010e775681861aea. [160] S.P. Maity, A. Phadikar, M.K. Kundu, image error concealment based on Qim, Int. J. Wavelets, Multiresolution Inf. Process. 10 (2012) 1–30, doi:10.1142/ S0219691312500166. [161] M.R.P. Md. Asikuzzaman, A Blind and Robust Video Watermarking Scheme in the DT CWT and SVD Domain, in: Proceedings of the ThePicture Coding Symposium (PCS), 2015, pp. 277–281. [162] P.T. Sudha, Comparison of wavelet based watermarking techniques Using SVD, in: Proceedings of the 2017 3rd International Conference on Pattern Recognition and Image Analysis (IPRIA), 2012, pp. 1–6. [163] G. Bhatnagar, Q.M. Jonathan Wu, Biometrics inspired watermarking based on a fractional dual tree complex wavelet transform, Futur. Gener. Comput. Syst. 29 (2013) 182–195, doi:10.1016/j.future.2012.05.021. [164] S. Singh, T.J. Siddiqui, Robust image steganography using complex wavelet transform, in: Proceedings of the International Conference on Multimedia, Signal Processing and Communication Technologies, 2013, pp. 56–60. [165] E.J. Candès, M.B. Wakin, An introduction to compressive sampling, IEEE Signal Process. Mag 25 (2008) 21–30. [166] D.L. Donoho, Compressed sensing, IEEE Trans. Inf. Theory 52 (2006) 1289–1306. [167] E. Candes, J. Romberg, Sparsity and incoherence in compressive sampling, Inverse Probl 23 (2007) 969. [168] A.V Sreedhanya, K.P. Soman, Ensuring security to the compressed sensing data using a steganographic approach, Bonfring Int. J. Adv. Image Process. 3 (2013) 1. [169] M. Xiao, Z. He, High capacity image steganography method based on framelet and compressive sensing, in: Proceedings of the MIPPR 2015 Multispectral Image Acquis. Process. Anal., 2015, p. 98110Y. [170] S. Shafee, B. Rajaei, A secure steganography algorithm using compressive sensing based on HVS feature, in: Proceedings of the 2017 Seventh International Confernce on Emerging Security Technology (EST), IEEE, 2017, pp. 74–78. [171] Y. Zhang, L.Y. Zhang, J. Zhou, L. Liu, F. Chen, X. He, A review of compressive sensing in information security field, IEEE Access 4 (2016) 2507–2519. [172] A. Orsdemir, H.O. Altun, G. Sharma, M.F. Bocko, On the security and robustness of encryption via compressed sensing, in: Proceedings of the MILCOM, Military Communications Conference, IEEE, 2008, pp. 1–7. [173] J.-S. Pan, W. Li, C.-S. Yang, L.-J. Yan, Image steganography based on subsampling and compressive sensing, Multimed. Tools Appl. 74 (2015) 9191–9205.

