1 Introduction

The Internet of Things (IoT) achieved rapid growth in modern society because of its powerful communication technologies. According to the reports for 2021 over 28 billion smart devices are connected across the world (Shafique et al 2020). Here more than 14.6 billion dives are connected to machine-to-machine. IoT platform provides the connection to the internet and devices as well as collects the surrounding environmental data through the devices. Some of the edge devices in IoT cannot analyze these data. So the edge devices are connected to the cloud for data analysis, but this creates network traffic along with data transfer latencies (Ghosh and Grolinger 2020). Among many of the applications of IoT, smart homes are the killer application because of their consumer accessibility and marketability (Shin et al 2018). The data gathered through these smart devices are used in various industries and businesses for decision-making purposes (Saqlain et al 2019). The increase in the population in cities leads to smart homes (Wang et al 2020). Nowadays cities are more concerned with adopting smart home technology to increase their quality of life. But the construction of an IoT architecture is a difficult task. This IoT system has been used in various environments in recent years. The floating city concept and smart city concepts are among them (Kirimtat et al 2020). The IoT architecture is made up of three layers. They are the perception layer, the Network layer, and the application layer. The perception layer is the lowest layer which contains all IoT devices such as sensors, actuators, and Consumer Electronics (CE) devices. The network layer is responsible for transmitting data from these devices to the application layer (Gu et al 2019). This application layer receives the data from the network layer and processes it. In this architecture, the web servers are present in the application layer to realize IoT applications such as smart cities, smart homes, schools, etc. The web servers are used to connect the applications and the IoT infrastructure to access the data (Purohit and Kumar 2019).

Semantic Web Service Ontology (SWSO) is used to define semantic web service which uses the Semantic Web service Language (SWSL). The non-semantic web servers are defined by Web Service Description Language (WSDL) which is an XML-based language (Tian et al 2018). Some of the features of WSDL are messaging, documentation, binding, services, types, and port types (Agarwal et al 2020). Because of these features, WSDL is popular in the software development industry. However, security, privacy, latency during communication, scalability, and centralization are the basic issues faced by the end users. Because these devices don’t have any managers to supervise and regulate their devices (Singh et al 2022). These devices can interact with each other through the gateways with the use of various communication protocols which may lead to the eavesdropping of an unauthorized person (Sardar and Anees 2021). Cybercriminals may access other people’s private data by obtaining information from some smart home appliances (Arif et al 2020). But the existing methods utilize the single classification method that provides only limited accuracy. Hence there is an opportunity to develop the accuracy by using the self-organized deep belief network (S-DBN) technique. The web services’ information is fetched from their WSDL file. The semantics of web servers are determined by messages, operations, mining service descriptions, and schemas that belong to the WSDL file. This research aims to provide a Starling Murmuration Optimized structural self-organized deep belief network (SMO-SSODBN) model for web services classification. The objective of this SMO-SSODBN technique is to enhance the classification’s accuracy.

To implement the idea of the smart city, numerous service providers have started to offer smart services. Effective web service selection is necessary to offer effective smart applications. Based on the quality of service (QoS) constraints, the web services are offered priorities and the service with the higher priority is often selected. Furthermore, choosing a web service based on QoS becomes time-consuming and challenging when there are numerous identical online services. Real-time provision of value-added services is defeated in this way. As a result, web service classification is utilized to choose web services quickly.

For web service classification, different Machine Learning (ML) methods are applied. The functional elements of web services are described utilizing the Ontology Web Language for Services (OWL-S) and the Web Services Description Language (WSDL), which are mostly utilized during service discovery. The QoS data helps choose web services. By separating these parameters utilizing the Feature-Selection (FS) approach, it is possible to determine the group of pertinent QoS parameters. Here, the SMO-SSODBN-based FS method addresses the issue of information overfitting and helps shorten training times while also simplifying the model. In addition, only one S-DBN classifier is utilized to construct the classification model. However, compared to a choice based on a single classifier, the classification decision made by the SMO-SSODBN is more accurate. So, along with FS, web service classification based on QoS criteria is discussed here.

The main contributions of this paper are presented as follows:

  • This paper presents an SMO-SSODBN architecture for smart web service allocation where the web services are classified based on the features present in the WSDL document. The SSODBN algorithm is mainly implemented in this work to minimize the test error and maximize the convergence.

  • The SMO-SSODBN architecture uses the service overlay concept to offer the appropriate services to the smart city users by extending the web of IoT devices concept for web service.

  • The blockchain defined network (BDN) is used to analyze the sensor node behavior and data transfer in the IoT network to improve the latency, privacy, and centralization issues.

  • The Starling Murmuration Optimization (SMO) algorithm is mainly employed to minimize the reconstruction error by tuning the parameters of the SSODBN architecture.

  • The efficiency of the proposed methodology is compared with different existing techniques in terms of latency, precision, recall, response time, accuracy, and F-measure.

  • The security of the smart services is improved using the Blockchain Distributed Network (BDN) and is tested using different security measures such as fairness, confidentiality, and anonymity.

The rest of this paper is arranged accordingly. Section 2 presents the related works and Sect. 3 presents the working of each step of the proposed methodology to achieve an efficient web service classification and secure smart city environment in detail. Section 4 presents the experimental outcomes achieved using the proposed methodology using different metrics when compared to the existing techniques and Sect. 5 concludes the paper.

2 Literature survey

Brincat et al. (2019) illustrated the Internet of Things in a real smart city scenario for an intelligent transportation system (ITS). The Internet of Things with ITS was used for improving data propagation and generating heterogeneous connectivity. The smart road scenario played a major role in the smart city environment. The cloud-based and mesh networks were applied for the propagation of intelligent localization and sensing data. The air pollution and fuel consumption were increased by using the intelligent transportation system. But it failed to predict the accurate localization of the network. Li et al. (2019) established the validation of security in the digital image watermarking process in smart cities using a CNN-based algorithm. The gray watermark images were processed by using this algorithm and then the watermark signals were embedded into the block discrete cosine transform (DCT). The algorithm used a neural network for the extraction and detection of watermarks. The suspected signals were considered as input signals and the output signal was the result of this process. The result showed 93.75% of classification accuracy. The digital watermarking in the smart city was easily hacked because it was less effective.

The securing spatial data infrastructure (SDI) for smart city services and applications was represented by Chaturvedi et al. (2019). This approach was used to integrate advanced standards like SAML, OAuth, and OpenID were connected for authentication and authorization. This approach was utilized for solving the difficulties of location-oriented problems and sharing their geographical information with others. Some data was not able to be accessed because of their data complexity. Kramer et al. (2019) elaborated a micro-service architectural design for secure applications in a smart city. The architectural design was used for making the applications flexible and scalable and the sensitive data in the cloud was secured by using a hybrid approach. The metadata with CP-ABE was encrypted for storing the data in a low-cost manner in the public cloud. This architectural design was utilized to generate scalability in smart city applications and the security approach helped to prevent the applications from unauthorized access. The communications between web services were complex.

