VisualNet: A Dual-Order Visual Cryptography Using Deep Learning Model

Visual Cryptography (VC) is a method employed to encrypt images in such a way that when the cipher is decrypted, it is easily decoded by the human eye, without any complicated mathematical computation needed [1, 2]. VC decryption is not based on computation but on visual perception, which is easy but strong. It is used to facilitate the reliable and highly secure sharing of images without a computational decryption algorithm or method, which reduces the computational complexity compared to traditional encryption-based algorithms. Thus, sensitive visual information can only be reassembled by summing up multiple shares, which boosts security and removes the overhead of key management [3-6]. The key differences between the VC and the conventional encryption-decryption algorithms have been clearly depicted in Table 1.

In a conventional image transmission system, the images are compressed and then transmitted to the remote unit through the noise enable channel environments [3]. Due to the restricted bandwidth of the wireless channels, the transmitted images should be compressed before they are transmitted into the wireless channels. Therefore, the compression algorithm is required to compress the encrypted images before they are directly transmitted into the bandwidth-limited wireless channels. The secret image is embedded into the source image with the aid of the compression technique, whereas the secret image is used as the key element for compressing the pixels in the source image. The type of compression technique is entirely based on the compression ratio [4].

The medical images are mostly required lossless compression technique, in which the loss of a single pixel is more important for analyzing the patterns in the secret medical images. Therefore, the conventional system requires a lossless compression technique for compressing the medical images which leads to a low compression ratio. The lossy compression technique has been used for compressing non-medical images which leads to a higher compression ratio [5, 6]. In this research article, the medical and non-medical images are encrypted using the proposed VisualNet model and Dual Order Compression-Encryption (DOCE) algorithm. The previous studies [7] are mainly related to lightweight semantic compression-based VC for IoT medical transmission, while this work proposes a shared generation framework based on deep learning (DL)and dual-order compression-encryption.

This research article is split into various sections. Section 2 states the traditional VC methods and approaches on various kinds of images along with their limitations and simulation results; Section 3 proposes a novel VisualNet DL model-based VC approach and DOCE algorithm for compressing the secret medical and non-medical images; Section 4 highlights the simulation results and their discussions, and finally, Section 5 states the conclusion with limitations of this proposed work.

Table 1. Key differences highlighted between Visual Cryptography (VC) and the conventional encryption-decryption algorithms

In this research article, the medical and non-medical images are encrypted using the proposed VisualNet model and DOCE algorithm. The secret medical and non-medical images are resized using preprocessing, and then they are fed into the proposed VisualNet DL model to generate the binary shares. These generated binary shares are embedded into the source images using the proposed DOCE algorithm to produce the encrypted images. The decrypted binary share images are generated and applied to the decryption system. It performs the addition operation to retrieve the original secret images. The performance of this proposed method has been estimated and analyzed with respect to the performance evaluation parameters PSNR, Mean Square Error (MSE), Pixel Difference Rate (PDR), and Mean Intensity Difference (MID). Two different medical and non-medical imaging datasets are used in this paper to analyze and compare the performance efficiency of the proposed VC algorithm with respect to recent methodologies.

Figure 1 shows the proposed VC system using double compression with the VisualNet model.

1.png

Figure 1. Proposed Visual Cryptography (VC) system using VisualNet model and Dual Order Compression-Encryption (DOCE) algorithm

3.1 Share generation using the proposed Convolutional Neural Network architecture

In this research work, the shares from the grey-scale images have been generated using the DL model. The traditional algorithm generates the shares as fixed patterns, and each fixed pattern is lower quality in nature. In the case of the proposed VisualNet model, the learned patterns (shares) are generated and their quality is higher. Hence, the reconstruction accuracy rate is higher in the proposed VisualNet model than the traditional share generation algorithm in the VC process. Figure 2 shows the proposed VisualNet model for the generation of shares.

