CN117580046A

CN117580046A - Deep learning-based 5G network dynamic security capability scheduling method

Info

Publication number: CN117580046A
Application number: CN202310680297.XA
Authority: CN
Inventors: 孙建国; 池剑磊; 田野; 李顺; 亓昆
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2023-06-09
Filing date: 2023-06-09
Publication date: 2024-02-20

Abstract

The invention belongs to the technical field of 5G network security, and discloses a 5G network dynamic security capability scheduling method based on deep learning, which comprises the following steps: and (3) data acquisition: collecting data generated by scheduling of security capability in a 5G network; step 2: data preprocessing: preprocessing the acquired data; step 3: modeling a neural network: establishing a multi-task neural network model, taking the interaction and correlation between security capabilities into consideration, gradually finding and utilizing an implicit data mode through network self-adaptive learning, and continuously optimizing a prediction effect; step 4: and (3) intelligent scheduling: and optimizing the use of the safety resources in the 5G network by applying the prediction result provided by the trained neural network model. The invention optimizes the security capability scheduling process by utilizing the strong computing capability and self-adaptability of the neural network in the deep learning, effectively improves the network performance, overcomes the security holes existing in the 5G network, and simultaneously ensures the network performance and reliability.

Description

Deep learning-based 5G network dynamic security capability scheduling method

Technical Field

The invention belongs to the technical field of 5G network security, and particularly relates to a 5G network dynamic security capability scheduling method based on deep learning.

Background

With the advent of the 5G network era, the characteristics of high speed, large connection, low latency, etc. brought by the system have greatly changed people's life and industry. At the same time, however, the security problem of 5G networks is becoming more serious.

The 5G network exhibits high isomerism, and there are many different types of devices, networks, protocols, applications, etc., including satellite communications, internet of vehicles, internet of things, edge computing, etc. This heterogeneity makes 5G networks vulnerable to a number of network security vulnerabilities.

1) Protocol loopholes

The 5G network supports various protocols such as IP protocol, TCP protocol, UDP protocol, HTTP protocol, etc. However, these protocols have vulnerabilities of different degrees, such as SYN attack of TCP protocol, flood attack of UDP protocol, injection attack of HTTP protocol, etc. An attacker can compromise the security of a 5G network by exploiting these vulnerabilities for network attacks.

2) Application vulnerability

The 5G network is widely applied to various scenes, such as intelligent home, remote medical treatment, intelligent transportation and the like. In these scenarios, there are a large number of applications that may also have security vulnerabilities, such as unauthorized access, overflow attacks, malware, etc. An attacker can exploit these vulnerabilities to steal sensitive data from the 5G network, such as user identities, passwords, credit card numbers, etc.

3) Hardware vulnerability

There are a large number of hardware devices in 5G networks, such as radio towers, routers, switches, etc. These devices may also present security vulnerabilities, such as unencrypted communications, vulnerable operating systems, unauthorized access, etc. An attacker can take possession of the 5G network by attacking these devices, compromising network security.

4) Intelligent device vulnerability

With the development of 5G networks, more and more intelligent devices are connected to the networks, such as intelligent home devices, intelligent watches, and the like. These devices may also have security vulnerabilities, such as lack of identity authentication, a vulnerable operating system, etc. An attacker can attack the 5G network by attacking these devices, compromising network security.

5) Network management vulnerabilities

5G networks involve a large amount of network management such as authentication, authorization, access control, etc. Security holes, such as unauthorized access, weak passwords, etc., may also exist in these network management processes. An attacker can master the 5G network by attacking these network management processes, jeopardizing network security.

In summary, the heterogeneity of 5G networks faces a large number of network security vulnerabilities. For these vulnerabilities, comprehensive security assessment, vulnerability remediation, and enhanced network security measures are required.

Disclosure of Invention

The invention aims to provide a 5G network dynamic security capability scheduling method based on deep learning so as to solve the technical problems.

In order to solve the technical problems, the specific technical scheme of the 5G network dynamic security capability scheduling method based on deep learning is as follows:

A5G network dynamic security capability scheduling method based on deep learning comprises the following steps:

step 1: and (3) data acquisition: collecting data generated by scheduling of security capability in a 5G network;

step 2: data preprocessing: preprocessing the acquired data;

step 3: modeling a neural network: establishing a multi-task neural network model, taking the interaction and correlation between security capabilities into consideration, gradually finding and utilizing an implicit data mode through network self-adaptive learning, continuously optimizing a prediction effect, and improving the accuracy and efficiency of perception, reasoning and decision; step 4: and (3) intelligent scheduling: and optimizing the use of the safety resources in the 5G network by applying the prediction result provided by the trained neural network model.

