An intelligent atrous convolution-based cascaded deep learning framework for enhanced privacy preservation performance in edge computing

Abstract

A system without any communication delays, called edge computing, has been introduced for nearer and faster services. The major concern in the edge computing scenario is its privacy risks. A user, as well as a cloud data preservation scheme, is the main aim of this paperwork. Test data is given by the user to access the cloud-based data processing framework. The training of the suitable model is carried out by utilizing the data stored in the cloud. The suggested model divides the entire model into two sections, namely, the untrusted cloud and the trusted edge. On the trusted edge side the data is directly provided to the developed advanced deep learning model called the Atrous Convolution based Cascaded Deep Temporal Convolution Network (ACC-DTCN) for the data analysis process. However, instead of giving the whole data directly to the untrusted cloud side, the test data is protected on the cloud side by utilizing a hybrid encryption technique called the Optimal Hybrid Encryption Model (OHEM). Both Attribute-Based Encryption (ABE) and Homomorphic Encryption (HE) are utilized in the recommended OHEM scheme. The OHEM variables are tuned with the help of an advanced algorithm called the Enhanced Ladybug Beetle Optimization algorithm (ELBOA). The confidence score vector among the testing and training data is predicted by the implemented ACC-DTCN model by utilizing the encrypted data on the cloud side. The suggested privacy preservation scheme provides higher prediction accuracy and prevents interference attacks while contrasting it against conventional methods during extensive experimentations.

Keywords

Internet of Things cloud computing edge computing privacy and security of data Atrous Convolution based Cascaded Deep Temporal Convolution Network Optimal Hybrid Encryption Model Enhanced Ladybug Beetle Optimization algorithm

1. Introduction

The enhancement and advancement in the field of communication technologies and artificial intelligence led to the development of several sophisticated and intelligent devices that are capable of producing enormous volumes of data that are utilized for training deep learning models, given as input in prediction models, and are used in data mining approaches. This mass production of data paves the way for the development of edge computing technology [21]. The enhanced interaction with the customers and the contravention of the centralized networks make edge computing technology provide enhanced artificial intelligence services with rigid “Quality of Services (QoS)” [23]. This ability makes the edge computing technique applicable in various fields like home automation systems, individualized recommendation models, enhancement models, traffic prediction systems, and so on [12]. But, all the host system in the edge computing platform, which produces enormous volumes of data ranging from scanners, web services, security cameras, laptops, workstations, phones, printers, etc., has a limited amount of computing resources and memory space, which restricts the data processing and data storage needed to provide the required, artificial intelligence-based services [17]. Many research works have been developed for incorporating an efficient artificial intelligence approach to edge computing networks [42].

Security issues, along with delays and inefficient computation, occur while transferring a huge volume of data from the cloud server over the “Wireless Area Networks (WAN)” [19]. Higher privacy and security are offered in the case of decentralized networks. However, these approaches are dependent on the host device’s performance, and it takes more time to come to in agreement with a model [43]. By associating the edge devices with the cloud server, the private data transmission to train the deep learning models, like federated learning, is replaced by the intermediate result transmission. This enhances the security of the data being transmitted in the edge computing network [26]. The process of task offloading to execute the processing of one sensor by various other sensors is one of the notable advantages of the edge computing prototype [36]. This eases the processing and computational burden on each sensor connected to the edge computing system in real-time applications. But, while offloading a task of one sensor to another sensor, both time consumption and privacy issues arise. These two factors need to be considered while offloading the services of one sensor to another in the edge computing network [22]. Privacy affects the overall confidentiality of the data being transmitted in edge computing, while time affects the user experience. Time-stringent industries are the ones mostly affected by the time-consuming edge computing platform on the process of task offloading. Customer dissatisfaction arises as the edge computing server takes more computation and processing time [35]. Security of the transmitted data is at risk because they are more vulnerable to data theft during the task offloading process [46]. The higher processing time requirement and the lower security of the data are the two main constraints that need to be solved in the edge computing paradigm.

The time consumption and privacy-related issues in edge computing platforms can be resolved by integrating the benefits of both decentralized and centralized approaches on the basis of the federated learning concept [8]. Even though the integration of these decentralized and centralized approaches is beneficial, some imperfections still exist. The first disadvantage of including federated learning in the edge computing platform is that the integrity of the data given as input to the learning models generated by the host devices is not verified by the cloud server as it has no access to the host device’s data [9]. When forged or altered data is given as input to these federated models, then inappropriate prediction outcomes would be provided by these approaches. The second drawback is that only training for a few general tasks is provided to these models to perform in an automated manner without user interference [24]. But, individualized services are required by the host devices based on user preferences and requirements which require continuous training of the learning model with historical data generated by a variety of users [7]. It is also worth considering the privacy issues in the training model and the deviations in the training data. A scalable and generalized prediction model is regarded as a crucial digital asset as it can be used for a variety of applications. However, such a system development requires more resources and is usually expensive to develop. It is also important to note the privacy issues in such models as the actual data can be derived in a reverse manner by providing some additional detail to these models [2]. Hence an enhanced privacy and security preservation scheme using an advanced deep learning model is developed in this research work dedicated to the edge computing paradigm.

The key contributions of this research work on providing an enhanced privacy preservation scheme in edge computing networks are listed below.

To generate an efficient privacy preservation prototype for supporting both the cloud and user side data in the edge computing paradigm using an enhanced deep learning approach.

To deploy a novel deep learning technique named ACC-DTCN for performing the task provided to them in both the trusted and untrusted side of the edge-cloud computing environment effectively with the utilization of the optimization technique.

To execute a hybrid cryptography approach called the OHEM, which is developed by fusing the HE and ABE schemes for enhancing the security of the data while being processed in the untrusted cloud side of the edge computing network to avoid accessing of the data by unauthorized persons.

To implement an improved optimization approach called the ELBOA by modifying the conventional LBO algorithm for the purpose of generating the optimal keys required for performing the OHEM scheme to enhance the privacy of the data on the cloud side.

To examine the privacy offered by the suggested privacy preservation model in edge computing by comparing it with conventional mechanisms in order to show the effectiveness of the developed scheme.

A brief outline of this paperwork is provided below. A detailed introduction to edge computing and the requirement of a privacy preservation scheme for edge computing platforms is given in Section 1. A detailed survey of some of the existing privacy preservation schemes dedicated especially to edge computing platforms is given in Section 2. Section 3 provides the formulation of the intelligent privacy preservation framework in edge computing. This section provides with the implementation of the enhanced privacy preservation model in edge computing using advanced deep learning with the meta-heuristic algorithm is elaborated in Section 4. The development of the atrous convolution-based cascaded deep learning network-based privacy preservation in the cloud sector is elucidated in Section 5. The results and discussion on the experimentation conducted on the suggested deep-learning-based data preservation scheme in edge computing systems are provided in Section 6. The summary of this work is given in Section 7.

2. Literature survey

2.1. Related works

In 2021, Qiang et al. [29] executed a “Visual Geometric Graph (VGG-19)”-based privacy preservation model for both cloud-side and user-side data. The test data was provided from the user side, while the data for training was provided from the cloud. The entire VGG-19 model was divided and provided to both the edge and cloud sides. The plaintext was directly given to the edge side VGG-19 model, whereas the encrypted data was given to the VGG-19 network on the cloud side. A binary framework with binarized weight values was utilized on the cloud side. The HE scheme was adopted to encrypt the input data given to the VGG-19 model on the cloud side. This narrowed the gap between the confidence scores on the testing and training data to make the prediction easier. The detection of inference attacks was successfully predicted by the suggested VGG-19 model but with slightly lower accuracy. Experimental outcomes demonstrated that the utilization of HE schemes even provided enhanced security against exploratory and inference attacks was much slower in execution than the encryption-less machine learning prediction models.

In 2022, Zheng et al. [47] introduced an “Unsupervised Recurrent Federated Learning (URFL)” model for providing privacy against attacks on the “Industrial Internet of Things (IIoT)” scenario. A Markov chain-based popularity detection technique was adopted to determine the dynamic popularity of every user. The Markova chain helped in attaining the relationship between the global as well as local popularity of the user. The determined user was then given to the URFL model to identify the user’s distributed popularity along with providing security to the system. The aggregation of the parameters was done with the aid of the “Federated Loss-Weighted Averaging (FedLWA)” function. The non-authenticated user’s behavior was analyzed using this FedLWA approach. Experimental verifications showed that the “Root Mean Squared Error (RMSE)” of the suggested privacy-preservation approach was reduced by 61% to 69% than other distributed and centralized schemes. The implemented privacy-preservation scheme also provided privacy to the user data along with manual labeling of the user’s data to enhance security.

In 2022, He et al. [14] introduced a federated learning-based privacy-preservation scheme aided by a HEscheme to offer security to the end devices with low delays without actual data transmission between the edge nodes. The implemented scheme was known as PL-FedIPEC. A new encryption scheme called the Paillier method was incorporated into the developed privacy-preservation scheme. The generated model minimized the time required for data encryption. The encrypted ciphertext cannot be evaluated by unauthorized individuals in this encryption scheme. Experimental results showed that the security offered by the suggested privacy-preservation scheme was higher as this model was capable of generating accurate detection outcomes similar to that of baseline approaches with reduced computation and training time requirements.

In 2020, Qu et al. [30] introduced a federated learning model with an incentive mechanism for achieving an intelligent cloud-edge model. Then the HE scheme was introduced for providing security to the data being stored in the edge server. The feasibility, as well as the effectiveness of the promoted models, was proved by conducting extensive simulations by utilizing both real-time as well as synthetic data.

In 2022, Wang et al. [37] have deployed a model known as the PrivStream to prevent the privacy of the data streaming in the IoT platform. The proposed model prevented the linkage of the untrusted edge server with the IoT data. The filtration of the sensitive information was carried out using a deep learning model. The robustness of the developed privacy preservation scheme was validated by injecting noises into the system’s training stages. Simulation outcomes proved that the storage and computational overhead of this approach was in an acceptable range.

In 2023, Telikani et al. [34] implemented a “High-Performance Evolutionary Data Sanitization scheme for IoT (HEDS4IoT)” on two parallel “Graphical Processing Units (GUP)” in a real-time scenario. Privacy in transmitting streaming data was provided by this technique. The suggested scheme performed two parallel operations as a “Parallel Fitness Function Engine (PF2E)” and “Parallel Indexing Engine (PIE)”. The computation process was fastened by an evolutionary algorithm. The recommended HEDS4IoT was proven to be suited for both dynamic as well as big data environments.

