Modelling of Metaheuristics with Machine Learning-Enabled Cybersecurity in Unmanned Aerial Vehicles

Abstract

The adoption and recent development of Unmanned Aerial Vehicles (UAVs) are because of their widespread applications in the private and public sectors, from logistics to environment monitoring. The incorporation of 5G technologies, satellites, and UAVs has provoked telecommunication networks to advance to provide more stable and high-quality services to remote areas. However, UAVs are vulnerable to cyberattacks because of the rapidly expanding volume and poor inbuilt security. Cyber security and the detection of cyber threats might considerably benefit from the development of artificial intelligence. A machine learning algorithm can be trained to search for attacks that may be similar to other types of attacks. This study proposes a new approach: metaheuristics with machine learning-enabled cybersecurity in unmanned aerial vehicles (MMLCS-UAVs). The presented MMLCS-UAV technique mainly focuses on the recognition and classification of intrusions in the UAV network. To obtain this, the presented MMLCS-UAV technique designed a quantum invasive weed optimization-based feature selection (QIWO-FS) method to select the optimal feature subsets. For intrusion detection, the MMLCS-UAV technique applied a weighted regularized extreme learning machine (WRELM) algorithm with swallow swarm optimization (SSO) as a parameter tuning model. The experimental validation of the MMLCS-UAV method was tested using benchmark datasets. This widespread comparison study reports the superiority of the MMLCS-UAV technique over other existing approaches.

1. Introduction

Nowadays, unmanned aerial vehicles (UAVs), also called drones, are gaining in popularity. They can be used for various purposes with regard to everyday flying objects interconnected with the internet, and can find themselves connected to other devices by sharing data through smart gadgets such as tablets and smartphones [1,2]. Put simply, drones are flying objects that either fly with the help of human pilots or autonomously. UAVs can be used for rescue operations, aerial photography, agriculture, package deliveries, environmental management, monitoring, and other perilous applications [3,4,5]. The reliability of a UAV and its wireless communications are significant for crucial applications. Intrusion detection (ID) methods and security schemes have been employed for ensuring critical safety features. Drones may interact with terrestrial networks [6]. Due to the high probability of drones using ground-based line of sight (LoS) networks as well as the mobility, high elevation of the drones, and service quality needs, UAV-based wireless communication and cellular-connected UAV transmissions may vary from their terrestrial counterparts [7,8,9].

It is essential to frame data connection channels among UAVs to form a mobile self-organizing network if a greater number of drones cooperate to perform a task [10]. The drones in the network can realize real-time data sharing over this mobile network, which need not be sent by a ground station, and this could effectively enhance the combat ability and survivability of the drone group [11,12,13,14]. As the drone network is a subdivision of mobile ad hoc networks (MANETs), frequent attacks on the MANET can threaten the drone network. Due to the multiplicity of network access techniques and the openness of networks, drone networks have been covered by many security threats [15,16,17]. The defense function of conventional network security systems is generally passive, and it can be difficult to fight network attacks with changing technologies. An intrusion detection system (IDS) constitutes the inadequacies of conventional security technologies as an active defensive network security system. IDSs have acquired much interest from users, but certain issues exist that must be enhanced in real-time applications [18]. A classical IDS is inefficient and has an insufficient outcome, particularly in modern computer networks, including large traffic and a high bandwidth. Conventional IDSs cannot fulfill the demand of the present network security, which is very complex, distributed, and automated. Therefore, numerous authors have presented machine learning (ML) methods in the ID field and have made great achievements to diminish the false alarm rate and enhance the detection efficiency level of IDSs [19,20,21].

One of the main aspects of the design of a ML-based detection technique is the selection of a proper set of features to build the model. Provided with a dataset with a large number of features, the identification of proper features can considerably improve the classification performance. Generally, not all features are beneficial to the classification process and several features can be either treated as noisy, reducing the process performance, or highly correlated to one another and eliminated. Feature selection (FS) is a promising technique used to decrease the feature space and choose the most significant features. As an important pre-processing step in ML, FS has gained significance in network management, particularly in network intrusion detection. The inclusion of the FS process in the ML-based classification model reduces the computation complexity and increases the classification performance.

