The LSB technique is considered an extremely simple technique in its performance, and therefore, it represents one of those common spatial image steganographic techniques [20]. The least significant bits within an image only introduces weak information and small modifications in these bits are not detectable through the eyes of humans. The secret bits are directly embedded into the cover image based on applying an LSB-based spatial-domain technique by modifying the cover’s least significant bits with no any distortion for the visual quality of the cover image. However, the embedding procedure generates a noise of 50% derived from the average bit embedding rate (i.e., embedded bits per pixel). Previous studies in the LSB steganography approach [46,47] have only focused on designing a method for increasing the capacity of the payload based on the use of the cover pixels. Further, the domain of steganalysis turns to be more effective in breaking such methods when statistical analysis is applied.

To achieve effectiveness of this technique, several developed LSB based image steganography versions are taken into account. The most significant versions apply different LSB matching algorithms [48], Adaptive LSB embedding based on image features such as texture contents or nature of edge pixels [49,50], Optimized LSB substitution based on learning methods [51,52] and so on. Additionally, the LSB technique is expanded to a maximum of 4 LSB planes in order to raise the capacity of the embedding procedure according to the cost of the minimized imperceptibility [53]. In a recent study, the authors proposed an LSB object-oriented image steganography [27]. In the mentioned research work, the secret data is embedded in the skin region of the cover image where the skin objects are selected using a skin detection algorithm. The neural network approach is applied to find the largest object among selected skin-tone objects. While the proposed work gives high embedding PSNR, it uses more than one cover image to embed the secret data.

The major benefit of the LSB steganography related to its ease of the embedding and extraction procedures. Nonetheless, LSB techniques are vulnerable to different statistical attacks, with some manipulations within the stego-image. As the LSB steganography represents the way of modifying the cover’s pixel values, its performance of extracting of the embedded data relies on some factors such as the compression quantization, noise effect, and intruder attacks.

Wu et al. presented a novel embedding aspect that relies on the occurring difference among pixel values [54]. The cover image consists of non-overlapping blocks of two joined pixels where the difference found within every block is changed. A greater difference between the two pixels allows a greater change and thus higher payload can be embedded. The number of secret bits, which are allowed to be embedded, relies on whether the pixel exists into a smooth or an edge location. In the edge location, the differences among neighboring pixels are found to be more, while in the smooth location it is found to be less. Therefore, more data can be embedded into the edge-area pixels in comparison with the smooth area. Since this technique aims to embed the data by changing the difference value within the two neighboring pixels instead of directly changing pixel values, it offers more effective results based on its stego-image quality and imperceptibility in comparison with the LSB replacement technique. Many techniques that relate to the PVD technique have been produced in order to offer several secure communications and to defeat any statistical attack. For instance, Hussain et al. [55] introduced an embedding technique for enhancing security based on the use of two modes that depend on the embedding procedure. In fact, this procedure is enhanced and based on two techniques, namely, the improved Rightmost Digit Replacement (iRMDR) and the Parity-Bit Pixel Value Difference (PBPVD). A further example of the enhanced security is the histogram analysis based vulnerability (PVD) technique [56]. To link the benefits of different embedding techniques together, hybrid embedding techniques are produced (i.e., Steganographic techniques that apply the LSB and PVD [57,58]).

Many enhanced PVD based image steganography versions were researched in order to improve the efficiency of PVD. The most significant versions apply the Adaptive PVD block technique by using different pseudo-random number techniques for determining the blocks [90] and tackling the fall-off boundary problem in the PVD technique [59]. In addition, Swain [32] proposed an adaptive PVD-based hiding scheme. In the mentioned work, the cover image is divided up into 1 × 2 overlapped blocks of adjacent pixels. After that, the method uses modular arithmetic and adaptive quantization range table to embed the secret data. The findings reveal that the mentioned scheme has higher PSNR value and embedding capacity in comparison with existing PVD schemes. However, the proposed scheme increases the algorithms’ complexity of embedding and extracting the secret data [60].

