MOTH FLAME OPTIMIZATION WITH HYBRID LSB EMBEDDING FOR ROBUST IMAGE STEGANOGRAPHY IN SPATIAL DOMAIN
Abstract
Data security is a major challenge in the communication environment because of the large volume of data generation where there is no security of data among the users. Hence, image steganography is introduced to ensure the security of the user sensitive information during communication. Several researchers have tried to improve the performance of steganography methods by developing various algorithms for optimal pixel selection and enhancement of image quality algorithms. The previous researchers hide the sensitive data in the transformed image coefficients randomly. But the performance of random data hiding is low. This paper presents an approach for image steganography using a Moth Flame optimization algorithm (MFO) that selects the pixel effectively. Then, a hybrid LSB (Least significant bit) algorithm is applied for secret data hiding. Here, the average pixel value differencing is computed in the optimal pixels. The pixel is considered as it is in edge area when the value is greater than 15 else; it is in a smooth area. The LSB and EMD combination is used in the smooth area pixel embedding, and the combination of LSB and PVD is used in the edge area pixel embedding. The proposed image steganography is simulated and compared with existing works. The simulation results show the performance of proposed work as well as the existing work, which clearly describes that the proposed work is robust than the current works.
Keywords: Steganography, LSB, EMD, PVD, Moth flame optimization
- Introduction
The confidentiality of data among the network has a major concern due to the growth of data generation. For every second, there is a large volume of data transfer from one end to another end. For example, Facebook, Twitter, emails, and file sharing services. In this situation, there is a chance for malicious users to access user confidential data with bad intentions like corrupting and damaging in the data.
The presence of a message is covered in appropriate digital media, For example, image, audio, and video, and this are called steganography. The Greek words staganos and graphein derives the term “Steganography” [1]. The motive of steganography is that the hidden data is invisible to the person so that the third party or attacker cannot detect the hidden data. Therefore, the steganography approach is identified as a key by the trusted users to transmit their sensitive data from one location to another location [2].
Steganography plans to embed the data into an image by giving security and minimizing the embedding capacity for such databases. The database is created by every field with important details. The bank keeps its customer data, which contains PAN, account number, Adhar, and some other information, and these data are stored in a database. Likewise, educational environments such as schools, colleges, and medical environments collect their user multiple data and stored in a database. In the first place, the database is made sure about utilizing cryptography. On the off chance that by any stretch of the imagination, programmers can assault and get data about a person steganography gives the security at the following level [3].
A single image that is inserted with all data of an individual to such that only a photograph of an individual is obvious to an unapproved individual. It gives better security, and as the quantity of images is decreased to one, memory required for capacity is less contrasted with having three or four pictures. Numerous calculations are created in image steganography running from spatial-area to Transformation-space and stretched out to hybrid domain.
Spatial domain and frequency domains are the two groups in data hiding techniques. LSB and BPCS (Bit Plane Complexity Segmentation) are spatial domain data hiding. Several authors proposed various methods based on the transform domain and spatial domain over the previous years. The DWT domain improves the secret image quality as well as the confidential data embedding capacity. The plan of LSB’s is directly replaced by the LSB substitution approach in the cover image for hiding secret data without modification [4].
The result is better when the steganography is performed with cryptography. So the encryption algorithm is used to hide the data before it transmits from one location to another [5]. The major characteristics of cryptography are authentication, integrity, and confidentiality [6]. The encrypted data is hidden by steganography so that no one can identify the secret data. The steganography consists of host object, sensitive data, and stego object. If the data is embedded into audio means, the output is stego audio [7]. The Steganography approach seems to be a good one if it considers three parameters for the processing, which means capacity, security, and image quality [8].