This study proposes a new metaheuristics with machine learning-enabled cybersecurity in unmanned aerial vehicles (MMLCS-UAVs) method. The presented MMLCS-UAV technique initially designed a quantum invasive weed optimization-based feature selection (QIWO-FS) method to improve the ID results. Next, a weighted regularized extreme learning machine (WRELM) model was applied to detect and categorize the intrusions. Finally, swallow swarm optimization (SSO) was used as a parameter tuning model for the WRELM model. The analysis of the results of the MMLCS-UAV technique was tested using benchmark datasets. Briefly, the contributions of this paper are given below.

• An intelligent MMLCS-UAV technique encompassing QIWO-FS, a WRELM classification, and SSO-based parameter tuning is introduced to detect intrusions in a UAV network.
• A QIWO-FS technique is developed using a standard IWO algorithm and quantum computing, which helps to select the useful features from the dataset.
• SSO-based parameter tuning for the WRELM model is designed to eliminate the tedious trial and error parameter tuning process. This also helps to enhance the predictive outcomes of the proposed model for unseen data.

The remaining sections of the paper are as follows. Section 2 elaborates on a detailed survey of existing works and Section 3 provides the proposed MMLCS-UAV technique. Section 4 provides the experimental validation of the proposed approach and Section 5 provides the concluding remarks.

2. Literature Review

Shrestha et al. [21] developed a drone- and satellite-oriented 5G network security system that could harness ML to efficiently discover cyberattacks and vulnerabilities. The solution could be classified into two parts: one was the application of a ML-oriented algorithm into satellite or terrestrial gateways and the other was the creation of a model for ID with the help of numerous ML techniques. Using a real-time CSE-CIC IDS-2018 network database, the system could identify different forms of attacks. To categorize the malicious or benign packets and enhance the security in drone networks, this study illustrated that ML techniques were employed. In [22], an effective technique of an IDS was formulated to identify anomalies in a vehicular system. An in-vehicle network (IVN) transmission system that was a control area network (CAN) was used in this work. The method could classify various attacks, including fuzzing attacks, on vehicles into reconnaissance and denial-of-service (DoS).

Zhang et al. [23] introduced an advanced technology termed an unmanned aerial system multifractal analysis IDS (AMDES) to recognize spoofing attacks. In an earlier study, an IDS related to a multifractal (MF) spectral analysis was employed to render precise MF spectrum predictions of network traffic. Tan et al. [24] modelled an ID technique related to deep belief networks (DBNs), boosted by a particle swarm optimization (PSO) algorithm. First, to gain an optimum DBN structure, a classifier method that relied upon the DBN was framed, the PSO method was employed to augment the hidden layer nodes of the DBN. Praveena et al. [25] introduced a deep reinforcement learning (DRL) method, optimized by the Black Widow Optimization technique for a drone network (DRL-BWO). Moreover, the DRL included an enhanced reinforcement learning-oriented DBN for ID. The BWO method was enforced for the parameter optimization of the DRL methodology. It aids the optimization of the ID performance of drone networks.

Masadeh et al. [26] applied a reinforcement learning (RL)-oriented method to solve the issue by allowing a drone to independently study the dynamics of a target or intruder. To be specific, numerous design variants of the RL-oriented method were applied that varied in the employed temporal difference approaches (state-action-reward-state-action or Q-learning) and in the exploration approaches (greedy or convergence-related ϵ-). Kumar et al. [27] introduced a new secure data sharing structure for software drone environments that combined deep learning and a blockchain (BC).

3. The Proposed Model

In this article, a new MMLCS-UAV technique was developed to accomplish cybersecurity in UAV networks. The presented MMLCS-UAV technique mainly focused on the recognition and classification of intrusions in the UAV network. In the presented MMLCS-UAV technique, a series of processes were carried out; namely, a feature subset selection using the QIWO-FS technique, a WRELM-based classification, and SSO-based parameter tuning. Figure 1 shows the workflow of MMLCS-UAV approach.

