Ye et al. [7] recently used a traditional machine learning method for stego image classification. Four distinct embedding approaches F5, Least Significant (LSB) Replacement, and PVD were used for Steganalysis. SVM and SVM-PSO were the classifiers employed. Linear, radial, multi-quadratic, ANOVA and polynomial kernels were utilized in classification. The classifier was taught to analyze each coefficient as an independent unit, and the results of this analysis aid in determining the Steganalysis conclusion. PSO classifiers outperformed all other kernels, according to the results. Kang et al. [17] pointed out that shading image steganography weakened the link between the gradient amplitudes of different color channels and presented a color image steganalysis computation based on the channel angle relationship. Each color channel used the gradient amplitude of image channels to obtain gradient amplitude and steganalysis characteristics. It can be observed that the proposed steganalysis method beat existing steganalysis algorithms for both WOW and S-UNIWARD steganography.

The approach made by Chaeikar et al. [16] covered three areas: bunch S.W. steganalysis, outfit S.W. steganalysis, and gathering S.W. steganalysis. In the following section, an SVM classifier ponders the datasets to their associated reference profiles and uses a trapezoidal cushy cooperation ability to choose the degree of enlistment of given pixels. Consequently, six datasets contain all the image pixels’ apparent pixel classes and enlistment degrees. The results then diverged from steganalysis systems with humble examination viewpoints that were comparative. The social event S.W. approach given here achieved 86.771 percent accuracy. This accuracy improved to 99.626 percent, unquestionably the most anytime achieved by any low assessment angle approach. The discoveries of JPEG image appraisal in spatial and DCT changes were introduced [18], and an examination was made. The author utilized straight, rearranged, defined, and computerized inspecting techniques. Four feature extraction elements were considered: request highlights, second request highlights, broadened DCT elements, and Markov highlights. The impact of features has been contemplated, and Principle Component Analysis (PCA) has been used to eliminate unwanted attributes. The acquired attributes were characterized by utilizing two separate classifiers: SVM and SVM-PSO. When utilized in spatial change, the classification portion often gave a high arrangement rate. As we can see, most of the study’s authors have used statistical and textural features to get structural changes of an image to classify using SVM or SVM-PSO. Some do not consider evaluating low-payload data, and some who consider this problem have calculated high dimensional features, which will use high computation and storage capacity, which is not accessible to all concerned personnel. Table 1 shows different machine algorithms from recent research on color image steganalysis.

Reinel et al. [19] proposed better architecture than QIAN NET, SR-NET, XU-NET, YE-NET, & ZHU-NET named GBRAS-NET by using spatial details of an image. The accuracy achieved by this architecture with MiPOD, HILL, and HUGO through BOSSbase 1.01 was 68.3, 68.5, and 74.6 for 0.2 bpp. The data augmentation architecture showed the highest accuracy. While BOWS 2 showed the accuracy improvement for WOW and S-UNIWARD-0.7% and 2.2%, respectively. An approach based on Siamese CNN for the Steganalysis of images was proposed by You et al. [20]. Steganographic Noise was extracted from the images’ sub-regions of one Siamese network using each subnet. For feature extraction, down sampling was done with a stride value set to 2 on Block B. ROC showed the testing of BOSSbase 1.01 images 256 × 256 with 0.4 bpp. The proposed SiaStegNet was tested on BOSS-256 & BOSS-512. Comparing the proposed with SRNet, the accuracy achieved with S-UNIWARD was 91.89%, with HILL 85.97%. ALASKA#2 dataset was also used to demonstrate the proposed network.

Liu et al. [5] proposed an approach based on CNN techniques having diverse filter modules & squeeze excitation. After each convolution layer, a B.N. layer was added to avoid overfitting and fasten the learning. Using the WOW algorithm & payload of 0.2 & 0.4 bpp, both models were trained. Resized images of BOSSbase 1.01 with the size of 256 × 256 were implanted using 0.1, 0.2, 0.3 & 0.4 bpp on S-UNIWARD & WOW algorithms. The error rate of the proposed method was minimized to 8.5% compared to SRM+EC, and 6.7% error was minimized as compared to Xu Net, 3.9% to Ye Net, and 3% to Yedroudj Net. Singh et al. [21] proposed a solution for Steganalysis by dividing their work into two sections. It was necessary to accurately forecast the cover image from the associated stego image to see accurate noise residual. A denoising learning kernel was applied in this study to obtain a more precise noise. This method created a CNN-based Steganalysis detector, and the noise residual is used to train the detector for more accurate detection. The suggested approach outperforms the other Steganalysis schemes, according to experimental results.