Generally, the steganography methods conceal an equal number of secret bits into each image pixel of the cover image. Thus, the quality of an image is automatically reduced while performing an equal amount of bits changes in all pixels of the cover image [9]. To tackle these issues, different adaptive embedding processes have been introduced in the steganography process [10]. Thus, most of the researchers focused on these adaptive techniques. Furthermore, the quality of edge pixels is not much affected after making changes during the embedding process. Thus, the embedding procedure should embed the secret data by considering certain optimization problems in the embedding process. The bio-inspired algorithms are recently used by many researchers to solve the various problem in different domains such as engineering and science. Also, there are several evolutionary algorithms such as PSO (Particle Swarm Optimization), GA (Genetic algorithm) and ABC (Ant Bee Colony) optimization algorithm used in image steganography. The metaheuristic algorithms are used in most of the applications to solve optimization issue [11]. The contribution of the paper includes the following:
In this paper, image steganography for secure data transmission from one location to another is proposed. The idea of a chaotic neural network is applied for encryption to achieve more security. The encrypted image is transformed to the frequency domain by IWT. The cost function of the MFO algorithm chooses an optimal pixel from the cover image. The next stage is data embedding and here we use a hybrid Least Significant Bit (LSB) and Exploiting Modification Direction (EMD). Embedding in edge area is done using the combination of LSBPVD (Least Significant Bit and PVD), whereas embedding in the smooth area is done using LSBEMD (Least Significant Bit and EMD). The stego image is obtained after completing the above processes.
The structure of the paper is described as: The introduction is presented in section 1. The related optimization-based steganography works are explained in section 2. The MFO based pixel selection and hybrid LSB are explained in section 3. The simulation results and analysis are explained in section 4. The paper is concluded in section 5.
- Literature survey
The relevant works proposed by previous researchers are discussed in this section.
Uma maheswari et al. [12] used GA and PSO algorithm for image steganography to enhance the performance of the embedding approach. Here, the best coefficients were obtained by GA and PSO for QR (Quick Response) coded secret data hiding.
The hybrid PSO algorithm was proposed by Navdeep kaur et al. [13] gor steganography. The steganography approaches were suffered by security and privacy problems even they hide the secret data in the host image. So, the PSO based hybrid algorithm was introduced by the author for optimal pixel selection. The PSO uses ACO (Ant Colony Optimization) algorithm for security and image quality to maintain purpose. After completing the data hiding process, the work was compared with PSO, ACo, and hybrid PSO. The hybrid PSO showed better performance than the simple PSO and ACO.
An ABC (Artificial Bee Colony) based steganography was introduced by Anan Banharnsakun et al. [14] to improve the LSB image steganography. Here, the secret data (image) was hidden by the optimization of the ABC block assignment in the cover image. The ABC solutions were represented as a block assignment list.
A pixel prediction based steganography was proposed by V. K. Reshma et al. [15]. The author used the SVNN classifier, which was error dependent for pixel identification. The pixel features were extracted by the SVNN classifier to obtain proper pixels in the medical image. Then, the Genetic algorithm was used to train the SVNN.
A PSO based image steganography was proposed by A. H. Mohsin et al. [16]. Here, the PSO algorithm optimally selects the pixel for secret data hiding in the spatial domain. Vaclav Snasel et al. [17] proposed an advanced steganography algorithm for the image using the instruction of AVX. This AVX algorithm used to reduce the execution time on the test data.
A hybrid optimization-based approach was developed by Ambika et al. in [18]. The hybrid optimization is the combination of Elephant Herding optimization and Monarch Butterfly optimization (EH-MB) algorithm. Here, multilevel DWT was used for domain conversion. The cost function was used by the optimization to find out the pixel. A transform domain based approach was proposed by Mansi S. Subhedar et al. [19]. Here, a contourlet transform technique and three popular matrix factorization technique was used for efficiency.
The above mentioned strategies require extra upgrades that relate to the unique proof of included highlights separated from neighboring pixels that are used in the data embedding procedure. The proposed work aims to enhance the embedding capacity of cover image, to reduce the distortion rate when the noise occurs. Thus, the research intention is to meet these objectives and demonstrate that our suggested system is suitable contrasted and the past investigations.
