Encryption Techniques for Intelligent Transportation Systems via Deep Learning for IOV in Smart Cities

Now-a-days, there is exponential growth in the field of Wireless Sensor Networks. A connected car is an essential element of the Internet of Vehicles(IoV) vision that is a highly attractive application of the Internet of Things (IoT).The underlying technologies include Internet of Everything (IoE), artificial intelligence,machine learning, neural networks, sensor technologies, and cloud/edgecomputing. The connectivity between vehicles is through inter communicationbetween sensors and smart devices inside the vehicles, as well as smart systems inthe environment as part of the Intelligent Transportation Systems (ITS).In WSN’s security is a major concern, since most of communication happen through a wireless media hence probability of attacks increases drastically. Intrusion detection as well as prevention measures should be taken for secure communication, hence observations of intrusion detection and prevention techniques have taken immense precedence in the research field.With the help of intrusion detection and prevention systems, we can categorize the activities of user in two categories namely normal activities and suspicious activities.There is a need to design effective intrusion detection and prevention system by exploring deep learning for wireless sensor networks. This research aims to deal with proposing algorithms and techniques for intrusion prevention system using deep packet inspection based on deep learning. In this, we have proposed a deep learning model using a convolutional Neural Network classifier. The proposed model consists of two stages like intrusion detection and intrusion prevention. The proposed model learn useful feature representations from a large amount of labeled data and then classifies them. In this work, a Convolutional Neural Network is used to prevent intrusion for wireless sensor networks. To evaluate and test the effectiveness of the proposed system, a WSN-DS dataset is used, and experiments are conducted on the dataset. The experimental results show that proposed system achieves 97% accuracy and performs substantially better than the existing system. The proposed work can be used as a future benchmark for deep learning and intrusion prevention research communitiesin the smart cities now a days.


INTRODUCTION
Now a day, the internet and the networks are playing a major role in business development and at the same time the network attacks are continuously changing their nature. To deals with network attacks an intrusion detection system (IDS) and intrusion prevention system (IPS) is available. With the help of an intrusion detection system, we can monitor the network for malicious activity. And when any malicious activity is detected an intrusion detection system alerts the system administrator by generating alarm, but an intrusion detection system is not able to take any action against detected malicious activity [1]. Another solution we have is an intrusion prevention system.Many times peoples think that both intrusion detection and intrusion prevention system are same but IPS is one step ahead than IDS.Basically, intrusion prevention begins with intrusion detection [2]. With the help of IPS, we can detect malicious activity and at the same time detected malicious activity is prevented by giving commands to a firewall.Many organizations are using the traditional tools like firewall, antivirus etc. to secure network but unfortunately these tools are not able to detect unknown attacks. So it is very difficult to achieve the security of networks. In such scenarios, intrusion detection and prevention is the basic requirement to secure networks. Many intrusion detection systems are available such as network based and anomaly based to detect known attacks but the accuracy of same system is low to detect unknown attacks [3]. In network based intrusion detection and prevention system, information is collected and analyzed from the packet itself to detect malicious activity in the network traffic [4]. We needed some techniques to achieve the security of network attacks quickly and accurately. There is a need to secure network from Denial of Service attacks, black hole, grey hole and many more and we can achieve this by designing an intrusion detection and prevention system using deep learning techniques. In last few years, peoples are using deep learning widely for the purpose of classification as well as pattern recognition. While using deep learning, deep learning models are generated by applying many data processing layers. In this paper, we are using the concept of deep packet inspection for the purpose of intrusion detection and prevention in WSN. Basically, deep packet inspection is a form of packet filtering in which header as well as data is examined. For the purpose of intrusion detection and prevention, few researchers used deep learning techniques. There are many more techniques in deep learning, but among these convolutional neural networks is the effective technique which gives significant performance. By using convolutional neural network, we can achieve best accuracy in the field of intrusion detection and prevention for wireless sensor networks. In this paper, we have used convolutional neural network for intrusion detection and prevention to achieve best accuracy. The IoT paradigm promises the following benefits: 1. Allows connectivity of all kind of things to develop smarter environments 2. Allows automation to make peoples' lives easier and more comfortable 3. Enables organisations to automate, increase efficiency and reduce costs 4. Enables companies to cut down on waste and improve service delivery 5. Enables companies to integrate business models and enhance productivity. With the passing of time, new technologies and mechanisms appeared andmatured, e.g. wireless and sensor technologies, machine-to-machine communication,artificial intelligence, machine learning, neural networks, cloud storage, andbig data analytics. In the last two decades, the convergence of such technologiesand related frameworks has resulted in a number of highly attractive applications ofIoT, including the following: • collected via in-car smart devices to help decision-making. In general, IoT touches every industry including manufacturing, agriculture,transportation, logistics, education, healthcare, and retail. Nevertheless, connectivityof devices, especially when these are heterogeneous and their number in thenetwork is increasing with time, imposes numerous new challenges. For example:• as number of devices increase, amount of information flow increases, which inturn has implications with respect to data safety, security, storage, and dataanalytics, as well as, decision-making in real-time The rest of the paper is organized as follows: Section II presents literature survey. Section III presents proposed work. Section IV presents results and discussion by using convolutional neural network to detect and prevent intrusion for WSN. And lastly, section V presents conclusion and future scope.

