Secure Task Offloading and Resource Allocation Strategies in Mobile Applications Using Probit Mish-Gated Recurrent Unit and an Enhanced-Searching-Based Serval Optimization Algorithm

Abstract

Today, with the presence of 5G communication systems, including Internet of Things (IoT) technology, there is a high demand for mobile devices (especially smartphones, tablets, wearable technology, and so on). Regarding this proliferation and high demand, the massive adoption of mobile devices (MDs) has led to an exponential increase in network latency; the heavy demand for cloud servers causes the degradation of data traffic, which considerably impacts the real-time communication and computing aspects of mobile devices. As a result, mobile edge computing (MEC), an efficient framework capable of enhancing processing, optimizing energy usage, and offloading computation tasks, is considered a promising solution. In current research, numerous models have been implemented to achieve resource allocation and task offloading. However, these techniques are ineffective due to privacy issues and a lack of sufficient resources. Hence, this study proposes secure task offloading and resource allocation strategies in mobile devices using the Probit Mish–Gated Recurrent Unit (PM-GRU) and Entropic Linear Interpolation-Serval Optimization Algorithm (ELI-SOA). Primarily, the tasks to be offloaded and their attributes are gathered from mobile users and passed to a local computing model to identify the edge server. Here, the task attributes and the server attributes are compared with a cache table using the Sorensen–Dice coefficient. If the attributes match, then details about the appropriate edge server are produced. If the attributes do not match, then they are inputted into a global scheme that analyzes the attributes and predicts the edge server based on the Probit Mish-Gated Recurrent Unit (PM-GRU). Then, the server information is preserved and updated in the cache table in the local scheme. Further, the attributes, along with the predicted edge server, are inputted into a system for privacy-preserving smart contract creation by using Exponential Earth Mover’s Distance Matrix-Based K-Anonymity (EEMDM-KA) to develop a secure smart contract. Subsequently, the traffic attributes in the smart contract are extracted, and the request load is balanced by using HCD-KM. Load-balanced requests are assigned to the edge server, and the optimal resources are allocated in the cloud server by using the Entropic Linear Interpolation-Serval Optimization Algorithm (ELI-SOA). Finally, the created smart contract is hashed based on KECCAK-512 and stored in the blockchain. With a high accuracy of 99.84%, the evaluation results showed that the proposed approach framework performed better than those used in previous efforts.

1. Introduction

The demand for smart gadgets, including smartphones, tablets, wearable technology, and intelligent sensors, is rising quickly around the world due to recent advancements stemming from the implementation of 5G communication and Internet of Things (IoT) technologies [1]. At this stage, the massive usage of mobile devices (MDs) leads to high network latency and data traffic, affecting real-time communication and computing processes [2]. Moreover, the large amount of mobile data produces heavy demand for cloud servers, which degrades systems’ consistency [3]. Hence, to enhance computation and optimize the energy consumption of mobile applications, an effective framework named mobile edge computing (MEC) was developed [4]. MEC makes it possible to offload computation tasks to nearby MEC servers, which help upgrade the resource allocation [5]. In numerous applications, such as IoT, Internet of Vehicles (IoV), and ultra-dense networks, MEC plays a significant role [6].

The architecture of MEC contains important layers, such as end devices, access networks, edge networks, and core infrastructure. These layers facilitate resource allocation and task offloading on mobile devices [7]. However, security and privacy issues represent an important problem that affects the reliability of the model. Due to this situation, the integration of secure and reliable task offloading and resource allocation is seen as an optimal solution [8]. Generally, authentication techniques such as the advanced encryption standard (AES), elliptic-curve cryptography (ECC), the digital signature algorithm (DSA), and the Rivest–Shamir–Adleman (RSA) algorithm are utilized to ensure secure task offloading and resource allocation processes [9]. Of these techniques, ECC is one of the most promising in the authentication process because it is more efficient and has high cryptographic strength, low computing power, and latency [10]. Numerous research works have presented task offloading and resource allocation in mobile applications. Figure 1 presents a diagram of MEC.

Previous studies based their approaches on ML techniques, such as the support vector machine (SVM), Markov model (MM), federated learning (FL), Q-learning, and decision tree (DT), to achieve effective task offloading and resource allocation in mobile devices [11]. These systems made significant contributions; however, they still present some inefficiency due to their high latency and battery power consumption. Therefore, effective DL approaches, such as deep neural networks (DNNs), deep reinforcement learning (DRL), and the deep deterministic policy gradient (DDPG) [12,13], were developed to improve the TO and RA processes in MEC. Furthermore, Lyapunov optimization, convex optimization, heuristic techniques, and game theory were also used to perform resource scheduling on mobile phones [14]. However, these models were not regularized well to optimize energy consumption, and they retained poor learning performance [15]. To overcome these downsides and upgrade the MEC network performance in mobile applications, our study proposes a combined Probit Mish–GRU model with the ELI–SOA to implement secure task offloading and resource allocation strategies.

1.1. Problem Statement

Some of the problems noticed in the existing works are as follows:

• Most of the existing works performed task offloading and resource allocation by verifying the task features with the global scheme alone but failed to analyze the local computing features.
• The existing works did not focus on the traffic attributes for load balancing. This resulted in more energy consumption.
• The security of smart contracts was not ensured in the previous works, which lowered the trustworthiness of the networks.
• System consistency is a major challenge for storing hashed information in the cloud computing environment.

1.2. Contributions

The proposed framework addresses several limitations in the existing mobile network systems by introducing innovative approaches. It focuses on hybrid models, such as cache- and content-based secure task offloading on edge and cloud servers, thus significantly enhancing model performance. This is achieved through the integration of an SD-based similarity evaluation with a cache table inside the local scheme and the introduction of a novel integrated Probit Mish (PM)–GRU working inside the global scheme for content-based task offloading, thereby improving the dynamic network performance. Moreover, energy-efficient resource allocation is improved using HCD-KM for load balancing, thus reducing communication and storage burdens. To tackle high computing latency and energy consumption at the edge server, the proposed design implements local and global schemes to enhance the edge server’s quality. Privacy and security concerns are addressed by incorporating EMD based on k-anonymization, ensuring privacy-preserving smart contract creation and enhancing the model’s trustworthiness. Additionally, blockchain is utilized to store hashed information, improving system consistency despite the consumption of computing resources during blockchain consensus.

The remainder of this article is organized as follows: in Section 2, we explore related models through a review of the recent literature. Similarly, Section 3 describes the mathematical derivation of the proposed framework, and Section 4 provides the experimental results and performance evaluations. Finally, Section 5 concludes the study, provides future enhancement strategies, and presents perspectives for future work.

