CN114070775B - Block chain network slicing security intelligent optimization method for 5G intelligent networking system - Google Patents

Block chain network slicing security intelligent optimization method for 5G intelligent networking system Download PDF

Info

Publication number
CN114070775B
CN114070775B CN202111203106.8A CN202111203106A CN114070775B CN 114070775 B CN114070775 B CN 114070775B CN 202111203106 A CN202111203106 A CN 202111203106A CN 114070775 B CN114070775 B CN 114070775B
Authority
CN
China
Prior art keywords
network
data
blockchain
model
supervised learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202111203106.8A
Other languages
Chinese (zh)
Other versions
CN114070775A (en
Inventor
伍军
施远
李高磊
李建华
洪源
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Jiaotong University
Shanghai Intelligent and Connected Vehicle R&D Center Co Ltd
Original Assignee
Shanghai Jiaotong University
Shanghai Intelligent and Connected Vehicle R&D Center Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shanghai Jiaotong University, Shanghai Intelligent and Connected Vehicle R&D Center Co Ltd filed Critical Shanghai Jiaotong University
Priority to CN202111203106.8A priority Critical patent/CN114070775B/en
Publication of CN114070775A publication Critical patent/CN114070775A/en
Application granted granted Critical
Publication of CN114070775B publication Critical patent/CN114070775B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L45/00—Routing or path finding of packets in data switching networks
    • H04L45/12—Shortest path evaluation
    • H04L45/121—Shortest path evaluation by minimising delays
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L45/00—Routing or path finding of packets in data switching networks
    • H04L45/28—Routing or path finding of packets in data switching networks using route fault recovery
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/10—Protocols in which an application is distributed across nodes in the network
    • H04L67/104—Peer-to-peer [P2P] networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Abstract

The invention relates to a block chain network slice security intelligent optimization method for a 5G intelligent network connection, which comprises the following specific steps: step 1: establishing a mobile blockchain network based on a 5G slicing environment; step 2: obtaining an original data set of the mobile block chain network operation, wherein the original data set comprises data under the normal operation condition and data under the transmission link fault condition, and carrying out data preprocessing; step 3: establishing a federal semi-supervised learning model according to a link state inference algorithm based on machine learning and training; step 4: the invention has the advantages that compared with the prior art, the invention has the advantages of enabling the mobile blockchain network to converge rapidly, improving the reasoning speed obviously, enabling the blockchain nodes on different network slices in the Internet of things to transmit local perception data more effectively, and the like.

Description

Block chain network slicing security intelligent optimization method for 5G intelligent networking system
Technical Field
The invention relates to the field of intelligent network slice security optimization, in particular to a block chain network slice security intelligent optimization method for a 5G intelligent network system.
Background
The current blockchain technology is not only applied to the fields of finance and digital encryption currency, but also starts to be applied to the fields of network security, industrial Internet of things and the like, the current blockchain network is established on an IP network, the time delay of a transmission block between the blockchain nodes brings a lot of security threats to the blockchain network, and a lot of attacks to the blockchain network are realized by using the transmission time delay. In addition, the transmission delay of the block also affects the consensus speed of the blockchain network, but the transmission delay cannot be completely eliminated, and the transmission delay can be shortened as far as possible, so that minimizing the delay of the transmission block between the blockchain nodes has become a key point of the application scene of expanding the blockchain by related institutions and enterprises, which is also helpful to the blockchain security.