The Exploiting modification direction (EMD) method is a common method that keeps the increased fidelity pertaining to the stego-images protected [61]. In general, the secret digit is converted based on the (2n + 1)-ary system when the embedding procedure is taking place through this method, such that n denotes the number of the cover pixels. The range of the distortion’s highest pixel value is just (±1). In particular, the EMD method applies a particular base for selecting the local variation corresponding to the pixel intensity in the cover image. Thus, more message size can be hidden in the pixels that exist in high texture areas. In fact, the EMD method delivers an effective visual quality in comparison with PVD and LSB methods. The highest capacity of the EMD method reaches 1.16 bpp for n = 2. At the same time, the embedding payload is radically reduced when the selected pixels are incremented. Consequently, various EMD methods are produced in order to enhance the embedding payload [33,62–64].

Kuo et al. [33] proposed the Generalized Exploiting Modification Direction (GEMD) technique where the major aim is that the (n + 1)-ary binary bits are embedded through n adjacent pixels. The findings demonstrate that the technique can keep the embedding payload (1 + 1/n) along with a modified set of pixels. However, the technique does not have the ability of hiding secret bits that are exceeding two per every pixel [5]. At the same time, it adjusts the entire pixels of the set when the embedding procedure of the secret is taking place. In order to tackle the issue of pixel modification, Kuo et al. [34] produced a new technique that is called the Modified Signed-Digit (MSD) technique. This technique can only adjust n/2 pixels but has only 1 bpp embedding payload. The MSD technique can proceed towards the RS steganalysis [65] with an efficient imperceptibility. Kuo et al. [35] introduced an embedding technique that is based on a multi-bit encoding function, which is applied in order to enhance the embedding payload. This technique embeds up to (k + 1/n) pixels on average for every available pixel, where k is selected based on how much embedded bits are existing per each pixel. Furthermore, the technique can minimize the conversion of the secret data’s overhead and gives a simple relation among the adjacent pixels. In the meantime, the technique keeps its security in order to resist the RS and bit plain detection analysis. However, it suffers from low visual quality in comparison with many different available EMD based techniques.

A further spatial-domain embedding technique that relies on the Multiple Base Notational Systems (MBNSs) is proposed in order to transform the secret information through to the notational scheme prior to the embedding procedure [5]. In 2006, this technique was initially proposed to enhance the original LSB substitution technique such that bit planes are applied in order to hide the secret bits [66,67].

In several techniques related to the MBNSs, secret information is changed into symbols and re-expressed according to the used MBNS (e.g., octal, decimal and binary systems) [5]. Additionally, an embedding process is applied for such symbols to be embedded into the pixels’ intensities. In general, when the notational base symbol is large, the embedding rate gets also large. Several studies aim at enhancing the capacity of the embedding process through different MBNS based techniques. Zhang et al. [68] produced such steganography. Particular bases are selected based on the local variation’s degree pertaining to the pixel magnitudes within the cover image such that busy area’ pixels can carry more secret bits. High embedding payload is obtained by this method. Comparisons of the obtained findings by the MBNSs are conducted with the PVD technique where it can be inferred that it achieves a superior and an effective quality factor and PSNR.

In [36], an adaptive embedding technique is produced according to the Varying-Radix Numeral System (VRNS). This technique divides any secret data into different numerals, which contain a capacity of different amounts of variable information. This division relates to the tolerance of cover pixels when managing the highest adulteration of the greater secret data. It is found to be proven from the findings that the payload is large enough while keeping an acceptable imperceptibility. Additionally, it controls the way it maintains security towards the RS steganalysis [65]. However, embedding the payload is still limited to many different radix-based methods. Consequently, an enhancement of the method in [36] has been conducted in [69]. The enhanced method is called the VRNS method, which is based on a hidden information when applying the Adaptation and Radix (AIHR) algorithm. Nonetheless, such a method obtains a greater payload compared to other available VRNS systems. On the other hand, this method contains few ambiguities in proposed flow. For example, there might be a way a receiver and sender are synchronized based on determining their bases. Additionally, in the AIHR extraction procedure, the ambiguity of selecting the multiple M might lead to not recovering the full secret data. Chen et al. [37] produce a General Multiple-Base (GMB) embedding technique in order to convert the secret bits into a number of M-ary secret digits that belong to a pixel-cluster (i.e., n pixels). The multiple M is selected in an automatic manner based on the end user’s input function. It offers various styles of the multi-purpose embedding procedure leading to high embedding payload or high quality of the stego-image. At less than or equal to 1.0 bpp, the GMB method resists the non-structural SPAM features and the RS steganalysis [65].