Further, the step 1 comprises the following specific steps:

step 1.1: demand analysis: determining data to be acquired according to the requirements of the 5G security capability intelligent scheduling scene;

Step 1.1.1: confirming the type of data to be acquired;

step 1.1.2: the time range and the acquisition frequency of data are clearly acquired;

step 1.1.3: determining the purpose and specific application scene of data acquisition;

step 1.1.4: making a corresponding data acquisition plan and flow;

step 1.2: preparing equipment: preparing the equipment required for collection;

step 1.2.1: confirming the type of data to be acquired, and selecting a corresponding sensor and a corresponding collector;

step 1.2.2: determining the installation positions and the number of the equipment;

step 1.2.3: determining the installation mode and the arrangement mode of equipment;

step 1.2.4: determining a power supply mode of equipment;

step 1.3: and (3) data acquisition: after the acquisition equipment is installed and ready, executing a data acquisition program, and starting to acquire data;

step 1.3.1: installing acquisition equipment, and connecting corresponding sensors and collectors;

step 1.3.2: executing a data acquisition program, starting to acquire data, wherein the data acquisition program is performed in a manual input or automatic acquisition mode;

step 1.3.3: in the process of collecting data, the data collection time and the related information of the data source are recorded for subsequent data management and analysis.

Step 1.4: data preprocessing and screening: preprocessing and screening the acquired data, reserving useful data, and filtering useless or redundant data;

Step 1.4.1: performing preliminary cleaning on the acquired data to remove null values and abnormal values;

step 1.4.2: screening and sorting importance of the collected data according to the requirements and the application scene;

step 1.4.3: formatting and standardizing the screened data for subsequent data operation and analysis;

step 1.5: data storage and management: storing the screened data into a database and managing the data;

step 1.5.1: determining a data storage scheme and a database type, and selecting a corresponding data storage and management tool;

step 1.5.2: designing a database structure, and establishing a data table and an index;

step 1.5.3: archiving, backing up and cleaning the data to ensure the safety and availability of the data; step 1.6: data analysis and application: on the basis of data storage and management, analyzing and applying the data, constructing a model by using the data, and carrying out prediction and decision support;

step 1.6.1: according to the requirements and the application scene, using a data visualization tool, a database management tool, decision trees, clusters, a support vector machine and a naive Bayesian algorithm to perform data analysis; step 1.6.2: constructing a model by using the data, and carrying out prediction and decision support;

Step 1.6.3: and displaying the data analysis result through data visualization.

Further, the step 2 comprises the following specific steps:

step 2.1: when data cleaning is executed, selecting a data cleaning method according to the data type and the data quality;

step 2.1.1: checking the data, and removing the repeated data and the invalid data;

step 2.1.2: removing the missing value or using an interpolation method to fill in the missing value;

step 2.1.3: processing the data outliers;

step 2.1.4: carrying out data box division and discretization processing according to service requirements or data distribution characteristics; step 2.2: data conversion: the data conversion is carried out on the basis of data cleaning, and the data conversion comprises the conversion and normalization operation of the data;

step 2.2.1: adopting a tokenize, stemming, stopwords preprocessing method for the text data, and removing special characters in the text;

step 2.2.2: encoding and decoding the data;

step 2.2.3: normalizing or normalizing the numerical data;

step 2.3: counting data;

step 2.3.1: carrying out descriptive statistical analysis on the data;

step 2.3.2: drawing a data distribution diagram;

step 2.3.3: and carrying out hypothesis testing on the data, detecting whether the data meets the distribution hypothesis, and knowing the association relation among the variables.

Further, the step 3 comprises the following specific steps:

step 3.1: network topology: selecting ANN, CNN, RNN or GAN network for training by considering depth, width and structural factors of the network;

step 3.2: training a neural network; when a proper neural network topology model is selected, the factors of network layer number, node number of each layer and activation function are considered, the data quantity and complexity are considered in the node number of each layer, the data quantity is not large or the problem is simple, the node number is reduced to improve training speed and prevent model overfitting, and if the data quantity is large or the problem is complex, more nodes are added to better extract characteristics and generalize; selecting an activation function ReLU to accelerate the training process, and simultaneously reducing the gradient vanishing problem;

step 3.3: and (3) model tuning: performing model tuning according to the training result of the neural network model; step 3.4: model deployment: after modeling and tuning of the neural network are completed, the model is deployed into an actual 5G security scene, the neural network model is converted into an industrial-level language or frame during deployment, and security and usability tests are conducted during deployment so as to ensure reliability and stability of the model.

Further, the step 3.1 refers to the following basic components constituting the neural network when selecting a suitable network structure for training: input layer, output layer, hidden layer, activation function, loss function, optimizer.

Further, the step 3.2 comprises the following specific steps:

before training the neural network, training data is required to be prepared, wherein the data preparation comprises data cleaning, data conversion and data standardization processing steps, and a data set is divided into a training set, a verification set and a test set according to service requirements;

then, when a neural network topology model is selected, factors such as the number of network layers, the number of nodes at each layer, an activation function and the like are considered;

after the network topology structure of the model is determined, training of the neural network is started, the existing data set is used for carrying out back propagation and gradient descent algorithm, super-parameter optimization including learning rate, iteration times and other regularization parameters is carried out in the training process, after the training is completed, the performance of the model is evaluated through cross verification or a test set, and the training steps are as follows:

a. initializing network parameters: initializing parameters including weight, bias and the like;

b. Forward propagation: the input data is calculated and output forward through a network;

for a neural network with L layers, the input of the ith layer is xi, the output is yi, and the output of the L layer is yL;

z(i)＝W(i)x(i-1)+b(i)

y(i)＝f(z(i))

wherein f represents an activation function, and W and b represent weights and biases, respectively;

c. calculating a cost function: calculating the difference between the output and the real label;

d. back propagation: according to the cost function, the contribution of each weight and bias to the error is inversely deduced, and the parameters are updated;

calculating the gradient of the cost function on each weight and bias, taking the output layer as an example:

wherein J represents a cost function, and as such, by multiplying by element;

for the hidden layer:

δ(i)＝(W(i+1))Tδ(i+1)⊙f′(z(i))

where T represents the transpose of the matrix. The formulas are used for calculating the gradient of each weight and bias, and updating network parameters;

e. iterative optimization: repeating steps b to d until the cost function converges.

