2.1 Image steganography

Information hiding in images has gained popularity in recent times [35]. Images have become important carriers to hide secret messages without changing the visual features and/or properties. As a result, images have become popular and widely used for steganography due to the degree of redundancy associated with them [36]. All image file formats are suitable for image steganography. File format types including TIFF, JPEG, PNG, GIF, and BMP are all appropriate to use. [37]. It is worth noting that each image file format has its advantages and disadvantages when employed for steganography purposes. Given that pixel values are utilized for image steganography, variations in pixel intensities between the original cover image and stego-images are sometimes experienced. The intensity variation is nonetheless subtle such that the undetectability and imperceptibility to the human visual system is achieved [38,39].

The commonality of images for steganography has subjected images to several targeted cyber-attacks including visual and statistical steganalysis attacks [40]. These attacks possess the ability to unearth concealed messages within images using steganalysis algorithms. Statistical steganalysis capabilities aimed at revealing hidden data in images include detection, extraction, disabling, and destruction of hidden data [41]. Tools and techniques used for such capabilities include lossy compression, denoising, image enhancement techniques, image approximation techniques, and geometrical modification [35]. These tools and techniques expose the vulnerabilities of image steganography on the digital landscape, rendering images a prime target of cybercriminal activities.

Image steganography uses three main traditional approaches (i.e., spatial domain, transform domain, and adaptive domain) to embed data [42]. The spatial domain approach entails the direct embedding of secret messages into image pixel values. This approach encompasses numerous techniques including the least significant bit (LSB) insertion algorithm [43–45], quantization-based methods [46], histogram-based methods [47], prediction error [48], modulo operations [49], and many other variations. Spatial domain methods have the advantages of high visual quality with minimal distortion effects, and high embedding payload capacity [38]. However, the spatial domain is less robust, making it susceptible to various forms of manipulation and attacks [38].

Given the challenges associated with spatial domain approaches, transform domain techniques emerged as a compelling alternative for secret data embedding [50]. The transform domain utilizes frequency sub-band coefficients to insert the secret message bits [51,52]. Although the data embedding and extraction processes are intricate compared to the spatial domain, this approach bolsters system security [50]. This embedding technique possesses the capability to withstand data manipulation approaches such as cropping, scaling, compression, and rotation. Some existing transform domain algorithms include Discrete Cosine Transform (DCT) [51], Discrete Fourier Transform (DFT) [53], Integer Wavelet Transform (IWT) [54], and Discrete Wavelet Transform (DWT) [55] among others. This method offers competitive advantages over spatial domain approaches by enhancing the robustness of the steganographic applications. However, both spatial and transform domain approaches have limitations [56], particularly regarding the susceptibility of the cover image to data manipulation and modification. Notwithstanding these limitations, spatial domain methods such as LSB Insertion algorithm and Pixel Value Differencing (PVD) remain the most prevalent data embedding techniques for steganographic applications [57]. The spatial domain method alters the LSBs of the carrier image by directly replacing the LSBs of the original cover image with the secret message bits, while transform domain randomizes all the bits in the carrier image [58].

Considering the intricacies associated with spatial and transform domains, the adaptive domain method also known as the model-based method or masking has surfaced. This method employs dynamic techniques for pixel selection for data embedding and estimating an allowable number of bits that can be hidden within the carrier object [50]. Examples of this method include artificial intelligence, blockchain technology, machine learning, and genetic algorithms. Recent innovations have seen the implementation of biometric techniques and facial recognition technologies for image steganography, contributing to the security enhancement and robustness [59–63]. Adaptive techniques have a comparative advantage over spatial and transform domains due to their robustness and the ability to avoid detection by statistical steganalysis attacks. This method is also able to efficiently balance the tradeoffs between embedding capacity and security. The trade-off high embedding capacity on one side and security and robustness improvement on another side, remains a challenge in image steganography applications, for which constant innovations are required.

2.2 Statistical steganalysis attacks

Steganalysis techniques undermine the security capabilities of steganography, as they detect messages concealed in images to reveal the message and estimate the size/length. Given that image steganography has gained prominence for secret information hiding, image steganalysis emerges as a countermeasure. Image steganalysis exploits image processing techniques such as cropping, filtering, and blurring to detect, extract, disable, or destroy hidden information within cover objects [64]. Steganalysis algorithms are extant, some of which include pixel difference histogram (PDH) analysis, sample-pair analysis, RS analysis, and Chi-square (X²) analysis [58] among others. RS steganalysis can detect LSB-based substitution stego-images, whereas Chi-square analysis which is based on a statistical distribution of binary values (0s and 1s) can determine if the image intensities follow random or distributed patterns. Statistical steganalysis process extracts the statistical characteristics of an image to accurately detect and estimate the exact size of hidden messages within a stego image [65]. By so doing, the hidden information is unveiled, and their length estimated. This breaches the confidentiality requirement of data transmission. All types of steganalysis possess the capability to identify, detect, and extract secret information hidden within a carrier object. For instance, PDH analysis can analyze and detect PVD-based image steganography. The analysis focuses on searching for the algorithm employed for the secret message concealment.

Chi-Square (X²) statistical steganalysis was proposed by Westfeld and Pfitzmann [66] with the ability to detect sequentially embedded messages within an image. This approach, however, could not identify the presence of hidden messages based on random embedding. Notably, Provos [67] improved the technique proposed by Westfeld and Pfitzmann [61] to have the ability to detect and estimate both sequentially and randomly hidden messages. The sample-pair technique proposed by Dumitrescu et al., [68], is also another effective approach to detecting hidden messages based on LSB steganographic hiding process. Among the various types of statistical steganalysis, the RS attack developed by Fridrich et al. [69] is the most effective and well-known steganalysis technique which possess the capability to detect and reveal secret messages embedded within an image. RS steganalysis technique detects both sequential and random embedded secret messages. Statistical attack techniques adeptly differentiate stego-images containing secret messages from cover images. This is done by mathematically investigating the relationship that exists between adjacent pixel groups and the pixel values of the stego-image, and the cover image [70]. Following the earlier work by Fridrich et al. [69], several steganalysis techniques with improved performance and detection capabilities have emerged [65–69,71–77]. The growing sophistication, complexity, and accuracy performance of steganalysis techniques have meant that a more secure image steganography scheme is required.