- Image steganography using MFO optimization
This paper presents an efficient method for user sensitive data (i.e secret image) hiding into the cover image which process generates the stego image. Then this stego image is sent to another user (receiver) and the receiver extracts the hidden secret image using an extraction algorithm from the stego image. Here, we use the cost-based MFO algorithm and the MFO based image steganography analysis. For better understanding, the block representation of the proposed architecture is represented in figure 1.
Fig. 1: Proposed architecture
There are four steps used in the proposed work to achieve secure image steganography. The steps are image encryption, spatial to frequency domain conversion, embedding, and extraction. The image embedding process uses the cover image and secret image. The wavelet transform is applied into the stego key that is the cover image for spatial to frequency domain conversion then the pixel for secret data hiding is selected using MFO. The data embedding technique used here is a hybrid Least Significant Bit (LSB) and Exploiting Modification Direction (EMD). Embedding in edge area is done using the combination of LSBIPVD (Least Significant Bit and Improved PVD) whereas embedding in the smooth area is done using LSBIEMD (Least Significant Bit and Improved EMD). After completing the embedding process, The inverse IWT is applied in the hidden image for the spatial domain. The last process is secret image extraction using the reverse process of embedding.
3.1 Secret image encryption
A chaotic neural network [20] algorithm is used to work for secret image encryption in this work. The chaotic neural network algorithm is comprised of three different blocks, such as a key generator, neuron layer, and permutation of neuron layer. In the neuron layer, three input layers are presented, and the permutation neuron layer contains three output layers, also three neurons. These two layers are supported by a key generator block with corresponding weights and biases. The inputs (linear combination) are delivered by two nonlinear sections, such as nonlinear optimization and bitxor operations. After completing this operation, the input enters into the activation function. The diffused data is the output of this layer, which is the input of the permutation neuron layer. Here, two steps are used to complete the permutation. At first, linear permutation function is used in the secret data, and then the chaotic key generator generates the permutation matrix. Finally, the two dimensional Cat map permutation algorithm shuffles the permuted strings. The steps are repeated for more time to get high security.
3.2 Domain conversion using wavelet transform
The integer dataset is mapped into another dataset by using Integer to integer wavelet transforms. The exact original dataset is produced by this wavelet transforms perfectly. The 1-dimensional DWT is called as a frequent filter bank algorithm. Convolution with the synthesis filters are involved in the reconstruction and add the convolution results. The 2-dimensional DWT first applies the one-step of 1-dimensional transform for entire rows and columns. The rows and columns produce resultant coefficients for further process. Integer wavelet transform (IWT) [21] is employed to the cover image for the process of cover adjustment. Here, the integer coefficients are used by IWT to avoid the rounding error and it provides perfect reconstruction. A slight modification of DWT (Discrete wavelet Transform) obtains the IWT. Generally, the DWT uses three phases such as splitting, predication, and update and these steps are called a lifting scheme. Simple filtering operations are used to modify the sample odd and even sequences. The following expressions describe the modifications,
(1)
(2)
From the above condition, O[m] represents the odd sample and E[m] represents the even sample. Then the odd samples are converted into a high-frequency coefficient and even samples are converted into a low-frequency coefficient.
3.3 Embedding phase
The high frequency and low-frequency coefficients are extracted by IWT, and these extracted coefficients are then used for optimal pixel selection for embedding. Here, the cost function is used for fitness evaluation, which includes IEB (Intensity, Entropy, and Brightness) of the swarms. Moth Flame optimization is the nature-inspired population-based optimization that is applied in various applications when compared to the existing optimization algorithms.