In 2019, Xu et al. [45] utilized a “Strength Pareto Evolutionary Algorithm (SPEA2)”-based privacy-preservation scheme for providing security to the sensors in the edge computing system in order to generate a “Time-efficient Offloading (TEO)” scheme. Both the secret data’s offloading and the time taken to protect these data were evaluated in this approach by the formalized method. Both the joint optimization of the average privacy entropy and the consumed time were achieved by the SPEA2 scheme. The robustness and efficacy of the executed privacy-preservation scheme were verified by carrying out extensive experiments.

In 2020, Xu et al. [44] suggested a “Balanced Service Offloading Method (BSOM)” to offer security and privacy against the data stored in edge nodes. The resources utilized by the network were optimized by this method in an effective way, along with balancing the load in the entire network. Experimental analysis was carried out to prove the enhanced privacy offered by the suggested scheme along with providing a flexible operation with minimized time consumption.

In 2020, Zheng et al. [50] developed a spectrum interference-based two-level data augmentation approach in a deep structured approach for automated modulation categorization. The actual signal was initially converted into the frequency domain with the aid of a short-time Fourier transform. Inverse Fourier transform was introduced according to the interfered spectrum for reconstructing the augmented signals and then, the actual and augmented signals were subjected to the network. The data augmentation at both the testing and training phases can be utilized for enhancing the generalization performance of a deep structured approach. The simulation findings of the developed model have illustrated the model’s effectiveness and advancement.

In 2020, Zheng et al. [52] implemented the Manifold Regularization-aided Deep Convolutional Autoencoder (MR-DCAE) approach for unauthorized broadcasting recognition. In addition, an Autoencoder (AE) was tuned by entropy-stochastic gradient descent, and then the renovation of errors in the testing stage can be adopted to determine whether the received signals are authorized [51]. The simulation outcomes revealed that the designed model has attained elevated performance.

In 2023, Zheng et al. [49] promoted a Priori Regularization Method in Deep Learning (DL-PR) for guiding loss optimization. The main aim of regularizing deep structured approaches was to enhance the accuracy of the Automatic Modulation Classification (AMC) according to the distribution of samples [48]. Furthermore, it can be revealed that the priori regularization can be interpreted as implicit data augmentation. The experimental outcome of the designed model proved the high robustness and the interpretability of the model.

2.2. Problem statement

Table 1
Features and challenges of existing privacy preservation models for edge computing.

Author [citation] Methodology Features Challenges

Qiang et al. [29] VGG19
It provides successful performance over exploratory attacks.

It also provides faster encryption performance.

It shows a slight deviation in the accuracy performance.

Zheng et al. [47] URFL
It assures improved accuracy by minimizing the RMSE rate.

It is independent of privacy violations of the user and manual labeling.

It does not solve the data deficiency problem.

He et al. [14] PL-FedIPEC
It requires minimal time to train the model under different environmental settings.

It shows a slight deviation in performance when analyzing sensitive information.

Qu et al. [30] Federated learning
It assures model enhancement and end device protection even with the huge amount of edge server computation.

It contains higher computational costs because of the network structure for determining the activation functions.

Wang et al. [37] PrivStream
It gives better inference accuracy for the targeted tasks.

It attains enhanced efficiency on practical transmission along with high-speed processing while using streaming data.

It does not assure effective performance in terms of energy consumption.

Telikani et al. [34] HEDS4IoT and PF2E
It provides faster performance when compared to the conventional methods.

It shows lower performance in terms of convergence and resource utilization.

Xu et al. [45] SPEA2
It reduces the average time consumption and provides high robustness through this method.

It is not applicable in real-time applications.

It shows a slight deviation in the analysis of energy consumption.

Xu et al. [44] BSOM
It enhances the performance regarding resource utilization and load balancing.

It is not suitable for practical applications in order to support other specific fields.

Author [citation]	Methodology	Features	Challenges
Qiang et al. [29]	VGG19	It provides successful performance over exploratory attacks. It also provides faster encryption performance.	It shows a slight deviation in the accuracy performance.
Zheng et al. [47]	URFL	It assures improved accuracy by minimizing the RMSE rate. It is independent of privacy violations of the user and manual labeling.	It does not solve the data deficiency problem.
He et al. [14]	PL-FedIPEC	It requires minimal time to train the model under different environmental settings.	It shows a slight deviation in performance when analyzing sensitive information.
Qu et al. [30]	Federated learning	It assures model enhancement and end device protection even with the huge amount of edge server computation.	It contains higher computational costs because of the network structure for determining the activation functions.
Wang et al. [37]	PrivStream	It gives better inference accuracy for the targeted tasks. It attains enhanced efficiency on practical transmission along with high-speed processing while using streaming data.	It does not assure effective performance in terms of energy consumption.
Telikani et al. [34]	HEDS4IoT and PF2E	It provides faster performance when compared to the conventional methods.	It shows lower performance in terms of convergence and resource utilization.
Xu et al. [45]	SPEA2	It reduces the average time consumption and provides high robustness through this method.	It is not applicable in real-time applications. It shows a slight deviation in the analysis of energy consumption.
Xu et al. [44]	BSOM	It enhances the performance regarding resource utilization and load balancing.	It is not suitable for practical applications in order to support other specific fields.

Edge computing is performed together with AI, which enhances the accurate processing and evaluation of streaming data generated with IoT applications. On the other hand, it also leads to privacy risks owing to the data transmission among the untrusted edge servers and local devices. Therefore, various privacy preservation models are developed in the existing models for edge computing, and they are summarized in Table 1. VGG19 [29] provides successful performance over exploratory attacks. It also provides faster encryption performance. However, it shows a slight deviation in the accuracy performance. URFL [47] assures improved accuracy by minimizing the RMSE rate. It is independent of privacy violations of the user and manual labeling. But, it does not solve the data deficiency problem. PL-FedIPEC [14] requires minimal time for training the model under different environmental settings. On the other hand, it shows a slight deviation in performance when analyzing sensitive information. Federated learning [30] assures model enhancement and end device protection even with the huge amount of edge server computation. Yet, it contains higher computational costs because of the network structure for determining the activation functions. PrivStream [37] gives better inference accuracy for the targeted tasks. It attains enhanced efficiency on practical transmission along with high-speed processing while using streaming data. Still, it does not assure effective performance in terms of energy consumption. HEDS4IoT and PF2E [34] provide faster performance when compared to conventional methods. However, it shows lower performance in terms of convergence and resource utilization. SPEA2 [45] reduces the average time consumption and provides high robustness through this method. But, it is not applicable in real-time applications. Also, it shows a slight deviation in the analysis of energy consumption. BSOM [44] enhances the performance regarding resource utilization and load balancing. However, it is not suitable for practical applications in order to support other specific fields. On considering these conventional challenges in existing approaches, it is necessary to implement privacy preservation techniques for edge computing.

2.3. Research gaps

The existing challenges of the conventional privacy preservation model for edge computing are considered in this research work, and the research problems are summarized as the key points below.

Mostly, edge computing causes privacy leakage for the user data, and it is required to develop an efficient privacy preservation model against inference attacks.

The most important factor, like accuracy loss, affects the performance of the conventional privacy preservation models and, therefore, it is required efficient deep learning techniques for enhancing the accuracy rate.

In addition, the challenge regarding the computational overhead is required to be solved by splitting the data of the edge and cloud side that are given into the deep learning training.

The existing HE scheme faces high computational burdens and computational costs, and therefore, it is required to develop hybrid encryption techniques to solve the conventional challenges.

When considering the transmission time, it relies on the transmission delay and propagation delay, and so a concrete amount of data is necessary for giving into the encryption and transmission phase of the model.

3. Intelligent privacy preservation framework in edge computing

3.1. Developed privacy preservation model in edge computing

A privacy preservation scheme for both the user and the cloud side is recommended in this work to prevent security against the data stored in cloud-based platforms. The user defines the test data to train the suggested deep learning-oriented privacy-preservation model in the cloud environment. Rather than providing the entire test data directly into the classification framework, the suggested privacy preservation system is divided into two main segments. The first segment is the trusted edge computing part with users, and the second segment is the untrusted cloud part with unknown users [4]. Privacy is provided by feeding the untrusted cloud with encrypted data and the trusted edge with plain text data. The segmentation of the recommended model into two units, the overall efficiency and the secrecy provided by the executed deep learning-based privacy preservation approach, is enhanced. At first, the user-defined test data is taken as input on the edge side. These test data are gathered using edge computing devices. Privacy of the gathered data on the edge side is provided by the developed deep learning model [13]. The gathered test data (plaintext) is given directly to the implemented ACC-DTCN model. From the ACC-DTCN framework, the output data is generated on the edge side. To enhance privacy in the untrusted cloud environment, the test data are primarily encrypted using a hybrid encryption scheme known as the OHEM on the cloud side. This OHEM scheme encrypts the test data. The test data is initially given to the HE technique [32]. The keys required by the HE scheme for encrypting the data are optimally selected by the suggested ELBOA. The encrypted data obtained from the HE scheme is then given to the ABE scheme for further encryption of the data. The keys required by the ABE scheme for encrypting the data are also optimally generated by the implemented ELBOA. The optimal generation of the encryption keys helps in minimizing the memory and time requirement of the encryption process. The final encrypted data is generated from the OHEM scheme on the untrusted cloud side [1]. This encrypted data is given to the to the developed ACC-DTCN model on the cloud side. The confidence score vector among the testing and training data is predicted by the implemented ACC-DTCN model by utilizing the encrypted data on the cloud side. While simulating the deployed privacy preservation scheme against existing methods, higher prediction accuracy is provided by the recommended model, and enhanced preservation against interference attacks is given by the executed privacy-preservation scheme. In addition, the strategies employed to develop an efficient privacy preservation model against inference attacks are discussed as follows. Generally, security strategies are needed to mitigate these inference attacks. An advanced encryption of model inputs, secure authentication of all parties, execution integrity, and model confidentiality are significant strategies for promoting an efficient privacy preservation model against inference attacks. While taking privacy leakage for user data, the proposed data privacy-preservation model in the edge-cloud environment model gains higher bandwidth with the help of machine learning and deep learning models that help to transmit more data between end devices and edge servers. It allows more private information while transmitting the data. Edge computing faces challenges with computational overhead like bandwidth and latency limitations. The propagation delay caused by data transmission is denoted as latency and bandwidth challenges can impact the performance of the developed model. Here, the given developed model gains higher bandwidth with the help of deep learning models which helps to resolve the computational overhead effectively. The architecture of the recommended data privacy-preservation model in an edge-cloud environment is illustrated in Figure 1.