2. Literature Review

Shahidinejad et al. [16] addressed the challenges of meeting user QoS requests, including improving the quality of user experience with the purpose of increasing performance within the mobile communication environment and optimizing the quality of task offloading. In order to investigate the challenge of integrating energy consumption with quality assurance in MEC and to meet the demand of mobile users for low-energy consumption and low latency, in [17], Tao et al. introduced an approach to minimizing energy consumption in MEC. In addition, they optimized and solved the issue using the Karush–Kuhn–Tucker approach. Based on the results of numerical simulations, this approach outperformed complete offloading and local computing in terms of delay performance and energy consumption. Wang et al. [18] proposed an MEC integrated into a framework for MSP performance trade-offs and used the Lyapunov algorithm to optimize the framework. The simulation performance showed important improvements where mobile service provision reached an expected maximum level with the promise of guaranteeing system stability and minimizing energy consumption. In [19], Von et al. introduced an integrated framework to improve computational performance after evaluating the costs of local and distant computing. Additionally, an expanded examination of metrics for computational offloading in MEC was carried out, and task offloading optimization approaches based on several AI approaches were compared in multiple dimensions. Zhang et al. [20] examined multiple task-aware unmanned aerial vehicle (UAV)-enabled MEC networks using RA- and DRL-based trajectory planning. Here, the multi-agent DDPG technique was employed to design trajectories. Similarly, the Particle Swarm Optimization (PSO) algorithm was used to carry out the RA procedure. This model had a completion time of 10 ms, which showed the high feasibility and supremacy of the system. However, it could design poor trajectories due to the massive spaces between ground devices. On the other hand, in [21], Baidas showed a successful RA in MEC networks with clustered non-orthogonal multi-access (NOMA) functioning enabled. Techniques such as successive convex approximation, K-means (KM), and stable marriage matching (SMM) were utilized to achieve RA while maximizing the offloading efficiency. The model had a high potential to handle multiple sub-carriers and mitigate the transmission delay. Still, it retained a high energy consumption (EC) value due to its large bandwidth. In the same manner, in [22], Kuang et al. first explored a model using cooperative computation offloading and RA to reduce the transmission delay in MEC. The purpose of this study was to minimize the frequency of CPU cycles and the transmission power. A monotonic optimization method was established to minimize the transmission power. Second, the Sheng Jin formula (SJF) was incorporated to reduce the CPU cycle frequency. Similarly, to minimize latency and upgrade the RA, a combined iterative approach based on Lagrangian dual decomposition was employed. However, this system did not consider all of the EC factors in CS. Conversely, in [23], Liao et al. presented a joint offloading decision and RA for MEC-enabled networks. Initially, graph theory was employed to recognize the bandwidth of the network. Finally, to generate a nearly optimal solution with the lowest possible cost in MD, a low-complexity heuristic algorithm was devised. This system attained a total of 75 offloaded tasks and achieved better offloading efficiency. However, it had a high computational complexity due to the increased network capacity.

In [24], Pereira et al. incorporated a framework named the RELIABLE resource allocation scheme for 5G mobile networks using MEC. This system employed the RELIABLE mechanism to handle resource management challenges. The performance of the RELIABLE scheme was great when dealing with the 5G network. This method took service time, mobility prediction, and bandwidth into account when allocating resources in the 5G network. However, it took considerable processing time, which mainly limited the model’s consistency. In the same way, in [25], Tong et al. used an adaptive computation offloading and RA technique based on DRL in an MEC environment. First, the Poisson distribution technique was used to generate tasks; secondly, the DRL technique was used to perform effective offloading and RA in MD. This technique finally helped reduce the response time and EC, but it had memory requirements that affected the MEC network performance.

In [26], Ning et al. incorporated an intelligent RA scheme into mobile blockchain to achieve secure transactions using DRL. This framework consisted of the integration of a mobile blockchain module (MBM), task offloading module (TOM), and decision-making modules (DMMs). In the MBM, all of the MDs were constructed in a blockchain network. In the TOM, the task mining of the MD with nearby MEC servers was carried out. Finally, the DMMs made a decision about RA in MEC using the fused DDPG and PSO approaches. This model attained high stability and a better convergence rate. However, it could reach false positives due to the random selection process of the training samples. Conversely, in [27], Yu et al. developed a DL-based TO and RA strategy in an MEC framework. A dueling deep Q-learning network (DQN) was established to monitor the TO and handle resource management. The EC and completion time obtained with this approach were, respectively, 52.18% and 34.72%. Thus, the results of the analysis showed that the model helped mitigate the EC and completion time. However, this method failed because it lacked a useful objective function.

In [28], Xue et al. showed a mobile network identity management and authentication system based on blockchain technology. In this case, user authentication was ensured using the public key and self-sovereign identities (SSIs) under blockchain monitoring. Similarly, a chameleon hash algorithm was established to record unauthorized users. The model was generalized well to reduce network delay and storage space. However, this technique took too much time to complete the validation process. In the same context, in [29], Li et al. created a federated-reinforcement-learning-based workload offloading method in MEC. This framework encompassed the local scheme and the system scheme. Initially, the local scheme was trained using a double DQN (DDQN). Likewise, the global scheme was trained based on historical information. This technique achieved task offloading on the edge server and resource allocation on the cloud server based on federated learning. This system had low latency and superior performance. However, it needed a large number of resources, which could affect the accuracy of the system.

Based on the above analysis, we compared the research contributions, some of which are summarized in

Table 1. Analysis of the contributions in the literature.

Ref.	Main Contributions	Weaknesses	Advantages
[16]	This study addressed the challenges of meeting user QoS requests, including improving the quality of user experience with the purpose of increasing the performance of user experience with mobile communication environments.	The research provides a consistent survey of approaches to secure computational offloading based on AI technology.	There is a significant lack of details concerning resource allocation strategies, and IoT was not properly mentioned in the literature.
[20]	This study presented a series of investigations of task-aware multi-unmanned aerial vehicle (UAV)-enabled MEC networks using RA- and DRL-based trajectory planning; there was an important focus on mobile system implementation.	The proposed system might design poor trajectories due to the massive space between ground devices.	The authors adopted the DDPG approach to design trajectories. Similarly, the particle swarm optimization (PSO) algorithm was used to carry out the RA procedure. This model had a completion time of 10 ms, which was considered high performance compared with the results obtained in recent works in the same context. This result showed the high feasibility and supremacy of the system.
[22]	This study proposed an approach using cooperative computation offloading and RA to reduce transmission delay in an MEC system. The structure of the design was based on minimizing the frequency of CPU cycles and transmission power.	The proposed method did not consider all of the EC factors in the CS implementation.	The adoption of a monotonic optimization method to minimize the transmission power and the integration of the Sheng Jin formula (SJF) to reduce the CPU cycle frequency were important contributions that minimized the latency and upgraded the RA.
[28]	This study introduced an identity management and authentication scheme for mobile systems based on blockchain technology; user authentication was ensured and secured using a public key and self-sovereign identities (SSIs) under blockchain monitoring.	The proposed technique took too much time to complete the validation process.	The model was adopted and generalized well to reduce network delay and storage space and allow good management of resource allocations.

, and we found that some of the recent contributions did not provide enough attention, leaving two main weaknesses in the existing literature. The first is an important lack of study of the real-life impact on the IoT environment. Secondly, the situation of the real-time communication and computing aspects of mobile devices was not adequately developed or even mentioned in these studies. There were no contributions demonstrating effective work concerning the performance of offloading on nearby edge servers that allocate suitable resources to virtual machines on cloud servers.

3. Secure Task Offloading with PM-GRU and ELI-SOA

In this study, secure task offloading and resource allocation strategies in mobile devices using Probit Mish combined with a GRU and ELI-SOA are proposed. The proposed system effectively offloads tasks to a nearby edge server and successfully allocates suitable resources to virtual machines in a cloud server. The safety and privacy of users’ sensitive information are a focus of the proposed framework. In Figure 2, the suggested architecture’s design is displayed.

Figure 1 illustrates the reliable task offloading and effective resource allocation inside the MEC network system. This model includes a local scheme, global scheme, load-balancing model, resource allocation scheme, blockchain, smart contract creation, and task offloading model.