A web service selection approach using classification was evaluated (Purohit and Kumar 2018). The domain inputs and weight calculation were minimized by using a hybrid weighting scheme. The PROMETHEE plus and PROMETHEE method was employed for the selection approach of web service. The measurement of effectiveness was very difficult. Balaji et al. (2021) demonstrated an automated query classification-based web service by using the Machine Learning (ML) technique. Various standards such as BPEL, SOAP, UDDI, WSDL, and ebXML were used for the web services. Many web services were performed as a query-response model. The user has submitted a query based on standards and services were supported for natural language queries. The result showed a high recall value. The correct values were not able to be predicted because of their computational complexity.

Alkhammash, (2020). presented formal modeling of requirements based on Web Ontology Language (OWL) ontologies for the growth of a secured smart city system. The technique employs the OWL verbalizer, the OntoGraf tool, the Protégé-OWL editor, and the Rodin platform. The OWL verbalizer is utilized to generate Attempto Controlled English (ACE) requirements from OWL ontologies. As input requirements, ACE representation is utilized and altered into formal Event-B models. But the structure of the data is not reliable. Peneti, et al. (2021) developed the Grey Wolf optimized Modular Neural Network (GWMNN) method to organize security in a smart environment. The translation, application, and construction layers are formed throughout this process, and blocks based on user authentication are developed to manage the privacy and security properties. The improved neural network is then used to keep computing resources and latency usage in IoT-enabled smart applications. The system's efficiency is analyzed using simulation results, which show that the system has high security (99.12%) and low latency when compared to deep learning networks and multi-layer perception. Meanwhile, it does not automatically perform the classification process.

For time series classification Ji, et al. (2019) presented a Just-in-time Shapelet Selection Service (JSSS) that chooses shapelets as the features in an online time series classification system. The Fast Shapelet Selection (FSS) method underlies the JSSS. Initially, utilizing the subclass splitting technique a few time series were sampled by FSS from the training dataset. Then, the series among two various LFDPs was chosen from the sampling time series after the FSS found Local Farthest Deviation Points (LFDPs). In a time series categorization method for online data, these two stages ensure feature extraction and effective training by drastically reducing the number of shapelet candidates and the training time. But the continuity of the time series is not adapted.

A Wide and bidirectional Long Short-Term Memory (Bi-LSTM)-based Web service classification approach was developed by Ye et al. (2019). In order to execute the Web service category’s predicted breadth utilizing the wide learning model, every discrete feature in the Web service description documents was integrated. Then, the Bi-LSTM model was utilized to perform the category’s depth forecast for web services by mining the context data and word order of the terms in the description texts of Web services. As the outcome of the service classification, it integrates the depth and breadth prediction of Web service categories utilizing the linear regression algorithm. At last, the experimental findings demonstrate that the developed method outperforms six Web service classification approaches based on LDA, LSTM, Bi-LSTM, TF-IDF, WE-LDA, and Wide&Deep, in terms of accuracy. On the other hand, it failed to link information in large-scale service networks.

For web services, Ali, et al. (2020) developed the Gen-Fuzzy Based Strategy as a classification method. The two steps of the developed technique were the Classification Stage and the Outlier Rejection Stage (ORS). A fuzzy system that has been offered was presented in ORS. Cooling Force, Domain Belonging Degree, Distance to Class Center, and Class Belonging Degree, were the four criteria employed in ORS. A Distance-Based Ensemble Classifier (DBEC) was utilized that provided training for each classifier individually. According to experimental findings, DBEC is effective in terms of precision, recall, and accuracy. Meanwhile, it required more datasets for classifying the services.

Xiao et al. (2021) presented the LDNM framework, which is a detailed integration of unstructured elements and structured elements, as a universal Web service classification framework. Initially, two document representation techniques are utilized to turn every service document into a feature vector such as the Doc2vec neural-network-based document embedding model and the topic distribution based on Latent Dirichlet Allocation (LDA). Second, by utilizing Node2vec, it is possible to produce structured representation vectors that come from tagging graphs service invoking. At last, utilizing a Multi-Layer Perceptron (MLP) neural network, these features are merged to guide a service classifier. However, it was not suitable to perform classification in multiple learning models.

Natarajan et al. (2020) investigated a web service that allows customers to choose from a list of discovered services the one that meets their Quality of Service (QoS) requirements. Here, the main goal is to create a User Preferential Model (UPM) and clustering-based Semantic Service Selection Model (SSSM) to improve web services. The discovered services are mapped with the requested service’s functional needs, while the retrieved services’ QoS parameters from the description, integration, and discovery that are universal are mapped with the requested service’s nonfunctional requirements. The developed two-tier user preference approach has been used to tackle the quality attribute issues. But it is not feasible to retrieve the perception in web service applications.

Table 1 depicts the comparison table of existing works determining the merits and demerits of each work individually.

Table 1 Literary works of various authors
Full size table

2.1 Research gap

The classification of web services in IoT maximized the utilization of the internet in many organizations through smartphones and personal computers. But there are gaps in the literature on the performance of classification. Some of the gaps are as follows.

2.1.1 Vulnerable to the hacking process

The web services are estimated in a digital image watermarking process that stores multiple sensitive information but it is more vulnerable to attacks. The proposed method addresses this gap by developing the S-DBN method eliminating the irrelevant weights and offering efficient security for data.

2.1.2 Multiple web service classification

The structured and unstructured feature elements are associated with classification and not extracted multiple features. The proposed method addresses this gap by presenting an SMO approach that diminishes the reconstruction error and performs an effective weight optimization in multi-classifier performance.

2.1.3 Requirement of numerous data

The training of the features is performed based on the obtained information and needs more datasets to classify the data individually. The proposed method addresses this gap by introducing the SSODBN technique that performs the classification process in varied application fields in real-time applications.

2.2 Key consideration

The advancement of IoT increased the usage of networks for sharing and storing numerous sensitive data in web service applications. The web provides an easy search for the required data and inorder to perform the classification process the SSODBN method is proposed to perform multi-classifiers of data among various domains in real-world applications. The irrelevant weights are among feature detectors and regularized reinforced transfer functions. The behavior of the sensor node and transmission of data is estimated by the BDN model and the parameters to diminish reconstruction error. The security of web services enhanced the convergence rate with a better classification process.

Table 2 lists every abbreviation utilized in the developed infrastructure’s approach and algorithms and provides definitions for each one.