Fig. 1.

Architecture of the recommended data privacy-preservation model in edge-cloud environment.

3.2. Main components of the proposed system

The three major components in the suggested scheme are the cloud server, edge devices, and the edge server. The cameras, sensors, and other devices in the IoT system with lower computational ability serve as the edge devices. The processing of the devices collected from the edge devices is done by the cloud server, which is installed near the edge devices. Both securities, as well as efficient computation, are offered by the edge server. The pre-processing of the user data is done by the edge server. On the basis of its own dataset, the suggested model is trained using the cloud platform. The ACC-DTCN is used by the cloud to train the qualified model for preventing data privacy. The plaintext is processed locally, and then encryption takes place in order to prevent the privacy of the user data. On the basis of the encryption results, the remaining processing is completed by the cloud. The decryption of the data is not possible without the appropriate private key. Thus, enhanced privacy is provided by this model. The generated model is generic and suitable for performing any task.

4. Enhanced privacy preservation model in edge computing using advanced deep learning with a meta-heuristic algorithm

4.1. Deep learning on encrypted data

As the computation of the encrypted data is possible in the case of the HE and ABE schemes, it is an ideal choice for the “privacy-preserving deep learning” technique. The random processing of the ciphertext is carried out using “Fully Homomorphic Encryption (FHE)”. But, in practical application, this FHE scheme results in a larger cost. Hence, “Leveled Homomorphic Encryption (LHE)” like the SEAL is adopted. The SEAL is capable of performing both summation and product operation on a “depth-bounded arithmetic device”. However, it is also noted that only the processing of the polynomial operations is supported by the SEAL. The processing of the non-linear operations is carried out using linear operations. The implementation of deep learning techniques is usually carried out in the Python paradigm. So, the processing of these ciphertexts is complicated without any codes for processing. This issue is tackled by Intel’s graph compile, nGraph-HE scheme, which allows the processing of ciphertexts executed in any platform. Thus, the deep learning model can be trained into a suitable “privacy-preserving system”.

4.2. Use of edge computing

The resource and computing to the cloud from the source of data is described by edge computing. Edge computing is utilized because it prevents issues such as delay and data leakage while transmitting the data by the cloud server itself. The edges are implemented closer to the data sources. The computation of the data generated by the edge devices and sending the computed data to the cloud server is performed by the edge server. Thus the delays in data transmission are reduced, and at the same time, the privacy of the data is improved. The processing of the raw data securely and transmitting it without any data leakage to the cloud server is achieved by the edge server. The LHE is utilized in this work as the edge server.

4.3. Ladybug beetle optimization algorithm

The LBO [31] algorithm is designed by considering the migration patterns of the swarm of ladybugs. As a social insect, there is always an interaction with every ladybug in a swarm. It is also noted that the ladybug coordinates with one another in the swarm to spot a warmer place in winter. The swarm intelligence of the ladybug is utilized for determining the safe and warmer place in winter. While following the path towards the warmer place, some ladybugs may lose track and die out of cold weather. The number of ladybugs involved in the search process is decreasing. The population of the ladybugs in the swarm is determined. On the basis of the fitness value, the population of ladybugs is sorted. On every cycle, the number of ladybugs in the swarm is determined, and their fitness values are re-evaluated. This process is simulated to determine the number of ladybugs that sustain the cold weather and move toward the warmer place. Once the required cycles are executed, the optimal ladybug is then determined. The sole inspiration behind LBO is its movement toward a warmer location to escape from the cold weather. The initial ladybug swarm population has $A (0)$ a number of ladybugs. This ladybug population gets updated in every cycle to obtain the final ladybug population having $A (a_{mx})$ number of ladybugs. It should be noted that the initial population is always greater than the final population of the ladybug, as most of the ladybug lose track of the warmer location and dies out of cold weather. Therefore, $A (a_{mx}) ⩽ A (0)$ . Once the final population is determined, the best fitness value is then evaluated. The fitness value indicates the warmth of every location spotted by the ladybug swarm. The fitness value equals the heat value of the location determined by every ladybug. The location with the warmest temperature is taken as the best fitness value. But, in order to make the LBO algorithm a standardized approach, it is considered a minimization issue. Hence, every ladybug’s location is inversely proportional to its fitness value. Therefore, the fitness value is minimum for the warmest location. Three main steps are involved while modeling the LBO algorithm.

On the basis of the uniform distribution function, the initial ladybug population is distributed arbitrarily throughout the lookup area. The pre-defined fitness value is evaluated to determine the initial ladybug population, and it is saved in the memory. With the utilization of a coordinated swarm movement, the entire swarm moves toward the warmer location. Even in search of a warmer location, the entire ladybug population moves as a swarm because of its social nature. By following the signals emitted from the group member, every member in the swarm follows one another. So, every ladybug tends to follow the lady prior to it. The ladybug in the frontline of the swarm is more capable of determining the best location with warmer temperatures than the other ladybugs in the swarm. The balancing of the exploration and exploitation phases is achieved by incorporating a mutation process arbitrarily to some ladybugs in every cycle of the LBO algorithm. Therefore, the location of the ladybug in every cycle in the lookup area is determined either by the mutation process or by the location of the other ladybugs in the swarm. Every ladybug in the swarm updates its location in every cycle of the LBO algorithm. By combining the old as well as the new location of the ladybug, the optimal swarm is selected on the basis of its fitness value. In the upcoming cycle, the location of the new ladybug population is determined. A ladybug in one swarm is amended with respect to another ladybug in the same swarm. This process is explained as follows. Let us assume the ladybug that needs to be amended as c. For the purpose of amending this $c^{th}$ ladybug, the ladybug d of the same swarm is considered. The location of the selected $c^{th}$ ladybug belonging to the $a^{th}$ cycle is denoted as $b (a)$ . With the aid of three vectors, the locomotion of the $c^{th}$ ladybug in the upcoming cycle $(a + 1)$ is evaluated. The movement of the $c^{th}$ ladybug is such that it moves a little towards the $d^{th}$ ladybug and a little towards the $(d - 1)^{th}$ ladybug in the direction of d. Apart from these two locations, a third location is also considered to avoid local optima and to maintain a balance between exploration and exploitation phases. The present position’s coefficient is regarded as the third parameter. The locomotion in the above-mentioned three directions is multiplied with an arbitrary variable in order to maintain searching in arbitrary directions. In addition, the third locomotion has to be multiplied by the ratio of the ladybug’s individual heat to the sum of the heat of the entire ladybug swarm. This aids the ladybug to avoid local optima. The new location of the $c^{th}$ ladybug is mathematically represented by Eq. (1).

\begin{aligned} b_{c} (a + 1) & = b_{c} (a) + e \times [b_{d} (a) - b_{d} (a)] + e \times [b_{d} (a) - b_{d - 1} (a)] \\ + e \times | B_{c} |^{- (\frac{a}{A (a)})} \times b_{c} (a) \end{aligned}

(1)

The term $B_{c}$ in Eq. (1) is the parameter whose value is equal to the ratio of the heat of the individual ladybug to the overall heat of the entire population. This value at the $a^{th}$ cycle is computed using Eq. (2).

\begin{aligned} B_{c} = \frac{g (b_{c} (a))}{\sum_{p = 1}^{A^{a}} g (b_{p} (a))} \end{aligned}

(2)

The roulette wheel selection process is adopted in the conventional LBO algorithm to update the location of the ladybird. However, this step increases the computational complexity in the LBOA. Hence, it is removed in the suggested ELBOA.

The population of the ladybug gets updated by considering the mutation process, which is regarded as the crucial process in the LBO algorithm to avoid local minima and to explore more areas in the lookup region. Also, the convergence rate of the LBO algorithm is enhanced with the inclusion of this mutation process. Therefore, the amending of a ladybug with respect to the mutation procedure and with respect to other ladybugs is evaluated arbitrarily. Let us assume the ladybug that has to be mutated is at the $c^{th}$ position in the swarm. The decision variable that has to be mutated is evaluated on the basis of Eq. (3).

\begin{aligned} h_{l} = o (h * κ_{l}) \end{aligned}

(3)

In Eq. (3), the term h denotes the decision variable’s length, the term $κ_{l}$ indicates the rate of mutation, and the term o represents the rounding function. Thus, arbitrary selection of the $h_{l}^{th}$ number of h decision variables belonging to the $c^{th}$ ladybug takes place. The chosen location of the $c^{th}$ ladybug replaces the arbitrary variable in the lookup area. While searching for a warmer location, some ladybugs may lose track and get disappear. When the track of the ladybug gets lost, it annihilates in cold weather. The annihilation of the ladybug is formulated with the aid of Eq. (4).

\begin{aligned} A (a + 1) = o (A (a) - e \times A (a) (\frac{D}{D_{mx}})) \end{aligned}

(4)

In Eq. (4), the term $D_{mx}$ denotes the maximum Number of Function Evaluations (NFE), and the term D indicates the NFE. The above equation is regarded as the stopping criteria in the LBO algorithm. The new population of ladybugs when the cycle count is taken as the stopping condition is given by Eq. (5).

\begin{aligned} A (a + 1) = o (A (a) - e \times A (a) (\frac{a}{a_{mx}})) \end{aligned}

(5)

In Eq. (5), the term $a_{mx}$ represents the maximum number of cycles in the LBO algorithm. The pseudocode of the ELBOA is given in Algorithm 1.

The flowchart of the generated ELBOA is given in Figure 2.

Fig. 2.

Flowchart of generated ELBOA.