3.1. Algorithmic Design of the QIWO-FS Technique

The presented MMLCS-UAV technique introduced a new QIWO-FS technique to select the optimal feature subsets. As a population-based optimization technique, invasive weed optimization (IWO) can fulfill the remarkable outcome of the mathematical formula of randomization and adaptation of the weed colonies. The IWO technique is a powerful and new optimization algorithm that finds the global optimal solution of a mathematical function by mimicking the randomness and compatibility of a weed colony. This algorithm is utilized as an underlying structure for optimization approaches. It is a simple process with a significant convergence speed, a lower computational encumbrance, independence to the problem, and near-global solutions; its gradient-free nature makes it useful for the resolution of FS problems.

The significant rise of weeds (strong herbs) is presumed to be a serious risk to plant products. Weeds have a high resistance level against environmental and climate changes. Thus, a strong optimization technique is obtained based on its features. In the presented method, the weed community and its resistance, compatibility, and randomness attempted to overcome the challenges. IWO was developed according to the phenomenon of agriculture; viz., motivated by an invasive weed colony. As previously mentioned, weeds (as a plant) are able to grow accidentally. There are various benefits to the existence of weeds in urban spaces. However, the accidental growth of these plants can encompass severe harm to human or planet activities; thus, it is regarded as a “weed”. Although IWO is the simplest technique based on the concept, implementation, and structure, it is an efficient optimization technique to solve optimization problems [28]. The subsequent stages are essential to better understand the habitat behaviors and their weeds:

(1)

Initializing Population: several seeds are partially spread in the searching space.

(2)

Reproduction is initiated by pouring every individual plant into the flowering plant; then, the procedure can generate seeds that are worth their proportion. Later, the quantity of the plant seeds linearly decreases from $S_{\max }$ to $S_{\min }$:

$$n\left(w_i\right)=\frac{S_{\max }\left(\max fit-fit\left(w_i\right)\right)+S_{\min }\left(fit\left(w_i\right)-\min fit\right)}{\max fit-\min fit}$$

(1)

(3)

This stage is related to determining a novel location of the seed in the search space. Here, the child seed is positioned near to the parents.

(4)

Competitive removal is related to generating the optimum seeds. This happens whilst the amount of seeds obtains a certain range $\left(p_{\max }\right)$.

(5)

Finally, once the criteria are fulfilled, to finish the process, the second phase is repeated; otherwise, the procedure stops.

As a consequence of the IWO feature as well as the local and global potency for exploration and exploitation, and in addition to the prosperous results in vast quantities of applications, several effective ideas of the quantum concept have been applied to develop the performance of IWO. The fundamental method for QIWO is similar to IWO; certain advantageous variations are implemented to increase the exploration stage in a quantum searching space [29]. To optimize the stochastic module for initializing the seed (population), every seed can be determined by the one Q-bit, which is termed the $Q$-seed.

The state of the $Q$-bit $\left(\Psi \right)$ is defined by:

$$\begin{gathered} \Psi =\underset{j=1}{\overset{n}{{\displaystyle \cup }}}|{\psi }_j\left(t\right)\}={[{\alpha }_j {\beta }_j]}^T \\ j=1,2,\cdots ,n \end{gathered}$$

(2)

In Equation (2), $\alpha$ and $\beta$ characterize the arbitrary integers that characterize the state possibility. $\left|\alpha {|}^2 \mathrm{and} \right|\beta {|}^2$ correspondingly show the possibility that the $Q$-bit $\left|\psi \right\}$ depends on “0” and “1” states. Afterward, they fulfil the relationship $\left|\alpha {|}^2+\right|\beta {|}^2=1$. Initially, the mediocrity of range $\left[0, 1\right]$ is selected as the first population.

$${\psi }_1=\frac{1}{\sqrt{2}}\left[\begin{array}{cc}1& 1\end{array}\right]$$

(3)

After initialization, the arbitrarily generated seed is normalized within zero and one:

$$|\psi {\rangle }_{\left[0,1\right]}=\frac{|\psi \rangle -\min (|\psi \rangle )}{Max(|\psi \rangle )-\min (|\psi \rangle )}$$

(4)