2.png

Figure 2. Proposed VisualNet model for the generation of shares

The grayscale image is converted into a binary image and fed into the proposed VisualNet DL model to produce the variable share patterns. This proposed VisualNet model has been designed into four segments as depicted in Figure 2. Segment1 and segment2 have been used to compute the local feature maps from the binary secret image, and segment3 and segment4 have been used to compute the higher order edge feature maps, as illustrated in the following Eq. (1):

$\{Segments\} = \{segment1, segfment2, segment3, segment4\}$        (1)

The segment1 contains three modules: Convolutional layer (ConvA), Max pooling layer (MP), and Leaky Rectification Linear Unit (LR), which are given in the following Eq. (2-3):

$Segment 1=\left[\begin{array}{c}\operatorname{ConvA} \\ M P \\ L R\end{array}\right]$       (2)

ConvA $=(32$ filters and $5 \times 5$ filter kernel size $)$       (3)

The convolution function between the secret image and the filter kernels has been depicted in the below Eq. (4):

$C(i, j)=\sum_{p=0}^{M-1} \sum_{q=0}^{N-1} B(i+p, j+q) \times k(p, q)$        (4)

where, $B()$ represents the binary secret image, $k()$ is the filter kernel, M and N represent the width and height of the secret image and $C()$ is the feature map output.

The Max pooling layer (MP) and Leaky Rectification Linear Unit (LR) are given in Eqs. (5)-(6):

$M P=\{$Maxpooling with $4 \times 4$ window$\}$      (5)

$L R=\{$Leaky ReLU$\}$       (6)

The ReLU provides either 0 or 1 as output. Due to ‘0’ output, the neurons stop learning immediately, which is called the dying neuron problem. In order to eliminate such limitation, the LR activation function has been used in this proposed VisualNet model, and it is described by the following Eq. (7):

$L(y)=\left\{\begin{array}{c}y ; y \geq 0 \\ a \times y ; \text { else }\end{array}\right.$       (7)

The ‘a’ is a functional variable, and it is fixed at 0.01 after several trials. This LR also improves the gradient flow and computational efficiency compared to the traditional ReLU.

Segment2 contains three modules ConvB, MP, and LR, which are given in Eqs. (8)-(11):

Segment 2 $=\left[\begin{array}{c}\text { ConvB } \\ M P \\ L R\end{array}\right]$      (8)

ConvB $=(64$ filters and $5 \times 5$ filter kernel size$)$       (9)

MP $=\{$Maxpooling with $4 \times 4$ window$\}$        (10)

$L R=\{$Leaky ReLU$\}$       (11)

The computed local feature maps from segment1 and segment2 have been fed into Arithmetic Feature Adder (AFA), which adds these two local feature maps to enhance the local feature output. This output response is next passed through the ConvE, which is designed with 64 filter kernels and with the kernel size of 5 × 5. The gradient flow has been improved by eliminating the negativity in the ConvE layer output.

Segment3 contains three modules ConvC, MP, and LR, as presented in Eqs. (12)-(15):

Segment $3=\left[\begin{array}{c}\text { ConvC } \\ M P \\ \text { LR }\end{array}\right]$       (12)

ConvC $=(512$ filters and $3 \times 3$ filter kernel size$)$       (13)

MP $=\{$Maxpooling with $4 \times 4$ window$\}$        (14)

$L R=\{$Leaky ReLU$\}$       (15)

The computed local feature maps and edge feature maps from segment2 and segment3 have been fed into AFA, which adds these two feature maps to enhance the feature output. This output response is next passed through the ConvF, which is designed with 64 filter kernels and with the kernel size of 7 × 7. The gradient flow has been improved by eliminating the negativity in the ConvF layer output.