[174] L. Shang, D.-S. Huang, C.-H. Zheng, Z.-L. Sun, Noise removal using a novel non-negative sparse coding shrinkage technique, Neurocomputing 69 (2006) 874–877. [175] P. Sallee, Model-based steganography, in: Proceedings of the International Workshop Digital Watermarking, Springer, 2003, pp. 154–167. [176] N. Hamid, A. Yahya, R.B. Ahmad, O. Al-Qershi, Characteristic region based image steganography using Speeded-Up robust features technique, in: Proceedings of the 2012 International Conference on Future Communication Networks, 2012, pp. 141–146, doi:10.1109/ICFCN.2012.6206858. [177] N. Hamid, A. Yahya, R.B. Ahmad, O.M. Al-Qershi, A Comparison between using SIFT and SURF for characteristic region based image steganography, Int. J. Comput. Sci. Issues. 9 (2012) 110–116. [178] C. Balasubramanian, S. Selvakumar, S. Geetha, High payload image steganography with reduced distortion using octonary pixel pairing scheme, Multimed. Tools Appl. 73 (2014) 2223–2245. [179] W.-J. Chen, C.-C. Chang, T.H.N. Le, High payload steganography mechanism using hybrid edge detector, Expert Syst. Appl. 37 (2010) 3292–3301, doi:10. 1016/j.eswa.2009.09.050. [180] A. Ioannidou, S.T. Halkidis, G. Stephanides, A novel technique for image steganography based on a high payload method and edge detection, Expert Syst. Appl. 39 (2012) 11517–11524. [181] B. Li, C. Wang, D.-S. Huang, Supervised feature extraction based on orthogonal discriminant projection, Neurocomputing 73 (2009) 191–196. [182] M. Niimi, H. Noda, E. Kawaguchi, Steganography based on region segmentation with a complexity measure, Syst. Comput. Japan. 30 (1999) 1–9, doi:10. 1002/(SICI)1520- 684X(199903)30:3<1::AID- SCJ1>3.0.CO;2- M. [183] S. Mungmode, R.R. Sedamkar, N. Kulkarni, A modified high frequency adaptive security approach using steganography for region selection based on threshold value, in: Procedia Comput. Sci., 2016, pp. 912–921, doi:10.1016/j. procs.2016.03.114. [184] O.M. Al-Qershi, B.E. Khoo, Controlling hiding capacity using image characteristics with a 2D-DE data hiding scheme, AEU - Int. J. Electron. Commun. 68 (2014) 346–350, doi:10.1016/j.aeue.2013.09.008. [185] S. Khan, T. Bianchi, Ant Colony Optimization (ACO) based data hiding in image complex region, Int. J. Electr. Comput. Eng. 8 (2018) 379–389. [186] X. Zhang, S. Wang, Steganography using multiple-base notational system and human vision sensitivity, IEEE Signal Process. Lett. 12 (2005) 67–70. [187] Y. Li, H. Zhu, R. Yu, G. Yang, J. Xu, An adaptive blind watermarking algorithm based on DCT and modified watson’s visual model, in: Proceedings of the International Symposium on Electronic Commerce and Security, IEEE, 2008, pp. 904–907. [188] W. Hong, Human visual system based data embedding method using quadtree partitioning, Signal Process. Image Commun. 27 (2012) 1123–1133, doi:10.1016/j.image.2012.09.002. [189] T. Luo, G. Jiang, M. Yu, H. Xu, W. Gao, Sparse recovery based reversible data hiding method using the human visual system, Multimed. Tools Appl. (2017) 1–24. [190] J.-X. Du, D.-S. Huang, G.-J. Zhang, Z.-F. Wang, A novel full structure optimization algorithm for radial basis probabilistic neural networks, Neurocomputing 70 (2006) 592–596. [191] Z.-L. Sun, D.-S. Huang, Y.-M. Cheun, Extracting nonlinear features for multispectral images by FCMC and KPCA, Digit. Signal Process. 15 (2005) 331–346. [192] K.-H. Liu, D.-S. Huang, Cancer classification using rotation forest, Comput. Biol. Med. 38 (2008) 601–610. [193] D.S. Huang, The Study of Data Mining Methods for Gene Expression Profiles, (2009). [194] C.-H. Zheng, D.-S. Huang, L. Shang, Feature selection in independent component subspace for microarray data classification, Neurocomputing 69 (2006) 2407–2410. [195] D.