$\mathrm{The} \max (|\psi \rangle )$ and $\min (|\psi \}$ correspondingly signify the maximal and minimal limits. In the presented technique, to produce the $Q$-seed, every individual was accomplished:

$$\begin{gathered} \left\{\begin{array}{c}{\alpha }_j\left(t\right)=rand\\ \left|{\alpha }_j\left(t\right){|}^2+\right|{\beta }_j\left(t\right){|}^2=1\end{array}\right\} \Rightarrow \\ \{\begin{array}{l}\left|{\beta }_j\left(t\right)\right|=\sqrt{1-|{\alpha }_j\left(t\right){|}^2}\Rightarrow \\ {\beta }_j\left(t\right)>0\&{\alpha }_j\left(t\right)>0\end{array} \\ \therefore {\beta }_j\left(t\right)=\sqrt{1-{\alpha }_j\left(t\right){}^2} \end{gathered}$$

(5)

In the FS mechanism, once the size of the feature vector refers to $N$, the sum of the combination of dissimilar features is likely to become $2^N$, which is an enormous space for a comprehensive search. The presented hybrid technique was employed to dynamically search the feature space and produce an accurate combination of features. The FS fell within the multi-objective problems because it fulfilled more than one objective to obtain a better solution, which decreased the group of selected features and concurrently decreased the performance of the output for the given classifier.

As aforesaid, the fitness function determines solutions in these scenarios to accomplish a balance amongst those objectives.

$$itness=\alpha {\Delta }_R\left(D\right)+\beta \frac{\left|Y\right|}{\left|T\right|}$$

(6)

In Equation (6), ${\Delta }_R\left(D\right)$ characterizes the classification error rate. $\left|Y\right|$ represents the size of the subset that the process chooses and $\left|T\right|$ is the total quantity of the features encompassed in the current dataset. $\alpha$ correspondingly shows the parameter $\in \left[0,1\right]$ regarding the weight of the classification error rate and $\beta =1-\alpha$ indicates the significance of the reduction.

3.2. Intrusion Detection Process

For the intrusion detection, the MMLCS-UAV technique applied SSO to the WRELM model. The WRELM was introduced to enhance the standard extreme learning machine (ELM) and determine the optimum weight for all the instances [30]. We assumed that $N$ was a hidden node from the ELM and sample pairs $\left\{X_i,P_i\right\}$; the resulting ELM was formulated by Equation (7):

$$f\left(X_i\right)={\displaystyle \sum }_{n=1}^N{\alpha }_{n}H\left({\phi }_n,X_i,b_n\right),i=1,2,\dots ,N,$$

(7)

where $H(·)$ symbolizes the activation functions, ${\phi }_n$ indicates the input weight vector, ${\alpha }_n$ implies the $n^{th}$ resulting in weight, and $b_n$ denotes the equivalent bias.

$$\tilde{P}=H\alpha =\left[\begin{array}{cccc}H\left({\phi }_1,X_1,b_1\right)& H\left({\phi }_2,X_1,b_2\right)& \cdots & H\left({\phi }_N,X_1,b_N\right)\\ H\left({\phi }_1,X_2,b_1\right)& H\left({\phi }_2,X_2,b_2\right)& \cdots & H\left({\phi }_N,X_2,b_N\right)\\ ⋮& ⋮& ⋮& ⋮\\ H\left({\phi }_1,X_N,b_1\right)& H\left({\phi }_2,X_N,b_2\right)& \cdots & H\left({\phi }_N,X_N,b_N\right)\end{array}\right]·\alpha ,$$

(8)

Here, $\tilde{P}$ refers to the resulting vector.

To obtain the assessment of the $\alpha$ parameter, the major function was exploited by:

$$\mathrm{argmin} \Vert \tilde{P}-P{\Vert }_2^2=\mathrm{argmin} \Vert H\alpha -P{\Vert }_2^2,$$

(9)

In Equation (9), $P$ denotes the observation data. The weight-regularized ELM is formulated by [31,32]:

$$\arg \min C\Vert \sigma \epsilon {\Vert }_2^2+\Vert \alpha {\Vert }_2^2$$

(10)

Dependent on the state:

$$P=H\alpha +\epsilon$$

(11)

In Equation (10), $\sigma$ refers to the diagonal matrix, $\epsilon$ signifies the vector of regression error, and C epitomizes the regularized term. The Lagrange multiplier was introduced to resolve the abovementioned problem, and the corresponding solution to $\alpha$ is shown below:

$$\alpha ={\left(H^{\prime }{\sigma }^{2}H+\frac{I}{C}\right)}^{-1}{H}^{\prime }{\sigma }^{2}P$$

(12)

The SSO approach was utilized to adjust the parameters related to the WRELM model. A major role of the SSO approach was energized by swallow swarm optimization [21]. There exists three diversities of particles; namely:

• An aimless particle $\left(o_i\right)$;
• A leader particle $\left(l_i\right)$;
• An explorer particle $\left(e_i\right)$.

