Data analysis and spatial design study of children's activity sites in rural Gansu, China

Weidi Zhang; Runbo Liu; Weidi Zhang; Runbo Liu

doi:10.3934/urs.2023020

Urban Resilience and Sustainability

2023, Volume 1, Issue 4: 314-333. doi: 10.3934/urs.2023020

Previous Article Next Article

Research article

Data analysis and spatial design study of children's activity sites in rural Gansu, China

Weidi Zhang ^,,
Runbo Liu

Department of Environmental Design, Shaanxi University of Science and Technology. Xi'an City, Shaanxi Province, 710021, China

Received: 17 October 2023 Revised: 13 December 2023 Accepted: 15 December 2023 Published: 31 December 2023

This paper examined rural outdoor activity sites for children in Zhangye, Gansu Province, China. The study analyzed the characteristics of children's groups and their behavioral patterns during outdoor activities in rural areas, and explored the factors affecting children's engagement in outdoor activities. A questionnaire survey was conducted to assess the physical and mental health status of rural children in Gansu Province, the distribution of outdoor activity time, and the influence of activity venues on children. Based on the survey results in rural Zhangye, Gansu Province, activity venues and facilities suitable for rural children in Gansu were identified. Finally, five design principles were proposed (i.e., safety, fun, comfort, accessibility, and guiding) to cater to the activity preferences of rural children in Gansu, China and provide a nurturing environment that fostered their physical and mental growth.

Keywords:

Citation: Weidi Zhang, Runbo Liu. Data analysis and spatial design study of children's activity sites in rural Gansu, China[J]. Urban Resilience and Sustainability, 2023, 1(4): 314-333. doi: 10.3934/urs.2023020

Related Papers:

[1]	Yue Li, Wusheng Xu, Wei Li, Ang Li, Zengjin Liu . Research on hybrid intrusion detection method based on the ADASYN and ID3 algorithms. Mathematical Biosciences and Engineering, 2022, 19(2): 2030-2042. doi: 10.3934/mbe.2022095
[2]	Hao Zhang, Lina Ge, Guifen Zhang, Jingwei Fan, Denghui Li, Chenyang Xu . A two-stage intrusion detection method based on light gradient boosting machine and autoencoder. Mathematical Biosciences and Engineering, 2023, 20(4): 6966-6992. doi: 10.3934/mbe.2023301
[3]	Naila Naz, Muazzam A Khan, Suliman A. Alsuhibany, Muhammad Diyan, Zhiyuan Tan, Muhammad Almas Khan, Jawad Ahmad . Ensemble learning-based IDS for sensors telemetry data in IoT networks. Mathematical Biosciences and Engineering, 2022, 19(10): 10550-10580. doi: 10.3934/mbe.2022493
[4]	Zhongwei Li, Wenqi Jiang, Xiaosheng Liu, Kai Tan, Xianji Jin, Ming Yang . GAN model using field fuzz mutation for in-vehicle CAN bus intrusion detection. Mathematical Biosciences and Engineering, 2022, 19(7): 6996-7018. doi: 10.3934/mbe.2022330
[5]	Muhammad Hassan Jamal, Muazzam A Khan, Safi Ullah, Mohammed S. Alshehri, Sultan Almakdi, Umer Rashid, Abdulwahab Alazeb, Jawad Ahmad . Multi-step attack detection in industrial networks using a hybrid deep learning architecture. Mathematical Biosciences and Engineering, 2023, 20(8): 13824-13848. doi: 10.3934/mbe.2023615
[6]	Noor Wali Khan, Mohammed S. Alshehri, Muazzam A Khan, Sultan Almakdi, Naghmeh Moradpoor, Abdulwahab Alazeb, Safi Ullah, Naila Naz, Jawad Ahmad . A hybrid deep learning-based intrusion detection system for IoT networks. Mathematical Biosciences and Engineering, 2023, 20(8): 13491-13520. doi: 10.3934/mbe.2023602
[7]	Kunpeng Li, Zepeng Wang, Yu Zhou, Sihai Li . Lung adenocarcinoma identification based on hybrid feature selections and attentional convolutional neural networks. Mathematical Biosciences and Engineering, 2024, 21(2): 2991-3015. doi: 10.3934/mbe.2024133
[8]	Jing Zhang, Baoqun Yin, Yu Zhong, Qiang Wei, Jia Zhao, Hazrat Bilal . A two-stage grasp detection method for sequential robotic grasping in stacking scenarios. Mathematical Biosciences and Engineering, 2024, 21(2): 3448-3472. doi: 10.3934/mbe.2024152
[9]	Ziyu Jin, Ning Li . Diagnosis of each main coronary artery stenosis based on whale optimization algorithm and stacking model. Mathematical Biosciences and Engineering, 2022, 19(5): 4568-4591. doi: 10.3934/mbe.2022211
[10]	Natalya Shakhovska, Vitaliy Yakovyna, Valentyna Chopyak . A new hybrid ensemble machine-learning model for severity risk assessment and post-COVID prediction system. Mathematical Biosciences and Engineering, 2022, 19(6): 6102-6123. doi: 10.3934/mbe.2022285

Abstract

1. Introduction

With the development of internet technology, there is a tremendous increase in the dimensions of information that are generated, exchanged and processed. Almost in all the fields there is a difficulty in handling large amount of high dimensional data. These data become the target for illegal activities which impose a severe threat to the network security. The traditional security techniques such as antivirus, firewall, data encryption and user identification which acts as the first line of defense that alone is not sufficient to provide the better security to the network. The second line of security is highly recommended that can be provide by an Intrusion Detection System (IDS) ^[1]. Using these two lines of security enhances the overall network security. An Intrusion Detection System (IDS) identifies the intruders who are harmful to an organization. The main goal of an IDS is to monitor the network or system for abnormal pattern or traffic and prevent it from unauthorized access ^[2]. However, the key problem is identifying the unknown malicious traffic. The Intrusion Detection Systems is broadly classified into two types based on the source of data. First, Network Intrusion Detection System (NIDS) placed near the network points checks the network traffic from the routers and gateways for intrusion. It can detect attacks in real time. The main limitation of NIDS is that it can monitor the traffic passing through the specific network nodes. Second, the Host Intrusion Detection System (HIDS) scans the individual host for suspicious activities like an unwanted configuration change, deletion or modification of system files, or an unwanted sequence of system calls. If any of these activity occurs, it sends an alert to the administrator ^[3]. The main limitation in HIDS is that it cannot analyze the network behaviors. Next the Intrusion Detection System (IDS) uses three different detection mechanisms to detect the attacks and each detection method is further classified accordingly. Figure 1 shows the different detection mechanisms to detect the intrusions.

Figure 1. Types of detection mechanism.

DownLoad: Full-Size Img PowerPoint

1.1. Signature-based/misuse detection mechanism

Signature-based IDS uses the set of rules or predefined signatures for detecting the known attacks ^[3]. Misuse detection techniques are based on knowledge or based on Machine Learning (ML) methods. In a knowledge-based strategy, the network stream of traffic or host audit data is analyzed and compared with the predefined set of rules. There are three ways of applying the knowledge-based approach: signature matching, rule-based expert system, and state transition analysis ^[3]. The signature matching compares the incoming network traffic with the predefined attack signatures for intrusion. A rule-based expert system finds out the intrusion by comparing the traffic with the predetermined rules. Finally, state transition analysis maintains the state transition model for each known suspicious activity. Machine Learning (ML) based IDS offers learning based on model to detect the normal behavior from attack behavior. The ML model generates the representation for the known models. It uses supervised machine learning techniques such as SVM, Decision Tree to detect the known attacks more efficiently.

