Deep Learning-Based Privacy-Preserving Data Transmission Scheme for Clustered IIoT Environment

The Industrial Internet of Things (IIoT) has received significant attention from several leading industries like agriculture, mining, transport, energy, and healthcare. IIoT acts as a vital part of Industry 4.0 that mainly employs machine learning (ML) to investigate the interconnection and massive quantity of the IIoT data. As the data are generally saved at the cloud server, security and privacy of the collected data from numerous distributed and heterogeneous devices remain a challenging issue. This article develops a novel multi-agent system (MAS) with deep learning-based privacy preserving data transmission (BDL-PPDT) scheme for clustered IIoT environment. The goal of the BDL-PPDT technique is to accomplish secure data transmission in clustered IIoT environment. The BDL-PPDT technique involves a two-stage process. Initially, an enhanced moth swarm algorithm-based clustering (EMSA-C) technique is derived to choose a proper set of clusters in the IIoT system and construct clusters. Besides, multi-agent system is used to enable secure inter-cluster communication. Moreover, multi-head attention with bidirectional long short-term memory (MHA-BLSTM) model is applied for intrusion detection process. Furthermore, the hyperparameter tuning process of the MHA-BLSTM model can be carried out by the stochastic gradient descent with momentum (SGDM) model to improve the detection rate. For examining the promising performance of the BDL-PPDT technique, an extensive comparison study takes place and the results are assessed under varying measures. A significant amount of capital is required. It goes without saying that one of the most obvious industrial IoT concerns is the high cost of adoption. Secure data storage and management connectivity failures are common among IoT devices due to the massive amount of data they create. The simulation results demonstrate the enhanced outcomes of the BDL-PPDT technique over the recent methods. Despite the fact that the offered BDL-PPDT technique has an accuracy of just 98.15 percent, it produces the best feasible outcome. Because of the data analysis conducted as detailed above, it was determined that the BDL-PPDT technique outperformed the other current techniques on a range of different criteria and was thus recommended.


Introduction
Industrial Internet of ings (IIoT) employs actuators and sensors with communication and computation capabilities to transform the way the information is exchanged, analyzed, transformed, and collected into decisions [1]. is pervasive capability results in advanced Industry 4.0 (called Industrial Internet) application for enhanced efficiency and productivity in large industries like healthcare, energy, agriculture, mining, and transportation. e innovative Industry 4.0 features, namely, ML-based quality control and predictive maintenance and run-time reasoning, need to be simplified by distributed data acquisition [2]. In IIoT-based systems like open banking and smart healthcare, ML and data methods trained with the local boundary should be interacted with the branches or users to make organizationwider knowledge [3]. Often, vendors desired to limit their internal insight on product improvements and development with the organizational boundary for increasing business values against their contender. Furthermore, industries like open banking and smart healthcare are vastly convoluted with human-specific sensitive information [4]. ML model is trained on sensitive information that could expose confidential or private data to the attackers [5]. erefore, trustworthiness and privacy are the key elements of ML in IIoT system. e Internet of ings is one of the primary drivers of the Industry 4.0 movement, since it enables greater automation, data collection, and analytics, as well as workflow and process optimization. e intelligence enabled by the Internet of ings enables devices to work cooperatively to produce outputs on an assembly line. MAS as a popular technique to offer trusts with distributed and decentralized settings might be employed in lots of potential applications, namely, supply chain management, IIoT, and healthcare [6]. e Internet of ings is a major driver of the Industry 4.0 movement since it enables increased automation, data collection, and analytics, as well as workflow and process optimization.
e Internet of ings' intelligence enables devices to operate together on an assembly line to produce outputs. A multi-agent system (MAS or "self-organized system") is a computerized system that is built of numerous intelligent agents that communicate with one another. Multi-agent systems are capable of resolving problems that a solo agent or a monolithic system would find difficult or impossible to solve. Methodical, functional, or procedural techniques, algorithmic search, or reinforcement learning can all be considered forms of intelligence. Among them, it is the mainstream application field, where blockchain is regarded as enabling technology for various applications. IIoT setup is a well-developed and also comprehensive deployment that can increase multiple challenges involving ensuring confidentiality, improving data accountability, availability, integrity, and availability (CIA). Blockchain can address this requirement and acts as a significant role by providing secure and verifiable solutions to store and share data [7]. IIoT application has requirements of similar kinds to guarantee trust and data integrity between several shareholders related to dissimilar parts of the logistic chain (e.g., storage, acquiring raw material, processing to customer, transportation, and industrial deployment). Also, in this application, requirements such as monitoring and maintaining history of each procedure are vital. e conventional security method has a number of constraints and does not fit for intelligent grid systems; for example, secured end-to-end encryption method could produce higher false alarm rate and interrupt analytical approach [8].
ere is a considerable range of potential smart grid risks, like passive and active attacks. Another way of the attack is the smart grids, namely, sniffing the information from the CPS through open source data, and in active attack, the hackers can change the information through data poisoning attack or inference attack [9]. In data poisoning attacks, attackers attempt to change the standard information.
is article develops a novel MAS with deep learningbased privacy-preserving data transmission (BDL-PPDT) scheme for clustered IIoT environment.
is research proposes a unique multi-agent system (MAS) method for clustered IIoT environments using deep learning-based privacy preserving data transmission (BDL-PPDT). e BDL-PPDT technique's objective is to achieve secure data transfer in a clustered IIoT environment. e BDL-PPDT technique involves the design of an enhanced moth swarm algorithm-based clustering (EMSA-C) technique for cluster head (CH) selection. In addition, blockchain technology (BCT) is applied for accomplishing secure inter-cluster communication. Furthermore, a new multi-head attention with bidirectional long short-term memory (MHA-BLSTM) model is used to find intrusions. To increase the detection rate, the stochastic gradient descent with momentum (SGDM) model can be used to tune the MHA-BLSTM model's hyperparameters. Finally, the stochastic gradient descent with momentum (SGDM)-based hyperparameter tuning process takes place. To inspect the significant performance of the BDL-PPDT technique, a wide-ranging comparative analysis is made and the results are inspected in terms of different measures.