Every particle from the colony (each colony had a subcolony) was accountable for the whole performance; they guided the colony nearer to the optimal condition. The explorer particles included the population of the colony. First, it was responsible to explore the search space [33]. By properly achieving an extreme point, a varying sound guided the group near the location; when the location was led one from problem space, then these particles played as the head leader $\left(H{L}_i\right)$. When the particles were at a favorable location similar to the adjacent particles, it was selected as the local leader ($L{L}_i)$; otherwise, every particle $e_i$ with respect to $V_{LLi}$ (velocity vector of the particles near to the LL), $V_{HLi}$ (velocity vector of the particle near to the HL), and the ability of backlash of these two bearings produced indiscriminative changes.

An aimless particle then started from the search condition, which did not have an optimal location, and the count of $f\left(o_i\right)$ was poor; they had arbitrary and exploratory searches. It began to arbitrarily move and did not have anything implemented with the location of $H{L}_i$ and $L{L}_i$. In addition, there existed a particle from the SSO approach that was named the leader. The optimal leader, termed LH, was realized as a common leader in the colony. There also existed another particle, termed the LL [34,35].

$$V_{HLi+1}=V_{HLi}+{\alpha }_{HL} rand()\left(e_{best}-e_i\right)+{\beta }_{HL}rand()\left(H{L}_i-e_i\right)$$

(13)

$$V_{LLi+1}=V_{LLi}+{\alpha }_{LL}rand\left(\right)\left(e_{best}-e_i\right)+{\beta }_{LL}rand\left(\right)\left(L{L}_i-e_i\right)$$

(14)

where $V_{HL}=$ Velocity of the HL, $V_{LL}=$ velocity of the LL, $e_{best}=$ better location of the explorer particle, and $e_i\to$ presents the location of the explorer particle:

$$V_{i+1}=V_{HLi+1}+V_{LLi+1}$$

(15)

The particle values were upgraded by:

$$e_{i+1}=e_i+V_{i+1}$$

(16)

In order to attain improvised classifier results, the SSO algorithm derived a fitness function as given below.

$$fitness\left(x_i\right)=ClassifierErrorRate\left(x_i\right)=\frac{number of misclassified samples}{Total number of samples}\ast 100$$

(17)

4. Performance Validation

In this section, the performance validation of the MMLCS-UAV technique is briefly investigated. Figure 2 depicts the confusion matrices of the MMLCS-UAV method. The results indicated that the MMLCS-UAV approach properly recognized all the attacks that existed in the dataset.

Table 1. Overall results of MMLCS-UAV model on 80:20 of TR/TS data.

Class	$\mathit{A}\mathit{c}\mathit{c}{\mathit{u}}_{\mathit{y}}$	$\mathit{P}\mathit{r}\mathit{e}{\mathit{c}}_{\mathit{n}}$	$\mathit{R}\mathit{e}\mathit{c}{\mathit{a}}_{\mathit{l}}$	${\mathit{F}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}$	MCC
Training Phase (80%)
Botnet	99.24	97.01	98.40	97.70	97.25
Dos	99.69	99.25	98.92	99.09	98.90
Web	99.42	98.09	98.42	98.26	97.91
Infilteration	99.25	97.41	98.07	97.74	97.29
BruteForce	99.22	97.79	97.46	97.62	97.16
DDos	99.40	99.09	97.41	98.24	97.89
Average	99.37	98.11	98.11	98.11	97.73
Testing Phase (20%)
Botnet	99.50	98.10	99.04	98.56	98.26
Dos	99.72	98.98	99.32	99.15	98.98
Web	99.44	98.00	98.66	98.33	98.00
Infilteration	99.50	98.71	98.39	98.55	98.24
BruteForce	99.50	99.06	98.13	98.59	98.29
DDos	99.78	99.62	98.87	99.25	99.12
Average	99.57	98.74	98.73	98.74	98.48