(1)

The local scheme: The local scheme, which is primarily linked to mobile users’ data and is used to control all tasks and manage data coming from offloaded tasks, is a sub-part of the integral MEC system framework. It is also in charge of making updates to data when attributes match, and then details about the appropriate edge server are produced. If the attributes do not match, they are inputted into the server scheme, which predicts the edge server based on the Probit model and GRU.

(2)

The global scheme: A physical location that is close to the systems of mobile devices (apps) managing extracted information coming from a dataset and from the EEMDM-KA program is designated in our approach as the location for edge server prediction; this server is led by the proposed PM-GRU as the global scheme, and it manages datasets and feature extraction.

(3)

The load-balancing model: Defined as HCD-KM, this technique represents a method of proportionally distributing network traffic across a whole MEC system; most information, such as text, videos, images, and other data, is processed inside this component to allow mobile users to interchange information properly and efficiently. Its information passes through the local scheme, and EEMDM-KA passes through the program for traffic attribute extraction to arrive at the load-balancing scheme and, finally, be sent to the cloud server.

(4)

The resource allocation scheme: Resource allocation, which is the process of assigning and managing HCD-KM, delivers better ELI-SOA assets by providing complete cloud service in a manner that provides reliable profitability to the blockchain for KECCAK 512. This system includes managing tangible assets, such as hardware, to make the best use of softer assets, such as human capital.

(5)

The blockchain: The blockchain framework manages all processing to avoid any leakages of MEC data from the edge server while delivering secure and reliable task offloading.

(6)

Smart contact creation: Smart contracts are created and hashed based on the Keccak 512-bit Hash Generator. This system is adopted to automate the execution of any agreement so that all users can be promptly certain of the outcome without any intermediary’s involvement or loss of time.

(7)

Task offloading model: This model’s main role consists of transferring all resource-intensive computational tasks to other connected systems.

The mathematical derivation and a detailed explanation of the individual phases are presented in the following.

3.1. Mobile Cloud Resource Allocation

First, numerous mobile users send a request to access cloud resources. Second, a user request undergoes further processing before resource allocation. Finally, the mobile users $\left({\mathrm{M}}_p^{\Theta }\right)$ are initialized as

$${\mathrm{M}}_p^{\Theta }\to \left|{\mathrm{M}}_1^{\Theta },{\mathrm{M}}_2^{\Theta },\dots \dots {\mathrm{M}}_P^{\Theta }\right|, p=1,2,\dots P$$

(1)

where $P$ denotes the number of mobile users. Here, the crucial attributes of the mobile users are established to implement further computation. This is described in the following.

In this phase, the tasks to be offloaded (T_h) and their attributes are extracted from mobile users. The offloading of tasks refers to the transfer of data from user devices to another external device. Here, the attributes, such as the number of dimensions, speed, cost, weight, and data size, are extracted. This can be formulated as follows:

$$T_h=\left(T_1,T_2,\dots \dots T_H\right)$$

(2)

$${\alpha }_b\left(T_h\right)\to \left[{\alpha }_1,{\alpha }_2,{\alpha }_3,\dots \dots {\alpha }_B\right], b=1,\dots ,B$$

(3)

where $h=1,2,\dots H$ indicates the total of $T_h$, and $B$ defines the maximum number of extracted attributes. Finally, $T_h$ and ${\alpha }_b$ are given to a local computing system for edge server prediction.

In this LS, $T_h$ and ${\alpha }_b$ are subjected to an evaluation of their similarity with a cache table. For this purpose, the Sorensen–Dice formula (SD) is used. It compares the attributes with the cache information to select a suitable edge server. This process can be described as follows:

$$S^{\circ }\left|c^{\infty }\right|=\raisebox{1ex}{$2\cdot ‖\left(T_h+{\alpha }_b\right)\cap {\delta }_f‖$}\!\left/ \!\raisebox{-1ex}{$‖\left(T_h+{\alpha }_b\right)‖+‖{\delta }_f‖$}\right.$$

(4)

$$S^{\circ }\to \left\{\begin{array}{ll}{S}^{\circ }\left|c^{\infty }\right|=={\delta }_f,& m{d}^{\Phi }\\ else,& n{t}^{\Omega }\end{array}\right.$$

(5)

where $S^{\circ }$ denotes the similarity evaluation, $c^{\infty }$ indicates the cache table, ${\delta }_f$ represents the historical attributes stored in $c^{\infty }$, and $m{d}^{\Phi }$ and $n{t}^{\Omega }$ define the matched and unmatched outcomes, respectively.

If the attributes are matched, then the LS $\left(l^{\nabla }\right)$ provides the relevant edge server $\left({\zeta }_y\right)$ according to the tasks to be offloaded. Then, the attributes are inputted into the privacy-preserving smart contract creation model and updated in the cache table. This is represented as follows:

$$l^{\nabla }=m{d}^{\Phi }\leftrightarrow {\zeta }_y, y=1,2,\dots Y$$

(6)

where $Y$ represents the number of edge servers.

Thus, the LS proficiently provides a suitable edge server for the task attributes. On the other hand, if the attributes are not matched, they are handled by the global scheme. The entire scheme evaluates the attributes and predicts a suitable edge server while preserving the information. Further, they are used to create smart contracts, which are then updated in the cache table. The significance of the global scheme is described in the following.

Here, $n{t}^{\Omega }$ is the input for the global scheme, which is trained with historical information of tasks to be offloaded. The global scheme analyzes the attributes and provides the optimal edge server. The methods involved in the global scheme are described in the following.

Initially, past data are gathered from the publicly available resources and utilized to train the global scheme. The collected data $\left(d_e\right)$ are represented as follows:

$$d_e=\left\{d_1,d_2,\dots \dots d_E\right\}, e=1,2,\dots E$$

(7)

where $E$ represents the total amount of collected data holding historical information about the tasks to be offloaded.

After data collection, the significant features, such as CPU cores, CPU capacity, CPU usage, memory usage, and received throughput of the network, are extracted from $d_e$. Thus, the extracted features $\left({\chi }_r\right)$ are denoted as follows:

$${\chi }_r=〈{\chi }_1,{\chi }_2,\dots \dots {\chi }_R〉$$

(8)

where $r=1toR$ indicates the number of features extracted from $d_e$.

In this phase, ${\chi }_r$ undergoes the process of ES prediction based on the Probit Mish–GRU. A traditional GRU was selected due to its high flexibility with sequence data and because it requires less memory. However, it has a poor learning efficiency and convergence rate. To overcome these drawbacks, our proposed approach incorporates the Probit Mish activation function to optimize the convergence rate and improve the learning process. Figure 3 displays the block design of the proposed Probit Mish–Gated Recurrent Unit (PM-GRU).

The working process of the proposed Probit Mish–GRU is discussed in the following.

• Input: Here, the extracted features are considered as the input. ${\chi }_r$ indicates the inputs in the present time frame, and $hi{d}^{u-1}$ indicates the input from the previous hidden state. The inputs proceed to flow via the gates.
• Probit Mish Activation Function: The Probit Mish activation function (PMAF) is employed to improve the learning process and convergence rate. This is formulated as follows:

$${\infty }^{sig}\left({\chi }_r\right)=\sqrt{2}{e}^{-{\chi }_r}\ast \left(2{\varpi }_r-1\right)$$

(9)

where ${\infty }^{sig}$ denotes the Probit Mish activation function, and $e$ denotes the Euler number.