MFO based optimal pixel selection
The MFO optimization is the swarm optimization algorithm, and it is enthused with the moth’s navigation method during the night, which is known as transverse orientation. The main modules of MFO are moths and flames. These moths and flames differ from other optimization for updating in the iteration. The moths and moths position are supposed as candidate solutions and problem variables, respectively. The moths are fly with 1-dimension, 2-dimension, and hyper dimension by moving its position vectors. In MFO, the moths are called as the search agent and the moths optimum position are called flames. Every moth is searching and moving around a flame and it updates the position when it obtains better solution than the previous solution. The logarithmic spiral function is the major approach for solution updation.
Moths are normally insects that are most similar to butterflies. Larvae and adult are the two important periods of their lifetime. In the first period, the larvae is changed into moth in cocoons. The moths have the fascinating point that their distinct navigation way during night. The moths used the moon light to fly and transverse orientation approach for navigation at night. The moth can fix an angle for travel for flying by follows the moon in a straight way for long distances. When the light is very close, the moths flies spirally. The lights trap them and use the fixed angle. If the light is long distance, then the transverse function is used.
MFO algorithm [22] is applied here for finding the optimal pixel in this work. The stable angle is maintained by Moth on behalf of the moon named transverse orientation. The false lights trickle these moths and maintain the comparative edge to the light source. Moths are assigned by the MFO algorithm for different weights and every moth represents their fitness value. Also, Every moth has a glow that saves the optimal pixel setup.
In this algorithm, the population is randomly selected in the beginning. Then the position of moths and flames are updated by fitness function. From that, the best position is selected for the embedding process. The MFO steps are explained below based on the converging behavior towards an artificial light.
Step 1: The moths and flames position model the MFO algorithm. In the search space, these moths and flames have different dimensions by set several variables for every moth and flame. There are two types of matrix are used to represent the MFO. The moth position matrix is represented at first and the flames position matrix is represented at next.
(3)
Where Moth represents its position matrix and Flame represents its position matrix. The number of moths used here is ‘m’ and the dimension (number of variable) is represented as ‘e’
Step 2: The input of both moth and flame fitness evaluation is the position of moths and flames that provides fitness value to evaluate the moth as well as the flame. The fitness function is dependent on the cost function of MFO. The corresponding moth fitness values are stored in the below matrix ‘Ft_om’,
(4)
The corresponding flame fitness values are stored in the below ‘Ft_of’ matrix
(5)
Step 3: The MFO general surface consist of the following function,
(6)
The moth and flame initialization are done by the below condition,
(7)
In condition (7), u represent the count of moth and flame. V represents the dimension and R represents the random number which is generated in the period [0,1] in uniform distribution. After completing the initialization, the iterations are run-up to the termination condition. The moth position is updated for the corresponding flame using the below condition in the iteration.
(9)
The logarithmic spiral function is denoted as a spiral in equation (9). The and represent the ith moth and jth flame respectively. The below condition gives the logarithmic spiral,
(10)
In equation (10), denotes to the distance between the ith moth and jth flame, and it is computed as:
(11)
‘k’ is a constant in equation (10) for describing the spiral function, and the random number is denoted as ‘r’.
Step 4: Optimal solution determination: The flame position updation occurs when the moth is fitter than flame. The flames fitness values are evaluated by fitness function. The population is deliberated as and fitness function is described in equation 12,
(12)
The cost function is represented as and it is evaluated using the parameters mentioned above, such as intensity, entropy, and brightness. Therefore the cost function is,
(13)
The seed point intensity between the ith row and jth column is characterized as , and the seed point entropy between the ith row and jth column is described as and the seed point brightness between the ith row and jth column is characterized as . The equation (14), (15), and (16) compute the cost parameters intensity, entropy, and brightness.
(14)
(15)
(16)
The seed point cost parameters from the ith row and jth column are computed with defining eight adjacent seed points mean value. From the rule, the better solution is considered as the best pixel region if the criteria of iteration are reached. The best solution is found by searching the overall position, and once we get the best solution, it is considered as the best position, and it is taken for embedding purpose.