The significant drawback of Signature-based Detection is that it needs to update the signatures regularly for new attacks for which signatures are not there in the database. As a result, it generates more false alarms. It has to maintain a large signature database.

1.2. Anomaly-based detection mechanism

Anomaly-based detection uses a hypothesis to detect novel, unknown intrusion if any deviation or behaviour change occurs ^[3]. Anomaly-based detection comprises statistical techniques, Finite State Machine (FSM) and machine learning techniques. FSM generates the behaviour model that contains states, actions and transitions. Semi-Supervised and Unsupervised ML techniques such as clustering algorithms, one class SVM are mainly used for anomaly detection. The anomaly-based detection can detect both known and unknown attacks. The prime limitation is that anomaly-based detection suffers from high false positives ^[3].

1.3. Hybrid detection mechanism

Hybrid detection mechanism combines signature and anomaly-based detection to detect the intrusion ^[3].

1.4. Limitations of current IDS

Current IDS can detect more accurately and precisely the known attacks, which leaves the system more vulnerable to novel malicious attacks where predefined signatures are not available ^[4].

High False positives-The current IDS suffers from high false positives. False positives are the incorrect classification of a normal event as malicious events. The main aim of an IDS is to minimize the false positives as far as possible ^[4].

High False-negatives-The current IDS suffers from high false negatives. False negatives are the incorrect classification of malicious events as normal events. The main aim of an IDS is to reduce the false negatives as minimum as possible ^[4].

Data Overload-Millions of data are generated every day depending on the company's size and the IDS tools used, which leads to a data overload ^[4].

1.5. DataStream mining

DataStream mining is the continuous well-ordered sequence of data that arrives in a timely fashion ^[5]. Data streams are a constant flow of unlimited data with high speed, and data changes with time compared to the traditional databases ^[6]. The primary requirement of the stream data is to inspect each instance only once. Therefore, it should use a limited amount of memory and should give the outcome in a less amount of time. Data streams are of two types, namely streams that are static and streams that are evolving. The bulk arrival of data that won't change with time is called static stream. The data that arrives continuously and changes with time is called evolving data streams ^[6].

The paper is organized as follows: Section 2 presents the earlier work on various machine learning and feature selection techniques. Section 3 discusses the methodology used for preprocessing, feature selection and classification. Section 4 discusses the experimental setup, dataset details and the performance evaluation metrics used. Section 5 presents the result outcomes and its performance comparison with the state of art methods and finally Section 6 summarizes the research work undergone and future scope.

2. Related works

There has been quite a lot of research on intrusion detection using machine learning and deep learning techniques and the hybrid approaches. The related work (Table 1) presents the latest techniques used and its relevant advantages and disadvantages.

Table 1. Related works.

Authors	Algorithms Used	Computational Methods	Pros	Cons
X. Li et al. (2021) ^[7]	CMPSO, ACO, KH, IKH, LNNLS-KH	The NSL_KDD dataset is taken for intrusion detection. The proposed LNNLS-KH is compared with CMPSO, ACO, KH, and IKH algorithms and best features are selected. Then KNN technique is applied for further classification.	Good convergence speed. Lower false positive rate. LNNLS-KH gives an accuracy of 96.12%	Attack with fewer samples, an adversarial learning method can be used to create similar attacks.
X. K. Zhou et al. (2021) ^[8]	AdaBoost, LSTM, CNN LSTM, VLSTM	The authors proposed a variational long short-term memory technique that detects intrusion anomalies efficiently based on feature reconstruction.	The loss function helps to reconstruct the hidden variable into meaningful form. Proposed VLSTM gives the accuracy of 89.5%.	Imbalanced data is still a challenge in anomaly detection.
T. H. Hai et al. (2020) ^[9]	Novel architecture of storage tools and distributed log processing	The Novel storage with HBase or Apache Spark enhances NIDS data processing.	Processing time is reduced.	Takes more query time.
S. N. Mighan et al. (2020) ^[10]	SVM, SAE-SVM	Hybrid scheme that uses deep learning and machine learning method together (SAE-SVM) can detect intrusion attacks more precisely.	SAE-SVM shows higher accuracy of 95.98%.	Computational time taken is more.
T. Vaiyapuri et al. (2020) ^[11]	Stacked Autoencoder, Sparse Autoencoder, Denoising Autoencoder, Contractive Autoencoder, Convolution Autoencoder	All the methods mentioned are compared with contractive autoencoder. The contractive autoencoder gives 87.98% intrusion detection accuracy.	SAE 85.23%, SAAE 86.02% DAE 86.92% ContAE 87.98% CAE 81.07%	Reduced reconstruction ability further needs to be improved
C. F. Tang et al. (2020) ^[12]	Stacked Autoencoder Deep Neural Network (SAE-DNN), SAAE-DNN	SAAE selects the needed features from the intrusion dataset and initializes weight to DNN thus improves the intrusion detection accuracy.	SAE-DNN gives 82.23% accuracy. The proposed SAAE-DNN shows higher accuracy of 87.74% when compared to SAE-DNN.	The accuracy achieved can be improved further.
A. D. Jadhav et al. (2019) ^[13]	SVM, KNN, Decision Tree, Naïve Bayes classifier machine learning techniques to detect the attacks. Proposed a distributed and parallel approach.	The proposed distributed and parallel approach enhances the efficiency of detecting the intrusions faster.	Faster detection of intrusion.	Not applied in real time environment.
A. Muallem et al. (2017) ^[14]	Hoeffding Tree Restricted, Hoeffding Trees AUE2 with Buffer AUE2 without Buffer Hoeffding Adaptive Trees with DDM and ADWIN	The survey talks about the flexibility of the methods when used in different domains of streaming data. Combinations of technique solves the intrusion anomaly detection problem.	All the techniques surveyed gives good accuracy. Hoeffding Tree 93% Restricted Hoeffding Trees 92.15% AUE2 with Buffer 94.07% AUE2 without Buffer 94.06% Hoeffding Adaptive Trees with DDM and ADWIN 92%.	Improvement in accuracy is needed.
G. Kim et al. (2014) ^[15]	Hybrid approach combines C 4.5 algorithm and one class SVM approach	C 4.5 is used to build the misuse detection model and decompose the training data into smaller subsets. Multiple one class SVM models are built to enhance intrusion detection accuracy	Good detection accuracy for both known and unknown attacks. Low false positive.	Processing time is more for the proposed technique.
H. K. Sok et al. (2013) ^[16]	ADT algorithm	The alternating decision tree algorithm is used for knowledge discovery and effective selection of features.	Classification process simplifies.	Evaluation speed needs to be improved.
S. J. Horng et al. (2011) ^[17]	BIRCH hierarchical clustering technique, SVM	It combines BIRCH hierarchical clustering technique and SVM which gives good detection accuracy.	The training time is reduced and gives a good detection accuracy of 95.72%. It mainly detects DOS or Probe attacks.	It cannot detect U2L and R2L attacks.
Tavallaee et al. (2009) ^[18]	KDD CUP 99 dataset	Analysis on KDD CUP 99 dataset is made.	Good for signature-based detection.	Poor detection when used for anomaly detection. Contains many duplicate values leads to performance degradation.

| Show Table

DownLoad: CSV

3. Methodology

The proposed work uses the NSL_KDD benchmark dataset for intrusion analysis. The first phase focuses on feature selection using Darwinian Particle Swarm Optimization and selects key features that contribute to intrusion. The second phase emphasises applying the proposed Stacked Encoding Hoeffding Tree technique to classify the data based on performance metrics like accuracy, specificity, sensitivity, false-positive rate, false-negative rate and F1 score.