Potdar et al. [70] proposed the GLM technique in order to map the data based on changing the pixels’ gray levels (i.e., not embedding it). According to a few mathematical functions, a group of pixels is determined where the values of their gray levels are organized in the bit stream that relates to the secret message, which could have been mapped within the cover image [71]. This technique applies the even and odd numbers aspect in order to provide an effective way of mapping the data through to the cover image. For instance, number ‘1’ is mapped as an odd data value, while number ‘0’ is mapped as an even data value. The benefits of such a technique comprise the reduced computational complexity and increased embedding payload. The hybrid embedding technique relies on the GLM technique, which is produced by Safarpour et al. [72] for embedding more secret data, and consequently, increasing the embedding payload.

The Quantization Index Modulation (QIM) technique [73] is considered an effective data embedding technique in the field of digital watermarking where it is applied in different steganography domains. This technique is based on embedding the information into the cover medium by first performing a modulation of an index or a set of indices with the embedded data. After that, the host is quantized according to the involved quantizer(s). The technique has high embedding payload, and it aims at allowing the embedder to manage the robustness and distortion levels obtained during the embedding process. Chung et al. introduced an enhanced data embedding technique that relies on a Singular Value Decomposition (SVD) and a vector quantization. The findings showed a better compression ratio and a more effective image quality [74]. A lossless data-hiding algorithm that applies the Side Match Vector Quantization (SMVQ) and the Search Order Coding (SOC) is proposed in [75]. This algorithm performs a compression rate of 0.325 bpp including a 256 of codebook size. A reversible data-hiding technique that is applied for several VQ indices is discussed in detail in [76]. This technique enhances several techniques such as enhancing the proposed techniques by Tsai and Yang and Lin and Chang, which provides 0.49 bpp of a compression rate.

For enhancements based on the embedding payload and the reduction of distortion, several enhanced quantization-based steganography versions are researched. One of these versions utilizes the elastic indicators and adjacent correlation. Through this approach, the indexes are encoded based on the difference values that are derived from the neighbouring indexes and the elastic sub codebooks are applied to enhance the compression rate [77].

The Palette based steganography is proposed in [78] to utilize the palette-based images as cover images. Image formats such as TIFF, PNG and GIF are appropriate for such a technique. In palette-based steganography, the colour that has a similar parity of a secret bit within a palette is used for the embedding procedure. The major advantage of the palette-based steganography is that the entire distortion within the stego-image is seen to be smaller in comparison with other related spatial techniques. On the other hand, the major drawback of this technique refers to the demand of particular images, which have lossless compression formats.

Imaizumi et al. [40] introduced a dense embedding technique based on the use of the palette based steganography that maintains the visual quality within an adequate level for an untraceable communication. Multiple secret bits are embedded in one pixel once the difference is assessed according to the Euclidian distance measures, knowing that the majority of the palette-based methods follow a strategy of single bit per pixel. As a comparison with further palette-based methods, the embedding payload is marginally increased, and the visual quality is seen to be further increased through the PSNR value of ~40 dB.

The prediction based embedding technique has currently attracted many researchers [5]. In the prediction-based steganography, the embedding procedure is based on directly changing the pixel values, which causes a substantial distortion in the stego-image [1]. This leads to poor visual quality and a low embedding payload. In order to tackle such a problem, a predictive coding technique is provided such that existing pixel values are effectively predicted based on the use of a predictor rather than changing the pixel values. The Error Values (EVs) of prediction are modified for the purpose of embedding the secret bits. Referring to the international standards for lossless and near lossless image compression, the process of compression comprises two different steps, which include the prediction and entropy coding of predicting EVs. The predictive rule is expressed as follows [1]:

(3) X′={min(a,b),ifc≥max(a,b)max(a,b),ifc≤min(a,b)a+b−c,otherwise

During the prediction step, a predictor is employed to estimate the pixel values of the host image. After that, the entropy coder is used to compress the prediction EV. The Median Edge Detector (MED) technique and the Gradient Adjusted Prediction (GAP) techniques represent new predictors that are applied in several prediction-based image coding techniques. Many different reversible prediction-based embedding techniques are enhanced and highlighted in the literature. Every technique attempts at enhancing many available proposed existing techniques.