Related Works
Sodhro et al. [10] proposed sustainable, secure, efficient, and reliable blockchain-driven methods. e presented method handles key arbitrarily by presenting the chain of blocks with a smaller amount of cores, less power drain, computation bit, and transmission. Next is an analytic hierarchy process (AHP) based smart decision-making method for the blockchaindriven that is more secured, reliable, sustainable, interoperable, and concurrent IIoT. Rahman et al. [11] presented a blockchain-based architecture to provision a verifiable query and privacy-preserving facilities to end-user in IIoT system. e architecture employs blockchain to save broad information as off-chain data and to save IoT information as on-chain data and provision search service to the user by performing a query in off-chain and on-chain data as well as generate an effective result.
Zhang et al. [12] developed a medical data privacy protection architecture-based blockchain (MPBC). In this method, they secure confidentiality by including different privacy noises to federated learning. Additionally, the increasing amount of healthcare data can make blockchain storage challenges. us, a storage mode is presented for reducing the storage burden of blockchain.
e new information is locally stored and the hash values are estimated by IPFS and are saved in blockchain. Deebak and Al-Turjman [13] introduced a privacy-preserving smart contract with blockchain and artificial intelligence (PPSC-BCAI) architecture which facilitates system activities, human interaction, security risks, fraudulent claims, and service alerts. In order to examine the data sharing and transaction, an XGBoost is employed. 2 Computational Intelligence and Neuroscience Weng et al. [14] proposed a secure, fair, and distributed DL architecture called DeepChain to resolve this problem. DeepChain provides a value-driven incentive method based on blockchain for forcing the participant to perform properly. In the meantime, DeepChain ensures data privacy for all the participants and provides auditability for the entire training procedure. Arachchige et al. [15] presented an architecture called PriModChain which forces trustworthiness and privacy on IIoT information by amalgamating federated ML, differential privacy, smart contracts, and Ethereum blockchain. e possibility of PriModChain based on resilience, privacy, security, reliability, and safety is estimated by the simulation technologically advanced in Python with socket programming on a multipurpose computer.
Kumar and Tripathi [16] designed a deep blockchainbased trustworthy privacy-preserving secured framework (DBTP2SF) for IIoT. is architecture contains two-phase privacy-preservation model, anomaly detection module, and trust management module. In the two-phase privacy model, a BC-enabled improved proof of work method is concurrently employed with AE, to convert cyber physical information to a novel form which avoids poisoning and inference attacks.