• Reset Gate: This gate should decide how much irrelevant information needs to be eradicated from the gates. The reset gate $\left({\Re }_{\otimes }\right)$ is determined as follows:

$${\Re }_{\otimes }={\infty }^{sig}\left({\varpi }_r\cdot hi{d}^{u-1},{\chi }_r\right)$$

(10)

where ${\varpi }_r$ represents the weight value of ${\Re }_{\otimes }$.

• Update Gate: The update gate decides how much crucial information needs to be allowed through the gates for further computation. The update gate $\left({\aleph }_{\oplus }\right)$ can be written in a mathematical form as follows:

$${\aleph }_{\oplus }\to {\infty }^{sig}\left({\varpi }_u\times hi{d}^{u-1},{\chi }_r\right)$$

(11)

where ${\varpi }_u$ signifies the weight parameter of ${\aleph }_{\oplus }$.

• Candidate Hidden State: This gate considers the reset gate and determines what relevant information needs to be stored. This is formulated as follows:

$$hi{d}^{u^{\prime }}={\wp }^{\tanh }\left({\varpi }_{ch}\ast \left[{\Re }_{\otimes }\cdot hi{d}^{u-1},{\chi }_r\right]\right)$$

(12)

$${\wp }^{\tanh }\left({\chi }_r\right)=\left[\frac{\sin \left({\chi }_r\right)}{\cos \left({\chi }_r\right)}\right]$$

(13)

where $hi{d}^u$ denotes the candidate hidden state, ${\varpi }_{ch}$ indicates the weight value of $hi{d}^{u^{\prime }}$, and ${\wp }^{\tanh }$ represents the tangent activation function.

• Hidden State: Finally, the hidden state makes a decision about what information needs to be passed through the network as an outcome, and it is defined as follows:

$$hi{d}^u={\wp }^{\tanh }\left(1-{\aleph }_{\oplus }\right)\times hi{d}^{u-1}+{\aleph }_{\oplus }\cdot hi{d}^{u^{\prime }}$$

(14)

where $hi{d}^u$ indicates the present hidden state. Finally, the global scheme $\left(g^{\Xi }\right)$ predicts the appropriate ES. This can be expressed as follows:

$$g^{\Xi }\left(hi{d}^u\right)=n{t}^{\Omega }\leftrightarrow {\zeta }_y$$

(15)

The pseudo-code for the Probit Mish–GRU is given in the Algorithm 1.

Algorithm 1 pseudo-code for the Probit Mish–GRU.

Input: Extracted features ${\chi }_r$

Output: Edge Server Prediction ${\zeta }_y$

Begin:

Initialize $hi{d}^{u-1}$, ${\wp }^{\tanh }$, $hi{d}^u$, ${\aleph }_{\oplus }$, ${\Re }_{\otimes }$, iteration $ir$, and maximum iteration $i{r}^{\max }$

Set $\left[ir=1\right]$

While $\left|ir\le i{r}^{\max }\right|$

For 1 to $R$ number of ${\chi }_r$

Perform reset gate

${\Re }_{\otimes }={\infty }^{sig}\left({\varpi }_r\cdot hi{d}^{u-1},{\chi }_r\right)$

Apply Probit Mish activation function

${\infty }^{sig}\left({\chi }_r\right)=\sqrt{2}{e}^{-{\chi }_r}\ast \left(2{\varpi }_r-1\right)$

Implement update gate ${\aleph }_{\oplus }\to {\infty }^{sig}\left({\varpi }_u\times hi{d}^{u-1},{\chi }_r\right)$

Calculate $hi{d}^{u^{\prime }}={\wp }^{\tanh }\left({\varpi }_{ch}\ast \left[{\Re }_{\otimes }\cdot hi{d}^{u-1},{\chi }_r\right]\right)$

Execute hidden state: $hi{d}^u={\wp }^{\tanh }\left(1-{\aleph }_{\oplus }\right)\times hi{d}^{u-1}+{\aleph }_{\oplus }\cdot hi{d}^{u^{\prime }}$

End For

End While

Return $g^{\Xi }\left(hi{d}^u\right)=n{t}^{\Omega }\leftrightarrow {\zeta }_y$

End

3.2. Secure Resource Allocation in MEC

In this stage, the outcomes from the local scheme and global scheme are preserved in a secure manner, and they are utilized to create a smart contract. This smart contract represents a protocol that executes programmed contracts or agreements when they are made. Then, the created smart contract is stored as the $c^{\infty }$ function. In the proposed framework, the Exponential Earth Mover’s distance-matrix-based K-Anonymity (EEMDM-KA) algorithm is used to perform privacy-preserving smart contract creation. The existing KA algorithm is employed due to its advantageous nature, which is characterized by its better security and greater efficiency in preserving the sensitive information of users if the system fails when addressing the nature of a region, which leads to poor consistency. Hence, the proposed approach uses the exponential EMD to consider the region’s nature, which improves the system’s performance. The steps involved in the EEMDM-KA algorithm are described in the following.

Sensitive data identification: In this phase, sensitive data $\left(w^{sen}\right)$ and non-sensitive $\left(q^{non}\right)$ data are identified by using the EEMDM-KA method. The EEMDM-KA algorithm is used to ensure the nature of the attributes, and it is represented as follows:

$$L^{\omega }={\displaystyle \sum \frac{\left|l^{\nabla }-g^{\Xi }\right|}{{\exp }^{j^{\prime }}}\to w^{sen},q^{non}}$$

(16)

where $L^w$ represents the distance parameter, and $j^{\prime }$ denotes the domain of the attributes.

Suppression: Here, several attributes of a user’s sensitive information are preserved by replacing the asterisk “*” to ensure reliable communication.

$${\varphi }^{spr}=\left|w^{sen}\right|\stackrel{replace}{\to }“\ast ”$$

(17)

Generalization: In the context of generalization, certain attributes of the sensitive information are represented in a particular range category. This is formulated as follows:

$${\phi }^{gen}=\left[w^{sen}\right]\to a^1<{\partial }^{att}\le b^1$$

(18)

where ${\varphi }^{spr}$ and ${\phi }^{gen}$ denote the processes of suppression and generalization. ${\partial }^{att}$ represents the attributes, and $a^1$ and $b^{\prime }$ indicate random values used to set the boundaries.

Smart Contract Creation: The preserved information is used to create a smart contract $\left({ƛ}_x^{sn}\right)$ and is further updated in the cache table; this value is defined as

$${ƛ}_x^{sn}=‖{\varphi }^{spr},{\phi }^{gen}‖+c^{\infty }\to \left({ƛ}_1^{sn},{ƛ}_2^{sn},\dots \dots {ƛ}_X^{sn}\right)$$

(19)

where $x=1,2,\dots X$ indicates the number of smart contracts created.

Subsequently, the traffic attributes, such as the number of requests, network bandwidth, network latency, uplink, and downlink, are extracted from the mobile users. Then, the extracted traffic attributes $\left({\lambda }_t\right)$ can be represented in mathematical form as

$${\lambda }_t\to \left({\lambda }_1,{\lambda }_2,\dots \dots {\lambda }_T\right)$$

(20)

where $t=1toT$ represents the value of ${\lambda }_t$. Here, ${\lambda }_t$ is inputted to perform load balancing by using the Hamming Curtis Dot K-Means (HCM-KM) technique. The process of load balancing helps to mitigate the delay and burden, thus resulting in high efficiency. In our proposed approach, the conventional K-Means (KM) technique is introduced to balance the user requests. The KM technique can be outperformed when it deals with large data resources. However, the Euclidean distance does not consider prior information, which limits the clustering outcomes. Therefore, the proposed approach employs the HCM algorithm to estimate the distance value, which leads to accurate clustering results. Then, the HCM-KM algorithm can be mathematically described as follows:

• Step 1: The number of clusters $\left({\kappa }_f\right)$ is specified; then, the centroid point $\left(c^{\psi }\right)$ is randomly initialized by using Equation (21):

$${\kappa }_f={\lambda }_t\stackrel{random}{\to }{c}^{\psi }$$

(21)

where $f=1toF$ denotes the total number of clusters to be formed.