A practical CNN-based Steganalysis approach with a combined domain and nonlinear detection mechanism was introduced by Wang et al. [22]. The nonlinear detection process depended on the spatial rich model to present the maximum and lowest nonlinear residual feature acquisition approach. The author employed high-pass channels for spatial residuals in the joint location framework and used instances from the discrete cosine change remaining (DCTR) for change steganography. The findings also demonstrated that the model’s combined WOW and S-UNIWARD identification precision was higher than SRM+EC, Ye-Net, Xu-Net, Yedroudj-Net, and Zhu-Net is roughly 4 to 6 percent higher than the best Zhu-Net. Wu et al. [23] introduced multiple steganography for deliberate image downsampling based on CNN-based Steganalysis. The detection performance can be increased after training with shrunk images. Because the suggested method’s preprocessing was so simple, it did not significantly increase CNN training time.

The experiment’s findings demonstrated that accuracy is up to 34.8 percent higher than in the traditional approach, especially when the same steganography is integrated. Singh et al. [24] proposed a brilliant Normalization method named Shared Normalization (S.N.). CNN was improved by the expansion of a new convolutional layer named the S.N. layer, bringing about a CNN that was advanced for computerized image steganalysis. Broad investigations were done to show that the proposed S.N. layer and the new CNN model for image steganalysis were compelling. The recommended network was contrasted with the cutting-edge prosperous model procedure and with different as late proposed CNN models. The second approach for Steganalysis is deep learning, as discussed in the introduction section. Table 2 describes some deep learning-based methods to show contributions made in the last few years. The techniques discussed in Most of the research focused on preprocessing images. The proposed method is inspired by de-noising kernels to get embedded information and classify it using CNN. The proposed method used Curvelet transformation for feature extraction recently published for de-noising any image. So, noise residuals can be fetched through Curvelet features. Table 2 was based on CNN and used different methods and layers to classify stego images.

An image with steganographic content has a type of distortion in that image. So, modules like S-UNIWARD, WOW, HUGO, and MiPOD [19] are used to embed this distortion using spatial details of the image used in this experiment. Researchers in Steganalysis rely heavily on the massive and varied BOSSBase 1.01 dataset. Each of the 10,000 color images in BOSSBase 1.01 has a resolution of 256 pixels on the longest side and 256 pixels on the shortest side. The dataset was designed as a standard against which steganographic and Steganalysis methods might be measured. It features a variety of visual elements, such as landscapes, textures, and graphics, to depict various situations in the actual world. This dataset can be used to train and evaluate steganalysis machine learning methods. The dataset contains photos modified using several steganographic algorithms, making it a valuable tool for researching and developing new steganalysis methods. The database discussed is mined for the original images. For experiments, we made use of a variety of steganographic techniques to create stego images. Normal images were taken from a public database and created stego images using the steganography algorithms previously discussed. In subsequent experiments, a steganalysis framework has been used. The dataset has been separated so that each group of stego photos is coupled with a normal image.

Each dataset developed is further divided into two sets of training and testing datasets to do some experimentations using the proposed steganalysis framework. This table shows the total number of normal, payload, and stego images in various steganographic datasets. S-UNIWARD is the initial dataset; it consists of a thousand normal images with payloads of 0.1, 0.2, 0.3, and 0.4. There are 4,000 steganographic images in this data set. The WOW dataset has 1,000 normal and 4,000 stego images, but the payload values are unknown. HUGO, the third dataset, similarly has 1,000 normal images and 4,000 stego images, but again, details need to be given concerning the payload. The final dataset, MiPOD, consists of 1,000 normal and 4,000 stego images, although the payload values utilized still need to be discovered. The four datasets contain 4,000 reference images and 16,000 stego images, making them excellent tools for steganography study. These datasets help create and evaluate steganalysis algorithms, which can uncover secret content in digital photographs. Following the creation of steganographic images, we have split the dataset into subsets for practical purposes shown in Table 3.

Information embedded into images can reside in any part of the image, so images are read from the directory then block division is implemented using non-overlapping block division by using a script written in MATLAB R2021a to reduce error in getting interested features in the feature extraction process. The script gets an image of any size and augments it into equal blocks of 256 × 256. Each image is adjusted to size with respect to multiples of 8 through padding so it can be divided into blocks of size 256 × 256. After this process, an image is converted into blocks of 256 × 256 as shown in Fig. 2.