The Proposed Model
In this study, an effective BDL-PPDT technique has been developed to accomplish secure data transmission in clustered IIoTenvironment. e BDL-PPDT technique has presented a new EMSA-C technique to choose a proper set of clusters in the IIoT system and construct clusters. Next, the MHA-BLSTM with SGDM model is utilized for intrusion detection and the hyperparameter tuning process is made by the SGDM model resulting in improved detection performance. Figure 1 illustrates the overall process of BDL-PPDT manner.

Process Involved in EMSA-C Technique.
e nocturnal behaviors of moth are the motivation for the MSA [17]. In the model, the exploration and exploitation tradeoff considers a divider of candidate solutions generating the population: (i) Onlooker (to exploit the optimal region discovered by the prospector). (ii) Pathfinder (to explore novel region of the searching space). (iii) Prospector (to exploit the novel regions attained by the pathfinder).
With other meta-heuristics models, this one begins with population initialization: whereas u and l represent maximum and minimum bounds of the searching space, χ i denotes the candidate solution, n indicates the population size, d signifies the dimensionality of problems, and rand means an arbitrary number drawn from a uniform distribution. For generating pathfinder crossover, it is essential to estimate the variation coefficient and dispersal degree at iteration t: whereas n p represents the amount of pathfinders  Computational Intelligence and Neuroscience In the MSA, the crossover point represents minimum dispersal value, as follows: Form this, n c ∈ c p crossover point is employed for creating a novel sub-trial pathfinder vector For each of the independent variables [18], the variables L pl and L p2 are calculated using the Lévy stable distribution. ere should only be one set of indexes r selected from the pathfinder solution, in which L pl and L p2 represent independent variable calculated from the Lévy α-stable distribution [18]. e set of indexes r should be only chosen from the pathfinder solution, and position is upgraded by the mutated variable extracted from the sub-trail vector as follows: Lastly, MSA employs a selection approach among the original and trial pathfinders as follows: e possibility of choosing the next pathfinder is determined by at employs the luminescence intensity estimated as follows: From the pathfinder, n f individual is chosen as prospector; this value is modified dynamically as follows: where T represents the maximal iteration number. e MSA enables the moth to move in a spiral manner over a pathfinder using equation (12): Let θ ∈ [r, 1] be an arbitrary value employed for giving the spiral formation to the prospector path, while e onlooker is the moth with the minimum luminescent intensity moving toward the shiniest source of light; in MSA, the onlooker is employed for intensifying the exploitation process. Further, the onlooker is separated into Gaussian walk and associative learning using immediate memory. Initially, the onlooker in the real iteration is attained as follows: whereas ε 2 and ε 3 represent uniformly distributed random value, best g denotes the optimal candidate solution, n o � round(n u /2) indicates the amount of onlookers performing a Gaussian movement, n u shows the amount of onlookers, and ε 1 means an arbitrary value estimated by e behavior of the moth considered short-term memory and associative learning is upgraded as follows: with n m � n u − n o being the amount of onlookers performing short-term memory and associative learning; 1 − (g/G) indicates a cognitive factor, 2g/ G represent a social factor, besr p indicates the optimal light source from the pathfinder, and G ∼ N( To improve the performance of the MSA, the EMSA is derived by the use of OBL concept. e efficient implementation of OBL contributes approximation of the opposite and current populations in the same generation for identifying optimum candidate solutions of a given problem. Object-based learning (OBL) is a student-centered learning approach that uses objects to facilitate deep learning. Objects may take many forms, small or large, but the method typically involves students handling or working at close quarters with and interrogating physical artefacts. e OBL models have been efficiently used in different meta-heuristics employed for improving convergence speed. e models of the opposite amount should be described in OBL.
Consider N ∈ N[x, y] to denote real numbers. e opposite numbers N0 are given by In d-dimension searching region, the depiction may be extended as follows: Consider an IoT network of n sensor deployed arbitrarily. In order to be CH selective, the projected SSA executes squirrel population that was utilized by generating suitable clusters and maintaining the lower power employment of systems. Consider X � (X 1 , X 2 , . . . , X n ) stands for the population vector of IoT with n sensors, where X i (j) ∈ 0, 1 { }. e CH and normal nodes were signified as one and zero. e fundamental population of NP solution has inspired arbitrarily by representing 0 s as well as 1 s and representing as follows: where p opt stands for the recommended percentage of CHs and r and refers to uniform arbitrary values in zero and one. An arbitrarily located sensor node has been decided as K clusters: C 1 , C 2 , . . . , C K . e CH selective has responsible to decrease the cost of FF. erefore, FF to CH selective was showcased as follows: with 2 i�1 w i � 1. e maximum stability period is given by decreasing the Standard Deviation (SD) of the RE of node which is a significant issue. erefore, SD (σ RE ) is applicable to measure the control of uniformly distributed load in sensor node and illustrated as follows: where μ RE � (1/n) n i�1 E(node i ), E(node i ) stands for the RE of i th node, and n depicts the node count. A final objective was dependent upon clustering quality in that function of cluster isolation and cohesion has been implemented. Once the proportion of cohesion for separating was minimal, afterward optimum clustering was executed. It is accomplished by utilizing FF ratio of overall Euclidean distance of CH to CM and restricted Euclidean distance of 2 varying CHs.