5. Atrous convolution-based cascaded deep learning network-based privacy preservation in the cloud sector

5.1. Proposed atrous convolution-based cascaded deep temporal convolution network

The raw data collected from various online sources ${RD}_{rd}^{gat}$ are provided to the designed ACC-DTCN framework for data protection. A neural network that convolves in the temporal domain is known as the DTCN [11]. It is a varied version of the “Convolutional Neural Network (CNN)” with fewer memory needs and a more receptive field designed especially for solving time series issues. The updation of every time step in DTCN takes place parallelly. DTCN is a type of convolutional network. Since no access to the future is provided to the DTCN’s filters, dilated convolutions are included in the DTCN structure. By skipping a particular time step’s input, the convolution filter is provided for a vast receptive field rather than the size of the input mentioned. The receptive field is increased exponentially by stacking these dilated convolutions above one another. The two major units in the DTCN are the residual blocks and the “Dilated Casual Convolutions (DCC)”. The crucial section of the DTCN is its DCC. When an input $F = [f_{1}, f_{2}, . ., f_{J}]$ is considered, the DCC of the DTCN will generate output $i_{j}$ at the present time j by utilizing both the prior and the current inputs alone. No correlation is there between the output and the future inputs. The input and output lengths are maintained to be similar with the utilization of zero padding. By padding towards the left in an asymmetric manner, the causality is incorporated into the DTCN. The long-term dependency of the sequential data is solved by providing larger receptive fields. The receptive field’s size is increased by arranging the filters of larger sizes or convolutional layers of larger sizes above one another. But these two approaches seem to be computationally inefficient. This issue is tackled by including dilation into these filters. By including gaps between the inputs, the filter sizes in DTCN are expanded by the DCC. The prior inputs up to $m (u - 1)$ past steps can be determined by the utilization of u sized filter with m dilation factor. DCC are stacked to enhance computational efficiency and to attain an enhanced receptive field. Initially m is considered to be one which keeps on increasing exponentially as it succeeds to the next layers. This ensures that all the inputs come within the receptive field. The depth of the DTCN is proportional to the receptive field’s size. Therefore, the number of layers required by the DCC of the DTCN is less when compared with conventional convolutions. The receptive field’s size is evaluated using Eq. (6).

\begin{aligned} C = 2^{E} \times (u - 1) \end{aligned}

(6)

In Eq. (6), the term E indicates DCC’s depth. Because of the disappearing gradient issues, deep neural networks become hard to train. This issue is resolved by including residual blocks in the deep structures. The stability of the network while training is enhanced by incorporating skip connections. The addition of the skip connection on the residual units prevents the gradient’s norm. The residual block’s input is the output from the final DCC layer. The output from one residual unit is provided as input to the next residual unit. To maintain the size of the output similar to that of the input $1 \times 1$ convolutions are added in the skip connections. Two sub-units arranged over one another are present in every residual unit. A dropout layer, DCC layer, and “Rectified Linear Unit (ReLU)” unit are all present in a single sub-unit of the DTCN’s residual blocks. Along with it, batch and weight normalizations are also included next to the DCC layer. The final receptive field of TCN has a size that is determined in Eq. (7).

\begin{aligned} C = 1 + Q_{r} \times Q_{s} \times (u - 1) \times \sum_{q = 1}^{E} m_{q} \end{aligned}

(7)

The term $m_{q}$ in Eq. (7) indicates the $q^{th}$ layer’s dilation factor, the term $Q_{s}$ denotes the total count of DCC in the residual block’s sub-unit, and the term $Q_{r}$ represents the total count of residual units in the DTCN. The receptive field of the kernels of the DCC layers in the DTCN structure is enhanced by incorporating atrous [25] convolution into it. There are no changes made to the parameters in the DTCN. The information carried out by each DCC layer is enhanced by including this atrous convolution into it. The atrous convolution is computed using Eq. (8).

\begin{aligned} k (f, i) = η (\sum_{t, w} n (f + t * v, y + w * v) * x (t, w) + y) \end{aligned}

(8)

In Eq. (8), the term v indicates the degree of expansion of the convolutional layers. In between the successive filters, $z - 1$ zeros are added when an atrous rate of z is considered. The filter size therefore gets enlarged from $G \times G$ to $G + (G - 1) (z - 1)$ without complicating the computation process or without the addition of any additional parameters. Cascading of several atrous DTCN takes place to develop the ACC-DTCN model. After being processed by the final atrous DTCN unit of the ACC-DTCN model, task output is generated from the implemented ACC-DTCN framework. The number of kernel size is 2 and the number of filters is 64. The “relu” activation function is used here. The diagrammatic illustration of ACC-DTCN is shown in Figure 3.

Fig. 3.

Diagrammatic illustration of the recommended ACC-DTCN-based classification model for data analysis.

5.2. Data encryption using optimal hybrid encryption model

The encryption of the test data ${RD}_{rd}^{gat}$ is carried out with the aid of a hybrid encryption scheme called the OHEM, in which both HE and ABE cryptographies are utilized.

HE: The encryption scheme in which the data that needs to be protected is converted into a ciphertext, which can be evaluated and processed without revealing its original format, is known as HE [10]. Without compromising the encryption, HE allows complicated mathematical functions to be carried out on the ciphertext. It should be noted that there must exist a correlation between the ciphertext and the plaintext in order to implement mathematical functions in the case of a HE scheme. The major challenges for homomorphic encryption are computationally intensive and slow. The operations like encrypting, performing operations on ciphertexts, and decrypting need more resources and it takes more time. The data in its original format is denoted as plaintext, whereas the data being encrypted by the encryption algorithm is termed ciphertext. Multiple summation and multiplication operations are carried out in a HE scheme. On the basis of the mathematical operation implemented in the homomorphic scheme, it is categorized as “multiplicatively homomorphic, as well as additively homomorphic”. It is further classified into “somewhat homomorphic, fully homomorphic, and partially homomorphic” encryption schemes. A brief description of these types of HE schemes is provided as follows.

Additively Homomorphic: If the results obtained on summing two ciphertexts are equal to the results on encrypting the integration of two the plaintext, then the homomorphic scheme is said to be an additively HE scheme.

Multiplicatively Homomorphic: If the results obtained on multiplying two ciphertexts are equal to the secret key raised to the power of product between two plaintexts, then the homomorphic scheme is said to be a multiplicative HE scheme.

Somewhat Homomorphic: In this scheme, only a finite number of either of the one mathematical operations (multiplication or addition) is carried out. This scheme is complicated to design, unlike the other schemes with infinite operations.

Fully Homomorphic: In this scheme, infinite multiplication and addition take place on the ciphertext for encryption.

Partially Homomorphic: This scheme is the easiest of the entire types of homomorphic schemes as it performs an infinite number of defined operations (either multiplication or addition) on the ciphertext.

An encryption technique is regarded as homomorphic when the provided security key H satisfies the function given in below Eq. (9).

\begin{aligned} N (L_{1}) Θ_{γ} N (L_{2}) \to N (L_{1} Θ_{ϖ} L_{2}); \forall L_{1}, L_{2} \in ϖ \end{aligned}

(9)

In Eq. (9), the term Θ indicates the mathematical function. The symbol Θ is replaced by ⊕ in the case of additively HE as $N (L_{1}) \oplus_{γ} N (L_{2}) \to N (L_{1} \oplus_{ϖ} L_{2})$ , and the symbol Θ is replaced by ⊗ in the case of multiplicatively HE as $N (L_{1}) \otimes_{γ} N (L_{2}) \to N (L_{1} \otimes_{ϖ} L_{2})$ . These two operators get executed at the same time in case of ring homomorphism while satisfying the “ring axioms”.

ABE: The ABE [16] is a generalized version of the public key encryption protocol. With the execution of an authentication process, fine-grained access control over the data that is being encrypted is obtained in the case of ABE. Both the encrypted ciphertexts and the user’s secret keys are interrelated with the user attributes. Only when the ciphertext’s attributes when matched with the secret key attributes enables the decryption of the ciphertexts. The important security concern in the ABE is collision resistance property. Two ABE schemes exist. One is the Ciphertext-Policy (CP)-ABE, and the other is the Key Policy (KP)-AE. The main difference between the CP-ABE and the KP-ABE is that in CP-ABE, the data is encrypted using the access tree, while the secret keys for the user are obtained by utilizing a certain set of attributes; whereas in KP-ABE, the data is decrypted by utilizing a certain set of attributes and the secret keys for the user is obtained using access trees. Four major “polynomial-time” functions, such as the Setup, Key Generation, Encryption, and Decryption, are performed in both the CP-ABE and KP-ABE. These four mechanisms are provided as follows for both types of ABEs.

KP-ABE :

Setup Phase: A group of decryption policies ρ supported by the ABE scheme, a security variable ϕ, and attributes $ψ = {K_{1}, K_{2}, . ., K_{M}}$ are provided as input to the setup algorithm. The setup algorithm for KP-ABE is represented as $S e t U p (ρ, 1^{ϕ}, ψ)$ . Both the master key O and certain public variables P are provided as output by this step.

Key Generation Phase: Along with O and P obtained from the setup phase, a decryption policy $(δ, ν); δ \subset ψ & ν \subset 2^{δ}$ such that $ν \in ρ$ is given as input in this step to generate the private key $R_{(δ, ν)}$ . The key generation algorithm for KP-ABE is denoted as $K e y G e n (O, P, δ, ν)$ .

Encryption: The message S, a group of attributes $T \subset ψ$ , and P is inputted into the encryption algorithm as $E n c r y p t (S, T, P)$ . The encrypted ciphertext U is obtained as the output.

Decryption: The secret key $R_{(δ, ν)}$ , the group of attributes $T \subset ψ$ , the public variable P, and the encrypted data U are taken as input by the decryption algorithm as $D e c r y p t (R_{(δ, ν)}, T, P, U)$ . The original plaintext or the provided message $\tilde{S}$ is obtained as output.

The overall processes in the KP-ABE scheme are given below: i)

$S e t U p (ρ, 1^{ϕ}, ψ) \to O, P$

ii)

$K e y G e n (O, P, δ, ν) \to R_{(δ, ν)}$

iii)

$E n c r y p t (S, T, P) \to U$

iv)

$D e c r y p t (R_{(δ, ν)}, T, P, U) \to \tilde{S}$

CP-ABE :

Setup Phase: A group of decryption policies ρ supported by the ABE scheme, a security variable ϕ, and attributes $ψ = {K_{1}, K_{2}, . ., K_{M}}$ are provided as input to the setup algorithm. The setup algorithm for CP-ABE is represented as $S e t U p (ρ, 1^{ϕ}, ψ)$ . Both the master key O and certain public variables P are provided as output by this step. The setup phase is similar to that of the CP-ABE scheme’s setup phase.