Segment4 contains three modules, ConvD, MP, and LR, and these are shown in Eqs. (16)-(19):

Segment $4=\left[\begin{array}{c}\text { ConvD } \\ M P \\ \text { LR }\end{array}\right]$       (16)

ConvD $=(1024$ filters and $3 \times 3$ filter kernel size$)$       (17)

MP $=\{$Maxpooling with $4 \times 4$ window$\}$        (18)

$L R=\{$Leaky ReLU$\}$       (19)

The computed higher-order edge feature maps from segment3 and segment4 have been fed into AFA, which adds these two higher-order edge feature maps to enhance the feature output. This output response is next passed through the ConvG, which is designed with 128 filter kernels and with the kernel size of 5 × 5. The gradient flow has been improved by eliminating the negativity in the ConvG layer output.

The final output from ConvE, ConvF and ConvG have been given to the Arithmetic Feature Map Integrator (AFMI) to produce the integrated output responses. This integrated response is fed into Fully Connected Neural Networks (FCNN) layers. In this proposed model, two fully connected layers with 4096 (for layer-1) and 512 (for layer-2) neurons have been assigned, respectively. The FCNN layer-2 produces two shares as Share-1 and Share-2. The size of Share-1 and Share-2 images is equivalent to the size of the binary secret image. The 512-neuron layer provides latent representations, which are decoded through convolutional and upsampling layers to generate two binary shares. Rather, it functions as a latent representation that is decoded using several convolutions and upsampling layers to generate two output tensors with the same spatial dimensions as the input image. A sigmoid activation followed by thresholding at 0.5 converts the outputs into binary shares.

The functional representations of these layers have been illustrated in the equations below.

AFMI output = concatenation{ConvE, ConvF, ConvG}

The hyper-parameters are the functional parameters which controls the functional output of the proposed VisualNet model in this work. The experimental setup of these functional hyper-parameters has been given in Table 2. The entire functional operation of this proposed model has been controlled by the set of parameters learning rate, batch size, epochs count, optimizer, momentum, activation function, stride, dropout, regularization, and regularization. All these experimental setup values are given in Table 2 and the internal layer specifications are given in Table 3.

Table 2. Functional experimental values of the hyper-parameters for the proposed VisualNet model

Table 3. Layer specifications

Figure 3(a) shows the PSNR computational graph with respect to epoch count on a non-medical imaging dataset, and Figure 3(b) shows the PSNR computational graph with respect to epoch count on a medical imaging dataset.

Figure 4(a) shows the PSNR computational graph with respect to learning rate on the non-medical imaging dataset, and Figure 4(b) shows the PSNR computational graph with respect to learning rate on the medical imaging dataset.

3.png

Figure 3. (a) Peak Signal-to-Noise Ratio (PSNR)computational graph with respect to epochs count on non-medical imaging dataset, (b) PSNR computational graph with respect to epochs count on medical imaging dataset

4.png

Figure 4. (a) Peak Signal-to-Noise Ratio (PSNR) computational graph with respect to learning rate on non-medical imaging dataset, (b) PSNR computational graph with respect to learning rate on medical imaging dataset

3.2 Encryption system

Encryption is the process of scramble the pixels in the secret image which is mainly used for image or data security. In this work, the encryption algorithm has been used for grey scale medical as well as non-medical images. In the conventional method, the encryption algorithm requires a key to scramble and descramble the original secret image, which is identified as the main limitation and this method of encryption consumes more time for generating the secret key. Moreover, compression is the process of compressing the pixels in the image and the compression techniques are categorized into either lossy or lossless based on the compression procedure. Here, the secret image is embedded into the source image with the aid of the compression technique, whereas the secret image is used as the key element for compressing the pixels in the source image. The type of compression technique is entirely based on the compression ratio. Medical images mostly require a lossless compression technique in which the loss of a single pixel is more important for analyzing the patterns in the secret medical images. Therefore, the conventional system requires a lossless compression technique for compressing the medical images which leads to a low compression ratio. The lossy compression technique has been used for compressing the non-medical images which leads to a higher compression ratio. In this article, the source image is only compressed, and hence a lossy compression technique has been used. In order to reduce the loss of pixels in the proposed lossy compression technique, the quantization tables (QAT) with low and higher quantization values have been used. Due to the utilization of these low- and higher-order QAT, the source image has been compressed by a dual compression process, which increases the compression ratio and reduces the loss of pixels during compression. This is the main advantage of this work to embed the secret medical image into the source image by dual QAT.

