Secure Surveillance Systems Using Partial-Regeneration-Based Non-Dominated Optimization and 5D-Chaotic Map

: Due to Internet of Things (IoT), it has become easy to surveil the critical regions. Images are important parts of Surveillance Systems, and it is required to protect the images during transmission and storage. These secure surveillance frameworks are required in IoT systems, because any kind of information leakage can thwart the legal system as well as personal privacy. In this paper, a secure surveillance framework for IoT systems is proposed using image encryption. A hyperchaotic map is used to generate the pseudorandom sequences. The initial parameters of the hyperchaotic map are obtained using partial-regeneration-based non-dominated optimization (PRNDO). The permutation and diffusion processes are applied to generate the encrypted images, and the convolution neural network (CNN) can play an essential role in this part. The performance of the proposed framework is assessed by drawing comparisons with competitive techniques based on security parameters. It shows that the proposed framework provides promising results as compared to the existing techniques.


Introduction
Internet of Things (IoT) has changed the way of life by providing connectivity among network-enabled devices. Surveillance is a major advantage of IoT that helps people to keep an eye on the security of homes, business, and critical places. Due to this, a large volume of data is transferred over the network that is further saved on servers. These data can be in the form of images, audio, video, and text, and are further utilized in many applications, such as military communication, healthcare, identification of suspicious activities, remote sensing, etc. As these data contain valuable information, any leakage of the data may create security breaches, which can further hamper the working of government, intelligence agencies, and personal privacy. Therefore, it is necessary to create a secure surveillance framework for IoT systems. In this work, our focus is to secure the images during transmission and storage, because images play an important role in various multimedia applications. The image captured from IoT devices can be sent to a server or directly to concerned personnel for surveillance. If the images are in plain form, then they can reveal all the information to an attacker. Therefore, transmission and storage of the images should be secured. To secure the images, cryptography is the most suitable

•
The secret keys are generated by utilizing a 5D hyperchaotic map. The reason of selection is to generate more complex pseudorandom numbers that cannot be easily estimated.

•
To strengthen the security of the hyperchaotic map, the initial parameters are optimized using partial-regeneration-based non-dominated optimization (PRNDO). A multi-objective fitness function is designed using correlation coefficient and entropy. These two parameters were selected after analyzing their wide acceptability in the literature. Fitness functions based on these parameters provide better optimized solutions. • After the secret key's generation, permutation is performed to change the pixels' position. Thereafter, the diffusion process is performed on scrambled images to produce the encrypted images. • Finally, the performance of the proposed technique is assessed through comparative analyses with competitive techniques based on security parameters.
The rest of the paper is organized as follows: The related work is discussed in Section 2, background is presented in Section 3, and Section 4 elaborates on the image encryption technique for the secure surveillance framework for IoT systems. In Section 5, the simulation results and analyses are discussed, and Section 6 concludes the paper and presents the future research directions.