Key Generation Phase: Along with O and P obtained from the setup phase, a group of attributes $T \subset ψ$ is given as input in this step to generate the private key $R_{T}$ . The key generation algorithm for CP-ABE is denoted as $K e y G e n (O, P, T)$ .

Encryption: The message S, the decryption policy $(δ, ν); δ \subset ψ & ν \subset 2^{δ}$ such that $ν \in ρ$ and P is inputted into the encryption algorithm as $E n c r y p t (S, δ, ν, P)$ . The encrypted ciphertext U is obtained as the output.

Decryption: The secret key $R_{T}$ , the decryption policy $(δ, ν)$ , the public variable P and the encrypted data U are taken as input by the decryption algorithm as $D e c r y p t (R_{T}, δ, ν, P, U)$ . The original plaintext or the provided message $\tilde{S}$ is obtained as output.

The overall processes in the CP-ABE scheme are given below: i)

$S e t U p (ρ, 1^{ϕ}, ψ) \to O, P$

ii)

$K e y G e n (O, P, T) \to R_{T}$

iii)

$E n c r y p t (S, δ, ν, P) \to U$

iv)

$D e c r y p t (R_{T}, δ, ν, P, U) \to \tilde{S}$

Initially, the test data generated ${RD}_{rd}^{gat}$ is given as input to the HE cryptography in the OHEM. The half-encrypted data is generated from the HE scheme. This half-encrypted data is further provided to the ABE scheme. The final completely encrypted data is generated from the ABE scheme. The keys required for encrypting the protected data by the OHEM are optimally selected with the utilization of the ELBOA. The key in binary format, i.e., $[0 ~or~ 1]$ is selected optimally by the ELBOA for both HE and ABE schemes in the OHEM cryptography scheme. The optimal selection of the cryptography keys aids in the minimization of memory $ms$ and time $tm$ requirements for the encryption process. The process of attaining this goal is mathematically represented by Eq. (10).

\begin{aligned} oe = \arg min_{{{OK}_{ok}^{HE} {, OK}_{ok}^{ABE}}} (ms + tm) \end{aligned}

(10)

In Eq. (10), the term $wd$ indicates the minimized memory space, $rx$ represents the minimized time requirement, $oe$ denotes the objective function requirement, ${OK}_{ok}^{HE}$ represents the optimally selected keys for the HE scheme by the ELBOA, and ${OK}_{ok}^{ABE}$ indicates the optimally selected keys for the ABE scheme. The term $ok$ indicates the number of optimally selected keys by the ELBOA, which is 16 for both HE and ABE schemes separately. The memory size $wd$ is computed using Eq. (11).

wd = gj \times in

(11)

In Eq. (11), the term $gj$ represents the location of the data and $in$ denotes the size of the words. The computation time $rx$ is evaluated using Eq. (12).

rx = \frac{wv}{yh}

(12)

In Eq. (12), the term $yh$ denotes the speed and $wv$ indicates the distance. From the OHEM, the final encrypted data ${ED}_{ed}^{HAB}$ is obtained. The process of encrypting the protected data by the OHEM is depicted in Figure 4.

Fig. 4.

Encryption of the protected data using suggested OHEM.

5.3. ACC-DTCN ON encrypted data

The encrypted data ${ED}_{ed}^{HAB}$ is retrieved if it needs to be accessed. Then it is processed by the ACC-DTCN in the testing phase. But, the processing of the encrypted data directly by the developed model is a complicated task as it requires separate software, system, and decryption schemes for processing. So, before processing the encrypted data, it is decrypted initially. The retrieved encrypted data ${ED}_{ed}^{HAB}$ is primarily decrypted using the same OHEM model. At first, the encrypted data ${ED}_{ed}^{HAB}$ is given to the ABE scheme. From the ABE scheme, the half-decrypted data is obtained. Then the half-decrypted data is provided to the ABE to get the final entirely decrypted protected data ${PD}_{pd}^{ATCN}$ . This assures the security of the test data. The HE transform is able to convert the encrypted data into suitable data that is being processed by ACC-DTCN. The SEAL is the transform used by the HE-ABE scheme. Parameters like the coefficient module, which has a maximum length of sixty digits, and poly_modulus_degree, which should be a power of two, have to be provided to perform the SEAL operation in the HE transform. The selection of these parameters is crucial as it has an effect on the efficiency of the introduced encryption process. The non-linear operations in the ACC-DTCN are not processed by the HE transform. Hence, the “client-assisted technique” nGraph is adapted to processing these non-linear functions. The ciphertext W is sent by the server to be decrypted by the client $D e c (W) \to V$ . The functioning of the server-sent ciphertext $X (V)$ is carried out by the client in order to generate the encryption results $E n c (X (V))$ . This encrypted result is then sent back to the server by the client. This is the process of “client-assisted technique”.

5.4. Split of encrypted deep learning network

The computational burden reduction as well as the efficiency-boosting of the HE transform function, is obtained by offloading the tasks in the edge server. This also enhances the security of the data stored in the cloud server. However, the offloading of the tasks in the cloud to the edge server also leads to the leakage of the information to the edge devices. The proper splitting of the ACC-DTCN network’s data for offloading purposes without compromising privacy as well as efficiency by properly allocating the number of tasks to each edge device is a complicated task to achieve. The number of encrypted data also needs to be considered. The balancing of the transmitting and computation time also needs to be carried out. Hence, on the basis of the network’s structure, the amount of data that needs to be transmitted to the cloud from the edge devices is determined.

6. Result and discussion

6.1. Experimental setup

The implemented data privacy preservation scheme in edge-cloud computing system was developed and executed in Python platform with a maximum iteration count of 50, population count of 10, and chromosome length of $2 * 16$ . The evaluation of the recommended data privacy-preserving scheme was carried out by contrasting it against several conventional cryptography schemes such as “Elliptic Curve Cryptography (ECC) [18], HE [10], ABE [16], and HE-ABE [5]”, with other data analysis models like “Recurrent Neural Network (RNN) [27], Visual Geometry Group (VGG)-16 [6], Long Short Term Memory (LSTM) [33], and Gated Recurrent Unit (GRU) [20]”, and against existing algorithms like “Tunicate Swarm Algorithm (TSA) [41], Golden Eagle Optimizer (GEO) [28], Harris Hawks Optimization (HHO) [15], and LBOA [31]”, respectively.

6.2. Processing datasets

The experimental dataset used in this proposed system is gathered from standard databases. A detailed description of the gathered dataset is provided in Table 2. The gathered data is indicated as ${RD}_{rd}^{gat}$ .

Table 2
Dataset description.

Sl. No. Name Link Description

1) “Heart Disease Dataset” “https://www.kaggle.com/datasets/johnsmith88/heart-disease-dataset access date: 2023-08-22” A collection of around 76 attributes related to heart disease prediction for determining whether an individual has heart disease or not is provided in this dataset.

2) “Lung Cancer Dataset” “https://www.kaggle.com/datasets/yusufdede/lung-cancer-dataset access ate: 2023-08-22” Seven columns of attributes in one file regarding lung cancer are given in this dataset.

3) “Diabetes dataset” “https://www.kaggle.com/datasets/akshaydattatraykhare/diabetes-dataset access date: 2023-08-22” Data containing constraints regarding diabetes arranged in nine columns under a single file is present in this dataset.

Sl. No.	Name	Link	Description
1)	“Heart Disease Dataset”	“https://www.kaggle.com/datasets/johnsmith88/heart-disease-dataset access date: 2023-08-22”	A collection of around 76 attributes related to heart disease prediction for determining whether an individual has heart disease or not is provided in this dataset.
2)	“Lung Cancer Dataset”	“https://www.kaggle.com/datasets/yusufdede/lung-cancer-dataset access ate: 2023-08-22”	Seven columns of attributes in one file regarding lung cancer are given in this dataset.
3)	“Diabetes dataset”	“https://www.kaggle.com/datasets/akshaydattatraykhare/diabetes-dataset access date: 2023-08-22”	Data containing constraints regarding diabetes arranged in nine columns under a single file is present in this dataset.

6.3. Evaluation indices

The evaluation indices used for the validation of the recommended privacy-preservation scheme in a cloud-edge environment are listed below.

(a) The “Negative Predictive Value (NPV)” $gs$ is computed using Eq. (13).

gs = \frac{YZ}{YZ + VW}

(13)

The term $VW$ in Eq. (13) represents the false negative and $YZ$ denotes the true negative.

(b) The calculation of recall $rs$ takes place using Eq. (14).

rs = \frac{ZY}{ZY + VW}

(14)

The term $ZY$ in Eq. (14) represents the true positives.

A = \frac{YZ + ZY}{YZ + WV + ZY + VW}

(15)

The term $WV$ in Eq. (15) represents the false positive.

(d) The evaluation of the “False Positive rate (FPR)” $yd$ is carried out using Eq. (16).

yd = \frac{WV}{YZ + WV}

(16)

(e) The calculation of specificity $ik$ is based on Eq. (17).

ik = \frac{YZ}{YZ + WV}

(17)

(f) The determination of “False Negative Rate (FNR)” $ok$ takes place using Eq. (18).

ok = \frac{VW}{ZY + VW}

(18)

(g) The computation of “Matthews Correlation Coefficient (MCC)” $lr$ takes place utilizing Eq. (19).

lr = \frac{YZ * ZY - WV * VW}{\sqrt{(ZY + WV) (ZY + VW) (YZ + WV) (YZ + VW)}}

(19)

(h) The F1-score $rv$ is determined by Eq. (20).

rv = \frac{2 * ZY}{2 * (ZY + WV + VW)}

(20)

(i) The “False Discovery Rate (FDR)” $tb$ is computed using Eq. (21).

tb = \frac{WV}{ZY + WV}

(21)

(j) The value of precision $wf$ is determined using Eq. (22).

wf = \frac{ZY}{ZY + WV}

(22)

6.4. ROC determination of the implemented data analysis model

The ROC examination of the implemented data analysis model is illustrated in Figure 5. The “True Positive Rate (TPR)” of the implemented ACC-DTCN model is 13.4%, 32.99%, 46.39%, and 56.7% are improved than the GRU, VGG-16, LSTM, and RNN frameworks, respectively for dataset 1 at FRP of 0.1. Thus the performance provided by the implemented ACC-DTCN model is higher than many state-of-the-art methods.