-S. Huang, W. Jiang, A general CPL-AdS methodology for fixing dynamic parameters in dual environments, IEEE Trans. Syst. Man, Cybern. Part B. 42 (2012) 1489–1500. [196] D.-S. Huang, J.-X. Du, A constructive hybrid structure optimization methodology for radial basis probabilistic neural networks, IEEE Trans. Neural Networks. 19 (2008) 2099–2115. [197] J.-X. Du, D.-S. Huang, X.-F. Wang, X. Gu, Shape recognition based on neural networks trained by differential evolution algorithm, Neurocomputing 70 (2007) 896–903. [198] Z.-L. Sun, D.-S. Huang, C.-H. Zheng, L. Shang, Optimal selection of time lags for TDSEP based on genetic algorithm, Neurocomputing 69 (2006) 884–887. [199] D.-S. Huang, J.-X. Mi, A new constrained independent component analysis method, IEEE Trans. Neural Networks. 18 (2007) 1532–1535. [200] Z. Zou, P. Premaratne, R. Monaragala, N. Bandara, M. Premaratne, Dynamic hand gesture recognition system using moment invariants, in: Proceedings of the 5th International Conference on Information and Automation for Sustainability, IEEE, 2010, pp. 108–113. [201] J.-X. Mi, D.-S. Huang, B. Wang, X. Zhu, The nearest-farthest subspace classification for face recognition, Neurocomputing 113 (2013) 241–250. [202] J. Zhang, D.-S. Huang, T.-M. Lok, M.R. Lyu, A novel adaptive sequential niche technique for multimodal function optimization, Neurocomputing 69 (2006) 2396–2401. [203] C.-H. Zheng, D.-S. Huang, K. Li, G. Irwin, Z.-L. Sun, MISEP method for postnonlinear blind source separation, Neural Comput 19 (2007) 2557–2578. [204] D. Huang, Radial basis probabilistic neural networks: model and application, Int. J. Pattern Recognit. Artif. Intell. 13 (1999) 1083–1101.

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326 [205] F. Han, D.-S. Huang, Improved extreme learning machine for function approximation by encoding a priori information, Neurocomputing 69 (2006) 2369–2373. [206] Z.-Q. Zhao, D.-S. Huang, A mended hybrid learning algorithm for radial basis function neural networks to improve generalization capability, Appl. Math. Model. 31 (2007) 1271–1281. [207] P. Premaratne, Q. Nguyen, Consumer electronics control system based on hand gesture moment invariants, IET Comput. Vis. 1 (2007) 35–41. [208] P. Premaratne, S. Ajaz, M. Premaratne, Hand gesture tracking and recognition system using Lucas–Kanade algorithms for control of consumer electronics, Neurocomputing 116 (2013) 242–249. [209] P. Premaratne, Human Computer Interaction Using Hand Gestures, Springer Science & Business Media, 2014. [210] S. Yang, P. Premaratne, P. Vial, Hand gesture recognition: An overview, in: Proceedings of the 2013 5th IEEE International Conference Broadband Network Multimedia Technology (IC-BNMT)„ IEEE, 2013, pp. 63–69. [211] P. Premaratne, S. Yang, Z. Zou, P. Vial, Australian sign language recognition using moment invariants, in: International Conference on Intelligent Computing, Springer, 2013, pp. 509–514. [212] P. Premaratne, Q. Nguyen, M. Premaratne, Human computer interaction using hand gestures, in: Proceedings of the International Conference Intelligent Computer, Springer, 2010, pp. 381–386. [213] W.-B. Zhao, D.-S. Huang, J.-Y. Du, L.-M. Wang, Genetic optimization of radial basis probabilistic neural networks, Int. J. Pattern Recognit. Artif. Intell. 18 (2004) 1473–1499. [214] D.-S. Huang, S.-D. Ma, Linear and nonlinear feedforward neural network classifiers: a comprehensive understanding, J. Intell. Syst. 9 (1999) 1–38. [215] Y. Zhao, D.-S. Huang, W. Jia, Completed local binary count for rotation invariant texture classification, IEEE Trans. Image Process. 21 (2012) 4492–4497. [216] X.-F. Wang, D.-S. Huang, J.