2.3 Previous/Related works

Empirical studies providing systematic review on image steganography techniques and methods aimed at resisting statistical steganalysis attacks are limited. Existing studies have failed to provide detailed empirical-based discussions on issues related to image steganography techniques and lacked a standardized methodology for reviewing the selected publications/articles. Ashwin et al., [78] conducted a review of image steganography techniques as well as steganalysis techniques capable of detecting secret information hidden in images. The study identified research trends, challenges, methods, and techniques for image steganography. Although Ashwin et al., [78] study provided early perspectives to scholars on existing techniques for resisting steganalysis attacks, the study was limited to only two embedding process approaches (i.e., spatial and transform). The study failed to provide broader insights into other notable techniques and algorithms dominating the field. The study also failed to adopt a standardized methodology for conducting the literature review. Subhedar and Mankar [79] focused on the issues and challenges of image steganography. The study provided key insights on image steganography performance evaluation metrics and explored various challenges that confront image steganography whose data embedding processes are based on spatial and transform domains. The study identified steganalysis techniques as key issues affecting the efficiency of steganography and provided future research direction. This study was however not systematic, as methods for selecting literature were not defined. The study also failed to discuss how existing techniques have performed against universal statistical steganalysis such as RDH and RS attacks.

Kadhim et al., [80] provided a review of image steganography techniques. The study discussed performance evaluation metrics as well as future research trends in the field of image steganography. The study provided key insights to researchers on the trends of digital image steganography but failed to provide a broader and comprehensive systematic review of key algorithms dominating the field. Standard methods were not applied in the selection of literature for the survey review. Mandal et al., [81] provided a review of digital image steganography tools available for embedding secret messages. The survey provided some image steganography techniques including adaptive and deep learning techniques and offered some key examples of some popular steganography tools. Comparison of the various tools were provided. Challenges of deep learning-based steganography were also enumerated. The study failed to adopt a standardized methodology for conducting the literature review and did not provide a comprehensive insight into all existing image steganography techniques/approaches. The study was limited to spatial and transform domain methods. Perhaps, the most comprehensive study and closely related to this paper is a systematic literature review conducted by Kaur et al., [50]. Kaur et al., [50] adopted standardized systematic literature review guidelines and selected 61 pieces of literature from four key databases comprising Web of Science, IEEE, Wiley, and ACM. The studies selected were published from 2011 to 2022. The results of the study show that extensive milestones for image steganography techniques have been achieved. Progress in all three data embedding processes (ie spatial, transform, and adaptive approaches) has seen notable improvement. The study further revealed that future research could focus on enhancing and striking an adequate balance between embedding capacity and robustness.

Other existing reviews focused on some specific domains within image steganography, further limiting the scope of the application of techniques for resisting statistical steganalysis. For example, Hussain et al., [82] provided a review on image steganography focusing on spatial domain techniques. The study highlighted some novel spatial domain techniques for image steganography including challenges and trends. Girdhar and Kumar [83] also provided a review of steganography techniques based on 3D images. Various 3D domain techniques including topological, geographical, and representation domains were discussed and compared in terms of payload capacity, resistance to attacks, and reversibility. Meng et al., [84] reviewed deep learning algorithm-based image steganography techniques. Various deep-learning algorithms were surveyed and discussed. Deep-learning algorithms used for coverless information hiding, steganalysis attacks, and watermarking were extensively presented and discussed. Qin et al., [85] comprehensively reviewed coverless image steganography techniques. The review provided a framework description of methods and techniques for coverless image steganography, highlighted recent developments in the area, and concluded that coverless image steganography provides resistance against steganalysis attacks.

Also, Puteaux et al., [86] focused their survey on reversible image steganography techniques. Techniques and methods compared included pixel value differencing or histogram shifting, re-echoing-based steganography, public key cryptography-based methods, prediction-based methods, and image partition-based techniques. Aslam et al., [87] conducted a review LSB based image steganography techniques. The review sampled 20 research studies published from 2016 to 2020. The 20 articles were further scaled down to 17 for the review. 20 data sets were identified for the evaluation of image steganography techniques. All the domain-specific studies reviewed [82–86] could not be conveniently classified as a systematic literature review except Aslam et al., [87]. The studies failed the threshold for systematic literature review when compared to the guidelines provided by Kitchenham and Charters [88]. The methods adopted for the study selection including inclusion and exclusion criteria, datasets, databases, data extraction methods, and queries were not detailed.

The above review works discussed may not be exhaustive for review research on image steganography techniques capable of resisting statistical steganalysis. However, the extensive literature search conducted in the most relevant scientific databases and libraries provided little evidence of a systematic literature review for image steganography techniques. The identified knowledge gap and other germane issues are the focus of this review. This research, therefore, seeks to conduct investigations into the literature on image steganography techniques capable of resisting statistical steganalysis attacks. By so doing, the review brings to the fore relevant studies on image steganography methods for resisting statistical steganalysis to bridge and/or expose the knowledge gap.

3.1 Research approach

The PRISMA guidelines were chosen to ensure the review process is transparent, clear, and credible [92]. The processes involved in PRISMA include defining the systematic scoping review, identifying potential studies through literature searches in relevant databases and electronic libraries using predefined keywords, abstract screening, selecting papers based on inclusion and exclusion criteria, article characterization, and mapping based on keywords and meta-analysis [93]. Based on the PRISMA guidelines, a data selection, extraction, and classification taxonomy were developed and implemented. The taxonomy defined review questions, literature search strategy, eligibility criteria for inclusion and exclusion, data analysis framework, and criteria for resolving opinion disparities among researchers.