Literature Review
DNA and hyperchaotic maps (DNA-HC) were used to encrypt the images. In [6], a secure surveillance system was designed using the sine tent cosine approach for the transmission of images, while in [7], a secure communication system was developed using compressive sensing and chaotic map (CSCM) for IoT applications. The authors of [8] encrypted the images using permutation and diffusion operations. The permutation process was carried out using a Henon map, while a chaotic restricted Boltzmann machine was used to perform the diffusion operation on the images [8]. In [9], compressive sensing was used to encrypt the images, where a 3D chaotic map was used to generate the measurement matrix and a 1D chaotic map was utilized to diffuse the pixels of the scrambled image. The authors of [10] proposed an improved fractional 1D chaotic map to perform the image encryption. The permutation and substitution steps were combined to change the position of pixels and the values of an image [10]. The authors of [11] investigated and presented current trends on IP and DS in IOT.
The authors of [12] utilized a differential equation to generate the chaotic oscillator, which is used to scramble the pixels of an image to obtain an encrypted image. In [13], a DP model was proposed for security and privacy. The authors of [14] implemented an image encryption framework: Fresnel transform. In this technique, a lens is not used, and the image is divided into three channels and each is encrypted using a separate key [14]. In [15], a time delay chaotic system was used to encrypt the images. The authors of [16] applied coupled map lattices to generate the keys. In this technique, the random-based diffusion process is implemented to encrypt the images [16]. The authors of [17] developed an image encryption scheme using a 6D chaotic map. The position of pixels of images is permuted at the bit-level to enhance the security, and the pixels are finally diffused using DNA coding. In [18], the authors encrypted the images using tent-dynamics coupled map lattices, and the cyclic shift algorithm was utilized to shift the rows and pseudorandom sequences. The above-discussed techniques involved the use of chaotic maps to generate the secret keys, and although these techniques provide secure image transmission and storage, in most of the techniques, the initial parameters of chaotic maps were not selected optimally. Thus, using parameter estimation techniques, these techniques can be broken [18]. The authors of [19] presented a deep learning model for traffic flow prediction.
Some of the researchers have addressed the above-stated issue by implementing the meta-heuristic techniques in image encryption. The authors of [20] developed an encryption system using DNA and the 1D logistic map for secure transmission of images. Particle swarm optimization (PSO) was utilized to select the best encrypted image based on entropy and the correlation coefficient. The non-dominated sorting genetic algorithm-II (NSGA-II) was used in [21] to optimize the intertwining logistic map, the genetic algorithm (GA) was applied in [22] to optimize the beta-chaotic map, and the Pareto evolutionary algorithm-II was implied in [23] to optimally select the parameters of the 4D chaotic map. Adaptive differential evolution (ADE) was used in [24] for tuning the hyperparameters of the Lorenz chaotic map. Grey hole and black hole attacks were presented in [25], while the 5D chaotic map was tuned in [26] using NSGA. An intertwining logistic map was optimized in [27] using memetic differential evolution, the hyperparameters of the betachaotic map were tuned in [28] using differential evolution, and IoT-based risk assessments were presented in [29].

Five-Dimensional Hyperchaotic Map
The five-dimensional (5D) hyperchaotic map (5DHCM) provides significantly complex chaotic behavior [30,31]. The 5DHCM has five attributes, a cubic nonlinear product, and a better Lyapunov exponent (LE). The huge number of attributes improves the key space, the large LE provides a more dynamic 5DHCM, and the cubic nonlinear product achieves good resistance [32]. Thus, 5DHCM can achieve better encryption keys and can be computed as: (1) Here, in Equation (1), d 1, d 2, d 3, d 4, and d 5 are initial attributes required for 5DHCM, and q, r, s, t, u, v, and w define the control attributes of 5DHCM. Chaotic attractors are represented in Figure 1 and Figure 4 for q = 30, r = 10, s = 15.7, t = 5, u = 2.5, v = 4.45, and w = 38.5. Figure 1 demonstrates the chaotic attractor in the d1-d2 plane, whereas Figure 2 shows the chaotic attractor in the d1, d2, and d3 planes. The chaotic attractor in the y1-y2-y3 plane is presented in Figure 2. The time series analysis of the state attribute d1 utilizes Equation (1), as depicted in Figure 3. It is demonstrated that d1 achieves random behavior and is non-periodic in nature. System (1) contains 5 equilibrium points, such as (0, 0, 0, 0, 0), (2.

Partial-Regeneration-Based Non-Dominated Optimization
Partial-regeneration-based non-dominated optimization (PRNDO) [34] provides optimal solutions considering both the best and worst cases. This type of optimization is mainly used to solve the asymmetrical problems. In this study, a bottom-boosting methodology was utilized to reduce the number of function evaluations. It also provides a balance between better solutions and computational speed by developing a novel partialregeneration strategy and mutation operator. Therefore, it can generate efficient solutions with good computational speed. The outline of PRNDO is shown in Figure 4. It contains population initialization, minimum heap (min-heap) construction, bottom-boosting, partial-regeneration strategy, and termination condition. Initially, a population is randomly generated with the pair consisting of solution and scenario. The other parameters are also initialized, such as control attributes, solution and scenario dimensions, and maximum number of generations. Thereafter, by using the given population, a min-heap is built in the beginning of every generation. Every solution is treated as a node of min-heap, and the objective value of it acts as a key of the node. After that, scenarios of min-heap nodes are updated using the bottom-boosting scheme. The fitness function evaluations are also controlled by this scheme. Using fitness values of each individual, a sorted population is generated in increasing order. The trial scenario is produced from mutation and binomial recombination. The fitness value of the trial scenario is then calculated with respect to the original scenario. If the trial provides better fitness, then the original scenario of the individual will be upgraded. Otherwise, the old scenario will be used in the next generation. During the updating of scenarios, min-heap is also updated using the current population. Next, all the individuals are controlled and updated using the partial-regeneration strategy. The mutation and binomial recombination is also utilized. Finally, the termination condition (such as maximum number of generations evaluation) is tested, and optimal solutions are obtained as a result. IoT-based data analytics are presented in [34,35]. In [36] a comparative study is performed based on machine learning techniques.