offers the overall cybersecurity performance of the MMLCS-UAV technique on 80% of the training (TR) data and 20% of the testing (TS) data. Figure 3 reports an average classification outcome of the MMLCS-UAV methodology on 80% of the TR data. The results exhibited that the MMLCS-UAV method showed the maximum performance in each class. It was noticed that the MMLCS-UAV system attained an average $acc{u}_y$ of 99.37%, $pre{c}_n$ of 98.11%, $rec{a}_l$ of 98.11%, F_score of 98.11%, and a Matthew’s correlation coefficient (MCC) of 97.73%.

Figure 4 portrays the average classification outcomes of the MMLCS-UAV technique on 20% of the TS data. The results exhibited that the MMLCS-UAV method displayed the maximum performance in each class. Note that the MMLCS-UAV methodology gained an average $acc{u}_y$ of 99.57%, $pre{c}_n$ of 98.74%, $rec{a}_l$ of 98.73%, $F_{score}$ of 98.74%, and MCC of 98.48%.

Table 2. Overall results of MMLCS-UAV model on 70:30 of TR/TS data.

Class	$\mathit{A}\mathit{c}\mathit{c}{\mathit{u}}_{\mathit{y}}$	$\mathit{P}\mathit{r}\mathit{e}{\mathit{c}}_{\mathit{n}}$	$\mathit{R}\mathit{e}\mathit{c}{\mathit{a}}_{\mathit{l}}$	${\mathit{F}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}$	MCC
Training Phase (70%)
Botnet	99.37	97.98	98.17	98.07	97.69
Dos	99.32	98.08	97.79	97.94	97.53
Web	99.38	98.36	97.89	98.12	97.75
Infilteration	99.44	97.97	98.79	98.38	98.05
BruteForce	99.44	98.17	98.45	98.31	97.98
DDos	99.65	99.25	98.69	98.97	98.76
Average	99.43	98.30	98.30	98.30	97.96
Testing Phase (30%)
Botnet	99.70	99.14	99.14	99.14	98.96
Dos	99.52	98.90	98.25	98.57	98.28
Web	99.78	98.92	99.78	99.35	99.22
Infilteration	99.48	98.12	98.58	98.35	98.04
BruteForce	99.59	99.35	98.29	98.82	98.57
DDos	99.48	98.15	98.61	98.38	98.07
Average	99.59	98.76	98.77	98.77	98.52

presents the overall cybersecurity performance of the MMLCS-UAV technique on 70% of the TR data and 30% of the TS data. Figure 5 shows the average classification outcomes of the MMLCS-UAV approach on 70% of the TR data. The outcomes displayed that the MMLCS-UAV approach exposed the maximum performance for each class. It was noted that the MMLCS-UAV method achieved an average $acc{u}_y$ of 99.43%, $pre{c}_n$ of 98.30%, $rec{a}_l$ of 98.30%, $F_{score}$ of 98.30%, and MCC of 97.96%.

Figure 6 exhibits the average classification outcome of the MMLCS-UAV model on 30% of the TS data. The outcome exhibited that the MMLCS-UAV technique showed the maximum performance in each class. It was noticed that the MMLCS-UAV method achieved an average $acc{u}_y$ of 99.59%, $pre{c}_n$ of 98.76%, $rec{a}_l$ of 98.77%, $F_{score}$ of 98.77%, and MCC of 98.52%.

In

Table 3. Comparison study of MMLCS-UAV model with recent models [21,24].