Meanwhile, mobile communication technology has also undergone rapid iteration and development, so far, the 5G of the present day has been entered into the 5G age, and not only can the network transmission speed ten times of that of 4G be provided, but also the characteristics of high scalability and low latency can be configured and reasonably divided by the network slicing technology as required, more service scenarios are served, more orthogonal and diversified services are provided, therefore, in order to adapt to more special scenarios such as disaster management and battlefield investigation and mobile internet of things, etc., the blockchain network starts to migrate from the original IP network to the mobile communication network, and as such the blockchain network established on the mobile communication network is called mobile blockchain network, with the continuous penetration of 5G in the industrial internet of things, the open interconnection degree of industrial critical information infrastructure is rapidly promoted, in order to meet the dual requirements of intelligent network connection scenes in terms of low delay, reliable service support and privacy security protection, 5G and blockchain technologies are fused into global hot spots in the field, but mobile blockchain networks also face similar problems as the existing blockchain networks, in new generation mobile communication networks, in order to support different industry diversified service scenes, customized network functions are required to be provided for different application scenes according to different functions and requirements, therefore, network slicing technology is proposed, nodes distributed on the blockchain networks are likely to be located in different slices in the mobile blockchain networks, the stability and failure rate of data transmission links among the slices can have great influence on the consensus convergence of the mobile blockchain networks, the existing methods do not take into account the impact of transmission link failure between different network slices in a mobile blockchain network on the convergence of the blockchain network's consensus.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a block chain network slicing security intelligent optimization method oriented to a 5G intelligent network.
The aim of the invention can be achieved by the following technical scheme:
a block chain network slice security intelligent optimization method oriented to a 5G intelligent network system comprises the following steps:
step 1: establishing a mobile blockchain network based on a 5G slicing environment;
step 2: obtaining an original data set of the mobile block chain network operation, wherein the original data set comprises data under the normal operation condition and data under the transmission link fault condition, and carrying out data preprocessing;
step 3: establishing a federal semi-supervised learning model according to a link state inference algorithm based on machine learning and training;
step 4: and obtaining an optimized global model after training the federal semi-supervised learning model so as to rapidly predict the range of a fault transmission link in the blockchain network and realize predictive rapid consensus convergence.
In the step 1, the core architecture of the mobile blockchain network includes a 5G network access layer, a 5G network slice layer and a mobile blockchain network application layer that are set from bottom to top.
The network architecture of the 5G network access layer is a centralized wireless access network so as to greatly reduce the cost and energy consumption of communication equipment, and the 5G network access layer is used for displaying service users of the 5G access network, wherein the service users comprise personal users, cities, network operators and enterprises;
the 5G network slice layer is a combination of a technical premise of providing differentiated service requirements for different users by 5G and arrangement of 5G network functions, and is used for displaying topology formed by block chain network slices and block transmission among the network slices;
the mobile blockchain network application layer is used for showing the whole process from transaction creation to blockbroadcasting of the blockchain nodes, and the expansibility and the blocksynchronization speed of the mobile blockchain network application layer are greatly improved based on the low-delay and high-expansibility characteristics of the 5G network.
The block chain network slice comprises block chain links and block chain nodes, the block chain nodes realize specific network functions through network function virtualization running in a data center, the block chain link points comprise key characteristics of the block chain, the key characteristics of the block chain comprise anonymity, safety, random generation addresses, block chain data storage and a workload proving mechanism, links among the block chain nodes are established through an SDN controller based on global information of a network, an administrator is allowed to remotely configure a physical network, resources are reserved for the network slice with network resource requirements, and the network slices are communicated with each other and perform block transmission through a transmission link.
In the step 2, when a transmission link fault condition occurs, part of the affected retracted data is mixed in the normal data flow, the retracted data is collected to obtain a retracted data packet, the retracted data packet is marked and used as a labeled data set, and the rest of the data is used as an unlabeled data set.
In the step 3, the link state deducing algorithm establishes a federal semi-supervised learning model by using a federal semi-supervised learning algorithm, learns the characteristics of the retracted data packets, predicts the range of the fault transmission link, and therefore reroutes the data packets affected between the blockchain nodes, so that the connection between the blockchain nodes is quickly recovered, and the consensus convergence speed of the mobile blockchain network is improved.
The federal semi-supervised learning algorithm is used for applying a semi-supervised learning method to federal learning, the federal learning comprises a global model G, a local model set L and a server, each local model corresponds to one client, the clients correspond to network slices, the federal semi-supervised learning algorithm divides model training of a labeled dataset and a non-labeled dataset into two processes, a supervised learning mode is adopted for training of the labeled dataset, optimization of the model is guided according to a cross entropy loss function, and a consistency regularization method is adopted for training of the non-labeled dataset.