-X. Du, H. Xu, L. Heutte, Classification of plant leaf images with complicated background, Appl. Math. Comput. 205 (2008) 916–926. [217] X.-F. Wang, D.-S. Huang, A novel multi-layer level set method for image segmentation, J. Univers. Comput. Sci. 14 (2008) 2428–2452. [218] D.-S. Huang, X.-M. Zhao, G.-B. Huang, Y.-M. Cheung, Classifying protein sequences using hydropathy blocks, Pattern Recognit 39 (2006) 2293–2300. [219] X.-F. Wang, D.-S. Huang, H. Xu, An efficient local Chan–Vese model for image segmentation, Pattern Recognit 43 (2010) 603–618. [220] D.-S. Huang, A constructive approach for finding arbitrary roots of polynomials by neural networks, IEEE Trans. Neural Netw. 15 (2004) 477–491. [221] Z.-Q. Zhao, D.-S. Huang, W. Jia, Palmprint recognition with 2DPCA+ PCA based on modular neural networks, Neurocomputing 71 (2007) 448–454. [222] W. Jiang, D.-S. Huang, S. Li, Random walk-based solution to triple level stochastic point location problem, IEEE Trans. Cybern. 46 (2016) 1438–1451. [223] P. Premaratne, M. Premaratne, Image matching using moment invariants, Neurocomputing 137 (2014) 65–70. [224] X.-F. Wang, D.-S. Huang, A novel density-based clustering framework by using level set method, IEEE Trans. Knowl. Data Eng. 21 (2009) 1515–1531. [225] P. Premaratne, F. Safaei, Q. Nguyen, Moment invariant based control system using hand gestures, Intell. Comput. Signal Process. Pattern Recognit., (2006) 322–333. [226] C.-Y. Lu, D.-S. Huang, Optimized projections for sparse representation based classification, Neurocomputing 113 (2013) 213–219. [227] P. Premaratne, C.C. Ko, Retrieval of symmetrical image blur using zero sheets, IEE Proc. Vision, Image Signal Process 148 (2001) 65–69. [228] V. Pomponiu, D. Cavagnino, M. Botta, Data hiding in the Wild: where computational intelligence meets digital forensics, in: Surveill. Action, Springer, 2018, pp. 301–331. [229] H.A. Atee, R. Ahmad, N.M. Noor, A.M.S. Rahma, Y. Aljeroudi, Extreme learning machine based optimal embedding location finder for image steganography, PLoS One 12 (2017) e0170329. [230] L.S. Nair, L.M. Joshy, An improved image steganography method with SPIHT and arithmetic coding, in: Proceedings of the 3rd International Conference on Frontiers of Intelligent Computing: Theory and Applications (FICTA, Springer, 2015, pp. 97–104. [231] Y.-G. Fu, R.-M. Shen, L.-P. Shen, X.-S. Lei, Reliable information hiding based on support vector machine, Informatica 16 (2005) 333–346. [232] H.-H. Tsai, D.-W. Sun, Color image watermark extraction based on support vector machines, Inf. Sci. (Ny). 177 (2007) 550–569. [233] S. Uma Maheswari, D. Jude Hemanth, Performance enhanced image steganography systems using transforms and optimization techniques, Multimed. Tools Appl. 76 (2017) 415–436, doi:10.1007/s11042- 015- 3035- 1. [234] P.D. Shah, R.S. Bichkar, A secure spatial domain image steganography using genetic algorithm and linear congruential generator, in: Int. Conf. Intell. Comput. Appl., Springer, 2018, pp. 119–129. [235] R. Roy, S. Laha, Optimization of stego image retaining secret information using genetic algorithm with 8-connected PSNR, Procedia Comput. Sci. 60 (2015) 468–477. [236] D. Bandyopadhyay, K. Dasgupta, J.K. Mandal, P. Dutta, V.K. Ojha, V. Snášel, A framework of secured and bio-inspired image steganography using chaotic encryption with genetic algorithm optimization (CEGAO), in: Proceedings of the Fifth International Conference on Innovations in Bio-Inspired Computing and Applications, Springer, 2014, pp. 271–280. [237] A. Miri, K. Faez, Adaptive image steganography based on transform domain via genetic algorithm, Optik (Stuttg) 145 (2017) 158–168, doi:10.1016/j.ijleo. 2017.07.043.