Methods	$\mathit{A}\mathit{c}\mathit{c}{\mathit{u}}_{\mathit{y}}$	$\mathit{P}\mathit{r}\mathit{e}{\mathit{c}}_{\mathit{n}}$	$\mathit{R}\mathit{e}\mathit{c}{\mathit{a}}_{\mathit{l}}$	${\mathit{F}}_{\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}}$
MMLCS-UAV	99.59	98.76	98.77	98.77
GA-DBN	91.76	90.28	98.14	98.02
BOA-DBN	91.13	89.70	98.47	97.86
LR Model	91.65	90.51	98.63	96.33
LDA Model	92.00	90.89	98.41	97.19
KNN Model	97.80	97.85	95.34	95.39
DT Model	95.70	95.60	97.34	96.77
GNB Model	96.50	96.56	98.15	96.40

, a comparison study of the MMLCS-UAV technique with existing approaches [21,24] such as a genetic algorithm with DBN (GA-DBN), a butterfly optimization algorithm with DBN (BOA-DBN), k-nearest neighbor (KNN), logistic regression (LR), a linear discriminant analysis (LDA), Gaussian Naive Bayes (GNB), and decision tree (DT) is provided. Figure 7 examines the comparative $acc{u}_y$ examination of the MMLCS-UAV model with other existing models. The results signified that the GA-DBN, BOA-DBN, LR, and LDA models reached a minimal $acc{u}_y$ of 91.76%, 91.13%, 91.65%, and 92%, respectively.

Meanwhile, the DT and GNB models gained closer $acc{u}_y$ values of 95.70% and 96.50%, respectively. Although the KNN model attained a reasonable $acc{u}_y$ of 97.80%, the MMLCS-UAV model ensured a maximum $acc{u}_y$ of 99.59%.

Figure 8 inspects the comparative $acc{u}_y$ examination of the MMLCS-UAV model with the other existing techniques. The results signified that the KNN, DT, GA-DBN, and GNB models reached a minimal $acc{u}_y$ of 89.7, 90.28, 90.51, and 90.89%, respectively. Meanwhile, the LDA and BOA-DBN methods obtained closer $acc{u}_y$ values of 95.6 and 96.56%, respectively. Although the LR approach attained a reasonable $acc{u}_y$ of 97.85%, the MMLCS-UAV method ensured a maximum $acc{u}_y$ of 98.76%.

Figure 9 portrays the comparative $acc{u}_y$ examination of the MMLCS-UAV approach with the other existing techniques. The results signified that the KNN, DT, GA-DBN, and GNB models reached a minimal $acc{u}_y$ of 95.34, 97.34, 98.14, and 98.15%, respectively. Simultaneously, the LDA and BOA-DBN methods achieved closer $acc{u}_y$ values of 98.41 and 98.47%, respectively. Although the LR method attained a reasonable $acc{u}_y$ of 98.63%, the MMLCS-UAV algorithm ensured a maximum $acc{u}_y$ of 98.77%.

Figure 10 displays the comparative $acc{u}_y$ inspection of the MMLCS-UAV method with the other existing approaches. The results exhibited that the KNN, DT, GA-DBN, and GNB approaches reached a minimal $acc{u}_y$ of 95.39, 96.33, 96.4, and 96.77%, respectively.

In the meantime, the LDA and BOA-DBN techniques reached closer $acc{u}_y$ values of 97.19 and 97.86%, respectively. Although the LR technique reached a reasonable $acc{u}_y$ of 98.02%, the MMLCS-UAV method ensured a maximum $acc{u}_y$ of 98.77%. These results highlighted the supreme performance of the MMLCS-UAV model.

5. Conclusions

In this article, a new MMLCS-UAV technique was developed to accomplish cybersecurity in UAV networks. The presented MMLCS-UAV technique mainly focused on the recognition and classification of intrusions in a UAV network. To obtain this, the presented MMLCS-UAV technique introduced a new QIWO-FS technique to select the optimal feature subsets. For intrusion detection, the MMLCS-UAV technique applied SSO to the WRELM model. The design of the QIWO algorithm and SSO algorithm helped to accomplish an enhanced classification performance. A widespread comparison study reported the superiority of the MMLCS-UAV technique over other existing approaches, with a highest accuracy of 99.59%, precision of 98.76%, recall of 98.77%, and F-score of 98.77%. Thus, the presented MMLCS-UAV technique could be employed for maximum cybersecurity in UAV networks. In the future, deep learning-based classifiers can be introduced to improve secure UAV communications. Moreover, the proposed algorithm could be tested on a large-scale real-time dataset. Additionally, data clustering techniques could be involved to improve the classification performance and security of UAV networks.