Curvelet is a multiscale geometric analysis tool that efficiently takes image curve singularities. Each image block is passed to curvelet transform to get texture features of an image that will be used for differentiation and learning for normal, Stego, and Cover images. Unlike methods like Wavelet transformation, curvelet transformation can be used on images in numerous dimensions and scales. Compared to Wavelet transformation, Curvelet transformation is more advantageous because of its sparse representation of local features, edges, and textures and its ability to capture curved edges and finer details in an image. Images are described in a curvelet domain, where the basic functions can be orientated in many ways. Curvelet transformation may have been chosen for feature extraction in a specific study because of its improved ability to capture and express highly curved qualities like contours and textures. Since it faithfully encodes curved and anisotropic features, the Curvelet transform is a viable option for image analysis, object recognition, and denoising.

Scale- and orientation-invariant picture characteristics are particularly well captured by curvelet transformation. This opens the door to discovering hidden information by revealing hidden patterns and textures. The key to their interpretability is an insight into the correspondence between these multi-scale traits and certain steganographic methods or artifacts. Subtle alterations to an image’s texture are common in steganography. Using Curvelet features can bring out these subtle alterations in texture, revealing previously unseen details. The presence and type of steganographic content can be deduced by analyzing the unique texture patterns retrieved via Curvelet transformation.

Furthermore, its parabolic scaling capabilities provide nearly optimal sparse symbols of objects with C.S. (“curve singularities”). On the other hand, the curve singularities can be achieved by dividing the image into sub-images and then applying the R.T. (“ridgelet transform”) to all the sub-images obtained. C.T. (Curvelet transform) was the name for this block-based transform. However, the ridgelet’s imprecise geometry has hampered the first-generation curvelets’ use in various applications. Then, a second-generation curvelet transform was offered to solve the difficulties with first-generation curvelets. We employ quick DCT (“discrete curvelet transform”) instead of wavelets and their derivatives to capture more directional data because the texture information of an image includes a mess of curves and lines. The following are the basic mathematical initiations of the CCT (“continuous curvelet transformation”) and DCT (“Discrete curvelet transforms”). The CV (“Curvelet transform”) can be defined as an inner product for a given signal f as (1)C(j,l,k)=⟨f,φj,l,k⟩

We can express the inner product as an integral over the frequency plane by expressing it as an integral over the frequency plane. Where the curvelet bases function is φj,l,k And parameters are j, l, and k, which are scale, direction, and position are j, l, and k, respectively. The continuous curvelet transform is implemented in 2D, i.e., (R2) and can be displayed in the frequency domain using x as the spatial and frequency domain variables. In the frequency domain, r and are the polar coordinates. Given a pair of smooth, non-negative, real-valued windows W(r) and V(t), referred to as radial and angular windows, respectively, with r (12, 2) and t [1,1]. W and V will always meet the following requirements: (2)∑j=−∞∞W2(2jr)=1,r∈(34,32) (3)∑l=−∞∞V2(t−1)=1,t∈(−12,12)

The frequency window Uj in the Fourier domain is then given for each j ≥ j0. (4)cUj(r,θ)=2−3j4W(2−jr)V(2⌊j2⌋θ2π)where ⌊j2⌋ represents the integer portion of j2. As a result, the support of U_j is a polar wedge that is interpreted by the support of W and V, which are applied in each direction with scale-dependent window widths. To generate real-valued curvelets, consider the symmetric variant of U_j, i.e., U_j (r +θ) + U_j (r, θ +π). Let us describe the waveform φj(x) using its Fourier transform φj(ω) =Uj(ω), where Uj(ω1,ω2) is the polar coordinate system window. In the sense that all curvelets at size 2^-j are formed by rotations and translations of φj. It can be considered the mother curvelet. Curvelets can be generated at scale 2^-j, orientation. θl, and position xkj,l using the function x = (x1, x2) by (5)cφj,l,k(x)=φj(Rθl(x−xk(j,l)))where k=(k1,k2)ϵZ2 specifies the order in which the translation parameters are applied, θl=2π.2−⌊j2⌋.l,l=0,1,… such that 0 ≤ θl < 2π, and xk(j,l)=Rθl−1(k1.2−j,k2.2−j2). Rθ and Rθ−1 denote the rotation by θ radians and its inverse, respectively, and are defined as (6)cRθ=(cos⁡θ sin⁡θ−sin⁡θ cos⁡θ),Rθ−1=RθT=R−θ