325

[238] M. Kaur, M. Juneja, Adaptive block based steganographic model with dynamic block estimation with fuzzy rules, in: Innovations in Computer Science and Engineering, Springer, 2017, pp. 167–176. [239] I. Shafi, M. Noman, M. Gohar, A. Ahmad, M. Khan, S. Din, S.H. Ahmad, J. Ahmad, An adaptive hybrid fuzzy-wavelet approach for image steganography using bit reduction and pixel adjustment, Soft Comput 22 (2018) 1555–1567. [240] C. Vanmathi, S. Prabu, Image steganography using fuzzy logic and chaotic for large payload and high imperceptibility, Int. J. Fuzzy Syst. 20 (2018) 460–473. [241] K. Koptyra, M.R. Ogiela, Multiply information coding and hiding using fuzzy vault, Soft Comput. (2018) 1–10. [242] S. Sajasi, A.-M.E. Moghadam, A high quality image steganography scheme based on fuzzy inference system, in: Proceedings of the 2013 13th Iranian Conference on Fuzzy Systems, IEEE, 2013, pp. 1–6. [243] S. Alam, T. Ahmad, M.N. Doja, A novel edge based chaotic steganography method using neural network, in: Proceedings of the 5th International Conference on Frontiers in Intelligent Computing: Theory and Applications, Springer, 2017, pp. 467–475. [244] S. Husien, H. Badi, Artificial neural network for steganography, Neural Comput. Appl. 26 (2015) 111–116. [245] M. Islam, A. Roy, R.H. Laskar, Neural network based robust image watermarking technique in LWT domain, J. Intell. Fuzzy Syst. 34 (2018) 1691–1700. [246] A. Zear, A.K. Singh, P. Kumar, Robust watermarking technique using back propagation neural network: a security protection mechanism for social applications, Int. J. Inf. Comput. Secur. 9 (2017) 20–35. [247] A. Rai, H.V. Singh, Machine learning-based robust watermarking technique for medical image transmitted over LTE network, J. Intell. Syst. 27 (2018) 105–114. [248] R. Tanwar, S. Malhotrab, Scope of support vector machine in steganography, in: IIOP Conference Series: Materials Science and Engineering, IOP Publishing, 2017, p. 12077. [249] A. Rai, H.V. Singh, SVM based robust watermarking for enhanced medical image security, Multimed. Tools Appl. (2017) 1–14. [250] O. Fındık, M. Bayrak, I.˙ Babaog˘ lu, E. Çomak, Color image watermarking scheme based on efficient preprocessing and support vector machines, in: Proceedings of the International Conference on Intelligent Computing, Springer, 2008, pp. 398–406. [251] M. Islam, R.H. Laskar, Geometric distortion correction based robust watermarking scheme in LWT-SVD domain with digital watermark extraction using SVM, Multimed. Tools Appl. (2017) 1–28. [252] B. Li, D.-S. Huang, C. Wang, K.-H. Liu, Feature extraction using constrained maximum variance mapping, Pattern Recognit 41 (2008) 3287–3294. [253] Z.-L. Sun, D.-S. Huang, Y.-M. Cheung, J. Liu, G.-B. Huang, Using FCMC, FVS, and PCA techniques for feature extraction of multispectral images, IEEE Geosci. Remote Sens. Lett. 2 (2005) 108–112. [254] C.-C. Chang, Y.-H. Yu, Y.-C. Hu, Hiding secret data into an ambtc-compressed image using genetic algorithm, in: Proceedings of the 2008 Second International Conference on Future Generation Communication and Networking Symposia, IEEE, 2008, pp. 154–157. [255] C.-C. Chang, Y.-H. Chen, C.