Further, the step 3.3 includes the following specific steps:

step 3.3.1: data preprocessing: the method comprises data normalization, standardization and data balancing;

step 3.3.2: and (3) network architecture design: selecting a network architecture, inactivating, batch normalizing, convoluting the size of a kernel, the size of a pooling kernel, the depth of a network and the width of the network;

Step 3.3.3: parameter initialization: the method comprises the steps of random parameter, pretraining with good generalization and initializing in a smaller range;

step 3.3.4: optimization algorithm: the method comprises the steps of calculation graph construction, SGD optimization and derivation algorithm, and adjustment of Learning Rate and Momentum;

step 3.3.5: regularization technique: including L1, L2 regularization, dropout, early Stopping; step 3.3.6: setting learning rate and iteration times;

after model training and optimization, the model is tested and evaluated.

Further, the step 4 includes the following specific steps:

step 4.1: and (3) system architecture design: before intelligent scheduling of the neural network model is performed, system architecture design is performed, the system architecture design considers factors of model training, model scheduling, data transmission and resource management, and according to specific scenes, intelligent scheduling of the model is realized by adopting a joint learning, multi-model fusion or incremental learning mechanism;

step 4.2: model selection: when model scheduling is carried out, the differences among different models are considered, including model structures, model sizes and algorithm performances, different models are selected for scheduling according to different scene requirements, meanwhile, the cost of model scheduling and profits of the models are considered, an intelligent scheduling strategy is formulated according to the factors, and the overall benefit of the system is maximized;

Step 4.3: model monitoring and evaluation: in the model scheduling process, real-time monitoring and evaluation are carried out on the model, problems are found in time and corresponding measures are taken through the real-time monitoring of the model, meanwhile, the performance and benefits of the model under different scenes are determined through the evaluation of the model, and corresponding adjustment and optimization are carried out on the model operation according to the evaluation result;

step 4.4: model deployment and integration: finally, the deployment and integration of the model are required to be completed in the model scheduling, the deployment environment and mode of the model are considered in the model deployment, different models are fused together by the model integration, corresponding interfaces are provided, and in the process of the model deployment and integration, data security is ensured by processing data transmission and encryption of the model;

step 4.5: intelligent scheduler design and implementation: the intelligent scheduler is a core component of model intelligent scheduling, and the intelligent scheduler is realized to consider model selection, resource utilization and cost control, and improve the resource utilization rate to the maximum through a reasonable scheduling strategy.

Further, the step 4.1 comprises the following specific steps:

step 4.1.1: model training: in a 5G security intelligent scheduling scene, the model training needs to consider the privacy of data and the real-time property of the data, and a method capable of protecting the data security and realizing faster training is adopted;

Step 4.1.2: model scheduling: to realize the intelligent scheduling process, the model is dynamically scheduled, the model structure and super parameters are adjusted in real time according to the current task demands and system resources, and the flexible, extensible and dynamically optimized model scheduling is realized by adopting a distributed system and a model self-adaptive adjustment technology;

step 4.1.3: and (3) data transmission: data compression, distributed storage, data processing and transmission protocol optimization technology are adopted, so that data transmission efficiency and system performance are improved;

step 4.1.4: and (3) resource management: for large-scale 5G security data, a distributed system, a containerization technology, a resource pre-allocation technology and a load balancing technology are adopted to realize efficient management of resources.

Further, the step 4.2 includes the following specific steps:

step 4.2.1: determining scene requirements: defining tasks to be realized by the model, the type of data to be processed and the data processing efficiency requirements in the scene requirements;

step 4.2.2: evaluating performance indexes: the indexes for measuring the performance of the model comprise accuracy, recall rate, precision and F1 Score, and the indexes are selected and balanced according to scene requirements;

step 4.2.3: model selection and comparison: comparing each model or algorithm according to the requirements and performance indexes, and selecting an optimal model;

Step 4.2.4: experiment and evaluation: in a specific scene, performing experiments and evaluations on the selected model, verifying the effect and performance of the model, evaluating the performance of the model by using a cross verification, test set verification and confusion matrix method, and finding an optimal model;

step 4.2.5: model optimization: in experiments and evaluations, aiming at optimization and improvement of a selected model, performance and robustness of the model are improved through super-parameter adjustment, data processing, feature engineering and the like;

step 4.2.6: deployment and adjustment: after the model is successfully selected and optimized, the model is deployed and adjusted in an actual scene so as to adapt to different data distribution and processing requirements.