• Step 2: During the next part of the derivation, the value of the distance between ${\lambda }_t$ and $c^{\psi }$ is estimated by using the HCM technique. The HCM technique is used to effectively calculate the distance among the data points. The data points are then assigned to the cluster that has the nearest distance value. This can be expressed as follows:

$${\eta }^{dis}={\displaystyle \sum _{t=1}^T\frac{‖{\lambda }_t-c^{\psi }‖}{\left|{\lambda }_t+c^{\psi }\right|}}$$

(22)

$${\kappa }_f\stackrel{assign}{\leftarrow }\min \left({\eta }^{dis}\right)\ast {\lambda }_t$$

(23)

• Step 3: Finally, the centroid point is recalculated, and the clustering process is repeated until it converges. Thus, the load-balanced requests $\left({\rho }_j^{\upsilon }\right)$ are given as follows:

$${\rho }_j^{\upsilon }=\left\{{\rho }_1^{\upsilon },{\rho }_2^{\upsilon },\dots \dots {\rho }_J^{\upsilon }\right\}, j=1,2,\dots J$$

(24)

where $J$ denotes the number of load-balanced requests.

In this stage, ${\rho }_j^{\upsilon }$ is assigned to the next edge server, and the cloud server’s virtual machines are then assigned to the best resources by that edge server. Here, resource allocation is performed by considering the task features and the cloud features. In our approach, the ELI-SOA technique is adopted to perform resource allocation. The traditional SOA was chosen due to its dynamic hunting strategy and its good ability to provide appropriate solutions. However, the position of the SOA is randomly replaced with the observation of the degradation of the search process. Therefore, the proposed approach incorporates the ELPI technique to update the SOA’s position, which enhances the optimization performance. The enhanced-searching-based SOA process is discussed in the following.

• Input features: Here, the task features and cloud features are chosen to perform RA. The task features, such as the number of dimensions, speed, cost, weight, and data size, are extracted. Then, cloud features, such as size, RAM, bandwidth, length, file size, storage, and VM, are also taken into account to allocate the optimal resources. The $G$-many extracted features $\left({\beta }_g\right)$ are initialized as follows:

$${\beta }_g=\left({\beta }_1,{\beta }_2,\dots \dots {\beta }_G\right), g=1toG$$

(25)

• Population Initialization: First, the population of the SOA is initialized in the search space. Here, $\left({\beta }_g\right)$ is assumed to be the serval. This is structured in the form of a matrix as follows:

$$\mathrm{N}={\left[\begin{array}{c}{\mathrm{N}}_1\\ ⋮\\ \begin{array}{c}{\mathrm{N}}_g\\ ⋮\end{array}\\ {\mathrm{N}}_G\end{array}\right]}_{G\times v}\stackrel{initialize}{\to }{\left[\begin{array}{ccccc}{\beta }_{1,1}& \dots & {\beta }_{1,k}& \dots & {\beta }_{1,v}\\ ⋮& \ddots & ⋮& \ddots & ⋮\\ {\beta }_{g,1}& \dots & {\beta }_{g,k}& \dots & {\beta }_{g,v}\\ ⋮& \ddots & ⋮& \ddots & ⋮\\ {\beta }_{G,1}& \dots & {\beta }_{G,k}& \dots & {\beta }_{G,v}\end{array}\right]}_{G\times v}$$

(26)

$${\beta }_{g,k}=l{w}_k+rn{d}_{g,k}\cdot \left[u{p}_k-l{w}_k\right]$$

(27)

where $\mathrm{N}$ denotes the population matrix of the serval’s location, $k=1tov$ represents the number of decision variables, $l{w}_k$ and $u{p}_k$ are, respectively, the values of the lower and upper bounds of the decision variable, and $rnd$ is a random parameter that holds a value between 0 and 1.

• Fitness Function: In this step, the fitness value of ${\beta }_{g,k}$ is derived by considering the minimum response time $\left(r{s}^{tm}\right)$. This can be estimated by using Equation (28):

$${\tau }_{fit}=\min \left(r{s}^{tm}\right){\left[\begin{array}{c}{{\tau }_{fit}}_1\\ ⋮\\ \begin{array}{c}{{\tau }_{fit}}_g\\ ⋮\end{array}\\ {{\tau }_{fit}}_G\end{array}\right]}_{G\times 1}={\left[\begin{array}{c}{\tau }_{fit}\left({\mathrm{N}}_1\right)\\ ⋮\\ \begin{array}{c}{\tau }_{fit}\left({\mathrm{N}}_g\right)\\ ⋮\end{array}\\ {\tau }_{fit}\left({\mathrm{N}}_G\right)\end{array}\right]}_{G\times 1}$$

(28)

$${\partial }^{can}\left(\mathrm{N}\right)\iff Best\left({\tau }_{fit}\right)$$

(29)

where ${\tau }_{fit}$ depicts the vector of the fitness function, and ${{\tau }_{fit}}_g$ demonstrates the estimated fitness function value (FFV) of the $g^{th}$ serval. Then, the best candidate solution $\left({\partial }^{can}\left(\mathrm{N}\right)\right)$ is defined according to the best FFV and $Best\left({\tau }_{fit}\right)$ between the estimated FFVs. Moreover, the location of the serval, FFV, and the best candidate can be updated in the following iteration of the SOA.

• Exploration Phase (prey selection and attack): In this phase, an exploration is carried out according to strategies followed by the serval to perform prey selection and attack. The exploration stage can be briefly described as follows:

$${\beta }_{g,k}^{sge1}\to {\beta }_{g,k}+rn{d}_{g,k}\ast \left|y_k^{pos}-o_{g,k}\cdot {\beta }_{g,k}\right|$$

(30)

$${\mathrm{N}}_g=\left\{\begin{array}{ll}if\left({\tau }_{fi{t}_g}^{sge1}\right)<{\tau }_{fi{t}_g},& {\beta }_{g,k}^{sge1}\\ else,& {\mathrm{N}}_g\end{array}\right.$$

(31)

where ${\beta }_{g,k}^{sge1}$ denotes the new position of the serval in the first stage, $y_k^{pos}$ expresses the prey’s location, $o_{g,k}$ indicates values that are randomly selected from the set {1, 2}, and ${\tau }_{fi{t}_g}^{sge1}$ denotes the updated FFV of the serval in the first stage.