According to the consistency regularization method, input data are disturbed by adding counterexamples, so that the model has robustness to disturbance of label-free data, the output result of the original data and the data processed through random transformation is consistent, and the consistency regularization expression is as follows:
Figure GDA0004242983790000031
wherein p is θ (y|u) is a class probability function, representing a neural network model with a weight parameter θ, input as unlabeled data u, and output as y 2 Representing the L2 norm.
The flow of the federal semi-supervised learning algorithm is specifically as follows:
step 301: data collection is carried out through a 5G network slice layer, and an original data set is divided into a labeled data set and an unlabeled data set after the original data set is obtained:
D＝{d i },i＝1,2,…,N
S＝{x j ,y j },j＝1,2,…,L
U＝{u k },k＝1,2,…M
M+L＝N
wherein D is a known original dataset, S is a labeled dataset, U is an unlabeled dataset, D i Is the ith data of the original data set, x j For the j-th tagged data, y j For tagged data x j Corresponding tag, u k Is the j-th unlabeled data;
step 302: the method based on consistency regularization adopts a consistency loss function between clients to process the label-free data, and is defined as follows:
Figure GDA0004242983790000041
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure GDA0004242983790000042
global parameters selected for the server based on similarity of untrained local models in the network slices, representing consensus models with output as labels, KL representing the parameters being fixed, KL representing the relative entropy, quantized probability distribution differences, u representing all unlabeled data, y the predicted output>
Figure GDA0004242983790000043
For the output of the local model to the input u, representing the prediction result of the local model, and the consistency loss function between clients represents the difference between the prediction result of the local model and the labels provided by the consensus models;
step 303: the server selects H global parameters in each communication and broadcasts, and trains the unlabeled data by adopting a final consistency regularization loss function guiding model, wherein the expression is as follows:
Figure GDA0004242983790000044
Figure GDA0004242983790000045
where Φ (·) is the final consistency regularized loss function, Θ (·) represents generating a Shan Re tag with a given maximum normalized exponential function value, max (·) represents outputting a tag in a single thermal form on the class with the largest protocol to satisfy the computation of the cross entropy loss function,
Figure GDA0004242983790000046
representing the resulting class probability, pi (·) is a random transformation function, < >>
Figure GDA0004242983790000047
Is a pseudo tag based on protocol, i.eTags of the true class in the form of single heat, when generating pseudo tags, discarding low confidence prediction data below a confidence threshold τ, and then using the pseudo tags +.>
Figure GDA0004242983790000048
Performing standard cross entropy minimization, wherein cross entropy is a cross entropy loss function and represents cross entropy calculation on the class probability and the label in the single thermal form of the real class;
step 304: model neural network p θ The weights θ of (y|x) are divided into supervised learning parameters σ and unsupervised learning parameters ψ, which are used for supervised learning and unsupervised learning, respectively, to reduce the interaction between supervised learning and unsupervised learning, and during supervised learning with a labeled dataset, the unsupervised learning parameters ψ are fixed, i.e. no back propagation is performed, to minimize the loss term, and the corresponding training objectives are as follows:
Figure GDA0004242983790000049
wherein x and y are elements in the labeled data set S, x is all labeled data, y is labels respectively corresponding to all labeled data, and λs is a super parameter for controlling learning proportion among items;
step 305: unsupervised learning and training are performed on the unlabeled dataset by using a consistency regularization method, and a supervised learning parameter sigma is kept unchanged in the learning process so as to minimize loss, wherein corresponding training targets are as follows:
Figure GDA0004242983790000051
wherein lambda is I Is the parameter of the ultrasonic wave to be used as the ultrasonic wave,
Figure GDA0004242983790000052
for the L2 regular term, the L2 paradigm is represented to preserve knowledge learned from the supervised learning parameters sigma such that the supervised learning parameters sigma and the unsupervised learning parameters sigmaThe smaller the difference of the learning parameters psi is, +.>
Figure GDA0004242983790000053
For the L1 regularization item, the unsupervised parameter psi set contains a plurality of 0 items so as to improve the communication efficiency of federal learning and sparse under the condition of not affecting the previous learning effect;
step 306: and for each round of model training, training and updating the supervised learning parameters and the unsupervised learning parameters in the local model, after the updating is completed, transmitting the parameters of the local model to a server by each client, aggregating the local model by using a model aggregation method by the server, selecting a designated number of global variables, transmitting the global variables to each client again by the server if the performance of the aggregated global model does not reach the expectations, and repeating training and aggregating the model until the performance of the global model reaches the expectations.