The benchmark NSL_KDD dataset is taken for the analysis. The dataset which we have taken is standalone dataset in order to incorporate stream data, the dataset is streamed using the techniques in Matlab tool. The system object technique simplifies the streaming process in Matlab. The data is now continuous and it possess the characteristics of streaming data. The attacks are detected using the proposed Stacked Autoencoder Hoeffding Tree (SAE-HT) classification approach. The bio-inspired technique called Darwinian Particle Swarm Optimization (DPSO) enhances the performance of the SAE-HT classification technique. The distracting variance is removed from the data using the DPSO feature selection technique that enables the classifier to perform better, especially when dealing with the high dimensional features. Figure 2 shows the flow diagram of SAE-HT classification technique.

Figure 2. Flow diagram of SAE-HT classification technique.

DownLoad: Full-Size Img PowerPoint

3.1. Data Preprocessing

Data preprocessing is a vital step in machine learning. The raw data collected is made ready to be used by machine learning techniques to extract meaningful insights from the data. The NSL_KDD dataset taken from Canadian Institute for Cybersecurity is analyzed. If the downloaded dataset is in gz or Tcpdump format, convert that to CSV file format and load the dataset into the environment. The proposed classification technique supports only the numeric data. Since most machine learning (ML) techniques use mathematical equations that only support the usage of numeric data, the conversion of categorical data into numerical data using data conversion functions should occur. The one-hot encoding technique converts the categorical data to numerical data, thus making it convenient to apply machine learning techniques to the dataset.

3.2. One-hot encoding technique

The one-hot encoding technique is the most effective encoding technology to deal with the conversion of numeric to categorical features ^[19]. It can convert the categorical features to a binary vector. The vector holds Zeroes, and One's as values. The vector holds only one element with the value one and other values corresponding to Zero. An element with value one indicates the occurrence of the possible values against the categorical features. The NSL_KDD dataset contains three categorical features such as protocol_type, service and flag. For example, the protocol_type consists of three attributes: ICMP, TCP, UDP. Using one hot encoding technique, ICMP can be encoded as (1, 0, 0), TCP can be encoded as (0, 1, 0), UDP can be encoded as (0, 0, 1). Similarly, categorical features service and flag are also encoded into one-hot encoding vectors.

3.3. Feature selection

Feature selection mainly reduces the features by removing the insignificant or less significant ones. There are many feature selection techniques. This paper uses the Bio-Inspired feature selection technique called Darwinian Particle Swarm Optimization (DPSO) to select optimal features. The main goal of the DPSO is to find the non-redundant and highly correlated features, thus eliminates the least correlated features ^[20]. DPSO is the extension of the PSO technique with the basic principle of survival of the fittest. The major drawback of PSO and other bio-inspired feature selection techniques is that it gets trapped in the local optimum. No long-term memory effect leads to premature convergence to the local optimum DPSO overcomes the above drawback. DPSO consists of multiple swarms in which each swarm performs like an individual PSO. All the swarms run simultaneously towards the local optimum, and each swarm is compared. The best swarm gets the extended life, and the stagnated or insufficient swarm got deleted.

DPSO is an effective evolutionary algorithm that searches the population of individuals for local optimum. The population represents the "swarm", and individuals represent the "particles". Throughout the evolutionary process, every particle updates its moving direction according to the position. There are two positions, namely local best and global best position ^[21]. The local best position (pbest) is the particle's position among all the particles that are visited so far. The global best position (gbest) is the best fitness achieved among all the visited particles so far. In each iteration, the particle updates its velocity and position, which is given by Eqs (1) and (2) ^[22].

${V}_{j}\left(t+1\right) = {\omega V}_{j}\left(t\right)+{c}_{1}{r}_{1}\left({P}_{j}-{Y}_{j}\left(t\right)\right)+{c}_{2}{r}_{2}\left({P}_{g}-{Y}_{j}\left(t\right)\right)$

(1)

${Y}_{j}\left(t+1\right) = {Y}_{j}\left(t\right)+{V}_{j}(t+1)$

(2)

${Y}_{j} = ({Y}_{j}^{1}, {Y}_{j}^{2}, \dots., {Y}_{j}^{D})$ denotes the particle position at generation j in a D-dimensional search space. V_j (t + 1) is the velocity produced at time t + 1. P_j is the best position of the particles found so far (pbest). It denotes the cognitive component of Eq (1). P_g is the best global position found so far (gbest). It represents the social component of Eq (1). ω signifies the inertia weight, c signifies a constant called local and global weight, r signifies random variable which ranges between (0, 1). The searching process keeps on going until the predefined threshold reaches. DPSO is more efficient than the original PSO and thus prevents premature convergence to a local optimum. Rearrange the velocity function in Eq (1), and it is given by Eq (3) ^[22].

${V}_{j}\left(t+1\right) = \alpha {V}_{j}\left(t\right)+\frac{\alpha }{2}{V}_{j}\left(t-1\right)+\frac{\alpha \left(1-\alpha \right)}{6}{V}_{j}\left(t-2\right)\\ +\frac{\alpha \left(1-\alpha \right)\left(2-\alpha \right)}{24}{V}_{j}\left(t-3\right)+{c}_{1}{r}_{1}\left({P}_{j}-{Y}_{j}\left(t\right)\right)+{c}_{2}{r}_{2}({P}_{g}-{Y}_{j}\left(t\right))$

(3)

The left side of Eq (3) gives a discrete version of the derivative of velocity D𝛼[V𝑡 + 1] with order α = 1. Finally, the Grunwald-Letnikov derivative expresses the discrete-time implementation using Eq (4) ^[22].

${D}^{\alpha }\left[{V}_{t+1}\right] = \frac{1}{{T}^{\alpha }}\sum \limits_{k = 0}^{r}\frac{{\left(-1\right)}^{k}\rm{\Gamma }\left(\alpha +1\right)v(t-kT)}{\rm{\Gamma }(k+1)\rm{\Gamma }(\alpha -k+1)}$

(4)

Here T represents a sample period, and r represents the truncate order. Repeat Eq (3) to update every particle velocity. The different values generated to control the convergence speed of the optimization process.

3.4. Proposed classification algorithm (SAE-HT)

Classification is the supervised machine learning technique that divides the set of data into different classes. As the data generation is continuous, it is impossible to store a massive amount of data. So, the data needs to be analyzed as it comes in. The classification techniques are very much helpful in classifying the streaming data ^[23]. This paper proposes a classification technique for stream data mining called the Stacked Autoencoder Hoeffding Tree approach. Even though stacked autoencoder gives good accuracy on its own but in order to improve it further we applied hoeffding tree approach. The stacked autoencoder is an unsupervised learning technique that maintains three layers: input, output, and hidden layers ^[24]. An encoder takes the input and maps it to the hidden representation. Finally, the decoder reconstructs the input. Equation (5) gives the encoder process ^[24].