-C. Lin, A data embedding scheme for color images based on genetic algorithm and absolute moment block truncation coding, Soft Comput 13 (2009) 321–331. [256] A.K. Alvi, R. Dawes, Image steganography using fuzzy domain transformation and pixel classification, in: Proceedings of the SEKE, 2013, pp. 277–282. [257] S. Kiani, M.E. Moghaddam, Fractal based digital image watermarking using fuzzy C-mean clustering, in: Proceedings of the 2009 International Conference on Information Management and Engineering, IEEE, 2009, pp. 638–642. [258] N.N. El-emam, Embedding a large amount of information using high secure neural based steganography algorithm, Int. J. Inf. Commun. Eng. (2008) 4. [259] D.C. Lou, J.L. Liu, M.C. Hu, Adaptive digital watermarking using neural network clustering technique, in: Proceeding 16th IPPR Conference on Computing Vision, Graph. Image Process., 2003. [260] H.A. Ghaleb Al-Jbara, L.B. Mat Kiah, H.A. Jalab, Increased capacity of image based steganography using artificial neural network, in: AIP Conference Proceedings, AIP, 2012, pp. 20–25. [261] S. Haykin, Neural networks: a Comprehensive foundation, 1999, Mc Millan, New Jersey, 2010. [262] D.-S. Huang, H.H.-S. Ip, K.C.K. Law, Z. Chi, Zeroing polynomials using modified constrained neural network approach, IEEE Trans. Neural Networks 16 (2005) 721–732. [263] D.-S. Huang, W.-B. Zhao, Determining the centers of radial basis probabilistic neural networks by recursive orthogonal least square algorithms, Appl. Math. Comput. 162 (2005) 461–473. [264] L. Shang, D.-S. Huang, J.-X. Du, C.-H. Zheng, Palmprint recognition using FastICA algorithm and radial basis probabilistic neural network, Neurocomputing 69 (2006) 1782–1786. [265] F. Han, D.-S. Huang, A new constrained learning algorithm for function approximation by encoding a priori information into feedforward neural networks, Neural Comput. Appl. 17 (2008) 433–439. [266] F. Han, Q.-H. Ling, D.-S. Huang, Modified constrained learning algorithms incorporating additional functional constraints into neural networks, Inf. Sci. (Ny). 178 (2008) 907–919. [267] D.-S. Huang, Systematic Theory of Neural Networks For Pattern Recognition, Publ. House Electron. Ind., China, Beijing, 1996, p. 201.

326

I.J. Kadhim, P. Premaratne and P.J. Vial et al. / Neurocomputing 335 (2019) 299–326

[268] V. Hajduk, D. Levický, Cover selection steganography, in: Proceedings of the ELMAR - International Symposium Electronic Mar., 2016, pp. 205–208, doi:10. 1109/ELMAR.2016.7731787. [269] R.-E. Yang, Z.-W. Zheng, W. Jin, Cover selection for image steganography based on image characteristics, Guangdianzi Jiguang/Journal Optoelectron. Laser. 25 (2014) 764–768. https://www.scopus.com/ inward/record.uri?eid=2- s2.0- 84899711801&partnerID=40&md5= 27c4b2bd3c298987ff3279e014929b6a. [270] H. Sajedi, M. Jamzad, Cover selection steganography method based on similarity of image blocks, in: Proceedings of the 2008 IEEE 8th International Conference on Computer and Information Technology Workshops, 2008, pp. 379– 384, doi:10.1109/CIT.2008.Workshops.34. [271] M.S. Subhedar, V.H. Mankar, Curvelet transform and cover selection for secure steganography, Multimed. Tools Appl. (2017) 1–24, doi:10.1007/ s11042- 017- 4706- x. [272] S. Wu, Y. Liu, S. Zhong, Y. Liu, What makes the stego image undetectable? in: Proceedings of the 7th International Conference on Internet Multimedia Computer Service, ACM, 2015, p. 47. [273] V. Hajduk, D. Levicky, Accelerated cover selection steganography, in: Proceedings of the 2017 27th International Conference on Radioelektronika, RADIOELEKTRONIKA 2017, 2017, doi:10.1109/RADIOELEK.2017.7937591. [274] A. Nissar, A.H. Mir, Classification of steganalysis techniques: a study, Digit. Signal Process. A Rev. J. 20 (2010) 1758–1770, doi:10.1016/j.dsp.2010.02.003. [275] R. Chandramouli, M. Kharrazi, N. Memon, Image steganography and steganalysis: Concepts and practice, Lect. Notes Comput. Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics). 2939 (2004) 35–49. https://www.scopus.com/inward/record.uri?eid=2-s2. 0-33847692468&partnerID=40&md5=2497e8ae07c9cdce8082d05991799893. [276] H. Zong, F. Liu, X. Luo, Blind image steganalysis based on wavelet coefficient correlation, Digit. Investig. 