Fig. 2: Convergence analysis
Figure 2 illustrates the convergence rate. This convergence comparison uses two standard images and compared with the different optimization algorithm. This optimization runs many times until convergence with convergent parameter set. The Baboon image got the best value got in the 18th iteration and the Lena image got the best values around 20th iteration. The existing optimization algorithm found bet fitness in the 32, 35, 37 iteration.
Combination of LSB and PVD approach for embedding
Bit 1 substitutes the first LSB of pixel which is the sign for hidden image extraction. The 2 data bits substitute another 2 LSBs. Therefore, a new pixel value is found. Let us consider is the decimal value of three LSBs is and is the decimal value of the three LSBs is . From that a difference value is calculated as and optimization is done by the below condition,
(17)
Here, three difference values are computed for , . This computed difference value falls into the range. This range decides how many bits can be hidden. Then the decided bits are converted into decimal values . Then the difference is computed for this new value. From that, two values are selected by the below condition,
(18)
LSB and EMD combination based embedding
The bit 0 replaces the first LSB bit of pixel which is the identifier of the extraction process. Then two data bits replace the other two LSBs . Therefore, a new value is obtained . The decimal value of the three LSBs of is and the decimal value of the three LSBs of is . A difference value is computed as and is optimized by (17)
3.4 Extraction phase
In the extraction process, the secret data is extracted from the considered cover image. Here, the hidden image uses IWT to represent the frequency coefficients of the image. The stego image is the input of the data extraction process. Generally, IWT represents the spatial domain image into a frequency domain by providing low and high frequency coefficients. The hidden image extraction process includes the same steps as embedding but in reversible order. At fisrt, the extraction procedure uses both the combination of LSB+PVD and LSB+EMD based on the condition. The LSB bit is 1 means the PVD is joined with LSB, and if it is zero means EMD is combined with LSB. This process is continued upto extract the ‘k’ pixels. Then, the decryption process is performed which is the reverse process of encryption.
- Simulation settings and results
In this section, the setup used for simulation, the dataset used and results are discussed for the proposed MFO and the results are compared with the ABC algorithm [21], SVNN algorithm [22], PSO algorithm [23], EH-MB [25]. The performance metrics are evaluated and compared with the mentioned existing algorithms. Figure 3 shows sample images used in the simulation. The first column denotes the cover image, the second column denotes secret image and the last column denotes the stego image. We planned for assessing the arrangement got from our MFO based steganography approach as far as the image quality and the twisting resilience of hidden images. The entire strategies right now executed in Matlab tool and all examinations were executed on a PC with an Intel Core i7 CPU, 2.8 GHz and 16 GB of memory. In the MFO, the quantity of moths and flames was set and the quantity of iterations was set. In the execution, the cover iamge resolution is 512×512 and the secret image resolution is 256×256. The input image, as well as the cover image, are grayscale images. It is utilized and the example pictures are spoken to in figure 3. The Lena image and Baboon images are taken as a sample images in which Cameraman and Jet images are hidden and finally, the stego image is displayed. The figure 3 (a) represents the cover image, 3(b) represents the hidden image, and 3(c) represents the stego image.
- (b) (c)
- (b) (c)
Fig.3 Sample images
In objective evaluation, the performance metrics for different images are executed to find the distortion of the stego-images using PSNR, MSE, and payload capacity. To assess the quality degree of the outcomes acquired from our proposed approach and the other previously mentioned strategies, the PSNR was utilized as the parameter for comparison.