Proposed Image Encryption Technique for Secure Surveillance Frameworks
An image encryption technique for secure surveillance frameworks for IoT systems is shown in Figure 5. The obtained images from surveillance cameras are initially encrypted by the proposed technique, then transmitted over the network. The obtained image is firstly decomposed into red (R), green (G), and blue (B) channels. The secret keys, such as d 1, d 2, d 3, d 4, and d 5, are obtained from 5DHCM (Algorithm 5). The required initial parameters of 5DHCM are optimized using PRNDO (Algorithm 2). All three channels are permuted using Arnold transform. Then, pixels of permuted channels R, G, and B are diffused utilizing d 1, d 2, and d 3, with the (⊕) operator, respectively. The obtained images are δR, G , and B . G and B are further diffused using ⊕ operation using δR, and produce δG and δB. Two more keys (d 6 and d 7) are generated. d 6 is generated from d 1 and d 2 using ⊕ operation, d 7 is generated from d 3 and d 4 using ⊕ operation, and d 4 is then utilized to update d 5 and d 6. Encrypted color channels, i.e., R, G, and B, are obtained by diffusing δR, δG, and B" with d 5 , d 6 , and d 7, and the η operator, respectively. Finally, the encrypted image (EM) is obtained by concatenating the R, G, and B.
The step-by-step approach is shown in Algorithm 1.

PRNDO-Based Optimal Parameters
The parameters of 5DHCM are d1, d2, d3, d4, d5, q, r, s, t, u, v, and w, and they need to be initialized. Manual selection of the above parameters is a computationally expensive task. For better key generation, these parameters are selected optimally by using PRNDO optimization. The procedure to generate the optimal parameters for 5DHCM using PRNDO is outlined in Algorithm 2.

Algorithm 2 PRNDO-based optimal parameters' generation
Input: Solution dimension Sol o , scenario dimension S o , maximum generation N h , and population size M l Output: Optimized population: Set generation u = 0; //population initialization P a = {(z 1,u , I 1,u ), (z 2,u , I 2,u ), . . . , (z M ,u , I M,u )}, where z k,u = {z 1 , S 2 , . . . , S Sol d } for y = 1 to M s do fitness of individual (z k,u , I k,u ) is evaluated using Algorithm 6; end for u = 1 to N h do Min heap is built for each individual in q u ; Use (Algorithm 3) to implement the bottom-boosting approach on min-heap; for k = 1 to M l do //Individual (z e , I e ) stored as a root node in the min-heap zk,u = ze, Ik,u = Ie; Extract root node from min-heap; end zbest = z1, u; Use partial-regeneration strategy for D u updating (Algorithm 4); D u+1 = D u ; end In Algorithm 2, firstly, the parameters' solution dimension Sol o , scenario dimension S o , maximum generation N h , and population size M l are initialized. Using normal distribution, initial population D u is generated randomly. Each individual (zk,u, Ik,u) represents a set consisting of solution and scenario. Fitness of (zk,u, Ik,u) is evaluated using Algorithm 6. A min-heap is built using D u . The scenarios are updated using Algorithm 3. The current population is updated by extracting the root node of min-heap. Then, the nodes are arranged in increasing order on the basis of fitness. The initial solution of a given population is utilized as a global best solution. Thereafter, a partial-regeneration strategy is utilized to further update the current population (Algorithm 4). The entire steps are iteratively implemented until the stopping criteria are not met.