3.2 Review research questions and protocol

Kitchenham and Charters [88] argued that review questions and review protocols are important components of the systematic literature review process as they reduce the researcher’s biases and provide a critical framework to guide acceptable systematic reviews. Review questions are formulated during the initial stages of study planning to situate the study goals as the foundation upon which the study hinges [93]. This study adopted the Goal-Question-Metric approach suggested by Caldiera and Rombach [94] (See Table 1 for the Goal-Metric Questions). This Goal-Question-Metric has previously been used by Lun et al., [95] and Wiafe et al., [96] as an efficient and effective approach for deriving systematic review objectives.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Adopted Goal-Question-Metric [94].

https://doi.org/10.1371/journal.pone.0308807.t001

Statistical steganalysis attacks are growing at a tremendous pace. As such, techniques and methods for steganography that could withstand such attacks have become topical. Questions such as the most used image steganography techniques for resisting steganalysis attacks, the performance and security impact of image steganography techniques, and future scope and research direction for techniques within the image steganography domain remain critical and unanswered concerns that require addressing. These knowledge gaps need to be addressed. The review questions, the reason behind the questions, and the research approach to achieve the questions are listed in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Formulated review questions and motivation.

https://doi.org/10.1371/journal.pone.0308807.t002

Following the formulation of the research questions and to further avoid biases in the literature search strategy, search terms and keywords, and study selection, the review protocol was separately developed by each of the members of the research team. The individual protocols were merged and further refined by the research team in a protocol development meeting. The merged protocol was refined, and the final protocol was adopted after an extensive review process and corrections. Fig 1 provides a detailed diagrammatic representation of the final protocol adopted for the study demonstrating the main review processes followed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Adopted review protocol for methodological analysis.

https://doi.org/10.1371/journal.pone.0308807.g001

3.3 Literature strategy

Brereton et al., [83] identified seven electronic databases as key for conducting exhaustive literature searches for studies within the information technology landscape and for software engineers specifically. These databases are IEEExplore, ACM Digital Library, Google Scholar, Citeseer Library, INSPEC, ScienceDirect, and EI Compendex. SCOPUS, Wiley Online, Web of Science (WOS), and Springer Link are also considered relevant electronic libraries [83]. Before the actual search, a preliminary search was conducted on Google Scholar, Citeseer, and SCOPUS to identify the most appropriate databases, search terms, and search period. Based on the preliminary search, four (4) databases (i.e., IEEE, ACM Digital, ScienceDirect, and Wiley Online) were chosen. These electronic databases and libraries were chosen because they had the most relevant published studies on image steganography techniques. The keywords and search terms used for the database searches were made up of two categories. The categories were Steganography and related words (steganography, image, image steganography) and Steganalysis and related words (Steganalysis, statistical steganalysis, RS steganalysis). The search phrases were developed by combining words from both categories using the “AND” Boolean Operator. After several searches in databases by the researchers, five search terms were perceived as appropriate based on the results from the preliminary search. These terms were (i) “Steganography” and “Steganalysis” (ii) “Image Steganography” and “Steganalysis” (iii) “Steganography” and “RS Steganalysis” (iv) “Image Steganography” and “Statistical Steganalysis” and (v) “Image” and “Statistical Steganalysis”. The search period was limited to 2012 to 2023 inclusive.

3.4 Eligibility criteria

For a publication to form part of this review, clear inclusion and exclusion criteria were defined. To be included, publications should have been written in English. Also, publications should have discussed image steganography and/or steganalysis attacks performance evaluation metrics. That is, publications whose titles related to image steganography and/or steganalysis attacks were included. Further, papers published from 2012 to 2023 were considered. Apart from these, only peer-reviewed publications were accepted. For the exclusion criteria, non-empirical studies were rejected. This suggests that point-of-view papers, review papers, and reports were excluded. Also, only peer-reviewed journal and conference papers were included. Book sections, chapters, posters, and thesis were excluded from the review. Moreover, publication abstracts that showed no relationship with the search terms were excluded. Publications whose content did not discuss how image steganographic techniques are employed to resist steganalysis attacks were removed. Lastly, publications ranked as low quality as agreed by the review team were excluded.

3.5 Study selection

Based on the search criteria, two (2) members of the review team performed independent searches using the identified search terms on all four (4) databases. For all searches, the search period was limited to 2012 to 2023 inclusive. The two (2) independent results were merged into one dataset. A total of 5146 publications were compiled. The dataset (n = 5146) was then screened to remove duplicates. After the duplicates were removed, 1379 publications remained. Next, the titles of the publications were scanned to determine their relatedness to the objectives of this review. For example, studies whose titles did not suggest any relation to image steganography techniques were removed. Next, the dataset was examined to maintain only journal and conference papers. Book sections, chapters, posters, and thesis were removed. Further, all non-empirical papers were discarded. This process reduced the total number of publications to 902. Reports were sought for retrieval and 13 reports were not retrieved. A total number of 889 records were maintained. After assessing the papers for eligibility, 736 papers were removed.

Two (2) members of the review team separately read the abstracts of the remaining publications (n = 153) to determine their relatedness to the search terms. The separate reports from the two (2) members were discussed by all members of the review team and merged. In cases of any disparities, a vote was conducted to resolve the issue. This activity further reduced the number of publications to 136. Lastly, two (2) other members of the review team read the content of the 136 publications to assess their quality. Their reports were also discussed and debated. Based on these discussions, 125 publications were retained as appropriate for review. Fig 2 provides a detailed summary of the selection process for the identified publications. Thus, 125 papers remained as final papers included in the systematic literature review. Also, a summary of the number of papers selected from the various electronic databases and the search terms is shown in Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. PRISMA flow diagram for publication selection process.

https://doi.org/10.1371/journal.pone.0308807.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Detailed record of articles selected for the systematic literature review.