In the step 304, the relationship between the supervised learning parameter σ and the unsupervised learning parameter ψ is as follows:
θ＝σ+ψ。
compared with the prior art, the invention has the following advantages:
the invention provides a core architecture of a mobile blockchain network based on a 5G slicing environment, focuses on minimizing transmission delay between blockchain nodes when a transmission link between 5G slices fails, and then introduces a link state inference algorithm based on machine learning in the architecture, wherein the algorithm learns the characteristics of a data packet to be retracted by using a federal semi-supervised learning method to infer a failed transmission link, so that the data packet affected between blockchain nodes is rerouted, thereby enabling the connection between the blockchain nodes to be quickly restored, and further enabling the mobile blockchain network consensus to be quickly converged; in the method based on federal semi-supervised learning, only the characteristics of part of the withdrawn data packets are learned to predict the range of a fault transmission link, but not a specific fault link, and the reasoning speed is obviously improved by sacrificing a certain accuracy; the FSSL model obtained by the reasoning algorithm can be deployed on any 5G network slice to rapidly predict the fault link range, and has good compatibility; under the support of a scheme of predictive fast consensus convergence, the blockchain nodes positioned on different network slices in the Internet of things can more effectively transmit local perception data.
Drawings
Fig. 1 is a schematic diagram of a mobile blockchain network framework based on 5G slice internet of things.
Fig. 2 is a schematic flow chart of FSSL.
Fig. 3 is a graph of loss function convergence rate versus time.
Fig. 4 is a convergence time chart of different schemes.
Fig. 5 is a graph of convergence time for different network scales.
Detailed Description
The invention will now be described in detail with reference to the drawings and specific examples. It will be apparent that the described embodiments are some, but not all, embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present invention without making any inventive effort, shall fall within the scope of the present invention.
Examples
The invention provides a block chain network slice security intelligent optimization method for a 5G intelligent network system, which is divided into a core architecture of a mobile block chain network based on a 5G network and a link state inference algorithm design based on a federal semi-supervised learning algorithm (FSSL).
A first part:
as shown in fig. 1, the core architecture of the mobile blockchain network includes a 5G network access layer, a 5G network slice layer, and a mobile blockchain network application layer.
5G network access layer:
the mobile communication network evolves from 4G to 5G, the network architecture of the access network is greatly changed, the network architecture of the access network is changed from the original distributed radio access network (D-RAN) to a centralized radio access network (C-RAN), the cost and the energy consumption of the communication equipment are greatly reduced by a centralized method, and the service users of the 5G access network, including personal users, cities, network operators and enterprises are shown in the figure;
5G network slice layer:
network slicing technology is a key technology of 5G, which is a combination of 5G providing differentiated service requirements to different users and the orchestration of network functions of network slicing 5G, network slicing can be regarded as a virtual network, including virtual blockchain links and virtual blockchain nodes, which are a subset of a physical network, the virtual blockchain nodes can implement specific network functions (such as routers and firewalls), in existing networks, these functions are implemented by devices in the physical network, in 5G networks, by network function Virtualization (VNF) running in a data center, virtual blockchain links between virtual blockchain nodes are implemented by multiple physical links, and the physical links on each physical path need to meet the bandwidth requirements of the virtual blockchain links, and the SDN controller establishes the virtual blockchain links based on global information on the network, and allows an administrator to remotely configure the physical network in order to reserve slice resources for which network resources are required.
