Developing an Efficient Deep Learning-Based Trusted Model for Pervasive Computing Using an LSTM-Based Classification Model

Mobile and pervasive computing is one of the recent paradigms available in the area of information technology. The role of pervasive computing is foremost in the field where it provides the ability to distribute computational services to the surroundings where people work and leads to issues such as trust, privacy, and identity. To provide an optimal solution to these generic problems, the proposed research work aims to implement a deep learning-based pervasive computing architecture to address these problems. Long short-term memory architecture is used during the development of the proposed trusted model. The applicability of the proposed model is validated by comparing its performance with the generic back-propagation neural network. This model results with an accuracy rate of 93.87% for the LSTM-based model much better than 85.88% for the back-propagation-based deep model. The obtained results reflect the usefulness and applicability of such an approach and the competitiveness against other existing ones.


Introduction
e recent advances in information technology has made the world to shift from big desktop computers and have a tendency to powerful and smaller devices to facilitate with large computational and heterogeneous wireless communications interfaces. Weiser [1] referred the concept of pervasive/ubiquitous computing which is the most recent paradigm in the world of computers. Ubiquitous/pervasive computing offers a number of advantages such as making life easier with the support of digital infrastructure and mobile devices accomplished by offering the services distribution to the people. Some of the pervasive and ubiquitous computing facing security-related issues are open for researchers to answer. Devices inside the pervasive systems are embedded and invisible which are operating in pervasive surroundings. e pervasive devices are performing mutual interaction without any identity in advance. So, it becomes complicated for the users to know where such devices are present and to exchange personal information. e traditional security of computing is mostly relying on the techniques of access control and authentication which provide access to the only registered users of the system. Ubiquitous and pervasive computing systems are very scalable and flexible due to which it is not suitable to adopt such services. e main characteristic of pervasive computing is the development and design of efficient services to the user who sends request for the services and in the context from which the request of service is sent. e contribution of the proposed research is to develop an efficient deep learning-based model to ensure the generic security issues such as trust, privacy, and authenticity over the Internet. From the selected dataset, only 12% of the data is used for the experimental purposes, and it results in a high accuracy rate of 93.87%. is high accuracy rate for such a small amount of data shows that if the training set increases, then the proposed model will provide more prominent accuracy rates. e paper is organized as follows: Section 2 represents the related work to the proposed research. Section 3 briefly shows the materials and method followed for conducting the proposed research. Section 4 shows the experimental results of the proposed study. e paper is concluded in Section 5.

Related Work
Different approaches and techniques are proposed by researchers for pervasive computing. Chen et al. [2] proposed an infrastructure of data-centric design to support the applications of context-aware. eir proposed middleware treats sources of contextual data as stream publishers. e system is robust to support self-organizing peer-to-peer overlay to support the services of data-driven. Katsiri and Mycroft [3] proposed a system for simulating pervasive systems through estimated knowledge about its situations and entities involved. e research has improved AESL with the function of higher order predicated to denote estimated knowledge about the possibility of predicted instance with value True for a time reference. Padovitz et al. [4] presented a framework of ECORA for the computing of context-aware for reasoning context about uncertainty and tracing the issues of scalability, usability, communication, and heterogeneity.
e system considers an agent-oriented hybrid approach and combining service of centralized reasoning with context-aware, reasoning able mobile software agents. Ahamed et al. [5] presented the S-MARKS design and implementation which consists of resource discovery, device validation, discovery of resource, and privacy of the module.
Boukerche and Ren [6] proposed a system of security for management of trust involving development of a trust model, nodes credential assignment, private key updating, managing the trust value, and taking suitable decision about the rights of access nodes. e research demonstrated that a malicious node can efficiently be excluded from the environment of pervasive and ubiquitous computing. Yu et al. [7] surveyed the literature for the comparison and classification framework for the four dimensions of the design concerned application migration: spatial, temporal, entity, and other concerns. Bello Usman and Gutierrez [8] augmented the basic concept of pervasive and mobile computing from different studies of the literature which uses methods and generic conceptual phases for the management of trust. e study covered a wide range of methods, techniques, models, and applications of trust-based protocol. Carullo et al. [9] presented a new approach for the establishment of trust leveraging the profiles of users. e authors [10] presented an approach which is capable of judging the trustworthiness of a device which interacts and behavior of the device with little interaction experience. e existing research in the field of gesture recognition and facial expression in the perspective of intelligent tutoring is analyzed for facilitating educational societies in building an efficient tutoring system [11].
Kurniawan and Kyas [12] presented a statistical decision approach for trust-based access control through Bayesian decision theory for identity management in Internet of ings. Dangelo et al. [13] presented the generic issues of pervasive computing architecture. e system integrated the techniques of artificial intelligence for achieving similar resemblance with the decision making which is human-like. Apriori algorithm was firstly used to extract the behavioral patterns, and then, Naïve Bayes classifier was used for decision making for trustworthiness of users. Uddin et al. [14] proposed an approach for the detection of terrorist activities. e five models of the deep neural network are used to monitor the behaviors of terrorist activities. e approaches used logistic regression, Naïve Bayes, and Support vector machine algorithms. e authors [15] proposed deep learning and neural networks algorithms for prediction of the behavior of punctuality of employees at the workplace. Khan et al. [16] proposed a variant of SVM, a LinearSVC for answer classification. e chi-square and univariate methods are used for the reduction of the size of the feature space. Deep learning algorithms are used in a variety of problems such as for evolutionary computing models in computer vision [17] and deep ensemble learning for human action recognition in still images [18].