https://doi.org/10.1371/journal.pone.0308807.t003

4.1 Publication trends

The selected publications were analyzed to understand the publication trends. The information recorded for this analysis included the year of publication, publication outlet, publication type, geographic origination of corresponding authors, and number of citations. The results show that publications on image steganography techniques for controlling statistical steganalysis attacks have increased considerably. For the year of publications, the results show fluctuations in the number of publications per year from 2012 to 2017 (see Fig 3). Since 2017, the number of publications per year increased tremendously. Articles published from 2018 to 2023 represented 73% of the total number of publications reviewed. This suggests a growing interest in image steganography studies for combatting steganalysis attacks. The analysis also shows an interesting result for the post-coronavirus Pandemic era (COVID-19), as approximately 49% of all articles were published from 2021 to 2023. This shows tremendous development of techniques against statistical attacks, following the numerous cyber-attacks, data breaches, and data compromises that were experienced during the peak of the COVID-19 lockdowns and global work-from-home phenomenon.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Yearly publication trends of reviewed studies.

https://doi.org/10.1371/journal.pone.0308807.g003

The results also indicated a skewed interest in publishing outlets. From the total of 125 papers reviewed, 66 (53%) were published with IEEE and 46 (37%) by ScienceDirect. Fig 4 indicates the breakdown of the trend by publication outlet. Further, the analysis of the publication types revealed most of the reviewed publications were journals (57%) (n = 125).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Publication trend by publication outlet.

https://doi.org/10.1371/journal.pone.0308807.g004

Similarly, the results were geographically skewed. The affiliations of the corresponding authors at the time of publication were used to extract the geographic originations of the papers. The majority (86%) of the reviewed papers (n = 125) originated from Asia followed by Europe (8%). India (43 of 125) and China (37 of 125) recorded the highest number of publications respectively. Fig 5 shows a summary of the geographical locations of all corresponding authors for the selected papers used for analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Publication trend by geographic location.

https://doi.org/10.1371/journal.pone.0308807.g005

The number of citations per paper at the time of this review was also analyzed. Majority (107 of 125) of the papers had 50 or lesser citations and only 8 had 100 or more. S1 Appendix shows the detailed list of the reviewed studies.

4.2 Image steganography techniques and methods

The review analyzed the methods and techniques that have been utilized in image steganography to resist statistical steganalysis attacks. Over 57 image steganography techniques and methods were identified. However, the techniques that have dominated image steganography studies are Modified LSB (M-LSB), LSB Matching (LSB-M), PVD, Genetic Algorithm (GA), GAN, CNN, DL Neural Networks, Hamiltonian Path (HP), Adaptive Edge Detection (AED), RDH, Residue Number System (RNS), DCT, IWT, among many others have been identified in literature as improving the imperceptibility of image steganography. Some of these methods have been implemented alone or sometimes with a combination of two of the methods enumerated. Others combined the methods with LSB and cryptographic protocols such as AES, RSA, and Elliptic Curve Cryptography (ECC) for encryption and decryption to enhance data security. As a result, many combinations of the above-mentioned techniques exist. The techniques and methods showed the capacity to enhance the visual quality of the carrier image and proved to be secure against statistical steganalysis attacks.

Fig 6 shows that GAN (17) is the most adopted technique. This is followed by AED (14). A total of 20 studies implemented a version of LSB comprising M-LSB (4), LSB-M (10), and LSB plus others (6). GA, RDH, and PVD were each implemented in 9 studies. The techniques that were used by less than two publications were grouped as “Others”. Fig 6 gives details of the number of times other methods were utilized. Table 4 also gives a breakdown detail of the trend in publication year and techniques implemented. As already mentioned, the embedding process for image steganography techniques can be classified into three domains ie (i) Spatial Domain-Based Techniques, (ii) Transform Domain-Based Techniques, and (iii) Adaptive Domain-Based Techniques. The review results reveal that spatial domain-based image steganography techniques have attracted more attention, as approximately 43% of all the reviewed papers utilized spatial domain for the secret data embedding process. This is followed by adaptive techniques, where 38% of reviewed papers employed such techniques. The rest of the studies used transform domain image steganography techniques (19%) (See Table 4). Further analysis of the review was conducted to understand the application of the image steganography techniques and the primary embedding domain employed for data hiding. This was necessary to observe the trend of specific techniques within each domain of application.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Image steganography techniques and methods for resisting attacks.

https://doi.org/10.1371/journal.pone.0308807.g006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Publication and image steganography embedding domains (2012 to 2023).

https://doi.org/10.1371/journal.pone.0308807.t004

The results as presented in Table 5 show that the spatial domain was the primary data embedding process for M-LSB, LSB-M, PVD, HP, LSB+Others, and AED. Also, almost all papers whose techniques were based on GA, GAN, DL, and CNN utilized the adaptive domain as the primary process of data embedding. Similarly, for DCT and IWT techniques, the transform domain method was mainly used. For RNS and RDH techniques, the domain for data embedding process was varied, whereas most of the other studies employed spatial domain and adaptive domain for the embedding process. The implication is that the spatial domain has gained wide application in use for image steganography, perhaps due to its advantage of high embedding payload capacity. Table 6 shows the trends in the year of publication versus image steganography techniques.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Embedding domains verses image steganography techniques.

https://doi.org/10.1371/journal.pone.0308807.t005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Image steganography techniques for resisting steganalysis attacks (2012 to 2023).