${\rm{h}}_{\rm{n}} = \rm{f}({\rm{W}}_{1}{\rm{x}}_{\rm{n}}+{\rm{b}}_{1})$

(56)

where h_n represents the encoder, vector determined from x_n. Encoding function f, weight matrix of the encoder W₁, and bias vector b₁.

Equation 6 represents the decoder process. Where g denotes decoding function, W₂ denotes the decoders weight matrix, and b₂ denotes the bias vector ^[24].

$\widehat{{\rm{X}}_{\rm{n}}} = \rm{g}({\rm{W}}_{2}{\rm{h}}_{\rm{n}}+{\rm{b}}_{2})$

(6)

When the decoder reconstructs the input data, there is a possibility that it results in reconstruction error. Equation (7) minimizes the reconstruction error ^[24].

$\rm{ \mathsf{ ϕ} }\left(\rm{\Theta }\right) = \genfrac{}{}{0pt}{}{\rm{arg}\rm{m}\rm{i}\rm{n}}{\rm{ \mathsf{ θ} }, {\rm{ \mathsf{ θ} }}^{\rm{\text{'}}}}\frac{1}{\rm{n}}{\sum }_{\rm{i} = 1}^{\rm{n}}\rm{L}({\rm{X}}^{\rm{i}}, {\widehat{\rm{X}}}^{\rm{i}})$

(7)

where L denotes the loss function, and $L\left(X, \widehat{X}\right) = \parallel X-\widehat{X}{\parallel }^{2}$ represents the loss function.

Now Hoeffeding tree is used to classify the class labels. The Hoeffding tree is a kind of decision tree that consists of a root node, test node and leaf node. The leaf node holds the class prediction. The main requirement in streaming data is to classify the data in a single pass. The data represented as a tree structure using the Hoeffding tree technique when the model built incrementally. The main disadvantage in Hoeffding tree classification is that it fails to classify the data into a tree when a tie occurs. Equation (8) gives the formula for Hoeffding bound calculation ^[25].

$\in = \sqrt{\frac{\left({R}^{2}\rm{log}\left(\frac{1}{\delta }\right)\right)}{2n}}$

(8)

where R denotes the range of random variable, δ denotes the desired probability not within ε of the expected value, N denotes the number of instances collected at the node.

Algorithm: Stacked Autoencoder Hoeffding Tree Approach
Input: NSL_KDD benchmark dataset
Output: Classification results: Accuracy, Sensitivity, Specificity, False Alarm Rate, False Negative Rate and F₁ Score.
Procedure
1. Load the network intrusion benchmark dataset (NSL_KDD dataset)
2. Data preprocessing
3.Apply bio-inspired feature selection technique (DPSO) to select the significant features
4. Partition the dataset into training and testing data
5. Input data to the encoder and maps it to the hidden representation to obtain a learned feature vector.
6. The feature vector from the previous layer is the input to the next layer. This process repeats till the training ends.
7. The decoder reconstructs the input from the hidden representation. Thus, Eq (7) minimizes the reconstruction error.
8. for entire training data, do
9. Use the Hoeffding tree technique and sort the instances to $f$ leaf
10. Update the necessary statistics in $f$
11. Increment ${m}_{f}$ , for all the instances at $f$
12. if ${m}_{f}$ mod ${m}_{min}=\rm{ }0$ and instances at $f$ are of a different class, then
13. Calculate ${G}_{f}k\left(i\right)$ for each feature
14. The feature with the highest ${G}_{f}$ value represents ${k}_{q}$
15. The feature with the second-highest ${G}_{f}$ value represents ${k}_{q}$
16.Calculate Hoeffding Tree Bound using Eq (8)
17. If ${k}_{q}$ ≠ ${k}_{\rm{\varnothing }}$ and ( ${G}_{f}{(k}_{p})$ - ${G}_{f}{(k}_{q})$ > ∈ or ∈< τ) then
18. Replace $f$ with an internal node that splits on ${k}_{p}$
19. for entire branches, do
20. Add and initialize a new leaf with sufficient statistics
21. End
22.Returns the classification result.

| Show Table

DownLoad: CSV

The pseudocode of the stacked autoencoder hoeffdding tree explains the following: Line 1 loads and streams the NSL_KDD dataset using Matlab platform. Line 2 performs the preprocessing of the dataset. The categorical data in the dataset is converted to numeric data using the one-hot encoding technique. In line 3, the DPSO feature selection algorithm is applied, and the important features are selected based on the selection score generated by the algorithm. Line 4 the data is divided into test and train sets. Lines 5–7 explains the stacked autoencoder process where the input data to the encoder is mapped to the hidden representations to obtain the learned features. The learned feature vector is given as an input to the next layer. This process continues till the training ends. The decoder reconstructs the hidden representations. The reconstruction error is minimized by calculating the loss function. Lines 8–21 explains the procedure of the hoeffding tree. The output generated by the stacked autoencoder is fed as an input to the hoeffding tree technique. The input is taken and the root node is decided and initialized. The tree is constructed incrementally for each training data until suitable leaf arrives. Each node has enough information to make a decision. It uses information gain to make the attribute split. The best attribute is found at each node and test is performed to decide whether the attribute yields better results based on Hoeffding bound. The test is applied on the attributes to find out which attribute gives better results and split the node for the growth of the tree. Line 22 returns the classification results.

4. Experimental setup

The experiment was conducted with the proposed technique in MATLAB R2021a on Windows 10 64-bit operating system with Ryzen 7 processor and 16 GB RAM. The experiment uses a stream-oriented offline database for querying the network traffic data. It enables a natural data analysis within the IDS. The streaming architecture is used across multiple sites to process attack data to increase the performance in large scale systems because the data is processed during the natural flow and stored only for a limited amount of time for analysis. In the research work, we have applied a hybrid classification technique. The proposed hybrid classification technique's performance was analyzed by applying various performance evaluation metrics.

4.1. Intrusion dataset

The NSL_KDD dataset is used for experimentation of the proposed work. The KDD CUP'99 dataset is the popular benchmark dataset used for network intrusion detection system. The main limitation of the KDD CUP'99 dataset is that it contains a high number of redundant records that affect the effectiveness of the evaluated system ^[26]. The improved version of KDD CUP'99 is the NSL_KDD dataset in which the redundant records are removed. NSL_KDD dataset have approximately 125, 973 training data and 22, 544 testing data ^[27]. Similar to KDD CUP'99, the records in NSL KDD dataset are unique and labelled as normal and anomaly. It has 41 features that address four different categories of attacks. Table 2 shows the NSL KDD features ^[27].

Table 2. NSL_KDD dataset features.