The 5G network dynamic security capability scheduling method based on deep learning has the following advantages: the invention optimizes the security capability dispatching process by utilizing the strong computing capability and self-adaptability of the neural network in the deep learning, ensures the security of the 5G network, effectively improves the network performance, overcomes the security holes existing in the 5G network, and simultaneously ensures the network performance and reliability.

Drawings

Fig. 1 is a flow chart of a method for scheduling dynamic security capability of a 5G network based on deep learning.

Detailed Description

In order to better understand the purpose, structure and function of the invention, the invention relates to a 5G network dynamic security capability scheduling method based on deep learning, which is further described in detail below with reference to the accompanying drawings.

The invention adopts a load demand prediction algorithm based on a neural network to realize intelligent scheduling of 5G security capability. The algorithm utilizes a deep neural network model to generalize and analyze historical load data so as to predict the change trend of future loads and design a corresponding load scheduling strategy. Specifically, based on a feedforward neural network (Feedforward Neural Network) model, data is input to an input layer, and a prediction result is obtained through calculation of a plurality of hidden layers.

As shown in fig. 1, in a 5G security capability intelligent scheduling scenario, the method for scheduling 5G network dynamic security capability based on deep learning of the present invention includes the following steps:

1. and (3) data acquisition: a large amount of data generated by security capability scheduling in the 5G network, such as security traffic, attack type, resource utilization, etc., is collected. And different data needs to be acquired for different types of security capabilities. These data require high timeliness, as well as high quality intelligent mining and processing capabilities.

The data acquisition process under the 5G security capability intelligent scheduling scene is as follows:

1) Demand analysis

And determining which data need to be collected according to the requirements of the 5G security capability intelligent scheduling scene. The main steps of demand analysis include:

confirm the type of data that needs to be collected, such as network behavior data, device data, user behavior data, etc.

-time range and acquisition frequency of explicitly acquired data.

Determining the purpose and specific application scenario of data acquisition, such as data analysis, anomaly detection, etc.

-formulating a corresponding data acquisition plan and flow.

2) Device preparation

The equipment needed for acquisition is ready, including sensors, collectors, etc. The main steps of equipment preparation include:

-identifying the type of data to be acquired, selecting the corresponding sensor and collector.

-determining the installation position and number of devices.

-determining the installation and arrangement of the devices.

-determining the power supply mode of the device.

3) Data acquisition

After the acquisition equipment is installed and ready, a data acquisition program is executed, and data acquisition is started. When data acquisition is carried out, the integrity and the accuracy of the data are required to be ensured, so that the acquired data are prevented from being missing or wrong. To this end, we can take the following steps: the target and the range of data acquisition are defined, and unnecessary data acquisition or important data missing are avoided; the proper data source is selected, the authoritative data source can be selected, and the possible data quality problem of the informal data source is avoided; the efficiency and the accuracy of data acquisition are improved by using an automatic tool such as a network data grabbing tool (Scrapy, beautifulSoup, selenium and the like), a data crawler (Octoparse, dataddo, parsehub and the like), an automatic testing tool (UFT, selenium, appium and the like) and the like; using a variety of data verification techniques, such as data comparison techniques, data cleansing techniques, and the like; finally, by adopting a data backup and archiving mechanism, the collected data can be ensured not to be lost due to system faults or other problems, and meanwhile, the long-term storage and use of the data can be ensured. The main steps of data acquisition include:

-installing a collecting device, connecting the respective sensor and the collector.

-executing a data acquisition procedure, starting to acquire data. The data acquisition procedure may be performed by means of manual input or automatic acquisition.

-recording data acquisition time, data source, etc. related information during acquisition of data for subsequent data management and analysis.

4) Data preprocessing and screening

Preprocessing and screening are carried out on the collected data, useful data is reserved, and useless or redundant data is filtered. The main steps of data preprocessing and screening include:

-performing a preliminary cleaning of the collected data, removing nulls, outliers, etc.

-filtering and ranking the collected data according to the requirements and application scenario.

-formatting and normalizing the screened data for subsequent data manipulation and analysis.

5) Data storage and management

And storing the screened data into a database, and managing the data, including data archiving, backup, cleaning and other works, so as to ensure the safety and usability of the data. The main steps of data storage and management include:

-determining the data storage scheme and the database type, selecting the corresponding data storage and management tool.

-designing a database structure, creating a data table and an index.

Archiving, backing up and cleaning up data, ensuring the security and availability of the data.

6) Data analysis and application

And analyzing and applying the data on the basis of data storage and management. And a model is constructed by utilizing the data, prediction and decision support are carried out, and the safety performance and reliability of the 5G network are improved. The main steps of data analysis and application include:

data analysis using data visualization tools (Tableau, power BI, etc.), data mining tools (Tableau, power BI, etc.), database management tools (MySQL, postgreSQL, etc.), and algorithms such as decision trees, clusters, support vector machines, naive bayes, etc., according to the requirements and application scenarios.

-constructing a model using the data for prediction and decision support.

-presenting the data analysis results by data visualization.