Secure Inter-Cluster Communication via Blockchain.
Generally, blockchain is assumed as a collection of blocks; also, a single block comprises of hash value of the existing block, information about the transaction (Ethereum, bitcoin), timestamp, and previous block. Furthermore, blockchain is determined as common and distributed digital ledger utilized to save the transaction data under different points. erefore, when an attacker tries to derive information, it is not possible as every block has cryptographic value of the earlier block [19]. Now, each transaction is attained under the application of cryptographic hash values, viz. confirmed by all the miners. It consists of blocks of each transaction and captures same value of the comprehensive ledger. Figure 2 illustrates the framework of blockchain. e blockchain offers the facility to share detailed ledgers in protective, confidential, and shared manner. Decentralized storage is the other source in blockchain, and the massive number of information data is linked and stored from existing blocks to earlier blocks through smart contract code. LitecoinDB, Swarm, SiacoinDB, MoneroDB, BigchainDB, Interplanetary File System (IPFS), and various factors were employed for decentralized dataset.

Intrusion Detection Process.
During the intrusion detection process, the MHA-BLSTM with SGDM model is utilized. LSTM is a variant of RNN that could resolve gradient disappearance problems by presenting memory cell state, input gate i, output gate 0, and forget gate f. LSTM could enhance the memory model of NN for receiving training and input data that is appropriate to model time series data, such as text, owing to the design characteristics. BiLSTM is an integration of backward and forward LSTM. e greatest benefit of the model is that the sequence context data are taken fully into account. An LSTM unit contains controlling gate, along with IG i t , a forget gate f t , outcome gate 0 t , and a memory cell state c t , that affects the unit capacity to update and store data. e IG outcome value lies between 0-1 according to the input h t− 1 and w t . Once the outcome is 1, it implies that the cell state data are retained completely, and once the outcome is 0, it is abandoned completely. en, the IG determines which value needs updating, and the tanh layer creates a novel candidate value vector c t that is added to the cell state. Next, both are integrated for updating the cell state c t ; lastly, the outcome layer decides the outcome value based on the cell state. Among other, W f , U f , b f , W i , U i , b j , W c , U c , b c , and W 0 , U 0 , b 0 represent the internal training parameter in the LSTM, σ(·) indicates sigmoid activation function, and ⊙ implies dot multiplication.
Computational Intelligence and Neuroscience e abovementioned method is the computation method of LSTM. As previously mentioned, BiLSTM comprises backward and forward LSTM. LSTM ������→ in BiLSTM reads the input from w 1 to e n for generating h t → , and other LSTM ← reads the input from e n to w 1 for generating h t− 1 Lastly, the output [h 1 , . . . h i , . . . h m , l 1 , . . . l j . . . l n ] of the entire sentence is attained, whereas h i and l j are exploited to signify the output of emoticons and words, correspondingly. Furthermore, set each intermediate layer in BiLSTM for returning the comprehensive output sequence, thus ensuring that the output of all the hidden layers retains the longer-distance data as possible.