Mobile blockchain network application layer:
the mobile blockchain network is established on the 5G network, and the expansibility of the mobile blockchain network and the speed of block synchronization (consensus convergence) are greatly improved due to the low-delay and high-expansibility characteristics of the 5G network, and in fig. 1, the process from transaction creation to block broadcasting into the mobile blockchain network is shown, wherein the data of the transaction are not only transfer and payment information in the prior sense, but also any data information needing interactive transmission, including the data transmission between sensor nodes in the Internet of things and the interaction of industrial control information in the industrial Internet of things, and the essence of the transaction is information interaction.
As shown in fig. 2, the present invention adopts an inference algorithm based on federal semi-supervised learning (FSSL), in which a data set is generally given, and then the data set is further divided into a tagged data set and an untagged data set, which are the same as the data set used for semi-supervised learning, under the framework of federal learning, including a global model G and a local model set L, the untagged data set is propagated to K clients, for the tagged data set S, supervised learning can be directly used for training, and each network slice is used as a client, when a transmission link is normal, the data traffic transmitted between the network slices is also normal, but when the transmission link is faulty, a part of the data traffic affected by the normal data traffic is mixed, and the data is collected at an interface of the network slice, so as to obtain a retraction data packet, in order to meet the requirements of the collected data set, the retraction data packet is manually marked and used as the tagged data set S, and the rest is used as the retraction data set S, and the retraction data packet is classified as the normal data traffic, and the normal data traffic is the original data traffic.
After classifying the collected original data set, the labeled data set with labels does not need to be subjected to excessive processing, and more attention needs to be paid to the unlabeled data set without labels, so that in semi-supervised learning, the most common processing method for the unlabeled data set is a consistency regularization method, and more effective classification results are generated by using less labeled data.
The method is characterized in that a consistency regularization method is adopted to process the unlabeled data set, and the input data is disturbed by adding counterexamples, so that the model has robustness to disturbance of the unlabeled data, and the existing consistency regularized expression is as follows:
Figure GDA0004242983790000071
wherein p is θ (y|u) is a weight parameter theta, the input is a label-free data instance u, the output is a class probability function of y, the class probability function represents a neural network model, pi (·) is a random enhancement data function, and II is 2 Representing the L2 norm, the consistency regularization means that for both the original data and the data processed by the data enhancement means, the output results remain consistent.
Description of algorithm:
step 301: data collection is carried out through a 5G network slice layer, and an original data set is divided into a labeled data set and an unlabeled data set after the original data set is obtained:
D＝{d i },i＝1,2,…,N
S＝{x j ,y j },j＝1,2,…,L
U＝{u k },k＝1,2,…M
M+L＝N
wherein D is a known original dataset, S is a labeled dataset, U is an unlabeled dataset, D i Is the ith data of the original data set, x j For the j-th tagged data, y j For tagged data x j Corresponding tag, u k Is the j-th unlabeled data;
step 302: the method based on consistency regularization adopts a consistency loss function between clients to process the label-free data, and is defined as follows:
Figure GDA0004242983790000081
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure GDA0004242983790000082
global parameters selected for the server based on similarity of untrained local models in the network slices, representing consensus models with output as labels, KL representing the parameters being fixed, KL representing the relative entropy, quantized probability distribution differences, u representing all unlabeled data, y the predicted output>
Figure GDA0004242983790000083
For the output of the local model to the input u, representing the prediction result of the local model, and the consistency loss function between clients represents the difference between the prediction result of the local model and the labels provided by the consensus models;
step 303: the server selects H global parameters in each communication and broadcasts, and trains the unlabeled data by adopting a final consistency regularization loss function guiding model, wherein the expression is as follows:
Figure GDA0004242983790000084
Figure GDA0004242983790000085