https://doi.org/10.1371/journal.pone.0308807.t006

4.3 Performance evaluation metrics for image steganography techniques

The implementation of image steganography is aimed at achieving some key objectives. The key objective parameters are high embedding payload capacity, imperceptibility (visual quality of resulting stego-image), robustness (distortion resistance), and security (un-detectability) among others. However, there is a trade-off between the performance evaluation parameters as most of the parameters result in opposite impacts with each other. For instance, techniques proposed to achieve high hiding capacity result in image distortion that ultimately reduces security and data protection. To achieve the objectives of image steganography techniques, various evaluation metrics are utilized. To measure imperceptibility, many studies have used Mean Square Error (MSE) [97], Peak-Signal-to-Noise-Ratio (PSNR) [98,99], Segmented Signal-to-Noise-Ratio (SNRseg) [100] and/or Signal-to-Noise-Ration (SNR) [101]. Also, Pearson Correlation Coefficient (NC) [102], Correlation Factor (r) [103], and Structural Similarity Index Measure (SSIM) [104,105] are used to measure the similarity between the cover image and the stego image to determine the image quality matrix. Bit Error Rate (BER) [106] is often used to measure the image distortion resistance, whereas Regular-Singular (RS) analysis [107,108] has proven effective in analyzing the detectability of the image steganography techniques against steganalysis attacks. Given that high embedding capacity is a key evaluation metric for image steganography techniques, Bits Per Pixel (BPP) is often used [109]. The dominant performance evaluation metrics for the reviewed papers, are PSNR, MSE, NC, SSIM, BPP, and RS analysis. The most used evaluation metrics are discussed below. However, the performance metrics used by each reviewed paper will be reported to ensure standardization and quality metrics comparison.

Imperceptibility is an important criterion in steganography [50]. Distortions between the original cover image (CI) and the resulting stego image (SI) must be relatively low to ensure higher imperceptibility of the image against attacks. Image Quality Measurement (IQM) is a mathematical approach to determining the quality of SI. When a secret message is embedded in the original selected CI, changes are noticed in the pixel values of the CI. Such changes affect the quality of the resulting SI. It is important to measure the changes in pixel values to ensure the SI is imperceptible. PSNR measures the distortion between CI and resulting SI. PSNR is determined using Eq 1 written as [98]: (1)

MAX_I. represents maximum oixel value, whereas the MSE is Mean Square Error. The MSE measures of noticeable distortion between CI and SI. MSE is determined using Eq 2 [97]: (2)

M and N represent the image height and width respectively. The lower the values obtained for MSE, the less distorted the difference between the CI and SI. Also, the higher the PSNR value, the higher the visual quality, thus higher imperceptibility.

Robustness of image the steganography technique proves that it is distortion resistant. To ensure that the technique is resistant to distortion, the similarity between the CI and SI is checked to determine whether the image has been distorted after embedding the secret message. SSIM is an important metric to check the structural similarity between the original CI and the resulting SI. The SSIM metric is calculated using Eq 3, and written as [103]: (3) Where c₁ = (k₁, L)² and c₂ = (k₂, L)². μx and μy are the CI and SI mean intensity. The variances of x and y are represented ð²_x and ð²_y respectively, whereas ð_xy represents the covariance of x and y. the pixel values varying range is denoted by L, and the constant parameters are represented by c₁ and c₂. k₁ and k₂ values are always to taken to be 0.01 and 0.03 respectively. The NC also checks the distortion resistance between CI and SI. NC computes the degree correlation between the CI and SI, is determined using mathematical Eq 4 as [102]: (4) Where X is the CI, Y is the SI, Ẋ is the mean pixel intensity values for the CI, and Ȳ is the mean pixel intensity values for the SI. Fundamentally, the image steganography technique aims to avoid statistical steganalysis attacks. As a result, one key parameter in the design is undetectability. Steganalysis attacks can have access to the data in transmission, thereby breaking the data confidentiality parameter. As already mentioned, Regular-Singular (RS) attacks are some of the well-known attacks. RS analysis is therefore performed to ensure the technique developed can resist statistical attacks. RS analysis is defined over three kinds of block flipping. The block flipping are positive flippings (F₁), negative flippings (F_-1), and Zero (0) flippings (F₀). F₁, F_-1, and F₀ become flipping functions and form what is termed a flipped group. The flipped group results from applying the flipping functions on each divided image block pixel value. Eq 5 is for determining the flipped group function [70]. (5) Where M = M (1), M (2), …, M (n) represents the flipped mask, and M (i) has values indicating either 1, 0, or -1. G is regular if f (G) < f (F(G)) otherwise G is singular when f (G)>f (F(G)). The implementation requires first dividing the image into non-overlapping blocks and re-arranging each one of them into a vector G = (X₁, X₂, X₃, …Xn). The blocks are arranged in a zigzag scan order. The discrimination function of the pixel’s correlation is measured using Eq 6 [70]: (6)

The pixel values are represented by x and n is used to represent the number of pixels Also, f represents partial correlation between the adjacent pixels. A smaller f value means a stronger correlation exists between adjacent pixel values. Payload capacity is an important measure for image steganography techniques. An algorithm for image steganography should be able to embed maximum secret messages without noticeable distortion. The overall effect, embedding the maximum payload capacity within the pixel values of the selected CI must be possible without distorting the visual quality of the resulting SI. Basically, the number of secret bits that have been hidden in the CI is the embedding payload capacity, which is calculated using BPP as shown in Eq 7 and written as [108]: (7) Where M and N are the CI cardinality, and embedding capacity (EC) which refers to the number of secret bits that can be embedded within total CI pixel values is determined using Eq 8 [70]: (8)

4.4 Performance metrics analysis

The performance evaluation metrics for all 125 reviewed papers are provided. The analysis covers the techniques employed, strengths, limitations, and results obtained in each reviewed paper. The problems or issues often discussed in image steganography research are diverse. Concerns such as the tradeoffs between embedding capacity and security, statistical attacks against image steganography systems, stego image distortion, low embedding capacity, and low visual image quality of stego images remain some key challenges and issues that are generally raised and discussed within the image steganography domain. As a result, most techniques are proposed to address these challenges. The analysis also covers the issues and problems discussed by the various articles that warranted the proposed techniques and methods. The reviewed papers are grouped according to the primary embedding process adopted. Table 7 covers papers based on Spatial Domain-Based Techniques, Table 8 covers papers based on Transform Domain-Based Techniques, and Table 9 is based on Adaptive Domain-Based Techniques. The evaluation metric indicated in each reviewed paper is reported.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Spatial domain-based image steganography techniques.