The entire DOCE algorithm is explained in the following steps:

Dual Order Compression-Encryption (DOCE) algorithm

Inputs: Source image-1 (SI-1), source image-2 (SI-2); Binary share-1 (BS-1), Binary share-2 (BS-2);

Outputs: Encrypted source image-1 (EI-1) and Encrypted source image-2 (EI-2);

Start:

Step 1:

The SI-1 and SI-2 images are resized into fixed 512 × 512 size in order to avoid overlapping and padding issues.

Step 2:

They resized source images are split into 8 × 8 non-overlapping blocks (NOB), and hence 4096 blocks are generated for each source image, respectively.

Step 3:

Generate two QAT tables with respect to lower-order quantization values and higher-order quantization values and these QAT tables are denoted as QAT-1 and QAT-2 as depicted in the following Eq. (20):

$Gsenerated Q A T s=\{Q A T-1, Q A T-2\}$       (20)

In order to retain good reconstruction quality and minimize quantization distortion during the embedding process, QAT-1 uses comparatively low quantization settings. By using larger quantization values, QAT-2, on the other hand, achieves better compression efficiency, stronger coefficient suppression, and increased robustness of the embedded information against noise and typical image processing procedures.

Step 4:

The $Q A T-1$ is constructed with lower-order quantization values, and hence the image that uses this $Q A T-1$ for the quantization process is compressed in order to obtain a low compression ratio. The $Q A T-1$ which is used for a low compression ratio, is given in Figure 5 below. Each value in this $Q A T-1$ represents the quantization value which is used to quantize the pixels in the source image. All the quantization values in this $Q A T-1$ are set between 0 and 15 and hence this $Q A T-1$ is used to produce the lower-order compressed source image.

5.png

Figure 5. $Q A T-1$ table which contains lower quantization values for performing quantization of the pixels in the source image for obtaining the compressed source image

Step 5:

The $Q A T-2$ is constructed with higher-order quantization values, and hence the image that uses this $Q A T-2$ for the quantization process is compressed in order to obtain a high compression ratio. The $Q A T-2$ which is used for a high compression ratio, is given in Figure 6 below. Each value in this $Q A T-2$ represents the quantization value which is used to quantize the pixels in the source image. All the quantization values in this $Q A T-2$ are set between 70 and 99, and hence this $Q A T-2$ is used to produce the higher-order compressed source image.

6.png

Figure 6. $Q A T-2$ table which contains higher quantization values for performing quantization of the pixels in the source image for obtaining the compressed source image

Step 6:

During this step process, the encrypted source images are generated by embedding the secret image based on the compression process. The compression process is performed based on the quantization tables $Q A T-1$ and $Q A T-2$.

The Encrypted source image-1 (EI-1) is generated using all Non-Overlapping Block of SI-1 and Binary Share image $B S-1$ by the below Eq. (21):

$E I-1=\left\{\begin{array}{lr}\frac{N O B}{Q A T-1} ; & \text { if } B S-1=1 \\ \frac{N O B}{Q A T-2} ; & \text { else }\end{array}\right.$       (21)

The Encrypted source image-2 (EI-2) is generated using all Non-Overlapping Block of SI-2 and Binary Share image $B S-2$ by the below Eq. (22):

$E I-2=\left\{\begin{array}{lr}\frac{N O B}{Q A T-1} ; & \text { if } B S-2=1 \\ \frac{N O B}{Q A T-2} ; & \text { else }\end{array}\right.$       (22)

Step 7:

Perform the reverse operation with the aid of the quantization tables $Q A T-1$ and $Q A T-2$ to recover the binary share images 1 and 2, respectively.