• Exploitation Phase (chasing process): After the process of attacking the prey, the exploitation phase takes place based on the chasing strategy of the serval. Here, the entropic linear interpolation (ELI) method is used to update the serval’s position, and its value is defined as follows:

$${\beta }_{g,k}^{sge2}={\beta }_{g,k}+\frac{\Sigma \mathrm{P}\left(u{p}_k-l{w}_k\right)}{\Sigma \mathrm{P}\left(u{p}_k-l{w}_k\right)}\ast \log \left(u{p}_k-l{w}_k\right)$$

(32)

$${\mathrm{N}}_g=\left\{\begin{array}{ll}if\left({\tau }_{fi{t}_g}^{sge2}\right)<{\tau }_{fi{t}_g},& {\beta }_{g,k}^{sge2}\\ else,& {\mathrm{N}}_g\end{array}\right.$$

(33)

where ${\beta }_{g,k}^{sge2}$ symbolizes the updated position of the serval in the second stage, $\frac{\Sigma \mathrm{P}\left(u{p}_k-l{w}_k\right)}{\Sigma \mathrm{P}\left(u{p}_k-l{w}_k\right)}\ast \log \left(u{p}_k-l{w}_k\right)$ indicates the ELI factor, $\mathrm{P}$ denotes the probability factor, and ${\tau }_{fi{t}_g}^{sge2}$ represents the new FFV of the serval in the second stage.

The above derivations are repeated until the process converges. Finally, the optimal resources are allocated for each of the requests to the cloud server $\left({\mu }^{\vartheta }\right)$. The value of $Q$ for the allocated resources $\left({\gamma }_q\right)$ is denoted as follows:

$${\gamma }_q=⌊{\gamma }_1,{\gamma }_2,\dots \dots {\gamma }_Q⌋, q=1,2,\dots Q$$

(34)

$${\mu }^{\vartheta }\stackrel{allocate}{\leftarrow }\left({\mathrm{M}}_p^{\Theta }\right)\cdot {\gamma }_q$$

(35)

The pseudo-code for the ELI-SOA algorithm is presented in the Algorithm 2.

Algorithm 2 pseudo-code for the ELI-SOA algorithm.

Input: Extracted features (${\beta }_g$)

Output: Allocated resources $\left({\gamma }_q\right)$

Begin:

Initialize $\mathrm{N}$, ${\tau }_{fit}$, ${\beta }_{g,k}^{sge1}$, $\mathbf{}{\tau }_{fi{t}_g}^{sge2}$, iteration $w$, and maximum iteration $w_{mx}$

Set $\left(w=1\right)$

While $\left[w\le w_{mx}\right]$

For 1 to G of ${\beta }_g$

Initialize population in the search space

${\beta }_{g,k}=l{w}_k+rn{d}_{g,k}\cdot \left[u{p}_k-l{w}_k\right]$

Estimate the FFV ${\tau }_{fit}$

Implement Exploration Phase

${\beta }_{g,k}^{sge1}\to {\beta }_{g,k}+rn{d}_{g,k}\ast \left|y_k^{pos}-o_{g,k}\cdot {\beta }_{g,k}\right|$

If $\left(\left({\tau }_{fi{t}_g}^{sge1}\right)<{\tau }_{fi{t}_g}\right)$

Update the serval’s position to ${\beta }_{g,k}^{sge1}$

Else

No need to update the serval’s position

End If

Perform Exploitation phase

${\beta }_{g,k}^{sge2}={\beta }_{g,k}+\frac{rn{d}_{g,k}\ast \left(u{p}_k-l{w}_k\right)}{itr}$

Update FFV and serval position

End For

End While

Return $\left({\gamma }_q\right)$

End

Finally, the created smart contract ${ƛ}_x^{sn}$ is hashed by using the Keccak 512-bit Hash Generator and then stored in the blockchain. The hashing demonstrates the ability of KECCAK-512 to work against attacks. KECCAK-512 is referred to as a sponge function because its functioning is based on the construction of a sponge. Finally, the different phases and steps of the KECCAK-512 algorithm are detailed in the following.

Padding phase: First, the input ${ƛ}_x^{sn}$ is padded by using a padding rule $\left({\wp }^{\epsilon }\right)$ that gives the size of the output, which is multiplied by the bit rate $\left(b^{rt}\right)$, and this is defined as follows:

$${\wp }^{\epsilon }={ƛ}_x^{sn}\times b^{rt}$$

(36)

Block division: Second, the padded input is further divided into several blocks of length $b^{rt}$. Then, each block is processed using an XOR operation with a bit rate. Moreover, the output is concatenated with the capacity $\left(c^{\Pi }\right)$, which is written as follows:

$$a_{m^{\prime }}=\left(a_{1^{\prime }},a_{2^{\prime }},\dots \dots a_{M^{\prime }}\right), m^{\prime }=1to{M}^{\prime }$$

(37)

$$xo{r}^{\oplus }=a_{m^{\prime }}\left(b^{rt}\right)\ast XOR+c^{\Pi }$$

(38)

where $M^{\prime }$ denotes the number of divided blocks $\left(a_{m^{\prime }}\right)$, and $xo{r}^{\oplus }$ represents the outcome of the XOR operation.

Absorbing phase: Here, $xo{r}^{\oplus }$ is inputted into the absorbing phase. In this phase, the function $\left(f_{\eta }\right)$ is applied, and the following steps are repeated for the number of rounds.

• Theta Step:

• For $0\le v^{\prime }<5$ and $0\le w^{\prime }<{\hslash }^{wd}$,

  $$D^{\prime }\left(v^{\prime },w^{\prime }\right)=q^{\prime }\left|v^{\prime },0,w^{\prime }\right|\oplus q^{\prime }\left|v^{\prime },1,w^{\prime }\right|\oplus q^{\prime }\left|v^{\prime },2,w^{\prime }\right|\oplus q^{\prime }\left|v^{\prime },3,w^{\prime }\right|\oplus q^{\prime }\left|v^{\prime },4,w^{\prime }\right|$$

  (39)
• For $0\le v^{\prime }<5$ and $0\le w^{\prime }<{\hslash }^{wd}$,

  $$E^{\prime }\left(v^{\prime },w^{\prime }\right)=D^{\prime }\left\{\left(v^{\prime }-1\right)\mathrm{mod}5,w^{\prime }\right\}\oplus D^{\prime }\left\{\left(v^{\prime }+1\right)\mathrm{mod}5,\left(w^{\prime }-1\right)\mathrm{mod}{\hslash }^{wd}\right\}$$

  (40)
• For $0\le v^{\prime }<5$, $0\le x^{\prime }<5$, and $0\le w^{\prime }<{\hslash }^{wd}$,

  $$q{1}^{\prime }\left(v^{\prime },x^{\prime },w^{\prime }\right)=q^{\prime }\left(v^{\prime },x^{\prime },w^{\prime }\right)\oplus E^{\prime }\left(v^{\prime },w^{\prime }\right)$$

  (41)

Here, $D^{\prime }$ and $E^{\prime }$ represent the intermediate variables; $q^{\prime }$ and $q{1}^{\prime }$ denote the initial state and intermediate state, respectively; $v^{\prime }$, $x^{\prime }$, and $w^{\prime }$ denote the axis points; and ${ƛ}^{wd}$ represents the width.