No	NSL_KDD Feature Names	No	NSL_KDD Feature Names
1	duration	21	_is_host_login
2	protocols_types	22	_is_guests_login
3	services	23	_counts
4	flag	24	src_counts
5	source_bytes	25	srcerror_rate
6	dstn_bytes	26	srcs_error_rate
7	lang	27	rerrors_rate
8	wrong_fragments	28	src_rerrors_rate
9	urgent	29	simlr_srcs_rate
10	hot	30	diff_srcs_rate
11	numbr_failed_login	31	src_diffs_host_rate
12	usr_login	32	dstv_host_counts
13	numbr_compromises	33	dstn_host_src_count
14	numbr_root_shell	34	dstn_simlr_src_rate
15	numbr_attempts	35	dstn_diffs_srce_rate
16	numbr_of_roots	36	dstn_hosts_sim_srce_port_rates
17	numbr_files_creation	37	dstn_hosts_src_diffr_host_rates
18	numbr_offshell	38	dstn_hosts_srcerror_rate
19	numbr_access_usrfiles	39	dstn_hosts_srce_srerror_rate
20	numbr_outbounds_cmd	40	dstn_hosts_error_rates
		41	dstn_hosts_srce_error_rates

| Show Table

DownLoad: CSV

4.2. Performance evaluation

The study was conducted to measure the performance and compare the results based on the performance metrics such as accuracy, sensitivity, specificity, False Positive Rate (FPR), False Negative Rate (FNR). The feature selection technique selects the important features that play a significant role in the intrusion. Then the proposed classifier performance is determined with the help of performance evaluation metrics. Accuracy is the most common performance evaluation technique. Accuracy is the total number of instances correctly predicted as an attack from the ratio of all predictions made from the given dataset. Equation (9) shows the way of accuracy (ACC) calculation ^[28].

$A\rm{C}\rm{C} = \frac{\rm{T}\rm{P}+\rm{T}\rm{N}}{\rm{T}\rm{P}+\rm{T}\rm{N}+\rm{F}\rm{P}+\rm{F}\rm{N}}$

(9)

The confusion matrix in Table 3 defines the TP, TN, FP, FN

Table 3. Confusion matrix for IDS.

Actual Instance	Predicted Instance
Actual Instance	Normal	Anomaly
Normal	True Negative (TN)	False Positive (FP)
Anomaly	False Negative (FN)	True Positive (TP)

| Show Table

DownLoad: CSV

TN (True Negative): The actual and predicted instances both are classified as normal.

FP (False Positive): The actual normal instance is predicted as an anomaly by the IDS.

FN (False Negative): The actual anomaly instance is predicted as normal by the IDS.

TP (True Positive): The actual anomaly instance is predicted as an anomaly by the IDS.

Sensitivity (S) is the probability to identify an attack instance as an attack accurately. Recall or True Positive Rate (TPR) are the other names for sensitivity. Sensitivity can be calculated as in Eq (10) ^[29].

$\rm{S} = \frac{\rm{T}\rm{P}}{\rm{T}\rm{P}+\rm{F}\rm{N}}$

(10)

Specificity (SP) is the probability to identify a normal instance as normal correctly. Specificity can be named as True Negative Rate (TNR). Equation (11) shows the specificity calculation formula ^[29].

$\rm{S}\rm{P} = \frac{\rm{T}\rm{N}}{\rm{T}\rm{N}+\rm{F}\rm{P}}$

(11)

FPR is the chance that the normal instance incorrectly classified as an attack. FPR is also known as False Alarm Rate (FAR). FPR calculation is given by Eq (12) ^[30].

$\rm{F}\rm{P}\rm{R} = \frac{\rm{F}\rm{P}}{\rm{F}\rm{P}+\rm{T}\rm{N}}$

(12)

FNR is the chance that the attack instance incorrectly classified as normal. It is also named Miss Rate. FNR calculation is given in Eq (13) ^[30].

$\rm{F}\rm{N}\rm{R} = \frac{\rm{F}\rm{N}}{\rm{F}\rm{N}+\rm{T}\rm{P}}$

(13)

F₁ Score is the harmonic mean of precision and recall. F₁ Score serves as the derived effectiveness measurement. F₁ Score calculation is given by Eq (14) ^[30].

${\rm{F}}_{1}\rm{S}\rm{c}\rm{o}\rm{r}\rm{e} = 2\rm{*}\frac{\rm{P}\rm{r}\rm{e}\rm{c}\rm{i}\rm{s}\rm{i}\rm{o}\rm{n}\rm{*}\rm{R}\rm{e}\rm{c}\rm{a}\rm{l}\rm{l}}{\rm{P}\rm{r}\rm{e}\rm{c}\rm{i}\rm{s}\rm{i}\rm{o}\rm{n}+\rm{R}\rm{e}\rm{c}\rm{a}\rm{l}\rm{l}}$

(14)

We designed an overall system architecture for supporting the properties of Network IDS with the help of machine learning techniques. Figure 2 shows the flow diagram of system architecture consisting of network benchmark dataset, feature selection, proposed classifier, evaluation parameters and analysis of results.

5. Results and discussion

In this experimental analysis, the feature selection technique selects the optimal subset of informative features from the given dataset. We take the optimal subset of features for analyzing the impact of the proposed technique on how accurately it classifies the network traffic into normal or anomaly.

A comprehensive study was conducted to validate the impact of the proposed technique, such as performance metrics, feature selection analysis which is helpful in the prediction of anomalies in the network traffic. The main objective of this experimental study is that.

● To identify the optimal subset of informative features that contributes to intrusion from the intrusion dataset. Though the DPSO feature selection technique is not new, the main advantage of DPSO is that it won't get trapped in the local optimum that helps us select the best features for evaluation.

● To enhance the efficiency of intrusion detection using a hybrid classification technique, which is novel and is used to evaluate the performance and detection capabilities.

● To evaluate the proposed classification technique on the NSL_KDD dataset.

● To study the performance evaluation of the proposed classification technique by applying various performance measures such as accuracy, sensitivity, specificity, FPR and FNR.

● To compare the efficiency of the proposed technique with state-of-art methods.

5.1. Feature selection techniques analysis

A comprehensive study was performed between five different bio-inspired feature section techniques in terms of accuracy, detection rate and FPR. The proposed feature selection technique achieved a considerable performance improvement compared to other feature selection techniques. The results achieved after applying the feature selection and proposed classification technique to the NSL_KDD dataset is depicted in Table 4. It compares classification accuracy, classifiers detection rate and false positive rate of the proposed technique with the state-of-art techniques. In the compared feature selection techniques LHHLS-KH feature selection technique has given the best accuracy of 96.12%, detection rate of 96.48% and false positive rate of 4%. Now the performance of the proposed technique combined with feature selection gives an accuracy of 97.70 %, detection rate of 97% and false positive rate of 1.25 % which is 0.52, 1.58 and 2.75% higher than LHHLS-KH technique.

Table 4. Classification accuracy, DR and FPR of various feature selection techniques (NSL_KDD dataset).

Feature Selection	Number of Features	Selected features	Accuracy (%)	Detection Rate (DR) (%)	False Positive Rate (FPR) (%)
CMPSO ^[7]	33	2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 17, 18, 19, 20, 21, 22, 23, 25, 26, 27, 30, 32, 33, 34, 35, 37, 40, 41	80.91	83.01	24.70
ACO ^[7]	31	1, 3, 4, 6, 7, 8, 12, 14, 15, 16, 17, 19, 20, 21, 23, 24, 25, 27, 28, 29, 30, 33, 34, 35, 36, 37, 38, 39, 40, 41	84.03	87.15	19.30
KH ^[7]	26	2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 14, 15, 18, 19, 21, 22, 23, 24, 26, 28, 30, 31, 32, 40, 41	88.21	89.46	12.88
IKH ^[7]	25	2, 3, 4, 5, 6, 8, 10, 11, 12, 14, 17, 18, 20, 21, 22, 27, 28, 29, 30, 31, 34, 35, 36, 39, 41	91.22	91.75	7.34
LHHLS-KH ^[7]	19	2, 3, 4, 6, 8, 10, 11, 15, 17, 19, 20, 21, 29, 30, 33, 34, 36, 37, 40	96.12	96.48	4
DPSO	28	2, 3, 4, 6, 7, 8, 11, 12, 13, 15, 16, 17, 18, 19, 21, 23, 24, 25, 26, 28, 29, 30, 33, 34, 35, 36, 37, 40	97.70	97	1.25