Attention mechanism is used to improve the effects of RNN-based model, and also it consists of dot-product attention and additive attention [20]. e calculation of attention is separated into 3 stages. Initially, utilize F attention function to score key and query to get si; next, utilize softmax function to standardize the scoring results si, for obtaining the weight ai. Lastly, estimate attention that is the weighted average of each value and weight ai. Multi-head attention mechanism has enhanced the classical attention method; thus, all the heads could extract the features of key and query in distinct subsets. More precisely, this feature comes from Q and K that is the projection of key and query in the subspaces. Note that in the multi-head attention model, the attention functions can be the scaled dot-product function that is similar to the classical attention mechanism, excepting the regulating scaling factors [21][22][23][24][25][26][27]. In this work, h should be debugged continuously for determining the appropriate values. Lastly, the result, i.e., returned in every head, is linearly converted and concatenated to attain multihead attention. Eventually, transmit the vector from the preceding layer to the densely connected layer. ey utilize ReLU function as the activation function for completing the nonlinear transformation. Finally, execute the softmax function on the output of the preceding layer and attain intrusion detection output.
For optimally adjusting the hyperparameter of the MHA-BLSTM model, the SGDM is applied. SGDM is a first-order momentum depending on SGD. e 1st-order momentum represents the exponential moving of the gradient direction at all the moments, nearly equivalent to sum of the gradient vector at the current T j moment. And, the T j is denoted by In another word, the descendant direction at t time is described using the descending direction accumulated before as well as gradient direction of the existing point. e empirical value of β 1 is 0.9, which implies the direction of decline is particularly the before accumulated direction of decline.

Results and Discussion
In this section, a detailed experimental validation of the BDL-PPDT technique takes place under varying numbers of IoT sensor nodes and rounds. e results are examined in varying aspects. An extensive throughput analysis of the BDL-PPDT technique with other methods is given in Table 1 and Figure 3.   A brief comparative NLT analysis of the BDL-PPDT technique is illustrated in Table 2  A brief comparative number of alive sensor node (NASN) analysis of the BDL-PPDT technique takes place in Table 3 and Figure 7. From the results, it can be noticed that   Here, the intrusion detection performance analysis of the BDL-PPDT technique is provided in Table 5 and Figure 9 [21,22]. e results are tested using the KDDCup99 dataset

Conclusion
In this study, an effective BDL-PPDT technique has been developed to accomplish secure data transmission in clustered IIoT environment. e BDL-PPDT technique has presented a new EMSA-C technique to choose a proper set of clusters in the IIoT system and construct clusters. Next, the MHA-BLSTM with SGDM model is utilized for intrusion detection and the hyperparameter tuning process is made by the SGDM model resulting in improved detection performance. To inspect the significant performance of the BDL-PPDT technique, a wide-ranging comparative analysis is made and the results are inspected in terms of different measures. e experimental outcome pointed out the improved performance of the BDL-PPDT technique over the recent methods in terms of different measures. In the future,    hyperparameter tuning process of the MHA-BLSTM model can be done by the meta-heuristic algorithms to improve the overall performance. Even though the BDL-PPDT method has an accuracy rate of just 98.15 percent, it still gives the best possible result. Because of the data analysis above, it was found that the BDL-PPDT technique outperformed the other current techniques on a wide range of different factors, and so it was recommended that people use it. Meta-heuristic methods will be utilized in the future to modify the hyperparameters of the MHA-BLSTM model, resulting in an overall improvement in overall performance.

Data Availability
e article contains all of the data.

Conflicts of Interest
e authors state that they do not have any conflicts of interest.