The distortion rate and the similarity extent of an image can be quantified by MSE to measure the reliability as given below:
(19)
Fig. 4: MSE analysis
Where, represents pixel count in the cover image. The cover image and stego image are defined as C and S, respectively. The PSNR computation is expressed as follows:
(20)
Fig. 5: PSNR analysis
The volume of hidden data without destructing the cover image performance is called payload capacity, and it is computed by the below condition,
(21)
Fig. 6: Payload capacity analysis
Figures 4, 5 and 6 show the performance analysis of image steganography such as PSNR, MSE and payload capacity. The proposed work becomes a better performance in all the metrics. From the results, the proposed method PSNR is higher than the existing method for the images used for simulation. Also, the proposed method provides a minimum MSE value because it alters minimum bits in each smooth area pixels. However, the existing methods decrease the quality of the stego image while increasing the payload capacity. Instead, the quality of stego image obtained by the proposed method is very high for all the cover images. If the payload capacity is increased to 2.7bpp then, the PSNR of the proposed method reaches up to 60dB for all images. But, the PSNR of the existing methods is less than 45dB. Since the proposed approach embeds the secret bits into both smooth region and edge region by setting different parameter values in the steganography embedding function. The edges are less affected by the changes after hiding secret data. Hence, more data is embedded in the edge regions using higher parameter value. Therefore, all edge pixels are identified without losing any edge pixel. Thus, the payload capacity is automatically improved.
- (b) (c) (d) (e)
- (b) (c) (d) (e)
Fig. 7: Noise analysis
The proposed strategy got 2% impulse noise when the extraction of secret image from cover image. The proposed work is contrasted and different strategies have appeared in Figure 7. Figure 7(a) represents the proposed image noise analysis and (b), (c), (d) and (e) represent the exiting algorithm [21] [22] [23] [25] respectively. A more clear correlation of their presentation can be made by analyzing the contrasts between the extricated secret images and the implanted picture. As defined in figure 7, the noise level of proposed work is lower than the other approaches which is taken for comparison. Because the MFO creates better outcome by best position updates and iteration level. The extracted stego image results dependent on the BER assessment of image utilizing the proposed strategy are represented by charts in Figure 8. During the extraction stage, the embedding process is performed in reverse order to extract the encrypted hidden image from the received stego image. It does not require the cover image. Then, the extracted hidden image is decrypted.
Fig. 8: BER analysis
Figure 8 show that the average BERs of the proposed method were always better than of those obtained with other approaches. The average BER in our approach is lower than that of the different approaches in all sample host images. This optimal embedding process has been embedded the secret image by selecting optimal parameters for smooth and edge pixels. Thus, the overall security level of the proposed steganography scheme has been improved.
- Conclusion
This paper proposed a new image steganography approach to enhance the usual LSB approach. Here, the LSB is enhanced by obtaining the optimal best pixels in the cover image for embedding the secret image using the Moth Flame Optimization (MFO) algorithm. After finding the pixels, an efficient embedding algorithm integrates the hidden image in the corresponding edge and smooth area pixels. This optimal pixel, and embedding algorithm improves security as well as image quality. The simulation shows better performance in metrics such as image quality that is measured by PSNR, embedding capacity, and security. We would thus be able to infer that the MFO strategy is profoundly effective from the viewpoint of algorithm performance and solution quality. Also, accordingly, the MFO based proposed approach can fill in as an effective option in the image steganography area.
References
[1] Douglas, Mandy, Karen Bailey, Mark Leeney, and Kevin Curran. “An overview of steganography techniques applied to the protection of biometric data.” Multimedia Tools and Applications 77, no. 13 (2018): 17333-17373.
[2] Muhammad, Khan, Jamil Ahmad, Naeem Ur Rehman, Zahoor Jan, and Muhammad Sajjad. “CISSKA-LSB: color image steganography using stego key-directed adaptive LSB substitution method.” Multimedia Tools and Applications 76, no. 6 (2017): 8597-8626.
[3] Jiang, Nan, Na Zhao, and Luo Wang. “LSB based quantum image steganography algorithm.” International Journal of Theoretical Physics 55, no. 1 (2016): 107-123.
[4] Sharma, Vijay Kumar, Devesh Kumar Srivastava, and Pratistha Mathur. “Efficient image steganography using graph signal processing.” IET Image Processing 12, no. 6 (2018): 1065-1071.