The inner product of an element f and a curvelet is the curvelet coefficient. φj,l,k, i.e., (7)cC(j,l,k)=⟨f,φj,l,k⟩=∫R2f(x)φj,l,k¯dx (8)cCj,l,k=1(2π)2∫f^(ω)φ^j,l,k(ω)¯dω=1(2π)2∫f^(ω)Uj(Rθlω)ei⟨xk(j,l),ω⟩dω

The integral function combines the Fourier transform of the input image with the curvelet transform’s directional filters and scale functions to compute the coefficient value C(j,l,k) at the specified scale, angle, and position. This coefficient represents a localized feature or pattern in the image, capturing its multi-scale and multi-directional information. The linear-digital curvelet transform of a Cartesian input array f(t1,t2);0≤t1,t2<n is defined by a set of coefficients as (9)CD(j,l,k)=∑0≤t1,t2<nf[t1,t2]φj,l,kD[t1,t2]¯

Instead of focusing on creating the best possible classifier, we have chosen to improve feature extraction and use previously obtained findings to determine which classification algorithm to use. We rely on the Support Vector Machine, a machine learning algorithm with superior feature representation ability to power our classification system. Because of its robustness and classification accuracy, the SVM has shown to be an effective machine learning and data mining tool. The primary objective in detecting stego instances is high sensitivity and high throughput. The labels were then classified as either Stego cases or others. Different SVMs have been trained on various kernels, and the one with the greatest classification performance has been selected for this study. We have created 72 features for training data from all steganographic produced photos. For educational reasons, cross-validation is a popular technique. We have utilized a cross-validation scheme with 10 samples and 5 results. Seventy percent of the data was used for instructional purposes, while the remaining thirty percent was used for assessment purposes. The SVM model performs admirably on high payload data, such as that seen in photos.

Making a difference between cover and stego image is quite a difficult task for a human, so statistical tools like Peak Signal Noise Ratio (PSNR) and Structured Similarity are being used to check the difference between cover and generated stego image. (10)PSNR=10.log10(MAXIM2MSE)

Peak Signal-to-Noise Ratio (PSNR) is a metric used to evaluate the quality of a reconstructed or compressed image, and the accompanying equation represents its calculation. The ratio between the original signal’s maximum power and the strength of any noise or distortion introduced during reconstruction or compression is known as the Signal-to-Noise Ratio (SNR). The maximum squared pixel value (MAX_IM2) is divided by the Mean Squared Error (MSE) to get PSNR for the reconstructed/compressed image. After that, a logarithmic scale with a base of 10 is applied to the final output. A higher PSNR number suggests a higher quality image since it indicates less distortion or noise in the reconstructed image. PSNR is commonly given in decibels (dB).

When considering steganalyzer, detection loss is the assessment parameter utilized most of the time. Many distinct measures, such as training accuracy, TPR, FNR, and AUC may be used to evaluate an SVM-based classification process. So, in this study, we will evaluate the proposed approach, which may be found below, using three different parameters. (11)Accuracy=TP+TNTP+TN+FP+FNwhere Accuracy is the total number of correct guesses divided by the total number of correct forecasts, then multiplied by 100 to get a percentage, the percentage of correctly identified samples in the true positive rate is determined using. (12)TPR=TPTP+FN

The percentage of samples that are wrongly identified constitutes the false positive rate, which is computed using: (13)FNR=FPFP+TN

Here, the terms “T.P.,” “T.N.,” “F.P.,” and “F.N.” refer to the quantities of true positives, true negatives, false positives, and false negatives, respectively, as defined in the previous section. The area under the curve (AUC) is a robust metric for gauging a classifier’s efficacy in a two-class classification task. It is a crucial aspect of ROC analysis, which compares the true positive rate (TPR) against the false positive rate (FPR) at various cutoffs for making a correct classification. The area under the curve (AUC) can be viewed as a metric of the feature space’s capacity to distinguish between the two classes. If the AUC is greater, the classifier did a better job distinguishing between the two groups. It’s also a measure of the classifier’s ability to rank instances based on their likelihood of being in the right class.