Table 2 The list of abbreviations
Full size table

3 Proposed methodology

The proposed methodology mainly aims to offer web service architecture for IoT devices along with web service classification. Initially, the SMO-SSODBN architecture is formulated and it is applied to the web service architecture formation. The last section presents the BDN framework for securing web services in the IoT platform.

3.1 Formulation of the SMO-SSODBN architecture

The SMO-SSODBN architecture formation is shown in this section and the SMO optimization algorithm is mainly used to optimize the SSODBN network structure by minimizing the reconstruction error.

3.1.1 Network structure of structural self-organized deep belief network

The S-DBN structure is determined with four parts namely input, hidden, output, and tutor signal layer. The input signal function is used to receive the incoming signal and extract the features from the original signal. The hidden layer is determined with 2 to 4 layers that perform the training process based on the data scale. The previous information obtained in the network in which the hidden layer acquired the data information from the previous layer extracted more features. In unsupervised learning, the regularization and dropout strategy is performed and the restricted weights are optimized. While performing weight updation operation the removed data are returned to its original position. In weight exceeding process, the sensitive data are normalized in web applications.

Figure 1 depicts the self-organized deep belief network’s (S-DBN) structure. The network is divided into four layers such as the hidden layer, the tutor signal layer, the input layer, and the output layer. The visible layer (V) is the initial layer that received the original signal from the network and the hidden layer (1–4, L1–L4) is determined to extract the features by varying 2 to 4 layers in the data scale. The primary Restricted Boltzmann Machine (RBM) is represented by V and L1 [19, 20] while L2 and L1 create the second RBM. In unsupervised training, every RBM’s hidden layer gets the data information and extracts the characteristics denoted from it.

Fig. 1
figure 1

S-DBN’s structure

Full size image

The feature data is then sent to the S- DBN’s third component, in which the output layer is performed, based on the specified task, and the obtained neurons equal the classification. The training process of DBN is categorized into supervised and unsupervised learning models. The mathematical expression for RBM is formulated as

$$q\left( {l_{i} = 1} \right) = \frac{1}{{1 + f^{{ - c_{i} - \sum_{j} u_{j} x_{ji} }} }}$$
(1)

The above equation represents the RBM knowledge learning (feature extraction)

$$q\left( {u_{j} = 1} \right) = \frac{1}{{1 + f^{{ - b_{j} - \sum_{i} l_{i} x_{ij} }} }}$$
(2)

The above equation represents the RBM knowledge reasoning (data reconstruction). \(u_{j}\) represents the \(j\) neuron’s value in the visible layer, and \(l_{i}\) represents the hidden layer, the visible layer is indicated by c, and the deviation of the hidden layer is denoted by b. The weight between the input \(j\) and hidden layer \(i\) neurons is denoted as \(x_{ji}\). If, \(\theta = \left( {x,c,b} \right)\) according to the divergence algorithm and comparison algorithm, the unsupervised training deviation updating and weight formula are as follows,

$$\Delta \theta = \left\langle {l_{i}^{0} u_{j}^{0} } \right\rangle - \left\langle {l_{i}^{1} u_{j}^{1} } \right\rangle$$
(3)
$$\theta^{n} = \Delta \theta + \theta^{0}$$
(4)

where, \(\left\langle . \right\rangle\) represents the state posterior distribution's point multiplication average value, \(l_{i}^{0} u_{j}^{0}\), and \(\theta^{0}\) denotes the initial parameters and initial state.\(l_{i}^{1} u_{j}^{1}\) denotes the hidden layers and the visible multiplication. The parameter variation is represented by \(\Delta \theta\). The training of S-DBN is categorized into dual phases (Chen and Pan 2021). Firstly, the RBM is trained in the network based on an unsupervised greedy approach that modifies the dropout automatically and recognizes the feature detector decorrelation, models the transfer function of the regularization reinforcement and presents regularization reinforcement terms. The Contrastive Divergence (CD) is utilized for adjusting weight. Secondly, the neurons are fixed in the specific position that enhances the neural network, and the weights are optimized by the Starling Murmuration Optimizer (SMO) algorithm is utilized.

If the network scale is needlessly huge in the typical DBN training procedure, the trained network’s generalization is likely poor due to over trained network and the formation of errors in the model. However, because of the presence of explanations for phenomena, neurons frequently have an effect of synergistic. Hence, a neuron in the same layer has readily influenced another neuron. When a characteristic detector learns an invalid characteristic, this error characteristic will probably spread to other characteristic detectors, as a result of which testing suffers.

Regularisation reinforcement and dropout are included in the unsupervised training process that randomly eliminates the hidden neuron. However, the deleted neurons cannot able to execute the training process. The weight of each neuron is fixed by employing an L2 regularization norm and if the weights are overheaded then it is adjusted to obtain the best location regarding a faster learning rate. The mean value is utilized to validate the output of the forward network and the weight functions are determined in the given ratio.

In the unsupervised design process, the dropout's mathematical are,

$$l_{i} = l_{i} \bullet msE$$
(5)
$$msE = \left\{ \begin{gathered} 1\,\,\,\,\,Qe \ge s_{ms} \hfill \\ 0\,\,\,\,Qe < s_{ms} \hfill \\ \end{gathered} \right.$$
(6)

\(i\) denotes the hidden layer unit's state value \(l_{i}\), and whether it is deleted in a given operation is firm as \(msE\).\(Qe\) indicates the possibility of a random choice, and \(s_{ms}\) represents the probability threshold. According to the above equations when \(Qe \ge s_{ms}\), the neuron \(i\) is reserved and its weight is updated, and when \(\,Qe < s_{ms}\), the neuron \(i\) is eliminated in this process. Dropout is used in each Contrastive Divergence (CD) operation to compute the intermediate state \(l_{i}^{0}\) and process \(u_{j}^{1}\).