where Φ (·) is the final consistency regularized loss function, Θ (·) represents generating a Shan Re tag with a given maximum normalized exponential function value, max (·) represents outputting a tag in a single thermal form on the class with the largest protocol to satisfy the computation of the cross entropy loss function,
Figure GDA0004242983790000086
representing the resulting class probability, pi (·) is a random transformation function, < >>
Figure GDA0004242983790000091
For protocol-based pseudo tags, i.e. tags of the true class in the form of single heat, the low confidence prediction data below the confidence threshold τ are discarded when the pseudo tag is generated, and the pseudo tag is reused->
Figure GDA0004242983790000092
Performing standard cross entropy minimization, wherein cross entropy is a cross entropy loss function and represents cross entropy calculation on the class probability and the label in the single thermal form of the real class;
step 304: model neural network p θ The weights θ of (y|x) are divided into supervised learning parameters σ and unsupervised learning parameters ψ, which are used for supervised learning and unsupervised learning, respectively, to reduce the interaction between supervised learning and unsupervised learning, and during supervised learning with a labeled dataset, the unsupervised learning parameters ψ are fixed, i.e. no back propagation is performed, to minimize the loss term, and the corresponding training objectives are as follows:
Figure GDA0004242983790000093
wherein x and y are elements in the labeled data set S, x is all labeled data, y is labels respectively corresponding to all labeled data, and λs is a super parameter for controlling learning proportion among items;
step 305: unsupervised learning and training are performed on the unlabeled dataset by using a consistency regularization method, and a supervised learning parameter sigma is kept unchanged in the learning process so as to minimize loss, wherein corresponding training targets are as follows:
Figure GDA0004242983790000094
wherein lambda is I Is the parameter of the ultrasonic wave to be used as the ultrasonic wave,
Figure GDA0004242983790000095
for the L2 regular term, the L2 paradigm is represented to preserve knowledge learned from the supervised learning parameters sigma such that the smaller the difference between the supervised learning parameters sigma and the unsupervised learning parameters ψ, the ++>
Figure GDA0004242983790000096
For the L1 regularization item, the unsupervised parameter psi set contains a plurality of 0 items so as to improve the communication efficiency of federal learning and sparse under the condition of not affecting the previous learning effect;
step 306: and for each round of model training, training and updating the supervised learning parameters and the unsupervised learning parameters in the local model, after the updating is completed, transmitting the parameters of the local model to a server by each client, aggregating the local model by using a model aggregation method by the server, selecting a designated number of global variables, transmitting the global variables to each client again by the server if the performance of the aggregated global model does not reach the expectations, and repeating training and aggregating the model until the performance of the global model reaches the expectations.
The L2 regularization directly adds the square sum of weight parameters on the basis of the original loss function, so as to solve the overfitting phenomenon in model training, and the weight is continuously attenuated by using a multiplicative factor to adjust the weight, and the weight is attenuated quickly when the weight is heavy, and the weight is attenuated slowly when the weight is heavy; the L1 regularization directly adds the absolute value of the weight parameter on the basis of the original loss function, and the attenuation is fast when the weight is small, and the reduction is slow when the weight is large, so that the weight of the final model is concentrated on the feature with high importance, and the weight quickly approaches 0 when the weight is unimportant, and the final weight becomes sparse.
In order to further improve the rapid prediction capability of the algorithm, in the method based on federal semi-supervised learning, only the range of the characteristic prediction fault link of the partially retracted data packet is learned, but not the specific fault link, the inference speed is obviously improved by sacrificing a certain accuracy, the efficiency and the practicability of the algorithm are improved, the inference algorithm can be deployed on any 5G network slice to rapidly predict the fault link range, and has good compatibility, because the algorithm performs inference prediction based on the characteristic of abnormal data traffic, the model obtained by the inference algorithm is not modified on the basis of the routing mechanism among the original network slices, the inference algorithm can be deployed on any 5G network slice without modification, and the inference algorithm can also be applied to the scene of the Internet of things based on the mobile blockchain network, so that the blockchain nodes positioned on different network slices in the Internet of things can more effectively transmit local perception data.