LITERATURE SURVEY
Similar work in the field of intrusion detection and prevention are presented in this section. A discussion on the existing literature related to the topic is presented below. Leila Mohammadpouret al have proposeda deep learning method to implement aneffective and flexible NIDS [5]. In this authors have used a convolutional neural network as an advanced deep learning technique. For implementation purpose authors used the NSL KDD dataset and the proposed system can detect network intrusion with detection rate of 99.97%. Yuchen Liu et al have proposed a Intrusion Detection Algorithm Based on Convolutional Neural Network [6]. For implementation, the authors used the KDD Cup dataset and the dataset consist of 10 test samples. In this authors have selected samples randomly. Basically, the proposed system can detect attacks with detection rate 97.66% and the proposed system is not able to prevent any type of attacks. SherazNaseer et al have proposed a deep convolutional neural network-based intrusion detection system [7]. Basically, the proposed system is trained and tested by using NSL KDD dataset and the proposed system can detect anomaly-based attacks. The accuracy of the proposed system is 85.22 %. Again the proposed system is not able to prevent attacks. AshimaChawla et al have proposedhost-based intrusion detection system with a combined convolutional neural network and recurrent neural network [8]. For the purpose of implementation, authors have used the ADFA dataset of system call traces. The proposed system can predict attacks. ZekiErdem et al have proposed a model network intrusion detection using deep learning [9]. For implementation purposes, authors have used the NSL KDD dataset and the proposed system can prevent zero-day attacks. Basically, to prevent zero-day attacks authors have monitored the system continuously.
FarrukhAslam Khan et al have proposeda novel two-stage deep learning (TSDL) model basedon a stacked auto-encoder with a soft-max classifier for efficient network intrusion detection [10]. Basically, the proposed system consists of two stages for the purpose of detection and each stage consists of two hidden layers and a soft-max classifier. In this, each hidden layer is separately pretrained on unlabeled data. The proposed system is trained and tested using the NSL KDD Cup99 dataset and the detection accuracy of the proposed system is 89.13%. Again the proposed system is not able to prevent any type of attack. Navapornet al has proposed an intrusion detection by deep learning with TensorFlow [11]. The proposed can classify different types of attacks and the proposed system requires only packet header information and not the payload. In this, authors have compared their results with a snort and the accuracy of the proposed system is 97%. Based on experimentation, it is concluded that deep learning is a good solution to detect all types of attacks. Chuanlong Yinet al have proposeda deep learning approach for intrusion detection using recurrent neural networks [12]. The proposed can detect attacks with high accuracy in binary and multiclass classification. For implementation, purpose authors have used the NSL KDD dataset. But the proposed system requires more time for training and in the future, the training time can be reduced. Wen Hui Lin et al have proposed a model for network intrusion detection for cyber threats using a convolutional neural network [13]. The proposed system can classify multiple threats with a higher detection rate of 95%. The proposed system can be used to enhance the precision of network intrusion detection. Again the proposed system is not able to prevent any types of attacks. Vinaykumar R et al have proposed a deep learning approach for intelligentintrusion detection systems [14]. The proposed model can detect and classify cyber-attacks. They have evaluated various datasets that are generated over theyears through static and dynamic approaches. The proposed model is going to learn the abstract and high dimensional featurerepresentation of the IDS data by passing them into multiple hidden layers. For implementation, purpose authors have used the KDD Cup99 dataset.Monika Roopak et al have proposed deep learning models for cybersecurity in IoT networks [15]. The proposed model doesn't require feature selection to be done before the classification learning and testing but with a large number of attributes in the datasets, the training time could be reduced by applying feature selection before training the model. For implementation, authors have used CICIDS2017 datasets and the proposed system can detect denial of service attack with accuracy of 97.16%. Amongst the many reasons that exist for the need of intrusion prevention systems, protection from denial of service attacks and protection from many critical exposures are the most important ones. More discussions have been made on intrusion detection system; however it is not effective enough as it deals with intrusions that have already taken effect on the system. When we want more security, then definitely we need strong system and we can achieve this by using a deep learning classifier. In our work, we are using convolutional neural network as a deep learning classifier. Our proposed system will be able to detect as well as prevent intrusion like Denial of Service, Black hole, Gray hole.