9 (2012) 58–68. [277] B.T. McBride, G.L. Peterson, S.C. Gustafson, A new blind method for detecting novel steganography, (2005). Inas Jawad Kadhim received her Bachelor degree in Electrical Engineering and Engineering Education in 2003 from University of Technology, Baghdad, Iraq. In 2005, she received her Master degree in Telecommunication Engineering from University of Technology, Baghdad, Iraq. She is currently pursuing the PhD degree with the School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, Australia. Her research interests include data hiding, image processing and machine learning.

Prashan Premaratne was born in Sri Lanka in 1972 and was awarded an Australia government scholarship to pursue undergraduate studies at the University of Melbourne, Australia in 1994. He graduated with a Bachelor of Engineering (Electrical and Electronics) from the Department of Electrical and Electronics at the University of Melbourne in 1997. He also won few scholarships for postgraduate studies and graduated with a PhD in Electrical and Computer Engineering from the National University of Singapore in 2001. He was a Software Engineer at the Fujitsu Singapore Limited from 1998 to 1999 and worked as a Research Engineer after completing his PhD in 2001. In September 2001, he joined the Corporate Research Centre for Sensor Signal and Information Processing in Adelaide, Australia as a Research Fellow to develop a Ship Classification Project for Australia Defence establishment. Since 2003, he has been an academic at the University of Wollongong, Australia and is a Senior lecturer at the School of Electrical, Computer and Telecommunications Engineering. Dr. Premaratne is a Senior Member of IEEE and is the author of the book “Human Computer Interaction Using Hand Gestures” published by Springer International in 2014. He is also an Assistant Editor of Springer Journal of Cognitive Science. Peter James Vial is currently a Senior Lecturer at the School of Electrical, Computer and Telecommunications Engineering at Wollongong University. Dr Vial completed his Bachelor of Electrical Engineering in 1986, working as an Electrical Engineer at the Port Kembla Steelworks in Wollongong (Australia) until 1991. In 1992 he became a Teaching Fellow at Wollongong University which was reclassified to Associate Lecturer. In 1996 he received his Masters in Telecommunications from Wollongong University, and in 20 0 0 he received a Diploma in Education (Mathematics). In 2004 he was promoted to Lecturer and in 2009 he received his PhD from Wollongong University. In 2016 he was promoted to Senior Lecturer. He has been involved in developing and teaching both postgraduate and undergraduate engineering laboratories. His main research interest is in wireless communications systems, especially related to Ultra Wideband systems. He maintains a keen interest in engineering education and is a Senior Member of the esteemed Institute of Electrical and Electronic Engineers (IEEE) based on his contribution to engineering education at Wollongong University since 1992. He has been nominated twice for the Vice-Chancellor’s Awards for Outstanding Contribution to Teaching and Learning at the University of Wollongong by his students. Brendan Halloran obtained his Bachelor of Engineering degree in 2016 from the University of Wollongong, where he is currently pursuing his PhD in distributed processing for robotic vision with the School of Electrical, Computer and Telecommunications Engineering. His main research interests are probabilistic graphical modelling and SLAM.