Secret Keys
The process of secret keys' generation using 5DHCM is elaborated in Algorithm 5. It uses the optimized parameters obtained from Algorithm 2. The keys such as d1, d2, d3, d4, d5, d6, and d7 are generated using Algorithm 5. To diffuse the image, these keys are utilized by Algorithm 1 to obtain the encrypted image.

Fitness Evaluation
To compute the individuals, the following fitness function is utilized Correlation and entropy are widely accepted parameters to obtain the optimal solutions because in image encryption, there is a need to reduce the relationship between the pixels. In the case of maximum entropy, each pixel carries the same amount of information. Both parameters hide the statistical behavior of the images from the attacker. The desired value of correlation and entropy should be near 0 and 8, respectively. The fitness function is described in Algorithm 6.
Apply inverse Arnold transform on R, G, and B; //Concatenate the decrypted channels DI = [R, G, B]; return DI

Image Decryption Algorithm
The encrypted image is decrypted using Algorithm 7 to obtain the original image. The initial values of d1, d2, d3, d4, d5, q, r, s, t, u, v, w, and η attributes are needed to decrypt the image.

Comparative Analysis
Simulations were performed using MATLAB 2020a with the help of the image processing toolbox. The proposed encryption model is compared with five well-known competitive approaches by using various performance metrics. The initial parameters of PRNDO are assigned as F = 0.6, Cr = 0.2, N = 50, KS = 170, and T = 5. Four surveillance camera images were used, with 256 × 256 size.

Visual Analysis
Input and the respective encrypted surveillance images are shown in Figure 6. It clearly indicates that the obtained encrypted surveillance images seem like purely noisy images in nature. Therefore, no attacker can obtain any kind of details of the encrypted images.

Image Decryption Algorithm
The encrypted image is decrypted using Algorithm 7 to obtain the original image. The initial values of d1, d2, d3, d4, d5, q, r, s, t, u, v, w, and η attributes are needed to decrypt the image.

Comparative Analysis
Simulations were performed using MATLAB 2020a with the help of the image processing toolbox. The proposed encryption model is compared with five well-known competitive approaches by using various performance metrics. The initial parameters of PRNDO are assigned as F = 0.6, Cr = 0.2, N = 50, KS = 170, and T = 5. Four surveillance camera images were used, with 256 × 256 size.

Visual Analysis
Input and the respective encrypted surveillance images are shown in Figure 6. It clearly indicates that the obtained encrypted surveillance images seem like purely noisy images in nature. Therefore, no attacker can obtain any kind of details of the encrypted images.

Entropy
Entropy [29] computes the degree of randomness in images. In encrypted images is desirable to have 8 degrees of randomness. An entropy analysis of the proposed secu surveillance system is depicted in Table 1. It shows that the entropy values of all the tec niques on surveillance images (SIk), where k = 1, 2, 3, or 4, were near to 8, but the propos secure surveillance system achieved higher values than the existing techniques.

Entropy
Entropy [29] computes the degree of randomness in images. In encrypted images, it is desirable to have 8 degrees of randomness. An entropy analysis of the proposed secure surveillance system is depicted in Table 1. It shows that the entropy values of all the techniques on surveillance images (SIk), where k = 1, 2, 3, or 4, were near to 8, but the proposed secure surveillance system achieved higher values than the existing techniques.

Mean Absolute Error
Mean absolute error (MAE) analysis is shown in Table 3. It revealed that the proposed model obtained significantly better values than the existing models.

Security Analysis
In this section, we will test the proposed secure surveillance system model against various security attacks. The objective was to evaluate whether or not the proposed secure surveillance system model can resist various security threats.

Histogram Analysis
To evaluate the performance of the proposed secure surveillance system model against an attack, we used histogram analysis. It was found that the histograms of the encrypted images should be uniform. Histogram analysis of the proposed model is shown in Figure 7, and it was found that the histograms of encrypted images obtained from the proposed secure surveillance system model were evenly distributed.