Mobile Ad Hoc Network (MANET) [2],[42][56][28]
MANET is used for communication between vehicles and roadside smart devices.In this network, also known as Wireless Ad Hoc Network (WANET) or Ad hocWireless Network, nodes (i.e. vehicles) are autonomous and mobile in nature andfree to move independently in any direction.
Nodes are, therefore, interfaceddynamically in an arbitrary fashion. Hence, the network uses ad hoc nature of routing protocols that are either table-driven or of on-demand variety.
MANETs have a self-configuring infrastructure-less network. These have thefollowing features: dynamic topology; variable capacity links; energy constrainedoperation; and limited physical security. Some MANET are generally restricted to alocal area of wireless devices, while others may be connected to the Internet. In MANET, every node is a potential router. Here, the routing protocols that areoften used include: ADOV (Ad Hoc On-demand Distance Vector), DSR (Dynamic1 Connected Vehicles in the IoV: Concepts, Technologies … 7Source Routing), OLSR (Optimised Link State Routing) and TORA (TemporallyOrdered Routing Algorithm).MANETs are decentralised and, therefore, typically more robust than centralized networks due to the multi-hop nature of data transmission. Since the MANETarchitecture evolves with time, it has the potential to resolve issues such as isolationor disconnection of nodes or the slowness of Internet data flows. Also, due to thefact that MANET topology changes, these networks are more flexible and scalable,with lower administration costs. However, with the time evolving nature of thenetwork and changing mobility patterns of vehicles and devices, there can bevariations in the network performance and quality of service. There are also concernsrelating to signal protection, reliability of nodes, limited processing power,and even the adequacy of power supply. Still, as said before, flexibility of MANETis a huge positive attribute.