[7] Tang, Hongmei, Gaochan Jin, Cuixia Wu, and Peijiao Song. “A new image encryption and steganography scheme.” In 2009 international conference on computer and communications security, pp. 60-63. IEEE, 2009.
[8] Pant, Vinay Kumar, Jyoti Prakash, and Amit Asthana. “Three step data security model for cloud computing based on RSA and steganography.” In 2015 International Conference on Green Computing and Internet of Things (ICGCIoT), pp. 490-494. IEEE, 2015.
[9] Garg, Priyansha, Moolchand Sharma, Shivani Agrawal, and Yastika Kumar. “Security on cloud computing using split algorithm along with cryptography and steganography.” In International Conference on Innovative Computing and Communications, pp. 71-79. Springer, Singapore, 2019.
[10] Balasubramanian, C., S. Selvakumar, and S. Geetha. “High payload image steganography with reduced distortion using octonary pixel pairing scheme.” Multimedia tools and applications 73, no. 3 (2014): 2223-2245.
[11] Rabie, Tamer, Mohammed Baziyad, and Ibrahim Kamel. “Enhanced high capacity image steganography using discrete wavelet transform and the Laplacian pyramid.” Multimedia Tools and Applications 77, no. 18 (2018): 23673-23698.
[12] Maheswari, S. Uma, and D. Jude Hemanth. “Performance enhanced image steganography systems using transforms and optimization techniques.” Multimedia Tools and Applications 76, no. 1 (2017): 415-436.
[13] Kaur, Navdeep, and Anu Garg. “Steganography Using PSO Based Hybrid Algorithm.” International Journal of Advanced Research in Computer Science and Software Engineering 4, no. 1 (2014).
[14] Banharnsakun, Anan. “Artificial bee colony approach for enhancing LSB based image steganography.” Multimedia Tools and Applications 77, no. 20 (2018): 27491-27504.
[15] Reshma, V. K., RS Vinod Kumar, D. Shahi, and M. B. Shyjith. “Optimized support vector neural network and contourlet transform for image steganography.” Evolutionary Intelligence (2020): 1-17.
[16] Mohsin, Ali Hadi, A. A. Zaidan, B. B. Zaidan, O. S. Albahri, A. S. Albahri, M. A. Alsalem, K. I. Mohammed et al. “New Method of Image Steganography Based on Particle Swarm Optimization Algorithm in Spatial Domain for High Embedding Capacity.” IEEE Access 7 (2019): 168994-169010.
[17] Snasel, Vaclav, Pavel Kromer, Jakub Safarik, and Jan Platos. “JPEG steganography with particle swarm optimization accelerated by AVX.” Concurrency and Computation: Practice and Experience (2019): e5448.
[18] Biradar, Rajkumar L. “Secure medical image steganography through optimal pixel selection by EH-MB pipelined optimization technique.” Health and Technology (2019): 1-17.
[19] Subhedar, Mansi S., and Vijay H. Mankar. “Image steganography using contourlet transform and matrix decomposition techniques.” Multimedia Tools and Applications 78, no. 15 (2019): 22155-22181.
[20] Wang, Xing-Yuan, Lei Yang, Rong Liu, and Abdurahman Kadir. “A chaotic image encryption algorithm based on perceptron model.” Nonlinear Dynamics 62, no. 3 (2010): 615-621.
[21] Alotaibi, Reem A., and Lamiaa A. Elrefaei. “Text-image watermarking based on integer wavelet transform (IWT) and discrete cosine transform (DCT).” Applied Computing and Informatics 15, no. 2 (2019): 191-202.
[22] El Aziz, Mohamed Abd, Ahmed A. Ewees, and Aboul Ella Hassanien. “Whale optimization algorithm and moth-flame optimization for multilevel thresholding image segmentation.” ExpertSystems with Applications 83 (2017): 242-256.