For example: it can be seen in fig. 1 that when slice 1 and slice 8 are communicating with each other for block transmission, the original transmission links are { slice 1, slice 2, slice 3, slice 4, slice 6, slice 8}, but for some reasons the transmission links { slice 4, slice 6} fail, slice 3 deduces the range of the failed transmission link based on the model deployed thereon, and the transmission links are re-selected, and data traffic is forwarded through the transmission links { slice 3, slice 5, slice 6, slice 8}, so that the communication connection from slice 1 to slice 8 is restored quickly.
An experimental environment is built for experimental verification, and the experiment is divided into two parts: the first step is to establish a P2P block chain simulation network to realize consensus simulation; and secondly, running a simulation blockchain network, representing a network slice by using a blockchain client side, establishing a data set by collecting transmission data of each blockchain node in an experiment, and then training a federal semi-supervised learning (FSSL) model to infer the state of a transmission link by using the data set, so as to finally obtain a converged global model, and rapidly predicting the range of a fault transmission link in the blockchain network by using the global model.
Establishing a blockchain simulation network:
the method comprises the steps of establishing a blockchain simulation network through python programming, wherein the topology of the blockchain network is generated according to the functions of the blockchain network, firstly establishing a blockchain node, the node comprises some key characteristics of blockchains, such as anonymity, security, randomly generated addresses, blockchain data storage, a workload proving mechanism and the like, establishing a plurality of nodes to form the blockchain network, simulating the operation condition of an actual network when the network topology is formed and data transmission is carried out among the nodes, and selecting kademlia discovery protocol as the node discovery protocol of a P2P network to meet experimental requirements, wherein the core idea is that the nodes nearby are discovered through calculating the logic distance among the nodes, so that the node search is converged, and the protocol is simplified and only three requests are realized for simplifying the experiment and clearly explaining the problem:
1. node handshake: the nodes interact with other nodes through handshake operation, request the states of the other nodes, compare the states with the states of the nodes, and update the states when the states expire;
2. generating blockchain data: the generated blockchain data comprises link discovery data, transaction data, link interruption information and other data generated when the simulation network operates, and communication among different network slices is marked in the subsequent model training;
3. transaction broadcast: the node handshake is sent as a heartbeat and block synchronization is performed based on the heartbeat information, and all transactions will be broadcast to each node in the blockchain network.
Training an FSSL model:
after the blockchain simulation network is operated, transaction data are transmitted among different blockchain nodes, then a data set is established by collecting flow data of a blockchain client, a transmission link is manually disconnected to simulate transmission link faults in an actual environment, so that the data set under the normal condition of the transmission link and the data set under the condition of the transmission link faults are obtained, the obtained original data set is an unlabeled data set, and in order to obtain a better training effect, labeled data in an FSSL model are shared among the blockchain Internet of things clients.
In the experiment, part of data of an original data set is manually marked, so that the original data set is divided into a labeled data set and an unlabeled data set, and then training is performed to achieve the expected effect.
As shown in fig. 3, the training speed and convergence of the FSSL training loss function is significantly improved when the labeled datasets are shared with each other.
After training is completed, the trained FSSL model is deployed on a blockchain node of a blockchain simulation network, the reasoning effect of the FSSL model is tested through a simulation experiment, a transmission link fault is simulated through manually disconnecting a certain transmission link in the network, and then the time from predicting the fault transmission link range, recalculating the path and finally route convergence of the global model is recorded.
As shown in fig. 4, the node adopts a scheme of predictive fast consensus convergence (PFCC scheme) and deploys an FSSL model, and performs 20 times of simulation test on a convergence time value, so as to emphasize experimental effect, two simulation experiments are performed on the same network topology structure, in the two simulation experiments, two routing protocols, RIP and OSPF, are respectively used, and as can be seen from the figure, compared with the existing routing mechanism, the routing convergence time of the scheme of predictive fast consensus convergence (PFCC scheme) is shorter (less than 2 seconds), and when the network faces transmission link interruption, the network can be quickly restored to be connected, so that the consensus convergence speed and robustness of the mobile blockchain network are improved.