https://doi.org/10.1371/journal.pone.0308807.t007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Transform domain-based image steganography techniques.

https://doi.org/10.1371/journal.pone.0308807.t008

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Adaptive domain-based image steganography techniques.

https://doi.org/10.1371/journal.pone.0308807.t009

In order to compare the superiority of each of the methods mentioned in Tables 7–9 over other methods listed and to demonstrate the efficiency of each method through the approved standards (ie Payload Capacity, measured in Bit Per Pixel (BPP) and Imperceptibility using Peak Signal to Noise Ratio (PSNR) and measured in decibel (dB)), a graphical representation is provided. See Fig 7.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Comparison of embedding capacity and security of image steganography techniques.

https://doi.org/10.1371/journal.pone.0308807.g007

4.5 Security analysis of image steganography techniques

The security impact analysis examines the various identified techniques using some key parameters. Section 4.4 has already provided a detailed review of all the 125 publications retained for this study, which are presented along with their strengths and limitations. However, some other key indicators are relevant to determine how the various existing techniques can provide robustness and resistance against attacks and their overall security. This will also enable comparison among the reviewed papers using common standard metrics and parameters. The indicators assessed in this section include the image dataset employed for the experiment, type of data embedding process, data embedding style, secret image type, real-time implementation of a proposed algorithm or technique, application of cryptography protocol (encryption), data compression, values obtained for the PSNR, robustness against steganalysis attacks and the overall security of each technique.

Table 10 provides a detailed comparison of the various existing techniques reviewed which used grayscale images for the experiment whereas Table 11 provides a detailed comparison of the various existing techniques reviewed which used color images. The reviewed articles show that four benchmark datasets consisting of BOSS base, USC-SIPI, Seam Carving Original Q75, and 24 KODAK image Databases have widely been used. These databases contained specific images. The specific image dataset used by each reviewed article is reported. The data-hiding process is divided into spatial, transform, and adaptive domains. The data embedding style is divided into random and sequential. For secret image type, the categorizations are color or grayscale. Yes or no is used to represent whether the respective technique implemented the algorithm in real-time, whether encryption was applied to the secret data, and whether the secret data was compressed. Robustness against steganalysis attacks is divided into high, medium, and low. The specific parameters considered for the robustness are embedding process and style, secret image type, and encryption. Techniques that fully satisfy the evaluation criteria of the researchers considering the key parameters are rated high, those that partially satisfy are rated medium and those that least satisfy are rated low. Security of the reviewed articles is divided into good, average, and low. The overall security is evaluated by taking into consideration all the parameters previously discussed, most importantly PSNR values, Encryption, Real-time implementation, Compression, and embedding process. Other parameters discussed in section 4 (4.4) were also taken into consideration. The techniques that satisfy the maximum parameters as determined by the researchers are rated good. Those that satisfy the parameters partially are rated average, whereas those that least satisfy the key parameters are rated low. To avoid bias, the Delphi Expert Method [110] was adopted to evaluate the studies culminating in the rating provided for the robustness against attacks and overall security. All five researchers acted as experts and evaluated each study against the set of key parameters separately. Thereafter, a meeting was called to consolidate each rating. Where individual opinions differ, the cycle of Delphi was reinitiated until a consensus was reached. The method was designed in such a way that the researchers provided reasoning for individual responses. This was to help confirm the plausibility and strength of the individual researchers’ evaluation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 10. Comparison of various existing image steganography techniques and methods for grayscale images.

https://doi.org/10.1371/journal.pone.0308807.t010

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 11. Comparison of various existing image steganography techniques and methods for color images.

https://doi.org/10.1371/journal.pone.0308807.t011

4.6 Future scope and research directions for image steganography

The challenge of image steganography remains to achieve high embedding payload capacity while maintaining robustness, distortion resistance, imperceptibility, and overall security (un-detectability). This challenge still exists in many of the reviewed works. The existing systems suffer from low embedding rate, low visual quality of stego image, image distortion, high computational complexity, performance accuracy, low throughput efficiency, as well as detection and modification of secret data. These gaps are largely due to the techniques employed by the existing works. Other identified gaps in most of the existing works are vulnerabilities such as double-frequencies, zero points, and non-accurate detection of statistical steganalysis results. These vulnerabilities have been extensively exploited by steganalysers.

Several of the reviewed works have no layer of protection against unauthorized access to secret data. This is because many of the existing works did not apply cryptographic protocols. Those that implemented cryptography for encryption and decryption are also based on the raster order LSB substitution method which is prone to RS statistical steganalysis attacks [111]. From Tables 10 and 11, only 34 out of the 125 reviewed papers employed encryption (cryptography). This represents 27% of all reviewed papers. The key aim of image steganography technique is to hide the existence of secret data using cover objects (audio, video, image, text, network) [112,113]. Also, for steganography to achieve its aim, the transferred message on the recipient side should be the same as the original message without noticeable suspicion by a third party [114,115]. Embedding secret data into the cover object does not provide the security needed [116–118]. This is because, an unauthorized person can read the message when the cover image is attacked, breaking the requirement for confidentiality of the message.

The analysis of the previous works has shown that there is a need to put in place appropriate corrective measures to strike an adequate balance between high payload and security against statistical steganalysis including RS attacks. Thus, techniques that achieve higher payload capacity and better-corrected pixels in ensuring enhanced security protection of secret data in storage and transmission are required. One key challenge of the image steganography embedding process is the secret message size [119]. This challenge could be overcome by employing lossless compression algorithm techniques to achieve higher payload capacity and high embedding rate [120]. From Table 10, only 13% of all the reviewed articles in this study implemented data compression. Compression reduces the secret data size before embedding process begins [121,122].

Clearly, this systematic literature review has shown that the research direction in image steganography has been broad and diverse since 2012. As challenges in image steganography continue, the research domain also continues to evolve. Aside from the traditional methods, researchers have begun experimenting other areas of application for image steganography. For example, Table 4 shows that 9 of the papers adopted other different techniques than the known traditional methods for steganography. This can be inferred that scholars within the image steganography domain are exploring newer and more innovative approaches.