Hong et al. [79] proposed an embedding technique that relies on the modification of prediction error (MPE), which adjusts the prediction errors’ histogram in order to select the unoccupied area for embedding the secret data. The visual quality pertaining to the MPE method ensures more than 48 dB. To acquire an increased embedding payload, the authors in [41] proposed a multiple predictor base steganographic technique, which is considered as an enhancement to the MPE technique without the need for adding any predictor overhead. Determining the accurate predictor in the embedding process is based on the predictor’s history. The produced technique demonstrates that the enhancement occurs for the visual quality and embedding payload where its security is not evaluated by any steganalysis technique.

Recently, Benhfid et al. [42] adopted the interpolation through multiple linear box-splines (MLBS) on three directional mesh in order to develop a reversible embedding technique. The secret data is embedded within the error between the interpolated and cover pixels. The secret data is initially put into the interpolated pixels as a particular error. The error is adopted to acquire several stego-pixels that are extremely close to the cover pixels. Furthermore, the LSBs pertaining to the secret data’s interpolated pixels are replaced into the cover pixels. After that, the Optimal Pixel Adjustment Procedure (OPAP) is used for the purpose of minimizing the difference between the original cover pixel and the stego-pixel. To compare with other studies in the literature, the findings reveal that the produced technique contains a great embedding payload when the PSNR values are retained at a good level. Additionally, the findings show that this technique achieves a low detectability rate when examined through different steganalysis attacks.

In recent years, the introducing of deep learning in steganography has shown a great improvement in the effectiveness of steganography methods. Deep learning steganography is learned from machine learning. Several deep learning steganography methods [43–45] have been developed to improve the imperceptibility and security of steganography. Baluja [43] designed a convolutional neural network (CNN) model based on an encoder-decoder structure. The encoder successfully conceals the secret image into a cover image of the same size of the secret image, while the decoder reveals the complete secret image. The proposed method has a large payload with a minimum degree of distortion to the cover image. It distributes the bits of secret image across all the available bits of the cover image. However, in terms of security, the generated stego-images are distorted in color. In addition, this model takes more time to embed the secret image and requires much more memory since it uses three networks in the embedding and extracting processes [80].

Zhu et al. [44] proposed a deep learning data hiding method, called HiDDeN, based on using Generative adversarial networks (GANs). This method consists of a stego-image generator, an attack simulator, and an extractor. Different noises, such as JPEG compression and Gaussian filter, were modeled in the simulator to train the network. The proposed HiDDeN method can extract the hidden bits with high accuracy even under different attacks, such as JPEG compression and Gaussian blur. However, while this method is resistance to a set of various noises, it requires excessive memory, and therefore cannot effectively embed large payloads [81].

Recently, Shang et al. [45] proposed a deep learning steganography method based on GANs and adversarial example techniques to enhances the security of deep learning steganography. This method consists of two phases, namely, model training and security improving. By the security improving phase (i.e., using adversarial example techniques), the generated stego-images can fool the deep learning steganalysis techniques and the extracted secret images are less distorted. The experiments reveal that the MSE values of stego-images are less than one percent. However, the proposed method has lower embedding payload (~0.4 bpp).

Many spatial-domain steganography techniques can achieve high payload, but they are susceptible to extremely few updates, which are likely to be encountered based on different image processing tasks (e.g., scaling, rotation, cropping, and so on). Moreover, these techniques recompense the image’s statistical features indicating a weak robustness towards image filters and lossy compression. As a summary, Tab. 4 provides number of significant comparisons for the merits and demerits pertaining to the well-known spatial-domain steganography techniques.