In the transfer function, the regularization strengthening method is used for controlling the overfitting and enhancing the efficiency of unsupervised training. The regularization enforcement factor and the transfer function are expressed as below,

$$E{}_{X} = \beta \,Q + \alpha \,F_{X}$$
(7)
$$Q = Q\left( {u,\,l} \right) \propto e^{{\left( { - F\left( {u,\,l} \right)} \right)}} = e^{{l^{A} Xu + c^{A} u + b^{A} l}}$$
(8)
$$F_{X} = \frac{1}{d \times n}\,\sum\limits_{i = 1}^{d} {\sum\limits_{i = 1}^{n} {v_{ji}^{2} } }$$
(9)

where \(E{}_{X}\), \(Q\), \(u_{j}\), \(l_{i}\) represents a new objective function, original objective function, value of neuron \(j\), and value of neuron \(i\) respectively. The term \(b\)\(c\) denotes the bias of hidden and visible layers. \(A\) is the matrix transposition. At the starting stage of training, assign \(\beta + \alpha = 1\). Here \(\beta\), \(\alpha\) is the parameter performance. The expression of the unsupervised learning transfer function is shown below,

$$E_{x}{\prime} = \frac{\partial \log Q(u,l)}{{\partial \theta }} + \frac{2}{d \times n}\sum\limits_{i = 1}^{d} {\sum\limits_{j = 1}^{n} {x_{ji} } }$$
(10)

The weights are upgraded with the CD algorithm as shown below expressions,

$$\Lambda x_{ji} = \lambda_{x} \left( {\left\langle {u_{j}^{0} - l_{i}^{o} } \right\rangle - \left\langle {u_{j}^{1} - l_{i}^{1} } \right\rangle } \right)$$
(11)
$$\Lambda c = \frac{{\lambda_{c} }}{{c^{2} }}\left( {\,\,\left\langle {\,l_{i}^{{o^{2} }} } \right\rangle - \left\langle {\,l_{i}^{{1^{2} }} } \right\rangle \,\,} \right)$$
(12)
$$\Lambda b = \frac{{\lambda_{b} }}{{b^{2} }}\left( {\,\,\left\langle {\,u_{j}^{{o^{2} }} } \right\rangle - \left\langle {\,u_{j}^{{1^{2} }} } \right\rangle \,\,} \right)$$
(13)

In the above expressions, \(\lambda_{x}\),\(\lambda_{c}\),\(\lambda_{b}\) are the weight’s upgraded rate and offsets. The biases and weights are upgraded by using the below equations,

$$x_{ji} (n) = x_{ji} (n - 1) + \Delta x_{ji} (n)$$
(14)
$$c(n) = c(n - 1) + \Delta c(n)$$
(15)
$$b(n) = b(n - 1) + \Delta b(n)$$
(16)

3.1.2 Supervised training algorithm

Reinforcing regularization and dropout is added into unsupervised training in S-DBN. In order to decrease overfitting and speed up training, a dropout strategy was formed that tested the sensitivity of the network. But in the absence of weight optimization, the hidden layer remains in an independent state and extracts a sparse activation state. Moreover, some phenomena can be explained, neurons frequently work in concert. As a result, a neuron in a similar layer easily affects another neuron. A dropout technique is put forth and utilized in supervised training. It prevents neural synergism during training and enhances numerous benchmark tests.

All hidden layer neuron is arbitrarily eliminated with a given probability. The connected weight and state value of the removed neuron won’t be changed because it won’t currently be a part of the neural network.

The final RBM output is received by the final layers of neural networks. The feedback errors are generated with the comparison of accurate signals from the output signals of networks. Initially, the data signals are transferred to the input layers, and later it is transferred to the output layer. Secondly, the BP errors are created that are utilized to tune the network metrics. The following phases are described below;

1. Initializing the forward neural networks and determining the step size.

$$z_{i} \left( h \right) = \sum {x_{ji} } z_{i} \left( {h - 1} \right)$$
(17)

2. Compute the feed-forward outputs, the output of the neuron is \(z_{i} \left( h \right)\) in layer 1.

$$z_{i} \left( h \right) = \sum {x_{ji} } z_{i} \left( {h - 1} \right)$$
(18)

3. Compute the output errors, the tutor signal is represented by \(S\), the output layer rate is denoted by \(Z\), and the error signal is denoted by \(f\).

$$f_{i} = S_{i} - Z_{i}$$
(19)

4. Employ SMO, for the neuron of output layers,

$$\beta_{k} \left( {m - 1} \right) = z_{i} \left( h \right)\left[ {1 - z_{i} \left( h \right)} \right].\sum {\beta_{i} } \left( h \right)x_{ji} \left( h \right)$$
(20)

5. Change the weights, the learning value is represented by \(\lambda\)

$$x_{ji} \left( h \right) = x_{ji} \left( h \right) + \lambda \beta_{i} z_{i} \left( h \right)$$
(21)

6. Conclude the training process.

3.2 Starling Murmuration Optimizer (SMO) algorithm for reconstruction error optimization in the S-DBN

The SMO optimization algorithm (Yu et al 2018) is mainly incorporated to minimize the reconstruction error. The reconstruction error estimates the performance of each layer on unsupervised models. The mathematical expression of reconstruction error is defined in the below equation.

$$\Re_{\varepsilon } = \frac{{\sum\nolimits_{x = 1}^{a} {\sum\nolimits_{y = 1}^{b} {\left( {P_{x,y} - D_{x,y} } \right)} } }}{a \times b \times Pz} \times 100\%$$
(22)

The term \(a\) depicts total training samples, \(b\) indicates pixel amount per sample, \(Pz\) represents pixel range, \(P_{x,y}\) implies estimated result, and \(D_{x,y}\) signifies actual value. The SMO algorithm is formulated using the following steps. The starling murmuration is one of the major natures magnificent which exhibits in the mass of various flocks including different starlings and above its roost it will dive with the sky for nearly half an hour. The starling's flocks are intermittently separated and the recombination is in a greatly synchronizing manner with murmuration. The direction changing, certain whirling, recombination, and splitting are distributed through the flocks with one starling to another utilizing optimized decision-making (Zamani et al 2022). The separation of a search system (diversity) phase in the SMO algorithm is described as follows. In murmuration, some of the starlings are commonly isolated with their flocks, which is considered a significant intention for designing. The mathematical equation for the separated population is expressed in the following equation;

$$Q_{S} = \frac{{\log \left( {u + E} \right)}}{{\log \left( {MAXIMUM\,Ju \times 2} \right)}}$$
(23)
$$Y_{j} \left( {u + 1} \right) = Y_{H} \left( u \right) + R_{1} \left( z \right) \times \left( {Y_{{s{\prime} }} \left( u \right) - Y_{s} \left( u \right)} \right)$$
(24)

The global position is represented by \(Y_{H} \left( u \right)\), the separated population, and proportions of the starlings are selected \(Y_{{s{\prime} }} \left( u \right)\), the randomly picked population is represented by \(Y_{s} \left( z \right)\). The new operator \(R_{1} \left( z \right)\) utilized the separated search method.


Step 1: Separation phase.

The separation phases are employed to maintain the diversity of the population according to the quantum harmonic oscillator. The mathematical equation for the separation phase is expressed in the following equation.

$$R_{1} \left( z \right) = \left( {\frac{\beta }{{2^{o} \times o! \times \pi^{\frac{1}{2}} }}} \right)^{\frac{1}{2}} I_{o} \left( {\beta \times z} \right) \times e^{{ - 0.5 \times \beta^{2} \times z^{2} }} ,\quad \beta = \left( {\frac{n \times l}{i}} \right)^{\frac{1}{2}}$$
(25)

The Hermite polynomial is represented by \(I_{o}\), the random number is represented by \(z\), the quantum harmonic oscillator is represented by \(\beta = \left( {\frac{n \times l}{i}} \right)\), particle mass is denoted by \(n\), the strength is denoted by \(l\), and the plank’s constant is denoted by \(i\).