Future research directions could enhance the security and robustness of image steganography applications by:

• Cryptographic protocols as a layer of security protection. Higher security and robustness in image steganography can be achieved using multiple encryptions to mask and scramble the content of the secret message before embedding. Encrypted embedded secret data have more ability to resist steganalysis.
• Future research could explore compression and image enhancement techniques to achieve a high payload while maintaining image visual quality. This could help solve the problem of balancing the tradeoff between security and embedding capacity
• Future research could utilize other novel techniques from domains that have the propensity to achieve computationally efficient, reduced computational complexity, improved performance, and undetectability which are the major issues advocated for by researchers within the image steganography domain. For instance, imperceptibility and security could be improved by employing emerging technologies such as Blockchain Technology [123]. Stego-images containing secret data are often transmitted over unsecured public networks, thereby making the secret data susceptible to many attacks including man-in-the-middle attacks, tampering, and eavesdropping [124,125].
• Blockchain technology could be employed in image steganography to ensure stego-images are more secure and authenticated [114]. This is because, blockchain has immutable properties, easy traceability, tracking capabilities, and transparency [50]. In addition, future research could rely on emerging artificial intelligence and machine learning power technologies such as ChatGPT to provide robust techniques against steganographic attacks.

5.1 Research trends in image steganography techniques

The review reveals an interesting result for image steganography research. Intriguingly, research on image steganography is skewed in terms of publication trends. The skewness in the publication trends for image steganography can be seen in analysis concerning the year of publication, publication outlets, country of origin of the corresponding author, and application domains for image steganography. The research shows that despite the growing interest in the research field in image steganography, research in the area took a sharp nosedive in 2020, but rather experienced astronomical expansion from 2021 to 2023. Approximately half of the papers studied in this research were published from 2021 to 2023. Indeed, cyber-attacks on organizations and individual data due to inherent vulnerabilities in network security protection were expanding [126], even before 2020. This might have contributed to the interest of researchers in this domain to find solutions to the ever-increasing threat. The volume of research conducted in this domain post-COVID-19 is not surprising, as the Coronavirus (COVID-19) pandemic resulted in an increased number and range of cyber-attacks resulting in personal and organizational data breaches and compromises [127]. The exponential increase in the research domain could be a direct response to the increasing trend of cyber-attacks during the COVID and the need for companies to work remotely as a means of cutting costs and making use of investments in technology during the pandemic. In terms of publication outlets, it is interesting to note that more than half of the articles reviewed were published in IEEE. The implication is that IEEE has become the destination of choice for researchers publishing studies on image steganography. This finding corroborates the study of Kaur et al., [50] where most of the reviewed papers were also published in IEEE. This brings to the fore the need to address the dominance in the publication of such crucial research areas by a particular publication house and expand the domain in other publication outlets. Although there are several publication outlets that publish research on image steganography, such outlets were dully not represented in this study. Given that image steganography techniques for resisting steganalysis have become a growing area of research interest, other publication outlets may put in place measures to attract researchers. This could include special issues concerning the domain and putting in place incentives to attract researchers. Surprisingly, despite the growing cases of cybercrimes in Sub-Sahara Africa [128], the interest of researchers in this geographic location is low. It must, however, be mentioned that researchers in Sub-Saharan Africa have begun showing interest in publishing in this area, as evidenced by a recent publication [70]. Developing research capabilities including collaboration with external scholars particularly those in India and China could ameliorate the low level of research by African scholars in this domain. The digital divide in Africa is growing. Internet penetration in Africa is also expanding, and as a result, digital crimes have increased. Developing research capabilities and acquiring the requisite technical knowledge to research image steganography techniques could prevent many of the data breaches and cyber-attacks as well as save African-based organizations from data breaches and compromises.

5.2 Image steganography techniques for resisting steganalysis

As observed in Fig 6 and Table 4, Generative Adversarial Neural Networks (GAN) is the most preferred image steganography technique for resisting steganalysis attacks. This finding supports arguments by Liu et al., [129] that GAN has seen increasing achievement in the field of image steganography, computer vision, and natural language processing. From the review, the application of GAN in image steganography witnessed exponential growth between 2018 and 2022. GAN was first proposed in 2014 [130] and has seen great application in many fields of Computer Science. In image steganography, it improves security by resisting cover modification, enhances the cover selection and synthesis processes, and achieves overall security protection against steganalysis attacks. The security capabilities of GAN are higher than other adaptive methods and traditional spatial and transform domain methods [131]. Quite interestingly, despite the complexity associated with GAN-based image steganography approaches, the technique has seen overwhelming applications. The increase in the use of GAN processes is attributed to recent developments in deep learning-based steganalysis [132–134]. GAN has the capability to resist state-of-the-art deep learning-based steganalysis [135]. GAN also can be used to improve the security performance of image steganography techniques in other domains including spatial domain applications. These capabilities make GAN a considerable option for image steganography regardless of the complexity associated with it.

The study further shows that machine learning-based algorithms are recently dominating image steganography research. This confirms the argument by Hussain et al., [82] on the growth of machine learning techniques including GAN, DL, CNN, and GA. These machine learning-based algorithms have emerged as powerful tools for image steganography capable of resisting steganalysis attacks. Subramanian et al, [131] argue that machine learning-based algorithms will continue to see greater applications in future image steganography works. DL, GA, and CNN like other machine learning algorithms including GAN are great techniques for fooling steganalysis and preventing them from detecting secret images hiding in cover images. In addition to machine learning-based algorithms, the study reveals that researchers are exploring many other areas of application for image steganography. At least 9 of the reviewed articles were based on other methods rather than known traditional steganography methods or machine learning methods.