2. Data preprocessing: the collected data is preprocessed, such as data cleaning, sampling, denoising, feature extraction and the like, so as to improve the data quality and effectiveness. The data preprocessing is an important step in a 5G security intelligent scheduling scene, and collected data can be more accurate, reliable and processable in the modes of data cleaning, data conversion, data statistics and the like, so that the method is more beneficial to subsequent data analysis and algorithm selection.

1) Data cleansing

In performing data cleansing, it is necessary to select an appropriate data cleansing method depending on the data type and data quality. For example, for digital data, the method of detecting abnormal value and missing value, smoothing data and the like can be adopted to clean the data; for text data, data cleaning may be performed by removing spaces, removing special characters, and the like. The main steps of data cleaning are as follows:

-checking the data, removing duplicate data and invalid data.

-removing missing values or saving missing values using interpolation methods.

Processing data outliers, e.g. removing extreme outliers, processing continuous outliers using mode, mean, etc.

And carrying out data processing modes such as data binning, discretization and the like according to service requirements or data distribution characteristics, so as to facilitate subsequent algorithm processing.

2) Data conversion

The data conversion is performed on the basis of data cleansing. The data conversion comprises the operations of converting and normalizing the data, so that the scale effect of the data and errors caused by the conversion of different data values are avoided. The main steps of data conversion are as follows:

-removing special characters in the text by preprocessing the text data using tokenize, stemming, stopwords or the like.

Encoding and decoding data, for example converting binary data into hexadecimal or converting string data into digital format.

Normalization or normalization (including linear and nonlinear transformations) of the numeric data, mapping the values of the data into a suitable range of intervals, avoiding the cumulative effect of data errors.

3) Data statistics

The data statistics is used for knowing the distribution condition of various indexes of the data, and is favorable for the selection of a subsequent algorithm. The main steps of data statistics are as follows:

descriptive statistical analysis of the data, such as calculating means, standard deviations, variances, etc.

Drawing a data distribution map, such as a histogram, a scatter diagram, a line diagram and the like, and a visual statistical method to help people to more intuitively know the distribution situation of the data.

-performing hypothesis testing on the data, such as chi-square testing, t-test, etc., detecting whether the data meets the distribution hypothesis, and knowing the association between the variables.

3. Modeling a neural network: a multi-tasking neural network model is built and interactions and correlations between security capabilities are considered. Through network self-adaptive learning, implicit data modes are discovered and utilized step by step, the prediction effect is continuously optimized, and the accuracy and efficiency of perception, reasoning and decision are improved. The neural network modeling under the 5G security capability intelligent scheduling scene needs to be subjected to the data preprocessing process, and a plurality of steps such as network topology structure selection, neural network training, model tuning, model deployment and the like. By repeating these steps, the model is optimized and improved continuously, so that a proper model can be obtained to cope with the problem of 5G network security.

1) Network topology

The selection of an appropriate network topology is one of the keys for neural network modeling. When the network topology structure is selected, factors such as depth, width, structure and the like of the network need to be considered, different network structures can be tried to train and compare which is more suitable for the scene. The following list several common neural network structures:

-a traditional Artificial Neural Network (ANN): this network architecture consists of multiple layers, with full connectivity between each layer. The input layer receives the original data, and the output layer outputs the final result. ANN is applicable to a variety of classification and regression problems.

-Convolutional Neural Network (CNN): CNN is a neural network structure applied to the fields of image processing and computer vision, and consists of a plurality of layers such as convolution layers, pooling layers and full connection layers. CNNs can learn the features of an image from an input image and classify or regress it.

-Recurrent Neural Network (RNN): the RNN is a neural network structure with memory capacity and is mainly applied to processing sequence data. The structure of the system comprises a loop connection, so that the neural network can consider past information when processing the sequence data.

-generating an antagonism network (GAN): GAN consists of two neural networks, namely a generator and a arbiter. The generator is responsible for generating the dummy data, and the arbiter is responsible for distinguishing between the true and false data. GAN is suitable for generative tasks such as image segmentation and image generation.

In addition to the above several common neural network structures, there are many variations, such as Deep Neural Networks (DNNs), convolutional Recurrent Neural Networks (CRNNs), and the like.

2) Neural network training

The training data needs to be prepared before the neural network training is performed. The data preparation comprises the processing steps of data cleaning, data conversion, data normalization and the like so as to ensure the processibility and the accuracy of the data. In addition, the data set is divided into a training set, a verification set and a test set according to the service requirement, so that the generalization capability and the expandability of the model are ensured.

And then when a proper neural network topology model is selected, the network layer number, the node number of each layer, an activation function and other factors need to be considered. Where the number of network layers affects the ability and complexity of the network model to learn data, neither too deep nor too shallow a network is generally optimal. In practice, some shallow neural networks, such as logistic regression or a simple neural network, can easily deal with some very familiar problems. When complex data is processed, the deep neural network is more suitable; the number of nodes per layer needs to take into account the amount of data and the complexity. If the data size is not large or the problem is relatively simple, the number of nodes can be reduced to increase the training speed and prevent the model from being over fitted. If the data volume is large or the problem is complex, more nodes are needed to better extract the characteristics and generalize; the choice of activation function also has a great influence on the performance of the neural network. Common activation functions include sigmoid, reLU, and tanh, etc., and appropriate activation functions may be selected according to the particular problem. In general, reLU is the most commonly used activation function in deep learning models, as it can accelerate the training process while reducing the gradient vanishing problem. Depending on the actual situation, convolutional neural networks, recurrent neural networks, deep neural networks, or other types of neural network models may be employed.