Step 2: Dynamic multi-flock phase.

The dynamic multi-flock phase is determined to design the starling behavior when the iterations are shifted the position. Inorder to explore and exploit the solution the starlings determined in search space are split into whirling, separating, and diving. The developed strategies are used to enhance the efficiency by reducing the complex problems. Initially, the starlings are chosen randomly and shifted to another position in the search space. The flock's quality is validated by computing search strategies and balancing the exploration and exploitation phase.

It is defined by utilizing the particular partition then the set \(Tg\) is divided with \(l\) non-empty flocks \(g_{1} .......g_{l}\).

$$Tg\left( u \right) = \left\{ {tg_{j} \left( u \right) \in T|tg_{j} \left( u \right) \le tg_{j + 1} \left( u \right)\,\,\,\,\,\,for\,j = 1,....O{\prime} } \right\}$$
(26)
$$S\left( u \right) = \left\{ {tg_{j} \left( u \right) \in Tg\left( u \right)\,\,\,\,\,for\,j = 1,....,l} \right\}$$
(27)
$$Q = T - S\,\;and\;Q = \bigcup\limits_{j}^{l} {Q_{j} } \cdot \left| {Q_{j} } \right| = \,\left| {Q_{k} } \right|\;for\;Y_{s} \left( z \right)\;j \ne k \in \left( {1,....,l} \right)$$
(28)

The portions of the starlings \(Q_{j}\) are designed to flock members \(g_{j}\), every flock \(g_{r}\) contains starlings \(\left( {o = \frac{O^{\prime}}{l}} \right)\), the representative set \(S\) is selected the representative \(\left( {S_{r} } \right)\), \(Tg\left( {u + 1} \right)\) set is arranged variously \(Tg\left( u \right)\). Each flock member and the representative of the multi-flocks \(g_{1} ,.....,g_{l}\) provided information among flocks through iterations. The flock quality is denoted by \(g_{l}\).


Step 3: Flock quality phase.

The flock quality contains different starlings in the iteration \(u\) indicated by \(R_{r}\) and it is expressed in the below equation

$$R_{r} \left( u \right) = \frac{{\sum\nolimits_{j = 1}^{l} {\frac{1}{o}\sum\nolimits_{k = 1}^{o} {tg_{jk} \left( u \right)} } }}{{\frac{1}{o}\sum\nolimits_{j = 1}^{o} {tg_{rj} \left( u \right)} }}$$
(29)

The fitness value in \(jth\) the starling of the subpopulation flock is represented by \(tg_{jk} \left( t \right)\),\(o\) determines the flock with various starlings,\(k\) which represents the flocks with murmuration \(l\). The diving exploration process is considered for exploring the effective search space. It contains the downward and upward dives of quantum and the quantum random dive (QRD) operator to select the dives of quantum. The qubit outcomes probability is denoted by \(\left| \beta \right|^{2}\) and is expressed as;

$$\left| \gamma \right\rangle = \cos \frac{\beta }{2}\left| 0 \right\rangle + \sin \frac{\beta }{2}e^{j\kappa } \left| 1 \right\rangle$$
(30)

The conditional shift operator is represented by \(T\), the qubit rotation matrix is represented by \(D\), and the rotation of the angle is represented by \(\gamma\) and \(\theta\).

$$D = \left[ \begin{gathered} e^{j\gamma } \cos \theta \,\,\,\,\,\,\,e^{j\mu } \sin \theta \hfill \\ - e^{ - j\gamma } \sin \theta \,\,\,\,e^{ - j\gamma } \cos \theta \hfill \\ \end{gathered} \right]$$
(31)

Step 4: Quantum random dive operator (QRD).

The two various quantum probabilities are \(\left| {\gamma^{V} \left( {Y_{j} } \right)} \right\rangle\) and \(\left| {\gamma^{E} \left( {Y_{j} } \right)} \right\rangle\) according to the unitary operator \(V\) select the downward or upward quantum dive.

$$QRD = \left\{ \begin{gathered} \left| {\gamma^{V} \left( {Y_{j} } \right)} \right| > \left| {\gamma^{E} \left( {Y_{j} } \right)} \right|\,\,\,\,for\,upward\,quantum\,dive \hfill \\ \left| {\gamma^{V} \left( {Y_{j} } \right)} \right| \le \left| {\gamma^{E} \left( {Y_{j} } \right)} \right|\,\,\,\,for\,downward\,quantum\,dive \hfill \\ \end{gathered} \right.$$
(32)

The steps in the Whirling search phase (exploitation) are explained in the following step. The flocks \(g_{r}\) have higher quality in the iteration \(u\), and the next location of various flocks in every starling \(t_{j}\) evaluated utilizing the whirling search phase. Then it is expressed as;

$$Y_{j} \left( {u + 1} \right) = Y_{j} \left( u \right) + D_{j} \left( u \right) \times \left( {Y_{SX} \left( u \right) - Y_{O} \left( u \right)} \right)$$
(33)
$$D_{j} \left( u \right) = \cos \,\left( {\sigma \left( u \right)} \right)$$
(34)

The present position of the starling is denoted by \(Y_{j} \left( u \right)\) and \(Y_{SX}\) is selected by the members of the flocks. The unrepeatable random neighbor in the starling is denoted by \(Y_{O} \left( u \right)\). The working of the proposed SMO-SSODBN architecture for web service classification is presented in Fig. 2 and the detailed description of the methodology is SMO-SSODBN in the upcoming sections.

Fig. 2
figure 2

Formulation of the SMO-SSODBN architecture for web service classification

Full size image

Figure 3 shows the SMO flow chart. In Fig. 2\(\tau_{R} \left( U \right)\) represents the average quality of flocks in iteration U.

Fig. 3
figure 3

SMO flow chart

Full size image

The pseudocode for the SMO-SSODBN architecture is presented in Algorithm 1.

Algorithm 1
figure a

Pseudocode of the SMO-SSODBN architecture

Full size image

3.3 SMO-SSODBN web service management platform

A web service in the Internet of Things (IoT) requires adaptive mobility, interoperability, privacy, network security, throughput, optimal routing, broadcasting ability, traffic management, etc. (Qamar et al 2016; Natarajan et al 2020). This paper proposes an SMO optimal SSODBN web service based on the cloud computing model for smart cities. The web of object (WoO) service on the IoT platform utilizes a web-based service structure to promote a smart city. The SMO-SSODBN architecture is comprised of three different layers in the information layer service profiles, user profiles, sensing values, and device metadata. In this, the functions of object management involve service federation, service support, service composition, user management, data support, and object management. The functions of an object gateway involve device control, device management, data management, etc. In specific, the presentation layer is a kind of application that offers web services to users through different network browsers. Meanwhile, the users can utilize service federation, service composition, and service recovery without any hardship with web forms.