| Show Table

DownLoad: CSV

To visualize the difference between classification accuracy, DR and FPR are shown in Figure 3. For the NSL_KDD dataset, the false positive rate of DPSO with the proposed classification technique is 1.25%. It reduces by 23.45, 18.05, 11.63, 6.09 and 2.75%, respectively, compared with Cross Mutation Particle Swarm Optimization (CMPSO), Ant Colony Optimization (ACO), Krill Herd (KH), Improved Krill Herd (IKH) and Linear Nearest Neighbor Lasso Step-Krill Herd (LHHLS-KH). Similarly, the detection rate of the DPSO with the proposed classification technique is 97% which is 13.99, 9.85, 7.54, 5.25, 0.52% higher than CMPSO, ACO, KH, IKH and LHHLS-KH. Similarly, the accuracy of DPSO with the proposed classification technique is 97.70% which is 16.79, 13.67, 9.49, 6.48 and 1.58% higher than CMPSO, ACO, KH, IKH and LHHLS-KH. In conclusion, the proposed classification technique with DPSO proves that it has higher detection accuracy, lower false positive rate and higher detection rate.

Figure 3. Comparison of classification accuracy, DR and FPR with FS techniques (NSL_KDD dataset).

DownLoad: Full-Size Img PowerPoint

5.2. Detection time analysis

The detection time was compared for further evaluation. Table 5 shows the detection time when applied to different techniques. The detection time represents the time taken from inputting the optimal subset of features into the classifier till the end of detection.

Table 5. Detection time of different feature selection techniques (NSL_KDD dataset).

Feature Selection Techniques	Detection Time (Sec)
CMPSO ^[20]	41.36
ACO ^[20]	41.39
KH ^[20]	37.78
IKH ^[20]	36.99
LNNLS-KH ^[20]	34.40
DPSO	41.56

| Show Table

DownLoad: CSV

Table 5 shows the time taken to detect the intrusion when different bio-inspired feature selection techniques are applied. The features selected by CMPSO, ACO, KH, IKH and LHHLS-KH are inputted to the individual classifier. In the existing systems compared, ACO feature selection took the highest detection time of 41.39 seconds with the accuracy of 84.03%, detection rate of 87.15% and false positive rate of 19.30%. Now the selected features with DPSO feature selection technique when given to a proposed hybrid classifier it took 41.56 seconds to detect the intrusion which is 0.41% higher than ACO feature selection technique. Even though it took slightly longer detection time there is a considerable increase in the performance. The DPSO detection accuracy is 97.7%, detection rate is 97% and false positive rate is 1.25% which when compared to ACO yields 13.67, 9.85% increase in accuracy, detection rate and 18.05% decrease in false positive rate.

5.3. Performance analysis of proposed classifier

The experimental study was conducted using the proposed stacked autoencoder hoeffding tree approach with and without feature selection. The proposed technique performance evaluation was based on accuracy, sensitivity, specificity, false-positive rate and false-negative rate. The results achieved with and without the feature selection of the proposed classification technique is given in Table 6.

Table 6. Performance comparison of the proposed technique with and without feature selection.

Performance Metrics (%)	SAE-HT without FS (41 Features)	SAE-HT with FS (28 Features)
Accuracy	55.93	97.7
Sensitivity (TPR)	100	97
Specificity (TNR)	28.57	98.75
False Positive Rate (FPR)	71.43	1.25
False Negative Rate (FNR)	0	3
F1 Score	37.6	98

| Show Table

DownLoad: CSV

The proposed SAE-HT was employed on the NSL_KDD dataset. Figure 4 demonstrates the performance comparison of proposed technique with and without FS. First, the proposed classifier was applied to all the 41 features of the dataset and found that the detection accuracy was only 55.93%. Moreover, the false positive rate was 71.43% and false negative rate was almost 0%. Our aim was to reduce the false positive rate to a greater extent and increase the accuracy as far as possible. Then the study was made and found that less relevant and unimportant features affect the performance of the proposed classifier. Second, the bio-inspired DPSO feature selection technique was applied and 28 important features that contributes to intrusion were selected. The selected features were given to the stacked autoencoder to learn the features so as to yield good performance and the output from the stacked autoencoder was given to hoeffding tree and found that the accuracy increased to 97.7%, false positive rate decreased to 1.25% which is 43.77% higher than SAE-HT without FS. Same way the false positive rate decreased by 70.18% when SAE-HT with FS was applied. Likewise, specificity was 97% and sensitivity rate was 98.75% when SEA-HT with FS was applied. Therefore, it proves that the proposed classifier with feature selection provides a good detection accuracy with less FPR and FNR.

Figure 4. Performance comparison of the proposed technique with and without feature selection.

DownLoad: Full-Size Img PowerPoint

The exact process is carried out for 29 features, and the accuracy is 97.3%. When applied to 30 features, the accuracy is 97.1%. From this, we conclude that the accuracy reaches saturation. There is no drastic improvement in accuracy when we include more features. As we increase the number of features once again, the accuracy starts decreasing. Therefore 28 features are the optimal features that contribute to intrusion. As the accuracy increases the false positive rate started decreasing.

Figure 5 depicts the performance of Stacked Autoencoder (SAE) and Stacked Autoencoder Hoeffding Tree (SAE-HT). The stacked autoencoder by itself gives a good accuracy, detection rate of 85.23%, 85.13% and reduced false alarm rate of 14.62% but achieves much better results in terms of accuracy, detection rate and false alarm rate when combined with the hoeffding tree. SAE-HT technique when applied it gave the accuracy of 97.7%, Detection Rate of 97% which is 12.47, 11.87% higher than SAE and false alarm rate of 1.25% which is decreased by 13.37% when compared to SAE.

Figure 5. Performance comparison: SAE Vs SAE-HT.

DownLoad: Full-Size Img PowerPoint

5.4. Comparison with state-of-art techniques

The proposed technique was compared with the state-of-Art techniques to prove its efficiency. Table 7 show the comparison of proposed with other state-of-art techniques. We have compared different conventional classification methods and deep leaning methods with our proposed technique.

Table 7. Comparison of proposed technique with state-of-art techniques.

Prediction Techniques	Accuracy (%)
VLSTM ^[8]	89.5
AdaBoost ^[8]	84.8
LSTM ^[8]	85.1
CNN LSTM ^[8]	83.3
SAE-SVM ^[10]	95.98
SAE ^[11]	85.23
SSAE ^[11]	86.02
DAE ^[11]	86.92
ContAE ^[11]	87.98
CAE ^[11]	81.07
SAE-DNN ^[12]	82.23
SAAE-DNN ^[12]	87.74
Hoeffding Tree ^[14]	93
Restricted Hoeffding Trees ^[14]	92.15
AUE2 with Buffer ^[14]	94.07
AUE2 without Buffer ^[14]	94.06
HAT + DDM + ADWIN ^[14]	92
Hierarchical Clustering and SVM ^[17]	95.72
Proposed SAE-HT	97.7

| Show Table

DownLoad: CSV

The SAE-HT classifier with feature selection gives the overall detection accuracy of 97.7 % which increases by 8.2, 12.9, 12.6, 14.4, 1.72, 12.47, 11.68, 10.78, 9.72, 16.7, 15.47, 9.96, 4.7, 5.55, 3.63, 3.64, 5.7 and 1.98% when compared to VLSTM, AdaBoost, LSTM, CNN LSTM, SAE-SVM, SAE, SAAE, DAE, Cont AE, CAE, SAE-DNN, SAAE-DNN, Hoeffding tree, Restricted hoeffding tree, AUE2 with buffer, AUE2 without buffer, HAT+DDM+ADWIN, Hierarchical Clustering and SVM.