After the network topology of the model is determined, training of the neural network is started. Training of neural networks requires the use of existing data sets for back propagation and gradient descent algorithms. Super-parameter optimization is needed in the training process, including learning rate, iteration times, other regularization parameters and the like. After training is completed, the performance of the model needs to be assessed by cross-validation or test set. The training steps are as follows:

a. initializing network parameters: including initialization of parameters such as weights, biases, etc.

b. Forward propagation: the input data is output through network forward computation.

For a neural network with L layers, the input of the ith layer is xi, the output is yi, and the output of the L layer is yL.

z(i)＝W(i)x(i-1)+b(i)

y(i)＝f(z(i))

Where f represents the activation function and W and b represent the weight and bias, respectively.

c. Calculating a cost function: the gap between the output and the real tag is calculated.

d. Back propagation: and according to the cost function, the contribution of each weight and bias to the error is deduced, and the parameters are updated.

Gradients of the cost function for the respective weights and biases are calculated. Take the output layer as an example:

where J represents a cost function, and by-element multiplication.

For the hidden layer:

δ(i)＝(W(i+1))Tδ(i+1)⊙f′(z(i))

Where T represents the transpose of the matrix. These formulas are used to calculate the gradient of each weight and bias and thus update the network parameters.

3) Model tuning

According to the training result of the neural network model, model tuning is needed. The main aim of model tuning is to improve the accuracy and generalization capability of the model and avoid the problems of over-fitting or under-fitting of the model.

The model tuning process of the neural network generally comprises the following steps:

-data preprocessing. Including Data Normalization, data Balancing, and the like.

-network architecture design. Including selecting an appropriate network architecture, deactivation (Dropout), batch normalization (Batch Normalization), size of convolution kernel, size of pooling kernel, network depth, width, etc.

-parameter initialization. Including random parameters, pretraining with good generalization, initialization in a smaller range, etc.

-an optimization algorithm. Including computational graph construction, optimization of SGD, various derivative algorithms, tuning of Learning Rate and Momentum, and the like.

-regularization techniques. Including L1, L2 regularization, dropout, early Stopping, etc.

-learning rate, number of iterations. Unreasonable learning rate and excessive training time can affect the performance of the neural network.

Taking a deep neural network as an example, the implementation process of model tuning is described below.

-data preprocessing: batch normalization and data enhancement

a. Batch normalization: the data standardization ensures that the data cannot deviate to a certain dimension, when the parameter training is used, the data input by each layer is standardized once, the influence of the gradient learned in the training on the updating amount of the layer is evaluated by the average value, the data standardization is performed, and the sensitivity of the network to the initial weight is reduced.

b. Data enhancement: data enhancement can be achieved by scaling, shearing, rotating, flipping, etc. the original data to generate a large amount of new data, thereby increasing the diversity of the data and improving model accuracy.

Network architecture design

a. Residual network (res net) was introduced: with ResNet, gradients can propagate directly across the entire network through identity functions, simplifying training in deep networks. The ResNet builds a residual network through a residual block, and the residual block can ensure that the residual network has better jump connection and information flow.

b. Increasing the Batch Size: because the neural network has large calculation amount, the method of Batch Size is selected to accelerate calculation and shorten training time. When the Size of the Batch is larger, more video memory and memory are needed, and better accuracy can be obtained.

Parameter initialization

a. Extensive random initialization: the initialization of neurons is progressively wider by defining upper and lower bounds for extensive random initialization. A number of random initialization exercises may be tried and then the best results selected.

b. Non-standard initialization: different random numbers of different samples are selected for initialization, namely random sampling. And searching for the optimal initialization by testing after a plurality of random number samples are acquired.

Optimization algorithm

adagrad: the idea of the algorithm is to adjust the learning rate according to the historical gradient and adapt the method.

Rmsprop: is an accelerated random gradient descent (SGD) algorithm. RMSprop adjusts the learning rate by calculating the mean gradient of the root mean square decay.

Adam: the adaptive learning rate optimization algorithm is used for estimating the efficient adaptation coefficient by calculating the motion variance and the motion average value of the gradient, so that a better effect is realized.

Regularization technique

Dropout: dropout is particularly effective for MLP and the like to generate overfitting neural networks, which can eliminate the memory of neurons in nonlinear activation tests.

Early stop: the error rate of the data of the verification set is obtained at any time in the training process, and when the error rate of the verification set is not reduced, the training is terminated in advance.

-learning rate and training iteration number

a. Different learning rates and epochs are tried to find the optimal value.

Based on the tuning techniques, we can train a deep neural network with better performance. After model training and optimization, the model needs to be tested and evaluated. In evaluating the model, various quality evaluation indexes such as accuracy, recall, F1 value and the like can be adopted. Selecting the appropriate test set and evaluation index can help us more accurately measure the quality and performance of the model.