The SMO-SSODBN architecture offers object web service or device with smart city applications using WoO gateway which drives the devices and offers internet access to non-internet protocol (IP) devices. The application of different overlay network service models provides mashup, a web-based application using service composition and federation. The SMO-SSODBN management platform is placed at the management layer of the proposed architecture with its diverse devices and web service objects namely device registration, metadata management, profile management, fault tolerance, and device control.

The application service operates as a service component of the management platform using registration and composition individually as that of web services. The sensor and actuator devices work cooperatively with the sensor data support unit directly or via a gateway. Furthermore, the web service on mobile devices is provided based on the users and devices that serve as an object. The generation of object-service features from the service composition is handled by the service federation. The working of the SMO-SSODBN web service management platform is presented in Fig. 4.

Fig. 4
figure 4

Outline of the SMO-optimized SSODBN architecture for efficient web service allocation in the smart city

Full size image

By the utilization of an object-service connection, the web service is formulated by the service composition unit. The function of the registry in service composition defines the service functionalities offered to the users; moreover, it discovers and publishes data about web services. The discovery and selection process includes service matching and service selection strategies for determining the web service according to their service criterion and for selecting a suitable web service from the generated services depending on the data ranking procedure. More specifically, the service deployment function offers the related services to the appropriate IoT devices and objects in real-time, it can be utilized as service and application descriptions. Moreover, the service management block is comprised of several components that are required to execute service operations. The object management component consists of multiple numbers of sensors devices, actuators, sensors, and users. The gateway in application service includes operations such as management of device value and metadata, device control, fault tolerance, device profile and registration, etc. The gateway acquires request messages from multiple devices and regarding user demands, the response messages are generated mechanically.

3.3.1 Smart WoO service model

The proposed SMO-SSODBN model is not just formed by the interpretation of the device as an object; it utilizes novel WoO and IoT services through the configuration of diverse objects and devices. The application of diverse overlay network service concepts provides new services via service composition and service federation methods. The SMO-SSODBN management protocol is comprised of four mechanisms that are described as follows: (i) the user constituent not only offers services but also conveys service and requests data to the object constituent, (ii) the object constituent virtualizes and objectifies real-world physical objects and physical devices using WoO service model, (iii) the service constituent integrates the services and other external services as service composition or federation, (iv) the physical constituent manages physical devices, actuators, and objects namely humidity, motion, door, window, humidifier, radiator, ventilator, cooler/heater, and temperature sensor, etc.

3.3.2 SMO-SSODBN gateway

The SMO-SSODBN gateway carries out various tasks such as real-time control, device web service scheduling, device fault tolerance, device metadata, and profile management, device registration, and device value management. The SMO-SSODBN gateway can capture different request messages and service retrieval from the user and then the response messages are generated spontaneously. The SMO-SSODBN gateway is placed between the SMO-SSODBN web service management platform and different devices. This gateway is applied with an embedded database to manage the device and store metadata information.

3.3.3 Determination of device and object

The object and device can create the non-IP device and the service composition and federation will provide device web service by using the SMO-SSODBN gateway. The object profile and metadata devices are modified for web service and the object information is for the objectification of the object and device. In the objectification of objects and devices, every object and device is using the SMO-SSODBN web service management platform which has an individual profile, metadata, control methods, messages, and other information.

3.4 Classification of web service

The web services are categorized into functionality-based categories during their phase. The web service classification process is delineated in Fig. 4. The web services are classified into a single category as Education, Smart electricity, Intelligent road networks, Health and social care, Sports, and Water and gas distribution. The information can be received from the WSDL documents such as schema, port type, messages, service document, and service name. The pre-processing method is used to the downloaded information for refining it. Each trained classifier is used to categorize the web service into a single category. The SSODBN classifier is mainly incorporated to determine the web service classification output.

3.4.1 Feature extraction of Web Service Description Language (WSDL)

The WSDL documents are designed to extract the relevant features. The components of the WSDL scheme contain attribute type and name. Few WSDL schemes such as name attributes of complicated types, documentation contents of schematic elements, and documentation contents of complicated types are extracted from WSDL contents. The following information is extracted from the components of each message such as name, and element. The information from the port variety is extracted from documentation content, the name of every operation as well as the name attributes of both output and input parameters in every operation. The information containing service documentation and service name is transmitted with text files. These picked contents of WSDL are utilized as the base for the classification of web services.

• Textual pre-processing

The objective of this strategy is to enhance information quality that is accessible for classification. The text file information is inconsistent and it leads to incorrect results in the mining process. Various pre-processing phases are used for the extraction of accurate values and consistent information. Various steps in pre-processing are described below;

Step 1: Tokenization.

The initial stage in morphological evaluation is tokenization. It is one of the methods used for separating the text stream into phrases, symbols, and words known as tokens. In machine learning format the texts are stored and the insignificant characters such as brackets, hyphens, and commas are required to be removed through tokenization.

Step 2: Word splitter.

According to the capitalization of the letters, the concentrated words are separated by employing a word splitter. The Validate Address Response is one of the operations in the web service and it is meaningless and split into address, response, and validate. The new processing phase is introduced by the word splitter that is not utilized through the existent web services clustering.

Step 3: Stop word removal.

The stop word removal with textual information is the next phase in the pre-processing. The stop list consists of articles and prepositions that are meaningless and it is easily eliminated from the documents.

Step 4: Function word removal.

The stop words are generally removed through the function words but some of the function words are non-stop words. The remaining functional words are removed by using this approach.

Step 5: Lemmatization phase.

Lemmatization is the last step in pre-processing. Vocabulary is used for reducing the words into valid dictionary forms and it performs the morphological evaluation of the words. The valid dictionary forms are named lemma. Let us consider the token saw, the stemming is reduced by a letter \(t\), and then the lemmatization is returned to see or saw which depends on the word utilized as the verb or noun. The lemmatization reduces the word lemma while the derivational words are minimized by using stemming. Web service providers normally release their services using their websites instead of using public registries. Hence more users are using search engines to retrieve an appropriate service of their choice. However, keyword-based search engines are not quite effective when providing users with appropriate web services due to the partially matched search queries. To achieve the appropriate web services, the users need to know the complete details of the keywords used. The proposed methodology mainly extracts the semantic information (device operation, service information, messages, etc.) from the WSDL document using the SMO-SSODBN classifier.