• Rho Step:

• For $0\le w^{\prime }<{ƛ}^{wd}$,

  $$F^{arr}\left(0,0,w^{\prime }\right)=q{1}^{\prime }\left(0,0,w^{\prime }\right)$$

  (42)
• Assume that $\left(v^{\prime },x^{\prime }\right)=\left(1,0\right)$.
• For $\left(0\le k^{\prime }\le 23\right)$ and for $\left(0<w^{\prime }<{ƛ}^{wd}\right)$,

  $$F^{arr}\left(v^{\prime },x^{\prime },w^{\prime }\right)=q{1}^{\prime }\left[v^{\prime },x^{\prime },\left(w^{\prime }-\left|k^{\prime }+1\right|\cdot \left|k^{\prime }+2\right|/2\right)\mathrm{mod}{ƛ}^{wd}\right]$$

  (43)

  $$\left(v^{\prime },x^{\prime }\right)\to \left(x^{\prime },‖2v^{\prime }+3x^{\prime }‖\mathrm{mod}{ƛ}^{wd}\right)$$

  (44)

• Pi Step:

• For $\left(0\le v^{\prime }<5\right)$, $\left(0\le x^{\prime }<5\right)$, and $\left(0\le w^{\prime }<{ƛ}^{wd}\right)$,

  $$F^{arr}\left(v^{\prime },x^{\prime },w^{\prime }\right)=F^{arr}\left[\left(v^{\prime }+3x^{\prime }\right)\mathrm{mod}5,v^{\prime },w^{\prime }\right]$$

  (45)

• Chi Step:

• For $\left(0\le v^{\prime }<5\right)$, $\left(0\le x^{\prime }<5\right)$, and $0\le w^{\prime }<{ƛ}^{wd}$,

  $$\begin{array}{l}{q}^{\prime }\left(v^{\prime },x^{\prime },w^{\prime }\right)=F^{arr}\left(v^{\prime },x^{\prime },w^{\prime }\right)\oplus \{NOT\left[F^{arr}\left\{\left(v^{\prime }+1\right)\mathrm{mod}5,x^{\prime },w^{\prime }\right\}\right]\\ AND{F}^{arr}\left(\left|v^{\prime }+2\right|\mathrm{mod}5,x^{\prime },w^{\prime }\right)\}\end{array}$$

  (46)

• Iota Step:

• For $\left(0\le w^{\prime }<{ƛ}^{wd}\right)$,

$$q^{\prime }\left(0,0,0\right)=q^{\prime }\left(0,0,w^{\prime }\right)\oplus r^{{\rm Z}}\left(w^{\prime }\right)$$

(47)

where $F^{arr}$ represents the permutation state array, $K^{‘}$ denotes the bit position, and $r^z$ defines the round constant. Finally, the smart contract is hashed and defined as (ζ^hash). After the process, it is stored inside the blockchain, allowing the proposed approach to work proficiently and allocate resources in a secure manner by using the PM-GRU and ELI-SOA.

4. Results and Discussion

This section aims to evaluate our proposed framework’s performance using multiple quality indicators by contrasting it with various established methodologies. Furthermore, Python is used as the working platform for the implementation of the proposed system. Here, in order to demonstrate the superiority and prominence of the suggested work, a performance evaluation and comparative analysis are conducted.

4.1. Dataset Description

The proposed system was assessed by using two datasets, namely, the Heterogeneous Computing Scheduling Program (HCSP) [30] and Grid Workloads Archive (GWA-T-12-Bitbrains) [31] datasets, which are mentioned in the reference section. The HCSP dataset contains significant information about day-to-day tasks and contains data on users and servers. This dataset includes information about 1024 tasks and 32 machines. The collected data were used to acquire the task features and to train the system. Moreover, the GWA-T-12-Bitbrains dataset encompasses the performance metrics of the 1750 virtual machines that were used to train the global scheme. In this dataset, 41 scenarios were recorded by constantly monitoring the task offloading and resource allocation processes of the virtual machines. Each of the scenarios contained plentiful task information. Of all of the data, 80% of the information was employed to perform training, while the remaining data were used for testing.

4.2. Simulation and Analysis Parameters

To assess its performance, the proposed system was executed using the PYTHON 3.7 software, which emphasizes code readability and integrates systems efficiently. Moreover, the hardware, which included the Windows 10 operating system, an Intel i5/core i7 processor, a system with 64 bits with a 3.20 GHz central processing unit (CPU) speed, and random-access memory (RAM) with 4 GB of storage space, was utilized to simulate the proposed framework.

4.3. Performance Analysis

This section describes the validation of the performance of the proposed ELI-SOA to depict the model’s consistency [32]. We conducted a comparative analysis between our proposed ELI-SOA and the traditional SOA, COA, BOA, and WOA techniques in terms of various metrics [33]. The analysis of the proposed Entropic Linear Interpolation–Serval Optimization Algorithm is discussed in the following.

The performance of the proposed ELI-SOA and other existing models, such as SOA, COA, BOA, and WOA, is compared in Figure 4 based on their resource utilization (RU). The proposed ELI-SOA incorporated the entropic linear interpolation technique to update the serval’s position, which enhanced its optimization performance. The proposed ELI-SOA achieved an RU of 97.65% for 100 tasks, whereas the traditional methods achieved an RU of approximately 92.72% for 100 tasks. In Figure 3, it is absolutely clear that the proposed ELI-SOA had greater RU and more optimal results than those of the prevailing models.

Figure 5 compares the proposed ELI-SOA approach in terms of the EC, response time (RT), and waiting time (WT). The proposed ELI-SOA had better search capabilities and was more efficient in identifying the optimal resources within a limited period of time. As shown in Figure 5, the proposed ELI-SOA had EC, RT, and WT values of 6545 mj, 3867 ms, and 1769 ms, respectively, for 200 tasks. However, the existing methods achieved an average EC, RT, and WT of 14,951 mj, 5038 ms, and 3072 ms, respectively, for 200 tasks. Thus, the global results for the performance support that the proposed ELI-SOA achieved better EC and lower time complexity than those of the other traditional methods.

Here, the performance of the proposed PM-GRU scheme was validated to prove the reliability of the system. In this phase, we also compared the proposed PM-GRU with some traditional classifiers, such as the GRU, bidirectional long short-term memory (Bi-LSTM), long short-term memory (LSTM), and recurrent neural networks (RNNs).

We evaluated the performance of the proposed PM-GRU in comparison with other existing techniques, such as the GRU, Bi-LSTM, LSTM, and RNNs, according to their accuracy, precision, sensitivity, and specificity. The proposed PM-GRU incorporated the Probit Mish activation function, which helped improve its learning process and convergence rate. Figure 6 demonstrates that the proposed PM-GRU obtained better accuracy, precision, sensitivity, and specificity of 99.84%, 99.87%, 99.55%, and 99.16%, respectively. However, the existing methodologies achieved an average accuracy, precision, sensitivity, and specificity of 94.40%, 94.48%, 93.83%, and 94.67%, respectively. The experimental outcomes allowed the conclusion that the proposed PM-GRU outperformed the other traditional methods.

The next figure illustrates an evaluation of the true positive rate (TPR) and true negative rate (TNR) of the proposed PM-GRU and other existing methods, such as GRU, BiLSTM, LSTM, and RNNs. The proposed PM-GRU effectively predicted the edge server with fewer false positives due to the improved PMAF. Figure 7 shows that the proposed PM-GRU achieved a TPR and TNR of 99.52% and 99.48%, respectively. However, the other existing methodologies had poor performances in terms of the TPR and TNR values, with 94% and 94.68%, respectively. So, the proposed architecture had an impressive performance and was proven to be a less error-prone model [34,35].

The proposed HCD-KM was finally compared with some traditional clustering algorithms, such as KM, Partition around Medoids (PAM), Balanced Iterative Reducing and Clustering Using Hierarchies (BIRCH), and Fuzzy C-Means (FCM), to demonstrate the importance of the suggested model. This is explained in more detail in the Figure 8.

Table 2. Makespan analysis.