Step 8:

Compute PSNR value between the original BS-1 and the recovered BS-1. If the obtained PSNR value is less than 30 dB, then repeat step 6 using the equations below.

The Encrypted source image-1 (EI-1) is generated using all Non-Overlapping Block of SI-1 and Binary Share image $B S-1$ by the below Eq. (23):

$E I-1=\left\{\begin{array}{lr}\frac{N O B}{Q A T-2} ; & \text { if } B S-1=1 \\ \frac{N O B}{Q A T-1} ; & \text { else }\end{array}\right.$       (23)

The Encrypted source image-2 (EI-2) is generated using all Non-Overlapping Block of SI-2 and Binary Share image $B S-2$ by the following Eq. (24):

$E I-2=\left\{\begin{array}{lr}\frac{N O B}{Q A T-2} ; & \text { if } B S-2=1 \\ \frac{N O B}{Q A T-1} ; & \text { else }\end{array}\right.$       (24)

End.

7.png

Figure 7. (a) Source image which is to be compressed by secret image, (b) grey scale secret image, (c) converted binary secret image, (d) generated BS-1 image, (e) generated BS-2 image, (f) encrypted EI-1 image by BS-1, (g) encrypted EI-2 image by BS-2, (h) recovered secret image (bit error rate is 0.0009)

8.png

Figure 8. (a) Source image which is to be compressed by secret image, (b) grey scale secret image, (c) converted binary secret image, (d) generated BS-1 image, (e) generated BS-2 image, (f) encrypted EI-1 image by BS-1, (g) encrypted EI-2 image by BS-2, (h) recovered secret image (bit error rate is 0.0008)

Figure 7(a) shows the source image which is to be compressed by secret image (tumor case brain MRI image), Figure 7(b) shows the grey scale secret image, Figure 7(c) shows the converted binary secret image, Figure 7(d) shows the generated BS-1 image, Figure 7(e) shows the generated BS-2 image, Figure 7(f) shows the encrypted EI-1 image by BS-1 and Figure 7(g) shows the encrypted EI-2 image by BS-2 and Figure 7(h) shows the recovered secret image.

Figure 8(a) shows the source image which is to be compressed by secret image (non-tumor case brain MRI image), Figure 8(b) shows the grey scale secret image, Figure 8(c) shows the converted binary secret image, Figure 8(d) shows the generated BS-1 image, Figure 8(e) shows the generated BS-2 image, Figure 8(f) shows the encrypted EI-1 image by BS-1 and Figure 8(g) shows the encrypted EI-2 image by BS-2 and Figure 8(h) shows the recovered secret image.

In this article, the VisualNet DL model-based encryption algorithm with a compression technique has been proposed to strengthen the image security system. The proposed system has been tested on two different datasets KBI and PR with respect to medical and non-medical images. This proposed system has been simulated, and its performance has been measured with respect to various parameter metrics PSNR, MSE, PDR, and MID, along with the encryption and decryption time period. The proposed VisualNet-based VC system obtains 37.7 dB PSNR, 12.12 MSE, 92.97 PDR, and 33.31 MID for the non-medical images on the KBI dataset. The proposed VisualNet-based VC system obtains 37.4dB PSNR, 12 MSE, 93.61 PDR, and 33.52 MID for the medical images on the PR dataset. The proposed DOCE algorithm consumed 0.25 ms for the encryption process and also consumed 0.21 ms for the decryption process on KBI dataset images. The proposed DOCE algorithm consumed 0.22ms for the encryption process and also consumed 0.27ms for the decryption process on PR dataset images. Even though this proposed method has advantages over conventional systems, it has certain limitations as stated below.

• This proposed VisualNet model has not been applied and checked with the frames in the video, which is important for telemedicine and its related medical applications.
• The security attack analysis has not been performed in this proposed method, which is important for strengthening the image encryption model, which will be considered as the future scope of this research work.