Different steganographic techniques were used to embed extra data in images so images could be classified accurately. Four types of steganographic methods named S-UNIWARD, WOW, HUGO, and MiPOD are used to embed Noise as information at the payload of 0.1, 0.2, 0.3, and 0.4. The Noise was added at different payloads using MATLAB R2021a, and PSNR was calculated for each image with a distorted image to evaluate each image after making it a stego image. In this case, Noise is considered as data embedded in an image as any changes occur, while steganography changes the structure of any image. So, Noise is also data added to standard images to change their structure to get structural changes from an image to classify any stego image. After adding Noise, some or much of the data of the image is changed. To check how many changes are made using a specific method PSNR is calculated between the cover and stego image. The peak signal-to-noise ratio between two images, measured in decibels, is computed by the PSNR block. This proportion is used to compare the original and distorted images’ quality. The quality of rebuilt image improves with increasing PSNR. Payload is the most critical measure to consider while embedding data in images. It can be observed through the table that as the payload increases from 0.1 to 0.4 for each steganographic method, PSNR also increases. This means that as data grows in images, PSNR shows that generated image signals have high-payload data. Table 4 shows the average PSNR between images.

The experiment used support vector machines with 10-fold cross-validation using MATLAB R2021a. Four steganographic methods were used to develop the steganalysis dataset. After steganography, we have five datasets, as described in Section 3.1. Features set of each dataset created using S-UNIWARD, WOW, HUGO, and MiPOD contains 1,000 samples for normal and 4,000 samples for stego. Each feature is divided into two sets of training and testing after splitting it into two parts by a ratio of 70:30, respectively. The model was trained with the SVM training parameters provided in Section 3.5, and it was exported using the SVM kernel with the most outstanding performance. After the training model test dataset was imported, the model was tested.

Classification results of the proposed method on different steganographic methods are shown in Table 5. It illustrates three performance measures Accuracy, TPR, FNR, and AUC. Accuracy indicates to what extent the SVM model accurately classifies normal and stego images. TPR is the proportion of accurate forecasts in predictions of the positive class. FNR is the proportion of positive data cases mistakenly labeled as negatives to all positive instances in the data. The likelihood that a random positive (green) example will be placed in front of an unexpected negative (red) example is represented by AUC. AUC has a value between 0 and 1. A model with 100% incorrect predictions has an AUC of 0.0, whereas a model with 100% correct predictions has an AUC of 1.0. Performance metrics show the highest performance on two steganography methods, S-UNIWARD and HUGO, 97.2% and 98.6%, respectively. On the other hand, we have evaluated the proposed method on the combined dataset by combining all the features. SVM shows the accuracy of 95.9%, as shown in Table 5.

Classifiers better identify positive and negative data when their AUC values are closer to 1. The AUC measures how likely the classifier will assign a higher score to a randomly selected positive example than a randomly selected negative one. The AUC for a perfect classifier is 1, while the AUC for a random classifier is 0.5. The suggested method achieved much better binary classification results than any other available datasets. The area under the curve (AUC) must first be calculated to evaluate the proposed model properly. Fig. 3 presents the receiver operating graph, which allows for visualizing the area under the curve. On the left graph, the area under the curve (AUC) for the normal class is 0.97, and the area under the curve for the stego class is presented on the right graph. Both of these values are relatively near to 1, which indicates that this model produced the best results when applied to the classification process.

Different techniques for steganalyzer were described in Section 2. Some are machine learning-based, and others are based on deep learning. To evaluate the efficacy of a new model, it is useful to compare it to others that have been tried and tested using the same data. As part of this research project, we compared our approach to similar current approaches. SRNET [25], GBRAS-Net [19], and LSCA-YeNet [26] were the methodologies examined and compared. SRNET [26] showed 89.18% accuracy at the payload of 0.4 using S-UNIWARD, GBRAS-Net showed 89.8% accuracy at 0.4 payloads using WOW, and LSCA-YeNet showed 84.44% of accuracy at 0.4 payloads using WOW as we know that payload is directly proportional to detection rate. As a result, our system outperforms other approaches in the comparison study while using a minimal set of characteristics. As a rule, more features are seen to be better. Yet, when sophisticated calculations are unavailable, the time required to complete a particular task rises because of the large number of characteristics. It is already a challenge in model training to accomplish classification jobs that a high feature count might increase classification time with massive datasets. Table 6 shows the proposed efficiency and accuracy among them.