4) Model deployment

After modeling and tuning of the neural network is completed, the model needs to be deployed into an actual 5G security scene. At deployment time, the neural network model needs to be converted into an industrial-level language or framework, such as c++, java, etc., in order to be practical. During deployment, necessary security and usability tests are also required to ensure reliability and stability of the model.

4. And (3) intelligent scheduling: and (3) optimizing the use of safety resources in the 5G network by applying a prediction result provided by the trained neural network model, and effectively aiming at sudden safety events and predictable safety threats. And the resource allocation is adjusted in real time through intelligent scheduling, so that the level and efficiency of network security are improved.

It should be noted that, in the 5G security capability intelligent scheduling scenario, the load demand prediction algorithm of the model needs to consider the real-time performance of the security event, so as to ensure that the system has the capability of high availability, high elasticity and dynamic adjustment, and simultaneously ensure user experience and privacy protection.

1) System architecture design

Before intelligent scheduling of the neural network model, a system architecture design is required. The system architecture design needs to take into account a number of factors, such as model training, model scheduling, data transmission, resource management, etc.

Model training: in a 5G security capability intelligent scheduling scene, model training needs to consider the privacy of data and the real-time property of the data. Therefore, methods such as federal learning (Federated Learning) and incremental learning (Incremental Learning) should be employed that can protect data security and enable faster training.

Model scheduling: to realize the intelligent scheduling process, the model needs to be dynamically scheduled, and the model structure, super parameters and the like are adjusted according to the current task demands and the real-time performance of system resources. Therefore, the flexible, extensible and dynamically optimized model scheduling can be realized by adopting the technologies of a distributed system, model self-adaptive adjustment and the like.

-data transmission: in a 5G scenario, the data transmission speed is high, but the problems such as packet loss or delay possibly occurring in the transmission process may affect the performance of the system. In order to solve these problems, techniques such as data compression, distributed storage, data processing, and transmission protocol optimization may be employed, thereby improving data transmission efficiency and system performance.

-resource management: for large-scale 5G security data, the system needs to consider an effective resource management mechanism to ensure the high efficiency and scalability of the system. The efficient management of resources can be realized by adopting a distributed system, a containerization technology, resource pre-allocation, load balancing and other technologies.

According to specific scenes, intelligent scheduling of the models can be realized by adopting mechanisms such as joint learning, multi-model fusion, incremental learning and the like.

2) Model selection

When model scheduling is performed, the variability among different models needs to be considered, including model structure, model size, algorithm performance and the like. Different models can be selected for scheduling according to different scene needs.

Assuming a deep neural network model as the best modeling scheme in a 5G security capability intelligent scheduling scenario, the following is the chosen procedure:

a. Determining scene requirements: in the scene requirement, a task to be implemented by the model, a type of data to be processed, a data processing efficiency requirement, and the like need to be defined. For example, if incremental learning is required, a model with online learning capability is required.

b. Evaluating performance indexes: the indexes for measuring the performance of the model comprise accuracy, recall rate, precision, F1Score and the like, and can be selected and balanced according to scene requirements.

c. Model selection and comparison: and comparing the models or algorithms according to the requirements and the performance indexes, and selecting an optimal model. Common models include Convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), long and short term memory networks (LSTMs), residual neural networks (res net), and the like.

d. Experiment and evaluation: in a specific scenario, experiments and evaluations of the selected model are required to verify its effectiveness and performance. The performance of the model can be evaluated by using methods such as cross-validation, test set validation, confusion matrix, and the like, and the optimal model can be found.

e. Model optimization: in experiments and evaluations, performance and robustness of the model can be improved through super parameter adjustment, data processing, feature engineering and the like aiming at optimization and improvement of the selected model.

f. Deployment and adjustment: after successful selection and optimization of the model, the model needs to be deployed and adjusted in the actual scene to adapt to different data distribution and processing requirements.

After the final model is selected, the cost of model scheduling, profit of each model and the like need to be considered. According to the factors, an intelligent scheduling strategy can be formulated, and the overall benefit of the system is maximized.

3) Model monitoring and evaluation

In the model scheduling process, the model needs to be monitored and evaluated in real time. Through real-time monitoring of the model, problems can be found in time and corresponding measures can be taken. Meanwhile, through the evaluation of the model, the performance and benefit of the model in different scenes can be determined. And according to the evaluation result, the model operation can be correspondingly adjusted and optimized.

4) Model deployment and integration

And finally, completing the deployment and integration of the model in the model scheduling. Model deployment requires consideration of the deployment environment and manner of the model. Model integration requires fusing different models together and providing corresponding interfaces. In the process of model deployment and integration, data transmission and encryption of the model are required to be processed to ensure data security.

5) Intelligent scheduler design and implementation

The intelligent scheduler is a core component of model intelligent scheduling, and an efficient, stable and extensible scheduler needs to be designed and realized. The implementation of the intelligent scheduler requires consideration of a number of factors, such as model selection, resource utilization, cost control, etc. The resource utilization can be improved to the maximum through a reasonable scheduling strategy.