Vehicular Ad Hoc Network (VANET) [36],[31],[45],[54],[60]
VANET is a subclass of MANET type of connectivity, applicable specifically tovehicles. It consists of groups of moving (and stationary) vehicles, connected via awireless network. Until recently, VANETs were used mainly to provide safety andcomfort to drivers in vehicular environments. Now, however, these networks areseen as infrastructure for an intelligent transportation system (ITS) for an increasingnumber of autonomous vehicles on the roads. One of the distinguishing characteristicsof VANET is the content-centric distribution-the content being important,rather than the source. This is in marked contrast to the Internet where an agentdemands information from a specific source.VANET communication protocols are similar to those used in wired networks,where each host has an IP address. However, here, assigning such addresses tomoving vehicles is far from trivial and often requires a Dynamic Host ConfigurationProtocol (DHCP) server, a heresy for Ad Hoc networks that operate without anyinfrastructure, using self-organisation protocols. One difference between VANETand MANET is that the routing protocols of MANET are not feasible to be used inVANET architecture. The new era of the IoT is driving the evolution of conventional VANET into thenew IoV vision. As such, IoV integrates VANET, IoT and the mobile CloudComputing. IoT and VANET have been discussed above; the Cloud paradigmprovides data storage and data analytics facilities, and provision of all kind ofservices, whether software related or virtualised hardware. In Vehicle-to-Cloud(V2C) communication, any vehicle in VANET can directly communicate with thecloud environment, and provision cloud-based services.IoV refers to the data interaction, in real time, between vehicles and roadsideunits using mobile-communication technology, vehicle navigation systems,smart-terminal devices, and information platforms. The objective is to enableinformation exchange and interaction and a driving-instruction-controlling networksystem. IoV can be regarded as an extension of Vehicle-to-Vehicle (V2V) connectivitythat enhances driving aids for full autonomous driving by furthering the vehicles'artificial intelligence awareness of the surrounding environment (including othervehicles and driving-related smart devices). The components of IoV include thefollowing [10]: Vehicles (as nodes of the ad hoc network), Roadside Units (RSU)and infrastructure (e.g. road and traffic related sensors, signals and smart objects) , personal devices (e.g. smart phones and PDAs), and humans (e.g. drivers).
Therefore, there are interactions such as the following: • Vehicle-to-Vehicle interaction • Vehicle and Infrastructure interaction • Vehicle and sensors interaction • Vehicle and RSU interaction • Vehicle and human interaction • Personal devices and vehicle interaction • Cloud and vehicle interaction • Device-to-device (D2D) interaction.

PROPOSED WORK
In this section, the proposed algorithm and the basic concepts of the proposed system are presented.
As it is presented in Figure 3.1, the Convolution layer, Max Pooling layer, Denselayer and Flatten layer is used. Figure 3.1 shows the intrusion prevention system architecture using CNN. The stages of the proposed architecture are as follows:  Raw Data In this step, we have provided dataset as an input for the purpose of preprocessing.

 Preprocessing
In this step, we have preprocessed input data.

 Test Data
In this step, we can apply test data as an input to the model. Again test data is nothing but the preprocessed data.

 Convolution Layer
Convolution layer is used to extract the features of the input data. In this work Conv1D is used, basically, it is the type of convolution and we can use it for the dataset with sequences.

 Max Pooling Layer
Max pooling layer is used to reduce the size of the representation to reduce the number of parameters as well as computation.

 Dense Layer
A dense layer is used to connect every neuron in one layer to every neuron in another layer. The dense layer uses softmax as an activation function for the classification of generated features of the input.

 Flatten Layer
Flatten layer is used, because we cannot provide rectangular or cubic shapes as a direct input to the system. Again it is in 1-dimensional form.

 Filtered Data
Lastly, we will get filtered data as an output.

Algorithm
The algorithm for the proposed system is presented and discussed below. Initially, we have provided a dataset as an input to a system. Then we have performed preprocessing on accepted input data to achieve the better results from the model. Next, we have trained the model by providing processed data as an input [16] [18], [41], [48], [51], [6]. In this, we have considered a Convolutional Neural Network model. Firstly, we have considered the convolution layer of the model. Here we have used Conv1D and basically, it works for sequential datasets [17] [19], [31], [45], [54], [60]. For the convolution layer, the size of the kernel is [4 * 4]. Secondly, we have applied the Max Pooling layer and for max-pooling, the size is [2 * 2] and basically, it is used to reduce the size of the representation and to reduce the computation in the network [18]. Thirdly, we have applied the Dense layer and basically, with the help of the Dense layer every neuron in one layer to every neuron in another layer is connected [19] [21], [26], [43], [59], [27]. The size of the dense layer is 50 and to prevent the overflow the value for dropout 0.2 is considered. For the purpose of optimization, an Adaptive Moment Estimation (adam) is used and the number of an epoch is set to 10. Again in this, the Rectified Linear Unit (ReLU) is used as an activation function [20] [22], [33], [47], [55], [61] and this ReLU uses the Softmax activation function. Fourthly, we have applied the Flatten layer. Basically, we cannot provide input to the system in rectangular or cubic shapes, so we have used the Flatten layer and it is in 1-dimensional form. Then, we have tested the proposed model and for testing, we have provided Test Data as input [23]. Initially, we have considered the value of Attack Type as Null and at the same time, we have initialized the value of Status equal to zero means they all attack type are considered as normal data [24]. Then, we have used to detect function to check the status. If the value of status is equal to zero then the data is considered as normal data and the message -Normal‖ is printed. If the value of status is not equal to zero then the data is considered abnormal and then we have checked the attack type of that data and the attack type of that data is printed. Lastly, we have evaluated the accuracy and the confusion matrix is printed.