Compared with the state-of-the-art techniques in the related work, the proposed SAE-HT classifier performs better in terms of accuracy, detection rate, and false alarm rate. Thus, proving that it has met the significant challenges by giving high accuracy of 97.7% and detection rate of 97% and well reduced false alarm rate of 1.25%.

6. Conclusions

An effective intrusion detection system was developed using the stacked autoencoder hoeffding tree technique for classification and the Darwinian particle swarm optimization method for feature selection. Using DPSO feature selection, the optimal subset of features contributing to intrusions are selected from the given intrusion dataset. The training and testing of data are given to our proposed classification technique to classify the network attacks. The main goal of the stacked autoencoder hoeffding tree technique is to increase the accuracy and reduce the false alarm rate. Moreover, we evaluated our proposed work on the NSL_KDD intrusion dataset. The dataset is streamed in a Matlab environment. Our proposed technique is applied and measures the classification performance using accuracy, specificity, sensitivity, detection rate, false alarm rate, false-negative rate. This classification performance is compared with other states of the art methods. The main challenge with the current intrusion detection is that it cannot detect the new unknown attacks. We proved that our proposed technique shows a robust significant amount of performance improvement in detection accuracy, thus reducing the false alarm rate to a greater extent. The supervised learning technique mainly depends on the predefined attack signatures that make new attacks goes undetected. The proposed stacked autoencoder hoeffding tree detects known attacks and, at the same time, detects unknown attacks to a greater extent. Results are compared with the state-of-art techniques and found that the SAE-HT technique gives higher accuracy of 97.7%, which increases by 8, 12.9, 12.6, 14.4%, 1.72, 12.47, 11.68, 10.78, 9.72, 16.7, 15.47, 9.96, 4.7, 5.55, 3.63, 3.64, 5.7 and 1.98% when compared to VLSTM, AdaBoost, LSTM, CNN LSTM, SAE-SVM, SAE, SAAE, DAE, Cont AE, CAE, SAE-DNN, SAAE-DNN, Hoeffding tree, Restricted hoeffding tree, AUE2 with buffer, AUE2 without buffer, HAT + DDM + ADWIN, Hierarchical Clustering and SVM and higher detection rate of 97%. Thus, reducing the false alarm rate to 1.25%. Further work can incorporate a study on a more effective method to improve the IDS. For the attack with fewer samples, an adversarial learning method can be used to create similar attacks, thus increasing the diversity of training examples that improve the detection accuracy.

Acknowledgments

This Publication is an outcome of the R & D work under MeitY's Visvesvaraya PhD Scheme, Government of India, implemented by Digital India Corporation.

Conflict of interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Funding

This research was supported by the Faculty of Management of Comenius University in Bratislava, Slovakia.