Materials and Methods
e following sections show the materials and methods used.

Dataset. Dataset (Dishonest Internet users dataset.txt)
[19] has been used in this study. e dataset has 322 instances and 5 attributes which is available on the UCI machine learning repository. Figure 1 represents the generic diagram of the intrusion detection system. e system of intrusion detection acts like a guard at the objective node to activate a firewall and to alert host devices when an unauthorized access or illegal traffic is detected. In our case, we have used the deep learning-based model to defend the unauthorized access and malicious network traffic.

Deep Neural Network Backward Propagation Model for the Classification of Trusted and Untrusted Internet Users.
To tackle the problem of perceptron, in 1986, Rumelhart et al. [20] defined a new supervised learning technique which is called the Back-Propagation Deep Neural Network (BPDNN) which is mostly used for classification problems. e BPNN is a supervised deep learning model where error variance between the calculated outputs and the desired outputs is back-propagated.
is process is repetitive through the learning process for minimizing the errors through weights thought the back-propagation of errors. As a consequence of weight regulations, a hidden unit sets their weight to signify significant features of the task domain. e BPDNN contains three layers which are the hidden layer, inputs layer, and outputs layer. Learning in the BPDNN is a two-step procedure [20,21].

Complexity
Step 1. Forward propagation: this step depends on the input and present weights, and the output is computed. For such computations, each hidden unit and output unit calculates net excitation which depends on these conditions: (i) Values of earlier layer units that are linked to the unit in deliberation (ii) Weights between the unit in consideration and the previous layer unit (iii) reshold value on the consideration unit is net excitation is accomplished by activation function returning the calculated output value for that unit. is activation function must be differentiable and continuous. Several activation functions can be used in the BPDNN. Sigmoid is an extensively used activation function.
Step 2. Backward propagation of error: in this step, the error is computed by variance between the actual output of each output unit and targeted output. is error is back-propagated to the former layer that is the hidden layer. Error at that node is calculated for each unit in the hidden layer N. In a similar way, error at each node of the previous hidden layer that is N-1 is calculated. Forward and backward steps are repetitive until the error is reduced up to the predictable level. e parameters of BPDNN are shown in Table 1.
e BPDNN graphically represented in Figure 2 contains three layers, the inputs layer, hidden layer, and outputs layer.

Cross Validation Method.
For data classification using hold-out methods: 70% for training and 73% for testing in this study.

Performance Evaluation
Metrics. Accuracy, model execution time, and ROC-AUC have been used as performance evaluation metrics to evaluate the performance of the model.

Experimental Results
e backward propagation deep learning-based different networks have been trained and tested with essential parameters and reported in Table 1 Figures 3 and 4, respectively. It is concluded from Figure 3 that the back-propagation neural network generates an accuracy rate of 85.88% for the proposed problem, while the LSTM-based classification and recognition model outperforms by generating an accuracy rate of 93.87% as depicted in Figure 5. e back-propagation neural network is good in sequence learning problems but fails in retaining the information used long before [22]. To address the problem of retention in back-propagation neural networks, Hochreiter and Schmidhuber proposed a modified version of the backpropagation neural network in 1997 known as long shortterm memory (LSTM) [23]. is model provided prominent results for many machine learning problems such as text recognition, speech recognition, network attack detection problems, and many others. is high applicability of the  Initial weights 3 Number of hidden units 4 Overtraining and initial stopping criteria 5 Number of instances 6 Function of activation 7 Inputs normalization Complexity 3 LSTM represents the outperformance to the vanilla recurrent neural networks (back-propagation, feed-forward propagation, and so on) significantly. e applicability of the proposed algorithm is also tested using the LSTM model. e performance results of the LSTM-based model are depicted in Figure 5, and it generates an accuracy rate of 93.87% for the proposed problem.
is high accuracy rate for the LSTM-based recognition model reflects the application of the proposed model for the said issue.
After testing the LSTM model for varying training and test sets, it is concluded from Figure 5 that the LSTM shows  an average highest accuracy rate of 93.87% for a training set of 70% and the remaining is selected as a test set. is high-accuracy value reflects the applicability of the LSTM model for the proposed problem. It is also concluded from Figure 5 that when the training set increases, the calculated time consumption also increases accordingly. e highest accuracy value of the LSTM model in Figure 5 reflects the solution to the nonretaining problem (forgetting/destroying the values used long before) of the back-propagation model.

Conclusions
Pervasive and ubiquitous computing is one of the advance paradigms in the area of information technology. e recent advances in information technology has made the world to shift from big desktop computers and have a tendency to powerful and smaller devices to facilitate with large computational and heterogeneous wireless communications interfaces. e role of pervasive computing is foremost in the field where it provides ability to distribute computational services to the surroundings where people work and leads to make issues such as trust, privacy, and identity. To provide an optimal solution to these generic problems, the proposed research work aims to implement a deep learning-based pervasive computing architecture to address these problems. Long short-term memory architecture is used during the development of the proposed trusted model. e applicability of the proposed model is validated by comparing its performance with the rival back-propagation neural network. is model results with an accuracy rate of 93.87% for the LSTM-based model much better than 85.88% for the back-propagation-based deep model. e obtained results reflect the usefulness and applicability of such approach and the competitiveness against other existing ones. In the future, the proposed research can be expanded for the recognition of unfair recommenders and its implementations on portable devices in order to validate it for real-world scenarios.
Data Availability e research has used the dataset which is available online.