The intelligent scheduler should take the following design measures:

a. and (3) data management: intelligent schedulers need the ability to achieve fast access and processing of data. In order to improve the data access speed and reduce the memory occupation, the data management can be optimized by adopting the technologies of memory sharing, data partition storage, data compression and the like.

b. Model management: the intelligent scheduler needs to manage various models and dynamically make model selection, tuning and combining. In order to realize flexible model management, the technology of incremental learning, self-adaptive adjustment, model integration and the like can be adopted, and meanwhile, the technology of distributed storage, calculation and the like is adopted, so that efficient management and scheduling of the model are realized.

c. And (3) task management: the intelligent scheduler needs to perform task management aiming at different tasks and scenes, and the effectiveness and the high efficiency of the tasks are ensured. Techniques such as automated task allocation, task priority, and real-time task scheduling may be employed to manage tasks.

d. System coordination: the intelligent scheduler needs to coordinate interactions and collaboration among multiple system components. Event driven, distributed communication, and message queuing techniques may be employed to optimize system coordination.

e. And (3) ensuring safety: the intelligent dispatcher needs to ensure the security of data and network, and prevent the problems of malicious attack, data leakage and the like. The security of the system can be ensured by adopting measures such as encryption communication, identity authentication, access control, security management and the like.

In summary, the neural network model intelligent scheduling in the 5G security capability intelligent scheduling scenario needs to undergo multiple steps such as system architecture design, model selection, model monitoring and evaluation, model deployment and integration, and intelligent scheduler design and implementation. By implementing the steps, the intelligent degree and the autonomous capability of the system can be improved, and the 5G network safety can be effectively ensured.

It will be understood that the invention has been described in terms of several embodiments, and that various changes and equivalents may be made to these features and embodiments by those skilled in the art without departing from the spirit and scope of the invention. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the invention without departing from the essential scope thereof. Therefore, it is intended that the invention not be limited to the particular embodiment disclosed, but that the invention will include all embodiments falling within the scope of the appended claims.

Claims

1. The 5G network dynamic security capability scheduling method based on deep learning is characterized by comprising the following steps of:

step 2: data preprocessing: preprocessing the acquired data;

step 3: modeling a neural network: establishing a multi-task neural network model, taking the interaction and correlation between security capabilities into consideration, gradually finding and utilizing an implicit data mode through network self-adaptive learning, continuously optimizing a prediction effect, and improving the accuracy and efficiency of perception, reasoning and decision;

step 4: and (3) intelligent scheduling: and optimizing the use of the safety resources in the 5G network by applying the prediction result provided by the trained neural network model.

2. The deep learning-based 5G network dynamic security capability scheduling method according to claim 1, wherein the step 1 comprises the following specific steps:

step 1.1.1: confirming the type of data to be acquired;

step 1.1.4: making a corresponding data acquisition plan and flow;

step 1.2: preparing equipment: preparing the equipment required for collection;

step 1.2.4: determining a power supply mode of equipment;

step 1.6.1: according to the requirements and the application scene, using a data visualization tool, a database management tool, decision trees, clusters, a support vector machine and a naive Bayesian algorithm to perform data analysis;

step 1.6.2: constructing a model by using the data, and carrying out prediction and decision support;

Step 1.6.3: and displaying the data analysis result through data visualization.

3. The deep learning-based 5G network dynamic security capability scheduling method according to claim 1, wherein the step 2 comprises the following specific steps:

step 2.1.3: processing the data outliers;

step 2.2.2: encoding and decoding the data;

step 2.2.3: normalizing or normalizing the numerical data;

step 2.3: counting data;

step 2.3.1: carrying out descriptive statistical analysis on the data;

Step 2.3.2: drawing a data distribution diagram;

4. The deep learning-based 5G network dynamic security capability scheduling method according to claim 1, wherein the step 3 comprises the following specific steps:

5. The deep learning-based 5G network dynamic security capability scheduling method according to claim 4, wherein the step 3.1 selects a suitable network structure for training, and refers to the following basic components that make up the neural network: input layer, output layer, hidden layer, activation function, loss function, optimizer.

6. The deep learning-based 5G network dynamic security capability scheduling method according to claim 4, wherein the step 3.2 comprises the following specific steps:

z(i)＝W(i)x(i-1)+b(i)

y(i)＝f(z(i))

wherein J represents a cost function, and as such, by multiplying by element;

for the hidden layer:

δ(i)＝(W(i+1))T6(i+1)⊙f'(z(i))

7. The deep learning-based 5G network dynamic security capability scheduling method of claim 4, wherein the step 3.3 includes the following specific steps:

step 3.3.5: regularization technique: including L1, L2 regularization, dropout, early Stopping;

step 3.3.6: setting learning rate and iteration times;

after model training and optimization, the model is tested and evaluated.

8. The deep learning-based 5G network dynamic security capability scheduling method according to claim 1, wherein the step 4 comprises the following specific steps:

9. The deep learning-based 5G network dynamic security capability scheduling method of claim 8, wherein the step 4.1 includes the following specific steps:

10. The deep learning-based 5G network dynamic security capability scheduling method of claim 8, wherein the step 4.2 includes the following specific steps: