1. Introduction
A vehicular ad hoc network (VANET) is a superior type of mobile ad hoc network (MANET).
In a VANET, automobiles as well as roadside units (RSUs) are connected to generate
a safe and effective driving atmosphere. A VANET contains three parts: i) on-board
units (OBUs), ii) roadside units (RSUs), and iii)a trusted authority (TA). Each vehicle
has an OBU that gathers, evaluates, and spreads information to other automobiles in
the area [1]. The RSU is connected with the roadside and communicates with vehicles, substructure,
and a TA [2].
Due to the decreasing amount of road accidents, a VANET provides an effective traffic
information system (TIS) that shares information about accident and traffic to vehicles.
The main goal of a VANET is to protect travelers, drivers, and a vehicle by means
of TIS [3]. However, a VANET faces several security issues, such as wormhole attacks, black
hole attacks (BHAs), and gray hole attacks (GHAs), which disturb usual routine of
networks. ABHA is a type of attack that aims to disturb the network communication
by fabricating several vehicle characteristics [4]. In a BHA, a malevolent node uses a routing protocol in order to advertise itself
for consuming the shortest path to a packet that it wants to interrupt. Thus, to manage
these security threats, intrusion detection systems (IDSs) were presented [5].
It is still difficult to formulate an IDS designed for a VANET with high vehicular
mobility [6]. For this, detection and prevention of a BHA in a VANET is most important [7]. Hence, several attack detection, attack reaction, and attack prevention methods
were suggested, but the existing methods do not provide sufficient attack detection
accuracy and have high computational time, which motivated us to do this work [8]. In this manuscript, a method for discovery and prevention of a BHA is proposed using
a tree hierarchical deep convolutional neural network and enhanced identity-based
encryption (BCMO-THDCNN-BHA-VANET) for accurately detecting BHA in a VANET [9].
Detection using clustering
VANET share wireless networks where vehicles communicate with each other to improve
traffic safety and efficiency. One of the important tasks in VANETs is to detect and
prevent accidents. Clustering is a technique used in VANETs to group vehicles based
on their geographical location, speed, and direction. Clustering also allows vehicles
to share information with their neighboring vehicles efficiently.
To use clustering for detecting accidents in VANETs, the following steps can be taken:
1. Cluster formation: Vehicles in the network are grouped into clusters based on their
geographical location, speed, and direction. This clustering can be done using different
approaches. K-means, hierarchical clustering, and density-based clustering are examples
of such approaches.
2. Cluster head (CH) selection: Each cluster has a CH, which is responsible for collecting
and disseminating information to other vehicles in the cluster. The selection of a
CH can be done using optimization algorithms.
Intrusion detection system (IDS)
A VANET IDS is a security device designed to identify and prevent harmful actions.
The IDS in a VANET monitors the network for unusual or suspicious behavior that might
signal an attack or intrusion. The IDS can aid in the prevention of numerous sorts
of assaults, such as denial of service, black hole, wormhole, and Sybil attacks.
The IDS in a VANET typically consists of two main components: anomaly detection and
misbehavior detection. Some common techniques used in IDSs for VANETs include the
following. Signature-related identification entails compiling a database of well-known
attack patterns or signature and matching them against the network traffic to detect
attacks. This approach analyzes network data and identifies deviations from usual
behavior. Reputation-based detection involves assigning a reputation score to each
vehicle based on its past behavior and using this score to identify malicious or selfish
behavior.
The IDS can operate in a centralized or distributed manner. In centralized IDS, a
central authority is responsible for monitoring the network and detecting attacks.
In distributed IDS, each vehicle is responsible for monitoring its own behavior and
detecting attacks. By detecting and preventing attacks, the IDS help to ensure the
reliability and effectiveness of vehicular communication and improve the overall safety
of the network.
The following is the work that was carried out,
· In this manuscript, detection with prevention of black hole attacks depending on
BCMO-THDCNN-BHA-VANETis proposed for accurately detecting a BHA in a VANET.
· The automobiles are initially organized in clusters using an improved k-means clustering
technique. The CH is chosen after the cluster formation using the balancing composite
motion optimization (BCMO) algorithm.
· The arrival of a malicious node in the cluster occurs after the selection of the
CH. THDCNN is proposed to categorize cluster nodes into two types: i) BHA nodes and
ii) normal nodes.
· If a BHA node is found, attack node information informs a specific CH to make a
final conclusion. Finally, the proposed method prevents the normal node data from
the attacker.
· A simulation of the model was done in a network simulator, and the performance metrics
were calculated.
· Finally, the proposed approach was analyzed with existing methods, such as generative
adversarial networks (GANs) for BHA detection (GAN-BHA-VANET) [10], deep neural networks (DNNs) with a thresholding algorithm (DNN-TA-BHA-VANET) [11], and random forest for BHA detection (RF-BHA-VANET) [12].
The rest of this manuscript is organized as follows. Related works are described in
section2, the proposed method is described insection3, the outcomes are demonstrated
in section4, and the conclusion is presented in section5.
3. Proposed Methodology
Ina VANET, identification with prevention of a BHA is done with the help of TCNN and
EIBEA. Fig. 1 shows a block diagram of the proposed method. In this figure, the VANET environment
is presented, including cars, RSUs, and internet and cloud servers. AnRSU is a fixed
communication point connected to the side of a road for vehicle-to-internet communication.
The proposed technique covers four processes: (i) cluster formation using improved
k-means clustering, (ii) CH selection uses balancing composite motion optimization
(BCMO), (iii) attack detection using the THDCNN method, and (iv) securing the VANET
using EIBEA.
Fig. 1. Block diagram of proposed BCMO-THDCNN-BHA-VANET methodology.
3.1 Improved k-means Clustering Method (IKMC)
Initially, the cluster formation of vehicles is arranged using the improved k-means
clustering method (IKMC). The $K$ value is typically a challenge to describe. The
selection of $K$directly defines which data clusters must be grouped into numerous
clusters. At the start of the algorithm, the value is expected to provide significant
benefits for cluster formation. The main idea behind IKMC is to generate a sequence
of values by taking the square of the distance among example points in each cluster
and the cluster centroid $K$values.
The sum of squared errors (SSE) is a performance metric, iterationon $K$value and
compute SSE. SmallSSE suggestsevery cluster is extra convergent than rest. SSE rapidly
decreases when the number of clusters is chosen to approximate the number of genuine
clusters. When the count of clusteris more than the count of actual cluster, SSE will
continue to fallat a slower rate. The$K$value can be found by using a K-SSE curve
andlocating a modulation point. As shown in Fig. 2, there is a clear modulation point at $K$= 2, so the clustering effect is strongest,
as shown in Fig. 3.
Selecting$K$ value lesser than actual value, value of costsignificantlycondensed for
each 1 growth of k; selected k is higher than true$K$, variation of cost value not
so clearat each 1 growth of$K$. Thus, an accurate$K$ value will be at the turning
point, as given, there is anactual clear point at $K$= 2
The built distribution's clustering results were compared to find the best number
of clusters.
where${E}_{n}^{*}\left(\log \left(W_{k}\right)\right)$ indicates $\log \left(W_{k}\right)$
expectations. $K$conforming to an extreme value of $O_{n}\left(K\right)$ is a better$K$
satisfied minimal$K$of $O_{n}\left(K\right)\geq O_{n+1}\left(K\right)$ as the optimal
number of clusters.
Initially, the vehicles are arranged in a cluster form using IKMC. The vehicles’ mobility
should be observed while clusters form in a VANET, and they are categorized by location
organized with the speedof vehicles. For the$j^{th}$vehicle with the $l^{th}$ centroid
($1\leq j\leq m,1\leq l\leq L$), the locationor position is denoted as$y_{j}\,\,and\,\,{y}_{l}^{g}$,
and the speeds are denoted as $s_{y,j}\,\,and\,{s}_{g}^{l}$. The Euclidean distances
between the present node position and upcoming node positions$Dis(j,l)$are expressed
in Eq. (2).
In Eq. (2), $y_{j}$represents the present node position, and the forecasted upcoming node position
of node $j$is denoted as$(y_{j})*$. The upcoming location of node $j$ is predicted
at $T_{e}$seconds. The objective function of the cluster is expressed in Eq. (4).
To derive the centroids’ optimum locations, the partial derivation of $Fun_{IKMC}$
is regarding the centroids’ location, explicitly${y}_{l}^{g}\,,1\leq l\leq L$.This
can be formulated with Eq. (5)
In Eq. (3), let $\frac{\partial e}{\partial {y}_{l}^{g}}$=0. Hence, the $l^{th}$ centroid optimum
position ${y}_{l}^{g}\,$can be found with Eq. (6).
Using IKMC, vehicles are precisely arranged in a cluster form based on Eq. (6). Every cluster contains a CH, which is chosen with the help of the BCMO algorithm.
Fig. 2. Optimal $K$ selection using IKMC algorithmand choosing the value of K.
Fig. 3. Clustering effect when $K$=2.
3.2 CH Selection using BCMO
Following the cluster formation, the CH is selected. BCMO is a method for optimizing
the CH selection process. In this paper, each node is approximated as a graph vertex,
and the distance between cars is the road vehicle’s setup and is taken, which is denoted
as boundaries. The distance between the two vehicles is evaluated with the help of
Eq. (7).
In Eq. (5), $a"and\,b"$ are coordinates of the vehicles, and$distance$ indicates the vehicles’
distance. Every vehicle chooses a CH to form a cluster group and computes the shortest
paths between the vehiclesby using the BCMO algorithm.Hence, the parameter of distance
is optimized with the BCMO approach.
The BCMO algorithm is a population-based optimization approach that was inspired by
the assumption that the solution space is Cartesian and that searchmotions of candidate
solutions arestabilized by global and local ones. Then, it optimally calculates the
shortest path between the nodes (vehicles) for efficient CH selection. The procedure
of BCMO is given below.
Step 1: Initialization
The BCMO population is initialized uniformly in the solution space based on Eq. (8):
where${x}_{i}^{L}\,and\,\,{x}_{i}^{U}$represent the upper and lower boundaries of
$i^{th}$ individuals, and $rand$denotes a $d$dimensional vector with a uniform distribution
in a range of [0,1].
Step 2: Random generation
Following the initialization process, BCMO input parameters aremaderandomly. The shortest
distance is selected based on the fitness function.
Step 3: Computingthe fitness function
A random solution countis made from the startingvalue, and the fitness function is
proportional to the distance between nodes. Eq. (9) computes the fitness function.
The smallest distance for optimum CH selection is found using Eq. (9).
Step 4: Updating position in solution space of BCMO to find the shortest path
To balance the capacity of every personalityin the search space to explore and exploitthem,
the chance of allocating positive and negative indicators $v'_{\frac{i}{j}}+v'_{j}$
in local and global search mustbethe same.The updated location of the $i^{th}$individual
at next groups formed in Eq. 10).
here$X{"}_{i}^{t'}$represents the input search, and the updated location will optimize
according to Eq. (10) distance parameter of the vehicles. Thus, the optimal CH will be obtained.
Step 5: Termination
We stop the procedure after finding the best answer or continue using Eqs. (8) to
(10) until the requirements are fulfilled. The result of the BCMO algorithm yields
the optimum CH, which is repeated until the halting requirement $t=t+1$is met.
3.3 BHA Detection with THDCNN Method
The arrival of amaliciousnode in cluster occurs after selectingthe CH. THDCNN is proposed
to categorize cluster nodes as either BHAnodes. Using THDCNN, each CH node examines
its neighboring node for maliciousness. The tested BHAnode and normal node will be
categorizedby using the training and testing dataset of THDCNN.
THDCNNcan be used for IDSs in VANETs, which involves detecting attacks and malicious
activities in the network. The structure of THDCNN for attack detection in a VANET
is similar to the architecture used for vehicular detection and classification.
The THDCNN for attack detection in a VANET typically includes the following components.
The completely linked layers are utilizedto achieve the final traffic data categorization.
The output layer produces the final classification result, indicating whether the
traffic is normal or an attack.
The THDCNN for attack detection in a VANET is trained using a supervised learning
approach, where the ground truth labels for each input traffic data are provided.
By detecting and preventing attacks in realtime, the THDCNN can help to ensure the
reliability and safety of the network. Fig. 4 shows the architecture of THDCNN.
The THDCNN model is inspired by hierarchical classifiers. The THDCNN model is made
up of several nodes that are linked in a tree-like fashion. In addition, except for
the leaf nodes, each node in the design comprise a deep convolution neural network
that has been trainedto categorize input examples to node to children nodesof the
tree. Prediction is carried out at the root node in this model since it is the greatest
node in the tree. Thus, the branch neural network's output node is derived from leaf
nodes, which are the outcome of a second phase branch.
Initially, the THDCNN model is trained using datasets of $D=\left\{1,2,3.......,n\right\}$
with n data points. The proposed method includes a neural network with numerous layersand
acts as a root node and multipleleaf nodes.Also, task defined to method to predict
LU/LC changes with $M$classes. The output node of THDCNN offers a three-dimension
matrix$D^{K\times M\times N}$, where $K$signifies the overall count of root node children,
$M$denotes the amount of novel samples, which is revealed$N$ for every class.
$D(k,m,n)$is the classified outcome of the $k^{th}$neuron for$n^{th}$ data in the$m^{th}$class.
Here, the value of $k^{th}$neuron belongs to [1,K], the $m^{th}$class belongs to [1,M],and
the $n^{th}$ data belongs to [1,N]. Thus, the average classified outcome of $N$data
is denoted as${D}_{avg}^{K\times M}$and is computed using Eq. (11).
Also, the probability from the softmax function is computed on${D}_{avg}^{K\times
M}$, and the probability matrix$R^{K\times M}$ is calculated with Eq. (12)
where$e^{{D_{avg}}(k,m)}$denotes the model's arithmetical information, which is employedto
improve the model's accuracy.
The ordered list $(L)$ of cluster nodes is made from$R^{K\times M}$, which has a property.
Here, the ordered list$(L)$ has examples of data$\left[N_{1},N_{2},N_{3}\right]$,
where every dataset contains $m$loads. Also, the output valuesof a cluster node are
organized in descending order$\left[N_{1}\geq N_{2}\geq N_{3}\right]$. As a result,
an arrangement is made to validate that the malicious nodes have a large probability
value, which is suppliedto the THDCNN model's leaf node. Finally, the THDCNN model
correctly identifies the cluster node as a BHA node or a regular node. If a BHA node
is discovered, the information about theattack node is forwarded to the appropriate
CH for a final decision. Otherwise, normal node data is encrypted and stored in the
cloud.
Fig. 4. Architecture of tree hierarchical deep convolutional neural network.
3.4 Security of VANET using Enhanced Identity-based EncryptionAlgorithm(EIBEA)
EIBEA is used to encrypt regular node data. A faster version of role-based categorization
is enhanced identity-based encryption.Users are allowedto access the information in
this method by means of their identity as a way of confirmation. This method was firstly
applied in a proxy serverto cancelillegal operators. The authorized users’ identities
are stored in proxy servers and are revoked ifthere is no identical key for that specific
identity whenitattemptsto use the server's service. As a result,in order to utilize
the service, each user must first register their identification. This consists of
four phases.
Setup phase: Initially,a public key and master key arecreated. A secret parameter
is selected from finite group$\zeta _{p}$. A random generator $gen$is selected from
cyclic group $G"$ , and then $gen\in G"$ , fix $gen_{1}=g^{\alpha }$, and choose $gen_{2}$in
$G"$. After selecting all security parameters, a random number $u_{1}$($u_{1}\in G_{1}$)
and a random $m$-length vector are chosen ($U_{1}=\left\{u_{1}\right\}$). Lastly,
$gen,gen_{1},gen_{2},u_{1}\,and\,U_{1}$are referred to as public keys, and is a master
key
Private key generation phase: Let $s$denote the $m$-bit identity of a user. The $j^{th}$
bit of$s$is $s_{j}$. Identity $s$ is formed by selectinga random measure, and $r$represents
a random number. A private key with personality is shown in Eq. (13).
Encryption phase: Let $"t"$be a random parameter selected in $\zeta _{p}$ and message$Mssg(Mssg\in
G_{1})$ . Then, the encryption key corresponding to the identity $s$can be expressed
in Eq. (14).
where$e$represents a bilinear map.
Decryption phase: Let $T=(T_{1},T_{2},T_{3})$ be the legal cipher text for message
$Mes$ with the operator identity$s$. Then, cipher text $T$can be decrypted by means
of $d"_{s}=(d"_{1},d"_{2})$as formulated in Eq. (15).
Finally, the EIBEA method prevents the normal node data from an attacker.
4. Results and Discussion
The efficiency of the proposed BCMO-THDCNN-BHA-VANET methodology was analyzed on the
basis of performance metrics. A simulation was completed with the NS-2 simulator,
which predicts network performance. It wasrun on a PC with the Windows 10 operating
system, 2GB of RAM, and an Intel i3 core CPU. Performance metrics such as recall,
accuracy, F-measure, and specificity were examined.
4.1 Dataset Description
The experiments were doneon a dataset that is publicly available. 50% of the datasetwas
used for training, and 50% was used for testing.
4.2 Performance Metrics
For identifying and categorizinga BHAnode and normal node, performance measures such
as precision, accuracy, specificity, recall, and F-scorewere examined.
· True positive ($A'A'$): The number of VANET network connections correctly identified
as BHAs.
· True negative ($N'N'$): The number of VANET network connections correctly identified
as normal connections.
· False positive ($A'N'$): The number of VANET network connections incorrectly identified
as BHAs.
· False negative ($N'A'$): The number of VANET network connections incorrectly identified
as normal connections.
Artificial datasets
To assess the efficacy of the proposed technique, certain fictitious datasets were
constructed and clustered using IKMC. Each dataset was considered to include three
clusters (A, B, and C) and varying vehicles within each cluster. The x-coordinates
and y-coordinates in cluster A were created individually from normal distributions
using mean ${\mu }_{x}^{A}$and standard deviation$\sigma ^{A}$, which were specified
as$N\left({\mu }_{x}^{A},\sigma ^{A}\right)$and $N\left({\mu }_{y}^{A},\sigma ^{A}\right)$,
respectively. Likewise, x- and y-coordinates in cluster B were created from $N\left({\mu
}_{x}^{B},\sigma ^{B}\right)$ and$N\left({\mu }_{y}^{B},\sigma ^{B}\right)$. However,
a vehicle in cluster C was madesomewhat differently. A proportion of vehicles in cluster
C were assumed to have a large standard deviation${\sigma }_{L}^{C}$, say, the standard
deviation of rest ($\sigma ^{C}$), while the means of x- and y-coordinates of every
vehicle in cluster Care similar(${\mu }_{x}^{C}$ and ${\mu }_{y}^{C}$). After 10%
of vehicles are stated, the x- and y-coordinates of vehicles in cluster C are made
from $N\left({\mu }_{x}^{C},{\sigma }_{L}^{C}\right)$ and$N\left({\mu }_{y}^{C},{\sigma
}_{L}^{C}\right)$, respectively, and the 2 coordinates of rest vehicles are in cluster
C are created from $N\left({\mu }_{x}^{C},\sigma ^{C}\right)$. The selected parameters
are given in Table 2.
To compare the performance of the IKMC method with WKM, MKHM, and KNN, the adjusted
Rand index was used. The adjusted Rand index was suggested by Hubert and Arabie and
is commonly utilized to compare cluster outcomes if an exterior principle or true
division is known. If $U$and $V$denote2 various partitions of vehicles under deliberation,$U$is
a true divider, and $V$is a cluster outcome, the adjusted Rand index for clustering
outcome $V$is computed using Eq. (17) as follows:
where$a$is the count of pairs of vehicles locatedin a similar class in $U$and in a
similar cluster in $V$, $b$is the count of pairs in an identical class in $U$but not
in an equal cluster in $V$, $c$is the count of pairs in a similar cluster in $V$but
not in a similar class in $U$, and $d$is the count of pairs in a dissimilar class
in $U$and various clusters in $V$.
Table 2. Parameters used when generating a vehicle.
|
|
Cluster A
|
Cluster B
|
Cluster C
|
|
Mean
|
${\mu }_{x}^{A}$=0, ${\mu }_{y}^{A}$=0
|
${\mu }_{x}^{B}$=6, ${\mu }_{y}^{B}$=-1
|
${\mu }_{x}^{C}$=6, ${\mu }_{y}^{C}$=2
|
|
Standard deviations
|
$\sigma ^{A}$=1.5
|
$\sigma ^{B}$=0.5
|
$\sigma ^{C}$=0.5, ${\sigma }_{L}^{C}$=2
|
Table 3. Adjusted Rand indices ofvarious clustering methods.
|
% Vehicles
|
WKM
|
MKHM
|
KNN
|
IKMC (proposed)
|
|
0
|
0.7903
|
0.7679
|
0.8231
|
0.9629
|
|
5
|
0.8376
|
0.7534
|
0.8198
|
0.9335
|
|
15
|
0.7957
|
0.7288
|
0.6782
|
0.9189
|
|
25
|
0.7708
|
0.7053
|
0.8123
|
0.9904
|
|
35
|
0.7595
|
0.7782
|
0.6673
|
0.9609
|
Table 3 compares the derived adjusted Rand index of each approach with a varied fraction
of noisy items included. Table 3 clearly shows that the suggested IKMC approach outperforms the WKM, MKHM, and KNN
clustering methods.
Performance of the proposed method according to initial K selection
The proposed IKMC technique includes one method for determining the initial K in Step
1. Obviously, the performance will differ depending on how the initial K is chosen.
Other options may include the following:
Method 1 (WKM): Choose k vehicles at random from all available vehicles.
Method 2 (MKHM): Sort all cars in the order of the variable's values,divide the values
into k equal intervals, and choose one car at random from each interval.
Method 3 (KNN): Take 10% of all cars at random as a sample and use the algorithm to
do preliminary clusteringon these sampled vehicles. The k clustering results are utilized
as the initial clustering.