Techniques/Makespan (ms)	Number of Tasks
	100	200	300	400	500
Proposed HCD-KM	2145	4123	6124	8245	10,227
KM	2867	4878	6713	8852	10,869
PAM	3446	5369	7423	9324	11,337
BIRCH	3845	5748	7812	9748	11,875
FCM	4428	6245	8354	10,278	12,434

compares the makespan of the suggested system with that of a few current models, including KM, PAM, BIRCH, and FCM. The proposed HCD-KM helped cluster the data based on priority, which definitely improved the computational efficiency. The proposed HCD-KM had a makespan of 2145 ms for 100 tasks. However, the traditional algorithms had a makespan of approximately 3646 ms for 100 tasks. By showing a comparison with the current methods, Table 2 clearly shows the high computational efficiency of the proposed system.

Then, the performance of the proposed EEMDM-KA and some existing methodologies, such as KA, L-Diversity, T-Closeness, and Randomization, was validated.

The proposed model was analyzed in terms of anonymization time (AT), as displayed in Figure 9. Here, the proposed EEMDM-KA was compared with traditional techniques, such as KA, L-Diversity, T-Closeness, and Randomization. The proposed EEMDM-KA was able to recognize sensitive information, and it was more effective in preserving user information. As shown in Figure 9, the proposed EMD based on k-Ano EEMDM-KA achieved an AT of 754 ms, whereas the other existing techniques had an AT of approximately 2011 ms. Thus, the evaluation outcomes showed that the proposed exponential EEMDM-KA had lower time complexity and more dominant performance than that of the existing techniques.

The following table compares the proposed EEMDM-KA with other techniques, such as KA, L-Diversity, T-Closeness, and Randomization, according to the privacy preservation rate (PPR). The proposed EEMDM-KA effectively preserved the users’ sensitive data by replacing them with an asterisk and defining them within a certain range category. As shown in

Table 3. Comparative analysis of the proposed model.

Methods	Privacy Preservation Rate (%)
Proposed: exponential EMD matrix based on k-Anonymization	97
KA	94
L-Diversity	90
T-Closeness	87
Randomization	83

, the proposed EEMDM-KA achieved a PPR of 97%, whereas the other traditional techniques had an average PPR of 88.5%. Thus, the proposed EEMDM-KA is more promising in privacy-preserving smart contract creation than the other conventional methods.

An analysis was carried out on the proposed system by comparing it with various related works using several quality metrics.

The results of the comparative analysis of the proposed framework are shown in

Table 4. Comparative analysis of the proposed work.

Works	Techniques	Accuracy (%)	EC (mj)	RT (ms)	RU (%)
Proposed work	PM-GRU and ELI-SOA	99.84	3965	1865	97.65
Wang et al. [36]	DCA	-	4500	-	-
Mahenge et al. [37]	PSO + GWO	-	4339	1982	81
Elgendy et al. [38]	MU-MT-COA	-	5089	2833	78
Nguyen et al. [39]	DRL	-	-	2600	-
Qi [40]	DNN	89	6873	-	-

. The data reveal the beneficial characteristics of the model. The proposed approach was compared with recent methods in the literature, such as those proposed by Wang et al. [36], Mahenge et al. [37], Elgendy et al. [38], Nguyen et al. [39], and Qi [40]. The proposed model used the PM-GRU and ELI-SOA to achieve task offloading and resource allocation in mobile applications. The proposed PMAF helped enhance the classification performance. Likewise, the proposed ELI-based position updating process led to optimal results. Similarly, the traditional systems utilized techniques such as the Difference of Convex Approximation (DCA), Particle Swarm Optimization (PSO), Gray Wolf Optimization (GWO), Multi-User Multi-Task Computing Offloading Algorithm (MU-MT-COA), DRL, and DNN to perform task offloading and resource allocation. As shown in Table 4, the proposed model had an accuracy, EC, RT, and RU of 99.84%, 3965 mj, 1865 ms, and 97.65%, respectively. However, the other existing methods had limited performance and high EC. Thus, the comparative analysis justified that the performance of the proposed framework was remarkably superior to that of the other traditional approaches.

This study assessed the proposed model in terms of various metrics while using two different datasets. Here, the dataset, namely, the Heterogeneous Computing Scheduling Program (HCSP), was used to obtain the task characteristics, which were shared between the user and server. Further, the Grid Workloads Archive (GWA-T-12-Bitbrains) dataset was used to train on the tasks obtained in the global scheme.

5. Conclusions and Future Work

5.1. Conclusions

In this study, a novel and promising approach to secure task offloading and resource allocation in mobile applications based on the Probit Mish–Gated Recurrent Unit (PM-GRU) and Entropic Linear Interpolation–Serval Optimization Algorithm is proposed. The proposed PM-GRU effectively enhances learning efficiency and classification performance. An effective Entropic Linear Interpolation–Serval Optimization Algorithm was proposed to allocate the optimal resources regarding the attributes of a task. This study developed the offloading of tasks with attributes by gathering them from mobile users and submitting them to local models, where the attributes were compared with a cache table using the Sørensen–Dice index (SDI) in two ways (when attributes matched and when they did not). Then, the information was preserved and updated in the cache table in the local scheme. Finally, an evaluation of this proposal was performed using multiple quality indicators by contrasting it with various established methodologies using smart contracts that were hashed based on the Keccak 512-bit Hash Generator and stored in the blockchain. This research also used the Heterogeneous Computing Scheduling Problem (HCSP) and Grid Workloads Archive (GWA-T-12-Bitbrains) datasets in a simulation of an implementation to contribute to the provision of datasets containing significant information about day-to-day tasks completed between the user and server. Up to 1024 tasks and 32 machines were incorporated into the program; these are essential data, especially for global server implementation.

The evaluation outcomes demonstrated that the proposed PM-GRU achieved a better accuracy, sensitivity, and specificity of 99.84%, 99.55%, and 99.16%, respectively, which proved the model’s supremacy. Likewise, the proposed Entropic Linear Interpolation–Serval Optimization Algorithm had an RU and EC of 97.65% and 3965 mj, respectively, which showed the consistency of the system. Then, the latency achieved by the proposed HCD-KM was 1875 ms, which depicted its better performance and lower time complexity. Moreover, the proposed EEMDM-KA obtained a high PPR of 97%, which justified the trustworthiness of the model. Ultimately, the comparative analysis and experimental findings demonstrated the superior efficacy and high significance of the suggested framework in relation to other widely used models. This approach is an essential contribution to the realization of these sorts of implementation scenarios, especially for scholars and institutions.

This work only concentrated on secure and reliable task offloading and resource allocation. However, it did not concentrate on edge server mobility, the mobility of terminal devices, or even the process of monitoring assigned tasks in a VM when a machine shuts down or some other issues arise.

Advantages and Drawbacks

As the cache information was updated in the cache table in the local computing system, the edge server was determined to be more appropriate for the task attributes. Further, the analysis of the task attributes aided in effective task offloading and load balancing. The preservation of smart contract privacy and load balancing prior to resource allocation provided optimal RA. Eventually, the created smart contracts were hashed using KECCAK-512 for storage in the blockchain to maximize the system’s efficiency. Even though secure task offloading and RA were the focus of this study, the tasks assigned in a VM if a machine disconnects or some other issues arise were not considered.

5.2. Future Work

In order to overcome the aforementioned limitation in the proposed framework, the assigned tasks in VMs will be monitored in future work. Hence, any machines that shut down and some other issues that arise in cloud environments can be identified to achieve enhanced task offloading and RA.