References

[1]	Gansu Province Bureau of Statistics (2021) Communiqué of the Seventh National Population Census in Gansu Province (No.6). Available from: http://tjj.gansu.gov.cn/tjj/c109465/202105/feef3c45ac4945528abf6c5d50dc3c2b.shtml.
[2]	Gansu Province Bureau of Statistics (2021) Communiqué of the Seventh National Population Census in Gansu Province (No.4). Available from: http://tjj.gansu.gov.cn/tjj/c109465/202105/d61df55e29384cf2b9cf79b26a4cfced.shtml.
[3]	Kou YW (2020) The interactive development of population urbanization and ecological urbanization in Gansu province - from the perspective of coordinated development of poverty alleviation and development. J University Sci Technol Beijing (Soc Sci Ed) 36: 55–61. http://dx.doi.org/10.3969/j.issn.1008-2689.2020.03.008 doi: 10.3969/j.issn.1008-2689.2020.03.008
[4]	Song Y, Liang JP, Ren ZL, et al. (2012) Research on sustainable development mode of rural left-behind children physical and mental health in western region of China - Taking left-behind children in Chongqing as example. China Sport Sci Technol 48: 100–108. http://dx.doi.org/10.3969/j.issn.1002-9826.2012.05.015 doi: 10.3969/j.issn.1002-9826.2012.05.015
[5]	General Office of the State Council (2011) Outline On the Development of Chinese Children (2011–2020). Available from: https://www.gov.cn/zwgk/2011-08/08/content-1920457.htm.
[6]	Duan CR, Lv LD, Guo J, et al. (2013) Survival and development of left-behind children in rural China: Based on the analysis of sixth census data. Population J 35: 37–49. http://dx.doi.org/10.3969/j.issn.1004-129X.2013.03.004 doi: 10.3969/j.issn.1004-129X.2013.03.004
[7]	Lu F, Fu SY (2022) The development of rural children's social development questionnaire. Stud Psychol Behav 20: 248–254. https://doi.org/10.12139/j.1672-0628.2022.02.015 doi: 10.12139/j.1672-0628.2022.02.015
[8]	Garau C, Annunziata A (2020) Supporting children's independent activities in smart and playable public places. Sustainability 12: 8352. https://doi.org/10.3390/su12208352 doi: 10.3390/su12208352
[9]	The Central People's Government of the People's Republic of China (2021) The prevalence of myopia among Chinese children and teenagers was 52.7 percent in 2020. Available from: https://www.gov.cn/xinwen/2021-07/13/content_5624709.htm.
[10]	Jiang CY, Jia YH (2022) The influence of skip-generation raising on the socialization of left-behind children in rural areas: Based on the baseline data of CEPS 2013–2014. J Yunnan Agr Univer (Soc Sci) 16: 28–35. http://dx.doi.org/10.12371/j.ynau(s).202206010 doi: 10.12371/j.ynau(s).202206010
[11]	Hazlehurst MF, Muqueeth S, Wolf KL, et al. (2022) Park access and mental health among parents and children during the COVID-19 pandemic. BMC Public Health 22: 800. https://doi.org/10.1186/s12889-022-13148-2 doi: 10.1186/s12889-022-13148-2
[12]	Lema-Gómez L, Arango-Paternina CM, Eusse-López C, et al. (2021) Family aspects, physical fitness, and physical activity associated with mental-health indicators in adolescents. BMC Public Health 21: 2324. https://doi.org/10.1186/s12889-021-12403-2 doi: 10.1186/s12889-021-12403-2
[13]	Gao J (2012) Research and practices on children outdoor playground in Japan. Landsc Archit 2012: 99–104. http://dx.doi.org/10.14085/j.fjyl.2012.05.024 doi: 10.14085/j.fjyl.2012.05.024
[14]	Si XP (2017) Research on the design of experiential children's public activity space environment. Master's thesis, TianGong University, Tianjin, 2017. http://dx.doi.org/10.7666/d.Y3193619
[15]	Huang Y (2016) Study on the mental health status of the left behind children in the rural areas after the extracurricular physical exercise program—By skipping, roller skating as an example. Master's thesis, Wuhan Sports University, Wuhan, 2016.
[16]	Li N (2008) Children's behavior psychology and children's park design. Master's thesis, Hunan University, Hunan, 2008. http://dx.doi.org/10.7666/d.y1260893
[17]	Ma HM (2019) A study on the optimization of urban and rural mass sports resources allocation in Gansu province from the perspective of healthy China. Master's thesis, Mudanjiang Normal University, Heilongjiang, 2019. http://dx.doi.org/10.27757/d.cnki.gmdjs.2019.000160
[18]	Liu W, Li C, Tong Y, et al. (2020) The places children go: Understanding spatial patterns and formation mechanism for children's commercial activity space in Changchun city, china. Sustainability 12: 1377. https://doi.org/10.3390/su12041377 doi: 10.3390/su12041377
[19]	Kemperman AD, Timmermans HJ (2011) Children's recreational physical activity. Leisure Sci 33: 183–204. https://doi.org/10.1080/01490400.2011.564922 doi: 10.1080/01490400.2011.564922
[20]	Pan NN (2018) Study on the Outdoor Activity site Design Based on the Characteristics of Children's Behavior. Master's thesis, Southeast University, Nanjing, 2018.
[21]	Zhou C, Long YL, Gan DX (2004) Discussion about application of "human-oriented" in the building of children's play space. J Hunan Agri Univer (Soc Sci Ed) 5: 84–85. http://dx.doi.org/10.3969/j.issn.1009-2013.2004.03.027 doi: 10.3969/j.issn.1009-2013.2004.03.027
[22]	Copperman RB, Bhat CR (2007) An analysis of the determinants of children's weekend physical activity participation. Transportation 34: 67–87. https://doi.org/10.1007/s11116-006-0005-5 doi: 10.1007/s11116-006-0005-5
[23]	Paleti R, Copperman RB, Bhat CR (2011) An empirical analysis of children's after school out-of-home activity-location engagement patterns and time allocation. Transportation 38: 273–303. https://doi.org/10.1007/s11116-010-9300-2 doi: 10.1007/s11116-010-9300-2
[24]	Hjort M, Martin WM, Stewart T, et al. (2018) Design of urban public spaces: Intent vs. reality. Int J Environ Res Public Health 15: 816. https://doi.org/10.3390/ijerph15040816 doi: 10.3390/ijerph15040816
[25]	Zhou Y, Wang M, Lin S, et al. (2022) Relationship between children's independent activities and the built environment of outdoor activity space in residential neighborhoods: A case study of Nanjing. Int J Environ Res Public Health 19: 9860. https://doi.org/10.3390/ijerph19169860 doi: 10.3390/ijerph19169860
[26]	Wang N (2017) Research on the interactive design of urban outdoor children recreation space. Master's thesis, Southwest University, Nanjing, 2017.
[27]	Gao L, Chen Y, Su HS (2016) Safety design of outdoor space for children in tropical cities: Taking outdoor children's playground of mid peninsula in Sanya for example. Decoration 2016: 104–105. http://dx.doi.org/10.16272/j.cnki.cn11-1392/j.2016.11.026 doi: 10.16272/j.cnki.cn11-1392/j.2016.11.026
[28]	Huang V, Zhang H, Song JQ, et al. (2014) The concept of "super flat" in contemporary commercial packaging design. Packag Eng 35: 37–40. http://dx.doi.org/10.19554/j.cnki.1001-3563.2014.12.009 doi: 10.19554/j.cnki.1001-3563.2014.12.009
[29]	Bao Y, Gao M, Luo D, et al. (2021) Effects of children's outdoor physical activity in the urban neighborhood activity space environment. Front Public Health 9: 631492. https://doi.org/10.3389/fpubh.2021.631492 doi: 10.3389/fpubh.2021.631492
[30]	Zhou WC (2018) Research on interesting landscape design for benefiting children's growth. Master's thesis, Wuhan Institute of Technology, Wuhan, 2018.
[31]	Sheng Q, Wan DY, Yu BY (2021) Effect of space configurational attributes on social interactions in urban parks. Sustainability 13: 7805. https://doi.org/10.3389/fpubh.2021.631492 doi: 10.3389/fpubh.2021.631492
[32]	Xie YY, Chen L, Zhou AH, et al. (2021) Accessibility analysis of children's service facilities at community scale in Beijing. World Geogr Res 30: 546. http://dx.doi.org/10.3969/j.issn.1004-9479.2021.03.2019548 doi: 10.3969/j.issn.1004-9479.2021.03.2019548
[33]	Tao C, Li JX, Lai DY (2021) Public activity space quality evaluation and optimization strategy. Planners 37: 75–83. http://dx.doi.org/10.3969/j.issn.1006-0022.2021.21.011 doi: 10.3969/j.issn.1006-0022.2021.21.011
[34]	Li A (2017) Research on the design of communication space for children in the city square. Master's thesis, Hebei University of Architecture, Hebei, 2017.
[35]	Gil-Madrona P, Martínez-López M, Prieto-Ayuso A, et al. (2019) Contribution of public playgrounds to motor, social, and creative development and obesity reduction in children. Sustainability 11: 3787. https://doi.org/10.3390/su11143787 doi: 10.3390/su11143787
[36]	Gao YJ, Wang CL, Huang ML, et al. (2022) A new perspective of sustainable perception: research on the smellscape of urban block space. Sustainability 14: 9184. https://doi.org/10.3390/su14159184 doi: 10.3390/su14159184
[37]	Yang MC, Tang M (2015) The methodology for fashion design based on shaping up human body. Creation Design 36: 70–77. http://dx.doi.org/10.3969/J.ISSN.1674-4187.2015.01.012 doi: 10.3969/J.ISSN.1674-4187.2015.01.012

This article has been cited by:

1.	V. Deeban Chakravarthy, K L. N. C. Prakash, Kadiyala Ramana, Thippa Reddy Gadekallu, 2023, Chapter 64, 978-981-19-3678-4, 733, 10.1007/978-981-19-3679-1_64
2.	Hao Zhang, Lina Ge, Guifen Zhang, Jingwei Fan, Denghui Li, Chenyang Xu, A two-stage intrusion detection method based on light gradient boosting machine and autoencoder, 2023, 20, 1551-0018, 6966, 10.3934/mbe.2023301
3.	Kadiyala Ramana, A. Revathi, A. Gayathri, Rutvij H. Jhaveri, C.V. Lakshmi Narayana, B. Naveen Kumar, WOGRU-IDS — An intelligent intrusion detection system for IoT assisted Wireless Sensor Networks, 2022, 196, 01403664, 195, 10.1016/j.comcom.2022.10.001
4.	Kranthi Kumar Singamaneni, Kadiyala Ramana, Gaurav Dhiman, Saurabh Singh, Byungun Yoon, A Novel Blockchain and Bi-Linear Polynomial-Based QCP-ABE Framework for Privacy and Security over the Complex Cloud Data, 2021, 21, 1424-8220, 7300, 10.3390/s21217300
5.	Olena Shlyakhetko, Michal Greguš, 2024, Chapter 5, 978-3-031-60814-8, 101, 10.1007/978-3-031-60815-5_5
6.	G. M. Nandana, Ashok Kumar Yadav, 2023, Chapter 8, 978-981-99-3314-3, 83, 10.1007/978-981-99-3315-0_8

Reader Comments

Your name:*

Email:*
© 2023 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)