COMPARASION WITH PREVIOUS WORKS
As the volume of the collected data increases, Deep Learning and Machine Learning (ML) techniques are applied to further enhance the intelligence and the capabilities of an application. The field of smart transportation has attracted many researchers and it has been approached with both ML and IoT techniques. In this, smart transportation covers route optimization, parking, street lights, accident prevention/detection, road anomalies, and infrastructure applications. The purpose of this paper is to make a self-contained review of ML techniques and IoT applications in Intelligent Transportation Systems (ITS) In Smart cities and obtain a clear view of the trends in the a fore mentioned fields and spot possible coverage needs that there is a great potential in Deep Learning coverage for the Smart Lighting Systems and Smart Parking applications so this proposed techniques is very useful in this applications. Wireless Sensor Networks (WSN), and Information and Communication Technology (ICT) have the potential to address some of the environmental, economic, and technical challenges as well as opportunities in this sector. As the number of interconnected devices continues to grow, this generates more big data with multiple modalities and spatial and temporal variations. Intelligent processing and analysis of this big data are necessary to developing a higher level of knowledge base and insights those results in better decision making, forecasting, and reliable management of sensor

RESULTS AND DISCUSSION
This section gives an idea about the experimentation. In this section, the Convolutional Neural Network classifier is implemented. For implementation, python is used as a scripting language and WSN-DS is used as a dataset. In this experiment, the Convolutional Neural Network is used for classification. The model has been implemented using python and Tensor-Flow on a system equipped with 16GB RAM and CPU Intel Corei7 and Windows 10 system. A WSN-DS dataset consists of 374000 entries and the total entries are divided into different labels as normal, black hole, gray hole, flooding, and TDMA. The proposed system can prevent attacks like a black hole, gray hole, flooding, and TDMA [23]. To check whether data is normal or attack the fields considered are as Is_CH, ADv_S, ADV_R, JOIN_S, JOIN_R, SCH_S, SCH_R, DATA_S, DATA_R, send code [16]. A WSN-DS dataset is shown in  Id-It is the node id that is used to distinguish the node from another node. Is_CH-It is the flag that is used to distinguish the node as a cluster head or normal node.Who CH-It is considered as a id of the current node.ADV_S-It is considered as a number of messages broadcast by cluster head to the normal node. ADV_R-It is considered as a number of messages received by cluster head from normal node JOIN_S-It is considered as a number of join request messages sent from the normal node to the cluster node. JOIN_R-It is considered as a number of join request messages received by cluster head from a normal node. SCH_S-It is considered as a number advertise schedule messages sent to the nodes. SCH_R-It is considered as a number advertise schedule messages received from the cluster head. DATA_S-It is considered as a number of data packets sent from a normal node to its cluster head. DATA_R--It is considered as a number of data packets received from the cluster head. Send Code-It is considered as a cluster sending code. Attack Type-It is considered as a type of attack or simply labels.