3.5 Secure IoT smart city environment based on optimized modular neural network

The developed IoT smart city can resolve communication latency, security, privacy, centralization, computational cost, and resource management. The IoT-enabled smart city has many advantages such as privacy, reliability, scalability, connectivity, and compatibility. The three layers of an IoT-enabled smart city are the construction, translation, and application layer. The environment of the edge network is continuously monitored by the many IoT devices and the data collected is transferred to the industries or smart applications, where effective connections are made between the IoT devices and the interface network. The collected data is transmitted to the translation layer due to the presence of fog nodes and distributed defined systems are presented together. Blockchain technology (Peneti et al 2021) is used to secure the fog nodes that enhance the service's network quality, security, and communication complexity. Each node in the translation layer contains actuators to execute the data conversion process which can enhance the total communication process of the smart city.

The blockchain offers a trustable IoT environment, secure data transmission, transparency, and so on leading the distribution system to choose the blockchain. The blockchain technique is combined with a software-defined network to develop BDN. The combination is done to define the effectiveness of the external entities, data plane separation, and transmission protocol. Figure 5 shows the BDN’s working process.

Fig. 5
figure 5

BDN for malicious behavior detection in the smart city

Full size image

Data gathering/acquisition is the first step to creating a safe and secure IoT environment. The information is gathered from the construction layer. Software configuration, manufacturing execution functions, and Enterprise Resource Planning (EPR) are utilized to collect information on the smart industry during the data-gathering process. Various IoT devices are used to gather different information such as information about the building, climatic conditions, health details, etc. The transfer of the gathered information is done through the Internet. The data are converted into blocks with the use of a data converter block during the data transfer process (Singh et al 2021). The collected data are converted into meaningful information by the data converter block. Then the gathered data are transferred to the transplant layer which provides security with the use of a blockchain approach.

The next step is security which is done in the translation layer to provide privacy. The transfer process is done by a combination of sensors and several blocks of the blockchain. The information transmission is executed based on the condition and the blocks are created. To manage the data privacy, security, trust, and reliability of BDN.

In the working process of BDN, with the verification of the blockchain address, a new node is created and the key information is generated. The registration process of a new user is also done with the verification of their address value and key. Next, the addition of a new block is executed by presenting an optimal key. The translation layer transmission is effectively executed to make sure the privacy; high cost, computing power, and storage problems are all resolved appropriately by security.

3.6 Device management controlling the SMO optimization and SSODBN architecture

In the application layer, propose an SMO-based SSODBN where SMO manages application services for the smart industry, including self-configuration, self-decision support, self-management, and self-distribution as well as centralization with the data centers. The use of SSODBN also helps to enhance output in the smart sector. In the smart industry, smart execution, smart control, and smart planning input data are denoted by \(\left\{ {E_{1} ,E_{2} ,...,E_{m} } \right\}\), \(\left\{ {C_{1} ,C_{2} ,...,C_{m} } \right\}\) and \(\left\{ {P_{1} ,P_{2} ,...,P_{m} } \right\}\) respectively, and \(\left\{ {S_{1} ,S_{2} ,...,S_{m} } \right\}\) represents production sequences. All sequence is made up of a feature vector set of production codes \(\left\{ {PC_{1} ,PC_{2} ,...,PC_{m} } \right\}\) denoted by \(PC_{j} \in Q^{n}\). The vector dimension length is n. \(\Delta_{t}\) represents the difference in time between the past and the present. \(\Delta F_{j}\) represents the time sequences for each production record. \(\left\{ {IC_{1} ,IC_{2} ,...,IC_{l} } \right\}\) represents the intervention codes.

All production’s SSODBN feature vector is selected by

$$V_{j}^{{\text{SSODBN }}} = \left\{ {y_{PCj} ,y_{ICj} ,\Delta F_{j} ,n_{j} } \right\}$$
(35)

The SSODBN computes the sequences of dispersed simultaneously states \(\omega 1,\omega 2,..\,...\,...\omega m\) and then verifies the aggregates of these states utilizing the weighted pooling function \(Z_{pooling} \left( {\omega 1,\omega 2,...\,...\,...\,\omega m} \right)\). It is done by the middle layer.

The outcome probability is calculated utilizing

$$F\left( {x|\omega 1,2,\,...\,...\,..m} \right) = F\left( {SSODBN\left( {Z_{pooling} } \right)} \right)$$
(36)

The probability outcome is \(F\left( {x|\omega 1,2,\,...\,...\,..m} \right)\) for the smart industry, and it is determined by the output and stored products in the database.

3.7 Security analysis

The blockchain and SDN approach is used as a conversion layer that includes the block in the smart city network. The security analysis is demonstrated with four properties and they are discussed below.

3.7.1 Anonymity property

While performing data transmission in web services the new user generation and actuator node is formed. Initially, the registration process is completed, and verifies the data for classification. The blocks are added to the network to determine the smart city functionality.

3.7.2 Fairness property

The SDN utilized a peer-to-peer network in blockchain that integrated all the sensors for providing information regarding the corrupted nodes. In this strategy when the new user registration is filed then the other cannot able to access the information. A secret code is provided that determines a complete verification is undertaken among all the transmitted data. Hence this function determines the fairness property.

3.7.3 Confidentiality property

Before transmitting the user initially encrypts the data by providing the private and public key individually. This provides a ciphertext in the web service that decrypts the receiver’s information. But it is significant to keep the sender’s and receiver’s keys as same. The confidentiality level is maximized while improving network security.

4 Result and discussion

In this section, the efficiency of the proposed method was determined by comparing the performance of BDN-GWMNN, OWL-ACE, CNN-DCT, and CPABE respectively. The proposed methodology overcomes the challenges of smart IoT in practical implication scenarios. The metrics such as accuracy, recall, F-measure, precision, response time, reproducibility, and reliability are measured and discussed in Sect. 4.2. The experimental settings are provided in Table 3. The dataset utilized in this work is the Quality of Web Service (QWS) dataset (Al-Masri and Mahmoud 2008; Al-Masri and Mahmoud 2007; Al-Masri and Mahmoud 2007) along with the different IoT real-time web service datasets collected from which 80% of the dataset is used for training while the remaining 20% is used for testing.