To compare the various approaches for determining starting K, a datasetwas created
in the same manner as before with 10% of cars in class C. Table 4 shows the findings for various total item counts and ways of picking the starting
K. It may be interesting to note that methods 2 (MKHM) and 3 (KNN) appear to be worse
than method 1 (WKM). Approach 3 was predicted to outperform the suggested IKMC approach,
and this is true when the number of vehicles is relatively big. It is possible to
infer that the suggested IKMC method's first K selection performs quite well as compared
with the WKM, MKHM, and KNN clustering techniques.
Table 4. Adjusted Rand indices withdifferent initial K selection.
|
# vehicles
|
Method 1 (WKM)
|
Method 2 (MKHM)
|
Method 3 (KNN)
|
IKMC (proposed)
|
|
300
|
0.8456
|
0.68002
|
0.71532
|
0.93927
|
|
600
|
0.82134
|
0.6562
|
0.78439
|
0.92889
|
|
900
|
0.81601
|
0.65237
|
0.70749
|
0.92832
|
|
1200
|
0.84926
|
0.63543
|
0.76650
|
0.93135
|
|
1500
|
0.81001
|
0.63680
|
0.75256
|
0.92939
|
|
1800
|
0.83955
|
0.63531
|
0.77791
|
0.93771
|
|
2100
|
0.79899
|
0.59487
|
0.72579
|
0.92736
|
|
2400
|
0.82880
|
0.67166
|
0.74584
|
0.93755
|
|
2700
|
0.78849
|
0.65068
|
0.73119
|
0.93284
|
|
3000
|
0.80911
|
0.65120
|
0.71507
|
0.92201
|
4.3 Performance Analysis
Figs. 5-10 show the performance analysis of the proposed BCMO-THDCNN-BHA-VANET technique.
Performance metrics like the accuracy, specificity, recall, precision, F-score, and
computation time were examined. The performance of the proposed approach was analyzed
with GAN-BHA-VANET, DNN-TA-BHA-VANET, and RF-BHA-VANET models.
Fig. 5 demonstrates the accuracy analysis. Here, the proposedapproach provides 23.05%, 32.05%,
and 32.10% higher accuracy for a normal node and 29.68%, 32.57%, and 44.28% higher
accuracy; Fig. 6 demonstrates the specificity analysis. Here, the proposed approach provides 32.05%,
40.28%,and 41.28% higher specificity for a normal node and29.65%, 30.24%, and 35.24%
higher specificity; Fig. 7 demonstrates the recall analysis. The BCMO-THDCNN-BHA-VANET method provides 30.27%,
37.59%, and22.05% higher recallfor a normal node and32.05%, 21.05%, and 23.05% higher
recall; Fig. 8 demonstrates the precision analysis. Here, the proposed approach provides 41.27%,
28.57%, and 34.20% higher precision for a normal node and23.58%, 27.38%, and 25.48%
higher precision for aBHA node compared to GAN-BHA-VANET, DNN-TA-BHA-VANET, and RF-BHA-VANET,
respectively.
Fig. 9 demonstrates theF-score analysis. BCMO-THDCNN-BHA-VANET method provides 29.65%, 32.07%,
and36.52% higher F-score for a normal node and30.24%, 22.15%, and 23.05% higher F-score;
Fig. 10 demonstrates the computation time analysis. The BCMO-THDCNN-BHA-VANET method provides
29.65%, 32.07%, and36.52% lower computation time compared to GAN-BHA-VANET, DNN-TA-BHA-VANET,
and RF-BHA-VANET, respectively.
Fig. 5 presents values for accuracy parameters for a BHA node as well as a normal node.
The maximum accuracy was around 99.66%.
It was observed that in the detection of aBHA node, the proposed BCMO-TCNN-BHA-VANET
approach’s performance continues to be substantially more accurate than other techniques.
When a node normally communicates, its communication troubled by attack node through
referringwrong appeal or via asks for the route by attacker node is precise route
or short way, andprecision is corrupted. The performance is better after finding a
BHA node. Specificity represents the ability of a detection mechanism to correctly
identify nodes that are not participating in the attack as being legitimate nodes.
Fig. 5. Analysis of accuracy.
Fig. 6. Comparison of specificity analysis.
Fig. 7. Comparison of recall analysis.
Fig. 8. Comparison of precision analysis.
Fig. 9. Comparison of F-score analysis.
Fig. 10. Comparison ofcomputation time analysis.