Fig. 5.

ROC assessment of the implemented data analysis model regarding “(a) dataset 1, (b) dataset 2, and (c) dataset 3”.

6.5. Validation of the developed data analysis model

The validation of the developed data analysis model in both cloud and edge sides is depicted in Figure 6. The accuracy of the developed ACC-DTCN model is 9.3%, 14.63%, 13.25%, and 17.5% is improved than the GRU, LSTM, VGG-16, and RNN frameworks, respectively for dataset 3 in the edge side. The accuracy of the developed ACC-DTCN model is 2.67%, 5.49%, 4.35%, and 5.49% is improved than the GRU, LSTM, VGG-16, and RNN frameworks, respectively for dataset 3 on the cloud side. Hence, it is proved that the developed ACC-DTCN model provides enhanced performance on the task assigned to them on the cloud side than edge side.

Fig. 6.

Validation of the performance of the developed data analysis model in terms of “(a) FNR, (b) recall, (c) accuracy, (d) Fi-Score, and (e) FPR”.

6.6. Confusion matrix examination of the generated data analysis framework

The confusion matrix examination of the generated analysis model is provided in Figure 7. The number of correct predictions made by the ACC-DTCN framework is $2.6 \times 10^{2}$ and $4.8 \times 10^{2}$ for dataset 1, $5 \times 10^{2}$ and $4.7 \times 10^{2}$ for dataset 2, and $4.6 \times 10^{2}$ and $4 \times 10^{2}$ for dataset 3, respectively. Accordingly, 95.31%, 95.02%, and 95.22% of accuracy are offered by the generated ACC-DTCN model. Hence, the enhanced performance is given by the ACC-DTCN model than the other existing approaches.

Fig. 7.

Confusion matrix examination of the generated data analysis framework regarding “(a) dataset 1, (b) dataset 2, and (c) dataset 3”.

6.7. Performance assessment of the recommended data privacy preservation scheme

The performance assessment of the recommended data privacy preservation scheme is illustrated in Figure 8. The memory size of the recommended ELBOA-OHEM scheme is 2.7%, 2.7%, 5.26%, and 26.15% is improved than the TSA-OHEM, GEO-OHEM, LBOA-OHEM, and HHO-OHEM schemes, respectively for dataset 2 at block size 4. Improved privacy for data in the edge-cloud environment is offered by the recommended ELBOA-OHEM scheme.

Fig. 8.

Performance assessment of the recommended data privacy-preservation scheme in terms of “(a) computation time, (b) encryption time, (c) memory size, and (d) decryption time”.

6.8. Cryptography scheme-based performance analysis

The cryptography scheme-based performance analysis of the generated data privacy-preservation model is given in Figure 9. The computation time required by the generated ELBOA-OHEM scheme is 28.13%, 34.29%, 11.54%, and 14.82% is enhanced than the TSA-OHEM, HHO-OHEM, LBOA-OHEM, and GEO-OHEM schemes, respectively for dataset 1 at block size 5. Thus, it is assured that the generated ELBOA-OHEM scheme offers more preservation for data in an edge-cloud environment.

Fig. 9.

Performance assessment of the recommended data privacy-preservation model in terms of “(a) computation time, (b) encryption time, (c) memory size, and (d) decryption time”.

6.9. Statistical validation of the developed data privacy preservation scheme

The statistical validation of the developed data privacy-preservation model is listed in Table 3. The mean value of the ELBOA-OHEM scheme is 8.25%, 9.05%, 8.45%, and 7.78% is higher than the TSA-OHEM, GEO-OHEM, HHO-OHEM, and LBOA-OHEM schemes, respectively for dataset 1. Thus, it is proven that more privacy is given by the developed ELBOA-OHEM scheme in preventing the data in an edge-cloud environment.

Table 3
Statistical validation of the developed data privacy-preservation model.

Model/ Indices TSA-OHEM [41] GEO-OHEM [28] HHO-OHEM [15] LBOA-OHEM [31] EBLOA-OHEM

Dataset 1

Standard Deviation 0.095 0.229 0.192 0.222 0.091

Worst 3.380 4.725 4.515 4.093 3.350

Mean 3.164 3.192 3.171 3.148 2.903

Median 3.113 3.128 3.144 3.090 2.884

Best 3.113 3.072 3.144 3.078 2.884

Dataset 2

Mean 3.236 3.347 3.069 3.179 2.978

Worst 3.470 5.145 3.111 3.923 5.376

Best 3.078 3.025 3.037 3.078 2.912

Standard Deviation 0.149 0.511 0.037 0.275 0.358

Median 3.296 3.086 3.037 3.078 2.912

Dataset 3

Median 3.060 3.112 3.067 3.024 2.897

Standard Deviation 0.417 0.392 0.496 0.646 0.307

Best 3.060 3.112 3.065 3.024 2.897

Mean 3.230 3.256 3.274 3.255 2.966

Worst 5.145 4.475 5.092 5.966 4.981

Model/ Indices	TSA-OHEM [41]	GEO-OHEM [28]	HHO-OHEM [15]	LBOA-OHEM [31]	EBLOA-OHEM
Dataset 1
Standard Deviation	0.095	0.229	0.192	0.222	0.091
Worst	3.380	4.725	4.515	4.093	3.350
Mean	3.164	3.192	3.171	3.148	2.903
Median	3.113	3.128	3.144	3.090	2.884
Best	3.113	3.072	3.144	3.078	2.884
Dataset 2
Mean	3.236	3.347	3.069	3.179	2.978
Worst	3.470	5.145	3.111	3.923	5.376
Best	3.078	3.025	3.037	3.078	2.912
Standard Deviation	0.149	0.511	0.037	0.275	0.358
Median	3.296	3.086	3.037	3.078	2.912
Dataset 3
Median	3.060	3.112	3.067	3.024	2.897
Standard Deviation	0.417	0.392	0.496	0.646	0.307
Best	3.060	3.112	3.065	3.024	2.897
Mean	3.230	3.256	3.274	3.255	2.966
Worst	5.145	4.475	5.092	5.966	4.981

6.10. Convergence evaluation of the proposed algorithm

The convergence examination of the proposed algorithm is shown in Figure 10. The cost function of the proposed ELBOA is 8.33%, 8.33%, 4.84%, and 4.84% more than the HHO, GEO, LBOA, and TSA, respectively, for dataset 2 at 40th iteration. The convergence curve of the loss function is mean-squared-error. The objective function of the convergence graph is how the target result is modified with each iteration of the solution. Here, improved performance is shown by the proposed ELBOA.

Fig. 10.

Convergence evaluation of the proposed algorithm regarding “(a) dataset 1, (b) dataset 2, and (c) dataset 3”.

6.11. Ablation experiments of the developed model

The ablation experiments of the developed data privacy-preservation scheme in an edge-cloud environment are shown in shown in Table 4. Generally, an ablation examination is a set of experiments in which modules of expert systems are replaced with the purpose of computing the efficacy of the system. Here, the ablation experiments of the designed approach revealed that it achieves enriched performance based on accuracy metrics.

Table 4
Ablation experiments of the developed model.

Methods LSTM [33] RNN [27] DNN [20] GRU [20] DenseNet [3] CNN [49] ACC-DTCN

Dataset 1

Accuracy 84.50 89.97 91.27 91.66 92.96 90.88 94.66

Dataset 2

Accuracy 82.92 89.75 91.02 91.90 83.41 92.0 95.31

Dataset 3

Accuracy 86.88 91.22 91.66 93.33 93.44 90.77 95.77

Methods	LSTM [33]	RNN [27]	DNN [20]	GRU [20]	DenseNet [3]	CNN [49]	ACC-DTCN
Dataset 1
Accuracy	84.50	89.97	91.27	91.66	92.96	90.88	94.66
Dataset 2
Accuracy	82.92	89.75	91.02	91.90	83.41	92.0	95.31
Dataset 3
Accuracy	86.88	91.22	91.66	93.33	93.44	90.77	95.77

6.12. Validation of computational time of the offered approach

The estimation of the computational time for the developed data privacy-preservation scheme in an edge-cloud environment approach is shown in Table 5. The computational complexity of the developed model is shown in Table 6. From Table 6, the term $MaxIter$ is defined as the maximum iterations, and also the variable $N_{pop}$ is denoted as the number of populations. Accordingly, the parameter $chlen$ is denoted as chromosome length. From Table 5, the simulation outcomes show that the designed model attains better computational time for all datasets.

Table 5
Evaluation of computational time of the developed model.

Algorithms TSA [41] GEO [28] HHO [15] LBOA [31] ACC-DTCN

Time (m) 20.08 17.24 15.23 14.49 13.35

Methods RNN [32] VGG16 [1] LSTM [31] GRU [11] ACC-DTCN

Accuracy 26.53 24.22 21.13 19.25 18.38

Algorithms	TSA [41]	GEO [28]	HHO [15]	LBOA [31]	ACC-DTCN
Time (m)	20.08	17.24	15.23	14.49	13.35
Methods	RNN [32]	VGG16 [1]	LSTM [31]	GRU [11]	ACC-DTCN
Accuracy	26.53	24.22	21.13	19.25	18.38

Table 6

Computational complexity of the developed model.

Algorithms	Computational complexity
TSA [41]	$O [MaxIter * N_{pop} + (N_{pop} * chlen)]$
GEO [28]	$O [MaxIter * N_{pop}]$
HHO [15]	$O [MaxIter * N_{pop} + N_{pop}]$
LBOA [31]	$O [MaxIter * N_{pop}]$
Proposed	$O [MaxIter * N_{pop}]$

6.13. Estimation of convergence curve of loss function

The estimation of the convergence curve of the loss function for the implemented model is shown in Figure 11. The accuracy score is the total count of correct predictions attained in implementation. Accuracy loss values demonstrate the difference from the preferred target states. While analyzing Figure 11, the model accuracy and loss are validated regarding the number of epochs.

Fig. 11.

Evaluation of convergence curve of loss function for the proposed model regarding (a) model accuracy and (b) model loss.