Correlation Analysis
Correlation can also be used to evaluate the statistical information of surveillance images. Thus, it is desirable that the correlation between the adjacent pixels vertically, diagonally, and horizontally should be minimum. The horizontal correlation analysis is depicted in Table 4. It shows that the correlation among horizontal pixels was minimum compared to the existing encryption models. Vertical and diagonal correlation analyses are shown in Tables 5 and 6. The minimum correlation values of the proposed secure surveillance system indicate that the proposed model can resist correlation attacks.  [2] 0.0090 0.0060 0.0040 0.0022 DFRT [3] −0.0050 0.0123 0.0090 0.0040 HDNA-HC [5] 0.0040 0.0070 −0.0090 0.0036 CSCM [7] 0.0072 0.0038 −0.0050 0.0055 Proposed −0.009 0.0016 0.0019 0.0021

Differential Analysis
The differential analysis can be computed using the Number of Pixel Change Rates (NPCR) and Unified Average Change Intensity (UACI). The NPCR and UACI analyses of the proposed secure surveillance system are depicted in Tables 7 and 8. It was found that the proposed model achieved better NPCR and UACI values than the competitive models. Thus, the proposed model is sensitive towards the tiny modification in input surveillance images. Robustness of the proposed secure surveillance system was evaluated against salt and pepper and Gaussian noise attacks with various densities of noise.

Key Analysis
To evaluate the performance of the proposed secure surveillance system against brute-force attack, key sensitivity and key space analysis were considered.

•
Secret key space The proposed secure surveillance system computes the secret keys by 5DHCM. It requires 7 state attributes (i.e., d1, d2, d3, d4, d5, d6, and d7) and 7 control attributes (i.e., q, r, s, t, u, v, and w). Assume that the precision of attributes is 10−16. Thus, each attribute has 1016 possible values. Thus, the proposed secure surveillance system has key space: Since the key space > 1030 (≈ 2100), therefore, the proposed encryption model can resist a brute-force attack.

•
Secret key sensitivity analysis It is desirable for the secret keys to be more sensitive towards tiny changes. Seven keys were evaluated to implement the proposed encryption. These 7 keys were computed by Equation (1) with the help of initial state (i.e., u1, u2, u3, u4, u5, u6, and u7) and control (i.e., s, r, t, i, q1, p, n, e, g, and q2) attributes. Assume that a tiny change is made in the bits of u1. To verify the key sensitivity, we will encrypt the surveillance images using an actual key and with a tiny change in the same key. The encrypted surveillance images with actual and modified keys are shown in Figure 11b,c. Figure 11d shows the error between Figure 11b and c. The computed error seems to be a completely random image. We have also decrypted Figure 11b by utilizing both keys. The evaluated results are depicted in Figure 11e,f. If the identical key is used for decryption, then the obtained image is similar to the actual image. The decrypted image obtained using a tiny change in the actual key seems to be a completely noisy image. Thus, the proposed model is very sensitive towards the encryption key.  Table 9 shows the error between a modified and an actual key-based encrypted image. It demonstrates that the proposed secure surveillance system is sensitive towards initial key values.  Tables 10 and 11. These tables reveal that the proposed model can encrypt and decrypt images at a higher speed, as it requires less time than the existing models.

Conclusions
An image encryption technique for secure surveillance for IoT systems has been proposed in this paper. The image was made secure before the transmission from the surveillance system. A hyperchaotic map was utilized to obtain the pseudorandom sequences. The initial parameters of the hyperchaotic map were produced using PRNDO. The permutation was carried out through Arnold transform. Thereafter, a diffusion process was implemented to generate the encrypted images. The performance of the proposed framework was assessed by drawing the comparisons with competitive techniques based on security parameters. It was found that the proposed framework provides promising results as compared to the existing techniques. The proposed technique is able to resist against different attacks, such as noise and enhancement attacks. Experimental analyses showed that the proposed technique outperformed the competitive techniques in terms of visual analysis, quantitative analysis, security analysis, and execution time. As the proposed technique can encrypt and decrypt images at higher speeds, therefore, the proposed technique can be used for real-time secure surveillance systems.