Table 3 Experimental settings
Full size table

4.1 Performance measures

The performance measures like accuracy, recall, F-measure, and precision are computed and their mathematical expressions are listed below. Here, the term \(t_{p} ,f_{p} ,t_{n} ,f_{n}\) represents the true positive, false positive, true negative, and false negative respectively.

$$Accuracy\,(A)\, = \,\frac{{t_{p} \, + \,t_{n} }}{{t_{p} \, + \,t_{n} \, + \,f_{p} \, + \,f_{n} }}\,$$
(37)
$${\text{Re}} call\,(R)\, = \frac{{t_{p} \,}}{{t_{p} \, + \,f_{n} }}\,$$
(38)
$$F - measure\,\,(F_{m} )\, = \frac{{t_{p} \,}}{{t_{p} \, + \,\frac{1}{2}\,(f_{p} \, + \,f_{n} )}}\,$$
(39)
$$\Pr ecision\,(P)\, = \frac{{t_{p} \,}}{{t_{p} \,\, + \,f_{p} \,}}\,\,\,$$
(40)

4.2 Performance analysis

Different smart city services are Education, Smart electricity, Intelligent road networks, Health and social care, Sports, and Water and gas distribution are tabulated in Table 4. The number of web services for each class is noted.

Table 4 Analysis of dataset categories
Full size table

The precision and reproducibility value of the proposed method and BDN-GWMNN for different smart city services are noted in Table 5. The proposed method has achieved a higher precision and reproducibility rate than BDN-GWMNN. The F-measure value of different smart city services is tabulated in Table 6.

Table 5 Comparative analysis of precision and reproducibility
Full size table
Table 6 F-measure analysis for different classes
Full size table

In Table 7, the metrics such as availability, reliability, and response time of the proposed method for smart city services such as Education, Smart electricity, Intelligent road networks, Health and social care, Sports, and Water and gas distribution are noted and examined. The Precision and Recall values of the proposed method with existing methods like BDN-GWMNN, OWL-ACE, CNN-DCT, and CPABE are tabulated and compared in Table 8. For various smart city services, the Precision and Recall values are determined. The table shows that the proposed method has a maximum precision and recall rate than the existing method.

Table 7 Availability, reliability, and response time analysis of the proposed method
Full size table
Table 8 Precision and recall analysis of various methods
Full size table

Figure 6 depicts the precision analysis of the proposed method with respect to the different number of services. The result shows that the proposed method has achieved a higher rate compared to other techniques. The highest precision rate of the proposed method is 91%.

Fig. 6
figure 6

Precision analysis

Full size image

Figure 7 shows the graphical representation of the analysis of recall value. Here, the graph is plotted between the different numbers of services and recall rate. The proposed method has attained a maximum recall value than BDN-GWMNN, OWL-ACE, CNN-DCT, and CPABE. The highest recall rate of the proposed method is 94%.

Fig. 7
figure 7

Recall analysis

Full size image

Figure 8 represents the F-measure analysis of the proposed methods with respect to the different number of services. The F-measure value of the proposed method is higher than existing methods such as BDN-GWMNN, OWL-ACE, CNN-DCT, and CPABE respectively. The obtained higher F-measure value of the proposed method is 90%.

Fig. 8
figure 8

F-measure analysis

Full size image

Latency analysis of different methods was graphically represented in Fig. 9. The proposed method is compared with BDN-GWMNN, OWL-ACE, CNN-DCT, and CPABE respectively. The evaluation result shows that the proposed method has achieved a lower latency compared to other methods. As the number of services increases, the latency of the proposed method also increases.

Fig. 9
figure 9

Latency analysis

Full size image

The reconstruction error curves for the proposed architecture are plotted via curves in Fig. 10a–c. The RE will start at 80% in the first RBM which is the highest value compared with the other two layers. The curve is rapidly down in the training which demonstrates the training process is very fast. The second curve’s starting point is lower compared to the first layer because this layer receives the learning information from the first layer. In the third layer, the starting point is lower than the second layer. In the training of each layer, the weights are given at random and the disturbances are increased which leads to the minimum point coverage in the third layer.

Fig. 10
figure 10

Reconstruction error for different layers a Layer-1, b Layer 2, and c Layer 3

Full size image

Figure 11 depicts the data confidentiality level validation for the proposed method. The confidentiality level determines the security level of data and protects it from unauthorized users. The personal information stored in the network is ensured in web services and safeguarded from illegal access. The performance of this approach is validated by comparing the proposed SSODBN with existing SDN-GWMNN, OWL-ACE, CNN-DCT as well as CPABE methods. The proposed method enhanced the protection level by attaining 89% and the existing techniques achieved 78%, 85%, 81%, and 76% respectively.

Fig. 11
figure 11

Confidentiality rate validation

Full size image

The security level evaluation to determine the efficiency of web services is delineated in Fig. 12. It measures the strength of the network and determines the unauthorized attacks detected in the network. The security level for the proposed method is maximized by 58 and the remaining state-of-the-art techniques diminished the security of data by 43, 51, 55 as well as 46 respectively.

Fig. 12
figure 12

security level analysis

Full size image

4.3 Discussion

The generation of web services in IoT networks has been maximized recently due to the usage of the internet in handheld devices. The researchers conducted various analyses regarding the secure classification of web services in the network. Some of the methods such as CNN-DCT, PROMETHEE plus, CPABE, BiLSTM, BDN-GWMNN, OWL-ACE, and LDNM methods are developed to estimate the real smart city scenario and web service classification. Heterogeneous connectivity is made throughout the network and analyzes the weighting scheme. However, it is more vulnerable to attacks and cannot able to predict the accurate localization of the network. Moreover, the updation of optimized weights is complex and thus fails to perform effective communication in web services. In large-scale networks, it does not link the gathered information due to the presence of limited data. But the proposed SSODBN method is determined to classify the web services among various organizations in real-time scenarios. It restricted the determination of unwanted weights and the SMO optimization diminished the reconstruction error in smart city architecture. The security of smart services is maximized by BDN which enhances the latency and centralization issues. Still, it generates a complexity while performing web service classification and it cannot safeguard the smart city services automatically.

5 Conclusion

This paper presents SMO-SSODBN architecture for web service classification and the experiments are conducted on two real-time IoT and web service datasets. The security of the smart city web services is improved using a BDN framework. The performance metrics such as accuracy, recall, F-measure, precision, response time, reproducibility, and reliability are computed. The proposed method has attained higher precision, recall, and F-measure rate than BDN-GWMNN, OWL-ACE, CNN-DCT, and CPABE. The latency of the proposed method is minimal compared to existing methods. The metrics are compared for different smart city services like Education, Smart electricity, intelligent road networks, Health and social care, Sports, and Water and gas distribution. The efficiency of BDN in enhancing the security of the smart city services is tested in terms of the anonymity, fairness, and confidentiality properties. The latency obtained by the proposed methodology is mainly within 110 ms for a total of 200 services. The proposed methodology offers an F-measure score of 90%, a recall value of 94%, and a precision value of 91% for web service classification. In future work, to create a massive service network that will make it easier to categorize Web services, we will investigate and utilize link information/ service relationships. Also, we plan to rate the service quality of each web service classified into four levels namely platinum, gold, silver, and bronze.