The overall implication is that previously preferred image steganography techniques particularly the least significant bit (LSB) insertion algorithms are becoming unpopular among data protection and information security researchers. This finding supports the assertion by Subramanian et. al., [131] that traditional algorithms like LSB are now receiving less attention in image steganographic applications. Between the spatial domain and transform domain, algorithms based on the spatial domain were more. This finding supports arguments by Hussain et al., [82] that the spatial domain methods for secret data embedding are more popular than the transform domain due to the easiness of embedding and extraction of data in the spatial domain. The spatial domain however suffers from less robustness. The major spatial domain methods include LSB, LSB-M, AED, PVD, and PH. The major transform domain methods identified were DCT and IWT techniques such as RDH and RNS however saw application across the various embedding domain processes (ie spatial, transform, and adaptive domains). Indeed, LSB is considered the fundamental and conventional steganography method capable of hiding a larger secret message in a cover image without noticeable visual distortions. Over time, different variations of LSB have been developed. The disadvantage of LSB is that an increase in payload reduces the overall visual quality making it an easy target for attacks. Given the challenges of LSB, Wu and Tsai [136] proposed PVD using the difference between two neighboring pixels to determine the number of secret bits to be embedded. Since then, many steganographic methods have been proposed to improve the initial PVD method. From the study, it can further be observed that AED is one of the prominent embedding strategies in the spatial domain. AED schemes have the capability to maintain minimum visual quality and are noted to provide higher imperceptibility when compared to other spatial domains [137]. From Table 4, AED recorded the second highest techniques for image steganography. Different hybrid edge-based methods including combining canny edge and fuzzy edge adaptors [138,139] were observed in the articles reviewed for this study. The study has revealed varied techniques for protecting data against attacks. However, more research investigations are required to identify how emerging technologies including artificial neural networks (ANN) could be explored to provide harmonized security capabilities against statistical steganalysis attacks.

5.3 Security performance of image steganography against attacks

The systematic review results revealed that the most significant contribution of steganography techniques is resistance against statistical detection analysis attacks such as Regular-Singular (RS) and Histogram analysis attacks. Adaptive embedding techniques such as GAN, GA, and CNN and transform domain techniques including DCT and IWT methods were hard to expose to such statistical detection analysis attacks. However, spatial domain techniques including LSB and PVD were easy to expose. Most of the studies reviewed reported improvement against RS and histogram analysis attacks, indicating continued research improvement in overcoming these types of attacks. Another key significance of existing steganographic techniques is resistance against non-structural detection attacks. Machine learning-based algorithms proved difficult to detect by non-structural detection attacks, whereas spatial domain and transform domain methods were easily detectable. In terms of geometric attacks, it was observed that adaptive embedding techniques such as GAN and CNN and techniques-based transform domain methods were resistant and hard to geometric attacks while spatial domain methods were vulnerable to such attacks.

The visual quality of adaptive-based methods and the undetectability of secret messages were high and robust against noise cropping and less prone to image rotation. However adaptive methods have limited embedding capacity. Even though spatial domains such as LSB, LSB-M, and PVD have higher payload capacity and visual quality, they are highly prone to noise cropping, and rotation. Overall, most of the reviewed studies reported higher SI visual quality, an important measure in ensuring the transmission of secret data is not detectable by the HVS. Transform domains such as DCT and IWT offered higher security considerations than spatial domain methods but were less effective when compared to adaptive embedding methods. Only a few of the techniques have also been implemented in a real-time application. When evaluation of image steganography is done using capacity, traditional embedding algorithms including the various variations of LSB offer higher embedding capacity than machine learning-based techniques such as CNN, GAN, and DL.

Despite the notable progress achieved in image steganographic techniques, computational complexity and time complexity were observed to be a major challenge in all the reviewed papers. Even though computational complexity is a generic challenge as most studies indicated, adaptive embedding techniques such as CNN, DL, and GAN were reported to have higher computational complexity results than both spatial domain and transform domain methods. This finding is, however, not surprising given that most of the adaptive embedding approaches were based on machine learning techniques. This is because, one key challenge associated with machine learning algorithms has been identified to be computational complexity [140,141]. The challenge of computational complexity is noted to significantly have a direct impact on image steganography techniques with respect to computational speed thereby having a tremendous impact on the performance of emerging image steganography applications. This notwithstanding, recent studies have reported measures to improve the computational complexity and time accuracy of machine learning algorithms [142]. This has occasioned the growing use of genetic algorithms (GA) in image steganography applications [70], as GA has been noted as reducing the computational complexities of machine learning-based algorithms.

From Tables 10 and 11, the results from the systematic review analysis have shown the positive effects of combining steganography and cryptography. The analysis further shows that image steganography studies that had implemented cryptography were rated high for robustness and good for overall security. The combined effects of cryptography and steganography provide an additional layer of protection for the privacy system against many security attacks [143,144]. Although the combination is noted as an extra payload on the time and space complexities of the application, it offers comparative advantages in terms of robustness, confidentiality, and privacy [145]. However, several techniques have recently been introduced to reduce the computational cost performance associated with the art of combining steganography and cryptography.

From Fig 7, Modified Least Significant Bits (M-LSB) had the highest PSNR value indicating the highest imperceptibility. This was obtained for RS 108. This was followed by RS111 with a PSNR value of 85, which utilized the LSB technique. For embedding capacity, the highest capacity recorded among the reviewed articles was 8.88BPP for the PVD technique. This was obtained in RS48. This was followed by RS26, a generative adversarial network (GAN) which obtained 5.61BPP. A careful examination of Tables 7–9 shows that Spatial domain techniques recorded the highest imperceptibility outcome. However, spatial domains are susceptible to steganalysis attacks. The average highest embedding capacity was recorded in the spatial domain and transform domain techniques. Genetic Algorithm (GA) and GAN applications under the adaptive domains showed the best results for balancing embedding capacity and robustness. This explains the growing use of GAN and GA algorithms. Even though, the high-capacity trade-off to security and robustness improvement remains a challenge [146–148], GAN, GA, and other emerging technologies such as generative artificial intelligence (AI) have the potential to overcome the challenge.