Normal:
To check whether the data transmitted is normal following two cases are considered. Case I: If the node is a cluster head then we have assigned value Is_CH=1. Then this node is going to broadcast an advertising message with value ADV_S=1 and at the same time, the value of ADV_R is null. After sending an advertising message this node receives join requests from other nodes like JOIN_R and then sends scheduling message SCH_S. Lastly, we have considered some value for DATA_R and DATA_S and send code is considered with the value null. Case II: If the node is a not cluster head then we have assigned value Is_CH=0. Then this node is going to receive an advertising message with value ADV_R=4 and at the same time, the value of ADV_S is null. After receiving an advertising message this node sends join requests to other nodes like JOIN_S and then sends scheduling message SCH_R. Lastly, we have considered some value for DATA_Sand send code and DATA_S is considered with the value null.
If the system can satisfy the above two cases then we can say that the transmitted data belongs to the category of the normal type.Both the cases for the normal category are shown in Table 4  In the black hole, the attacker node receives data from the source node and the received data is not forwarded to the destination node [24]. Simply attacker node drops all packets. To check whether the data transmitted is normal or not the following two cases are considered. Case I: If the node is cluster head we have assigned value Is_CH=1, then this node is going to broadcast advertising message to another node as ADV_S=1 and at the same time is going to receive advertising requests from other nodes as ADV_R=2. Again in this, we have considered the value of JOIN_S is equal to null as JOIN_S=0 and the value of JOIN_R is equal to non-null as JOIN_R=2. Then this node will send scheduling messages as SCH_S=1 and receives data message as DATA_R=5. And lastly, the value of send code =0. Case II: In this case, actually, the node is not the cluster head but the same node is telling identity as a cluster head. So we have considered the value of this node as Is_CH=1. Then this node is going to broadcast advertising messages to another node as ADV_S=1 and at the same time is going to receive advertising requests from other nodes as ADV_R=26. Again in this, we have considered the value of JOIN_S, as well as JOIN_R, which is equal to null as JOIN_S=0 and JOIN_R=0 respectively. Again in this the value of DATA_R and send code is considered as null as DATA_R=0 and send code=0. Both the cases for the blackhole category are shown in Table 4.3. In the gray hole, the attacker node receives data from the source node and the received data is not forwarded to the destination node [25] [26,[27. Simply attacker node drops some packets. To check whether the data transmitted is normal or not the following two cases are considered. Case I: If the node is cluster head we have assigned value Is_CH=1, then this node is going to broadcast advertising message to another node as ADV_S=1 and at the same time is going to receive advertising requests from other nodes as ADV_R=2. Again in this, we have considered the value of JOIN_S is equal to null as JOIN_S=0 and the value of JOIN_R is equal to null as JOIN_R=0. Then all other values are considered as scheduling messages as SCH_S=0and receives data message as DATA_R=0. And lastly, the value of send code =0. Case II: In this case, actually, the node is not the cluster head but the same node is telling identity as a cluster head. So we have considered the value of this node as Is_CH=1. Then this node is going to broadcast advertising messages to another node as ADV_S=1 and at the same time is going to receive advertising requests from other nodes as ADV_R=26. Again in this, we have considered the value of JOIN_S, as well as JOIN_R, which is equal to null as JOIN_S=0 and JOIN_R=0 respectively. Again in this the value of DATA_R and send code is considered as null as DATA_R=0 and send code=0. Both the cases for the grayhole category are shown in Table 4.4.

Flooding:
Basically in flooding attacks, the bogus message is generated to increase the network traffic for consuming server or network resources. To check whether the data transmitted is normal or not the following two cases are considered.

Case I:
If the node is a cluster head then we have assigned value Is_CH=1. Then this node is going to broadcast an advertising message with value ADV_S=1 and at the same time, the value of ADV_R is null. After sending an advertising message this node receives join requests from other nodes like JOIN_R =90 and then sends scheduling message SCH_S. Lastly, we have considered some value for DATA_S=1350 and DATA_S= 15 and send code is considered with the value null.