6.14. Comprehensive experimental evaluation regarding the number of epoches

The comprehensive experimental evaluation of the designed model is given in Figure 12. While taking Figure 12, the validation of the designed model shows a better accuracy rate regrading number of epoches. The elevated accuracy rate is attained while the count of epoches is in 250.

Fig. 12.

Comprehensive experimental evaluation of the designed model regarding (a) dataset 1, (b) dataset 2, and (b) dataset 3.

6.15. Comparison of the developed model using recent approaches

The comparison among the suggested data privacy-preservation schemes in an edge-cloud environment model using recent approaches is shown in Table 7. The DKN model attains a lower accuracy rate. Due to this, it cannot be suitable for the larger dataset. The CL model attains a second better performance. Finally, the better accuracy rate of the developed model attained as 95.78% for dataset 3. Therefore, it is revealed that the given developed model attains elevated performance.

Table 7
Comparison among the suggested models using recent approaches.

Methods DKN [39] DCRLS [40] CL [38] ACC-DTCN

Dataset 1

Accuracy 91.02 91.28 92.45 94.66

sensitivity 90.67 91.42 91.79 94.40

Specificity 91.20 91.20 92.80 94.80

Precision 84.67 84.78 87.23 90.68

FPR 8.80 8.80 7.20 5.20

FNR 9.33 8.58 8.21 5.60

NPV 94.80 95.20 95.47 96.93

FDR 15.33 15.22 12.77 9.32

F1-Score 87.57 87.97 89.45 92.50

MCC 80.66 81.28 83.64 88.40

Dataset 2

Accuracy 92.10 92.10 93.37 95.32

sensitivity 91.83 92.21 93.35 95.25

Specificity 92.38 91.98 93.39 95.39

Precision 92.71 92.38 93.70 95.61

FPR 7.62 8.02 6.61 4.61

FNR 8.17 7.79 6.65 4.75

NPV 91.47 91.80 93.01 95.01

FDR 7.29 7.62 6.30 4.39

F1-Score 92.26 92.29 93.52 95.43

MCC 84.19 84.19 86.72 90.63

Dataset 3

Accuracy 91.33 91.78 93.22 95.78

sensitivity 91.72 91.72 93.37 95.65

Specificity 90.89 91.85 93.05 95.92

Precision 92.10 92.87 93.96 96.45

FPR 9.11 8.15 6.95 4.08

FNR 8.28 8.28 6.63 4.35

NPV 90.45 90.54 92.38 95.01

FDR 7.90 7.13 6.04 3.55

F1-Score 91.91 92.29 93.67 96.05

MCC 82.58 83.49 86.38 91.52

Methods	DKN [39]	DCRLS [40]	CL [38]	ACC-DTCN
Dataset 1
Accuracy	91.02	91.28	92.45	94.66
sensitivity	90.67	91.42	91.79	94.40
Specificity	91.20	91.20	92.80	94.80
Precision	84.67	84.78	87.23	90.68
FPR	8.80	8.80	7.20	5.20
FNR	9.33	8.58	8.21	5.60
NPV	94.80	95.20	95.47	96.93
FDR	15.33	15.22	12.77	9.32
F1-Score	87.57	87.97	89.45	92.50
MCC	80.66	81.28	83.64	88.40
Dataset 2
Accuracy	92.10	92.10	93.37	95.32
sensitivity	91.83	92.21	93.35	95.25
Specificity	92.38	91.98	93.39	95.39
Precision	92.71	92.38	93.70	95.61
FPR	7.62	8.02	6.61	4.61
FNR	8.17	7.79	6.65	4.75
NPV	91.47	91.80	93.01	95.01
FDR	7.29	7.62	6.30	4.39
F1-Score	92.26	92.29	93.52	95.43
MCC	84.19	84.19	86.72	90.63
Dataset 3
Accuracy	91.33	91.78	93.22	95.78
sensitivity	91.72	91.72	93.37	95.65
Specificity	90.89	91.85	93.05	95.92
Precision	92.10	92.87	93.96	96.45
FPR	9.11	8.15	6.95	4.08
FNR	8.28	8.28	6.63	4.35
NPV	90.45	90.54	92.38	95.01
FDR	7.90	7.13	6.04	3.55
F1-Score	91.91	92.29	93.67	96.05
MCC	82.58	83.49	86.38	91.52

6.16. Case study

(i)
Data privacy for Privacy Preservation Performance in Edge Computing: The core aim of this work is to create an effective privacy preservation prototype using improved deep structured techniques. In addition, it protects the privacy data from inference attacks. Initially, the author has introduced the concept of membership inference attacks.
(ii)
Membership Inference Attacks: Each training procedure of the training record will impact the model parameters by the deep structured approaches to minimize the model’s training loss during the model training process.
(iii)
Privacy leakage Analysis: The major aim of this concept is a specific data record was utilized to train a target model or not. It can help to reflect the privacy leakage of a model.
(iv)
Model Performance Comparison: The performance validation of the developed model is conducted using standard datasets and diverse deep-structured approaches and heuristic algorithms. From the validation, the maximum accuracy rate of the developed model is 95.78% and also it attains a better convergence rate while comparing the other approaches.

7. Conclusion

An enhanced data privacy-preservation scheme in an edge-cloud environment was deployed successfully utilizing cryptography and deep learning techniques. The user data were gathered from the data source. The gathered data were provided to the ACC-DTCN model for performing the task assigned to them on the trusted edge side. On the cloud-side, in order to protect the test data, an encryption scheme was implemented. The test data was encrypted by the developed OHEM scheme. The encrypted data was then again provided as input to the ACC-DTCN model on the untrusted cloud side. Thus enhanced security of the transmitted data was provided in both edge and cloud-side environments. The experimental validation was carried out to prove the efficacy and privacy provided by the suggested data privacy-preservation scheme. The suggested ACC-DTCN model was 2.67%, 5.49%, 4.35%, and 5.49% more accurate than the GRU, LSTM, VGG-16, and RNN frameworks, respectively on the cloud side. Hence, enhanced privacy was proved to be provided by the suggested scheme on both users as well as the cloud side in the edge-cloud environment. However, need to include the concept of cooperative privacy preservation which will be perfectly suitable for smart health. In addition, the given proposed model needs more investigation of the security properties like mutual authentication, data confidentiality, and integrity, privacy preservation, and forward security. In the future, we will include the concept of cooperative privacy preservation for smart healthcare. Especially, the investigation of security properties will help to enhance the trustworthiness of the offered approach.

Footnotes

Conflict of interest

The authors have no conflict of interest to report.

References

Abdelhamid

A.A.

El-Kenawy

E.-S.M.

Ibrahim

Eid

M.M.

Khafaga

D.S.

Alhussan

A.A.

Mirjalili

Khodadadi

Lim

W.H.

Shams

M.Y.

, Innovative feature selection method based on hybrid sine cosine and dipper throated optimization algorithms, IEEE Access (2023).

Alsahli

M.A.

Alsanad

Hassan

M.M.

Gumaei

, Privacy preservation of user identity in contact tracing for COVID-19-like pandemics using edge computing, IEEE Access 9 (2021), 125065–125079. doi:10.1109/ACCESS.2021.3110762.

Amer

Liu

Velisavljevic

, A privacy-preserving approach to effectively utilize distributed data for malaria image detection, Bioengineering 11(4) (2024), 340.

Atteia

El-Kenawy

E.-S.M.

Samee

N.A.

Jamjoom

M.M.

Ibrahim

Abdelhamid

A.A.

Azar

A.T.

Khodadadi

Ghanem

R.A.

Shams

M.Y.

, Adaptive dynamic dipper throated optimization for feature selection in medical data, Computers, Materials & Continua 75(1) (2023), 1883–1900. doi:10.32604/cmc.2023.031723.

Brakerski

Cash

Tsabary

Wee

, Targeted homomorphic attribute-based encryption, in: Home Theory of Cryptography, 21 October 2016.

Campos-Leal

J.A.

Yee-Rendón

Vega-López

I.F.

, Simplifying VGG-16 for plant species identification, IEEE Latin America Transactions 20(11) (2022), 2330–2338. doi:10.1109/TLA.2022.9904757.

Chen

Y.-C.

Chang

C.-C.

Hung

C.-C.

Lin

J.-F.

Hsu

S.-Y.

, SecDT: Privacy-preserving outsourced decision tree classification without polynomial forms in edge-cloud computing, IEEE Transactions on Signal and Information Processing over Networks 8 (2022), 1037–1048. doi:10.1109/TSIPN.2022.3233193.

Cui

Nepal

, Achieving scalable access control over encrypted data for edge computing networks, IEEE Access 6 (2018), 30049–30059. doi:10.1109/ACCESS.2018.2844373.

Fitwi

Chen

Zhu

, Enforcing privacy preservation on edge cameras using lightweight video frame scrambling, IEEE Transactions on Services Computing 16(1) (2023), 276–287.

10.

Fontaine

Galand

, A survey of homomorphic encryption for nonspecialists, EURASIP Journal on Information Security 2007 (2007), 013801. doi:10.1186/1687-417X-2007-013801.

11.

Goay

C.H.

Ahmad

N.S.

Goh

, Temporal convolutional networks for transient simulation of high-speed channels, Alexandria Engineering Journal 74 (2023), 643–663.

12.

Gao

Wang

Jin

, Privacy on the edge: Customizable privacy-preserving context sharing in hierarchical edge computing, IEEE Transactions on Network Science and Engineering 7(4) (2020), 2298–2309.

13.

Hassan

Shams

M.Y.

Hikal

N.A.

Elmougy

, The effect of choosing optimizer algorithms to improve computer vision tasks: A comparative study, Multimedia Tools and Applications 82(11) (2023), 16591–16633. doi:10.1007/s11042-022-13820-0.

14.

Liu

Guo

Yang

, Privacy-preserving and low-latency federated learning in edge computing, IEEE Internet of Things Journal 9(20) (2022), 20149–20159.

15.

Heidari

A.A.

Mirjalili

Faris

Aljarah

Mafarja

Chen

, Harris hawks optimization: Algorithm and applications, Future Generation Computer Systems 97 (2019), 849–872. doi:10.1016/j.future.2019.02.028.

16.

Herranz

, Attacking pairing-free attribute-based encryption schemes, IEEE Access 8 (2020), 222226–222232. doi:10.1109/ACCESS.2020.3044143.