Case II:
If the node is a cluster head then we have assigned value Is_CH=1. Then this node is going to broadcast an advertising message with value ADV_S=7 and at the same time, the value of ADV_S is non-null. After broadcasting advertising message this node receives join requests from other nodes like JOIN_R and then sends scheduling message SCH_S. Lastly, we have considered some value for DATA_S= 238, DATA_S= 238and the value of send code is considered as null. Both the cases for the flooding category are shown in Table 4.5.     The table 4.9 shows the confusion matrix, it gives the value of precision, recall, F1-score, and accuracy. The overall accuracy of our proposed algorithm is 97%.  The figure 4.1 shows the graphical representation of attacks prevented using CNN. In the figure, precision, recall, and F1-score are given for attacks prevented using CNN. Accuracy for Normal, Blackhole, Grayhole, Flooding and TDMA are 99%, 94%, 89%,73% and 63% respectively. The figure 4.2 shows the results of attacks prevented using CNN with different data sizes. The figure shows the results of Precision, Recall and F1-Score for different data sizes. The proposed model is tested on the same dataset but every time size of the dataset considered is different. Based on results obtained, we can say that if we vary the size of the dataset the accuracy remains the same so that the size of data is not going to affect the proposed model.

CONCLUSION AND FUTURE SCOPE
From the discussions in the above sections, we understand the need to design a system that can prevent intrusions in a WSN. Due to anything, anytime, anywhere, the type of computing use of WSN has increased significantly. Looking at the various threats and several attacks in a wireless network,and security is the prime concern. Intrusion is one of the critical issues, and intrusion detection has already reached its saturation. Therefore, we need a scalable and attack resistance proactive intrusion prevention system using deep packet inspection in a wireless sensor network. Artificial Intelligence (AI), sometimes called Machine Intelligence, refers to intelligencedemonstrated by machines. The term is often used to describe machines(e.g. computers and cars) that mimic cognitive functions that humans associate withlearning and problem solving. Machine intelligence is gained by performing statisticaloperations on data to discover patterns therein, rather than intelligence beingexplicitly programmed within the systems. AI can be categorized into three different types, as mentioned below: • Analytical: this has characteristics consistent with cognitive intelligence,including learning based on prior experience • Human-inspired: this has elements from cognitive and emotional intelligencethat contribute to the decision-making processes. • Humanised: this has characteristics of several different types of competencies(e.g. cognitive, emotional, social, etc.) and has the ability to be self-aware.Machine Learning (ML), sometimes called Deep Learning, is the application ofAI. It enables a system to automatically learn and improve with experience, withouthuman intervention, beyond the initial programming that sets up the learningprocess. ML methods include: supervised ML algorithms, unsupervised ML algorithms,semi-supervised ML algorithms, and reinforcement ML algorithms. Somewell-known classes of algorithms include: genetic algorithms, rulebased ML,learning classifier systems, similarity and metric learning, clustering algorithms,Bayesian algorithms, neural networks, decision tree learning, etc. ML is currentlythe most promising path to strong AI [18] [29], [39], [44], [59], [51].In the context of connected vehicles, some useful features of AI-based systemsinclude speech recognition, driver monitoring, virtual driver assistance,camera-based vision systems, and radar-based detection (of other vehicle and roadside units) systems.
In this paper, Convolutional Neural Network algorithm is implemented to detect and prevent intrusion for Wireless Sensor Network. For implementation purposes, a WSN-DS dataset is used as an input to our system. At the time of implementation, different size of the dataset is used, and based on that, we can conclude like if we vary the size of a dataset, and accuracy of our algorithm not gets affected. In this paper, the results ofimplementation using Convolutional Neural Network with different data sizes are presented. So, Convolutional Neural Network algorithms are better to detect and prevent intrusion for WSN. Through the literature survey, we understand that there is a need to develop a scalable and attack resistance system for intrusion prevention using deep packet inspection in a WSN. A system is proposed to detect and prevent intrusion using the Convolutional Neural Network for the WSN and the accuracy of intrusion prevention with our system is 97%. Hence, the proposed system can be used to detect and prevent attacks in all types of networks where security matters.