17.

Jin

Zhang

Dong

Zhu

Bouguettaya

, Privacy-aware forecasting of quality of service in mobile edge computing, IEEE Transactions on Services Computing 16(1) (2023), 478–492.

18.

Lai

J.-Y.

Huang

C.-T.

, A highly efficient cipher processor for dual-field elliptic curve cryptography, IEEE Transactions on Circuits and Systems II: Express Briefs 56(5) (2009), 394–398.

19.

Song

Chen

, Privacy-preserving cost minimization in mobile crowd sensing supported by edge computing, IEEE Access 8 (2020), 121920–121928. doi:10.1109/ACCESS.2020.3007168.

20.

Liu

Zhang

Zhou

, Automatic modulation recognition based on CNN and GRU, Tsinghua Science and Technology 27(2) (2022), 422–431. doi:10.26599/TST.2020.9010057.

21.

Liu

Yao

Yang

Ning

, Cooperative privacy preservation for wearable devices in hybrid computing-based smart health, IEEE Internet of Things Journal 6(2) (2019), 1352–1362. doi:10.1109/JIOT.2018.2843561.

22.

Liu

Zheng

Huang

Quek

T.Q.S.

, Distributed reinforcement learning for privacy-preserving dynamic edge caching, IEEE Journal on Selected Areas in Communications 40(3) (2022), 749–760. doi:10.1109/JSAC.2022.3142348.

23.

Liu

Dai

Dou

, Privacy preservation for federated learning with robust aggregation in edge computing, IEEE Internet of Things Journal 10(8) (2023), 7343–7355.

24.

Liao

Lio

Hui

, Privacy-preserving asynchronous federated learning mechanism for edge network computing, IEEE Access 8 (2020), 48970–48981. doi:10.1109/ACCESS.2020.2978082.

25.

Liu

, Attention guided U-net with atrous convolution for accurate retinal vessels segmentation, IEEE Access 8 (2020), 32826–32839. doi:10.1109/ACCESS.2020.2974027.

26.

Miao

Liu

Choo

K.-K.R.

Yang

Wang

, Lightweight privacy-preserving medical diagnosis in edge computing, IEEE Transactions on Services Computing 15(3) (2022), 1606–1618.

27.

Zhang

Liu

, MM-RNN: A multimodal RNN for precipitation nowcasting, IEEE Transactions on Geoscience and Remote Sensing 61 (2023), 1–14.

28.

Mohammadi-Balani

Nayeri

M.D.

Azar

Taghizadeh-Yazdi

, Golden eagle optimizer: A nature-inspired metaheuristic algorithm, Computers & Industrial Engineering 152 (2021), 107050. doi:10.1016/j.cie.2020.107050.

29.

Qiang

Liu

Jin

, Defending CNN against privacy leakage in edge computing via binary neural networks, Future Generation Computer Systems 125 (2021), 460–470. doi:10.1016/j.future.2021.06.037.

30.

Wang

, Privacy-preserving model training architecture for intelligent edge computing, Computer Communications 162 (2020), 94–101. doi:10.1016/j.comcom.2020.08.007.

31.

Safiri

Nikoofard

, Ladybug beetle optimization algorithm: Application for real-world problems, The Journal of Supercomputing (2022).

32.

Samee

N.A.

El-Kenawy

E.-S.M.

Atteia

Jamjoom

M.M.

Ibrahim

Abdelhamid

A.A.

El-Attar

N.E.

Gaber

Slowik

Shams

M.Y.

, Metaheuristic optimization through deep learning classification of COVID-19 in chest X-ray images, Computers, Materials & Continua 73(2) (2022).

33.

Shu

Zhang

Sun

Tang

, Host–parasite: Graph LSTM-in-LSTM for group activity recognition, IEEE Transactions on Neural Networks and Learning Systems 32(2) (2021), 663–674. doi:10.1109/TNNLS.2020.2978942.

34.

Telikani

Shahbahrami

Shen

Gaydadjiev

Lin

J.C.-W.

, An edge-aided parallel evolutionary privacy-preserving algorithm for Internet of Things, Internet of Things (2023), 100831.

35.

Tian

Zeng

Bhuiyan

M.Z.A.

Tian

Cai

Chen

Zhong

, Trajectory privacy preservation based on a fog structure for cloud location services, IEEE Access 5 (2017), 7692–7701. doi:10.1109/ACCESS.2017.2698078.

36.

Ting

Qiu

Cao

Cheng

Wang

Shi

Wang

, Privacy-preserving participant grouping for mobile social sensing over edge clouds, IEEE Transactions on Network Science and Engineering 8(2) (2021), 865–880. doi:10.1109/TNSE.2021.3063959.

37.

Wang

Ren

Wang

Zhang

Shen

X.S.

, PrivStream: A privacy-preserving inference framework on IoT streaming data at the edge, Information Fusion 80 (2022), 282–294. doi:10.1016/j.inffus.2021.11.013.

38.

Xiao

Tong

Xing

Luo

Zhu

Song

Feng

, CapMatch: Semi-supervised contrastive transformer capsule with feature-based knowledge distillation for human activity recognition, IEEE Trans. Neural Netw. Learn. Syst. (2023).

39.

Xiao

Xing

Feng

, Densely knowledge-aware network for multivariate time series classification, IEEE Transactions on Systems Man and Cybernetics Systems, PP 99 (2024), 1–13. doi:10.1109/TSMC.2024.3416900.

40.

Xiao

Xing

Zhao

Luo

Dai

Zhu

, Deep contrastive representation learning with self-distillation, IEEE Transactions on Emerging Topics in Computational Intelligence 8(1) (2023), 3–15. doi:10.1109/TETCI.2023.3304948.

41.

Xie

Han

Zhou

Zhang

Z.-R.

Han

Tang

, Tuna swarm optimization: A novel swarm-based metaheuristic algorithm for global optimization, Artificial Intelligence and Machine Learning-Driven Decision-Making 2021, 9210050 (2021).

42.

Wan

Bhuiyan

M.Z.A.

, Joint optimization of offloading utility and privacy for edge computing enabled IoT, IEEE Internet of Things Journal 7(4) (2020), 2622–2629. doi:10.1109/JIOT.2019.2944007.

43.

Huang

Zhu

Sharma

Zhang

Bhuiyan

M.Z.A.

, Secure service offloading for Internet of vehicles in SDN-enabled mobile edge computing, IEEE Transactions on Intelligent Transportation Systems 22(6) (2021), 3720–3729. doi:10.1109/TITS.2020.3034197.

44.

Liu

Wang

Wan

Yang

, Joint optimization of resource utilization and load balance with privacy preservation for edge services in 5G networks, Mobile Networks and Applications 25 (2020), 713–724. doi:10.1007/s11036-019-01448-8.

45.

Liu

Jiang

Tang

, A time-efficient data offloading method with privacy preservation for intelligent sensors in edge computing, EURASIP Journal on Wireless Communications and Networking 2019(236) (2019).

46.

Zhang

Huang

, Optimal task assignment with delay constraint for parked vehicle assisted edge computing: A Stackelberg game approach, IEEE Communications Letters 24(3) (2020), 598–602. doi:10.1109/LCOMM.2019.2961663.

47.

Zheng

Liu

Huang

Zhang

Yang

, Unsupervised recurrent federated learning for edge popularity prediction in privacy-preserving mobile-edge computing networks, IEEE Internet of Things Journal 9(23) (2022), 24328–24345.

48.

Zheng

Tian

Jiang

Elhanashi

Saponara

, Application of wavelet-packet transform driven deep learning method in PM2.5 concentration prediction: A case study of Qingdao, China. Sustainable Cities and Society 92 (2023), 104486. doi:10.1016/j.scs.2023.104486.

49.

Zheng

Tian

Wang

Elhanashi

Saponara

, DL-PR: Generalized automatic modulation classification method based on deep learning with priori regularization, Engineering Applications of Artificial Intelligence 122 (2023), 106082. doi:10.1016/j.engappai.2023.106082.

50.

Zheng

Zhao

Wang

Yang

, Spectrum interference-based two-level data augmentation method in deep learning for automatic modulation classification, Neural Computing & Applications 33 (2020), 7723–7745. doi:10.1007/s00521-020-05514-1.

51.

Zheng

Zhao

Wang

Elhanashi

Saponara

, Fine-grained modulation classification using multi-scale radio transformer with dual-channel representation, IEEE Communications Letters 26(6) (2022), 1298–1302. doi:10.1109/LCOMM.2022.3145647.

52.

Zheng

Zhao

Zhang

Wang

, MR-DCAE: Manifold regularization-based deep convolutional autoencoder for unauthorized broadcasting identification, International Journal of Intelligent Systems 36(12) (2021), 7204–7238. doi:10.1002/int.22586.

An intelligent atrous convolution-based cascaded deep learning framework for enhanced privacy preservation performance in edge computing

Abstract

Keywords

1. Introduction

2. Literature survey

2.1. Related works

2.2. Problem statement

3. Intelligent privacy preservation framework in edge computing

3.1. Developed privacy preservation model in edge computing

4. Enhanced privacy preservation model in edge computing using advanced deep learning with a meta-heuristic algorithm

4.1. Deep learning on encrypted data

4.2. Use of edge computing

4.3. Ladybug beetle optimization algorithm

5.1. Proposed atrous convolution-based cascaded deep temporal convolution network

5.4. Split of encrypted deep learning network

6. Result and discussion

6.1. Experimental setup

6.2. Processing datasets

Table 5 Evaluation of computational time of the developed model. Algorithms TSA [41] GEO [28] HHO [15] LBOA [31] ACC-DTCN Time (m) 20.08 17.24 15.23 14.49 13.35 Methods RNN [32] VGG16 [1] LSTM [31] GRU [11] ACC-DTCN Accuracy 26.53 24.22 21.13 19.25 18.38

Footnotes

Conflict of interest

References

Table 5
Evaluation of computational time of the developed model.

Algorithms TSA [41] GEO [28] HHO [15] LBOA [31] ACC-DTCN

Time (m) 20.08 17.24 15.23 14.49 13.35

Methods RNN [32] VGG16 [1] LSTM [31] GRU [11] ACC-DTCN

Accuracy 26.53 24.22 21.13 19.25 18.38