As shown in fig. 5, to further test the effectiveness of the model, tests were performed at different network scales, from which two trends can be seen:
as the network scale is enlarged, the convergence time becomes longer, and meets the expectations, as the network scale is enlarged, the connection between nodes becomes more complex, and after the transmission link fault occurs in the network topology, the global model obtained by the reasoning algorithm can consume more calculation resources and time for reasoning the more complex topology;
with the expansion of the network scale, the increase of the convergence time is small, and although the network scale is expanded, because the model obtained by the invention predicts the range of the fault transmission link according to part of transmission data, the expansion of the network scale has little influence on the prediction speed of an inference algorithm, so the method provided by the invention is still effective for larger networks, and a scheme for representing the predictive rapid consensus convergence can be deployed on more network slices so as to support diversified services.
In addition, the FSSL model is used for predicting the transmission link state, so that the common knowledge convergence under the transmission link fault scene is improved, the block generation speed using the FSSL is compared with the block generation speed not using the FSSL in order to verify the effect, and the difference of the block generation speeds can reach the second level according to experimental records.
In summary, the invention provides a scheme for mobile blockchain predictive fast consensus convergence based on link state inference in a 5G slicing environment, in a built simulation experiment environment, a data set under the normal operation condition and a data set under the transmission link fault condition are obtained by operating a blockchain simulation network, then, part of the data set of an original data set is manually marked, the original data set is divided into a labeled data set and a non-labeled data set, finally, a federal semi-supervised learning reasoning algorithm is adopted, a flow data set is trained in a network slice, an optimized global model is obtained, the range of a fault transmission link is inferred and predicted, and an experimental result obtained through experiments achieves the expected effect.
While the invention has been described with reference to certain preferred embodiments, it will be understood by those skilled in the art that various changes and substitutions may be made without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the protection scope of the invention is subject to the protection scope of the claims.

Claims (6)

1. A block chain network slice security intelligent optimization method for a 5G intelligent network system is characterized by comprising the following steps:
step 1: establishing a mobile blockchain network based on a 5G slicing environment;
step 2: obtaining an original data set of the mobile block chain network operation, wherein the original data set comprises data under the normal operation condition and data under the transmission link fault condition, and carrying out data preprocessing;
step 3: establishing a federal semi-supervised learning model according to a link state inference algorithm based on machine learning and training;
step 4: obtaining an optimized global model after training the federal semi-supervised learning model so as to rapidly predict the range of a fault transmission link in the blockchain network and realize predictive rapid consensus convergence;
the federal semi-supervised learning algorithm is used for applying a semi-supervised learning method to federal learning, the federal learning comprises a global model G, a local model set L and a server, each local model corresponds to one client, the clients correspond to network slices, the federal semi-supervised learning algorithm divides model training of a labeled dataset and an unlabeled dataset into two processes, a supervised learning mode is adopted for training of the labeled dataset, optimization of the model is guided according to a cross entropy loss function, and a consistency regularization method is adopted for training of the unlabeled dataset;
according to the consistency regularization method, input data are disturbed by adding counterexamples, so that the model has robustness to disturbance of label-free data, the output result of the original data and the data processed through random transformation is consistent, and the consistency regularization expression is as follows:
Figure FDA0004052259310000011
wherein p is θ (y|u) is a class probability function, representing a neural network model with a weight parameter θ, input as unlabeled data u, and output as y 2 Represents an L2 norm;
the flow of the federal semi-supervised learning algorithm is specifically as follows:
step 301: data collection is carried out through a 5G network slice layer, and an original data set is divided into a labeled data set and an unlabeled data set after the original data set is obtained:
D＝{d i },i＝1,2,…,N
S＝{x j ,y j },j＝1,2,…,L
U＝{u k },k＝1,2,…M
M+L＝N
wherein D is a known original dataset, S is a labeled dataset, U is an unlabeled dataset, D i Is the ith data of the original data set, x j For the j-th tagged data, y j For tagged data x j Corresponding tag, u k Is the kth unlabeled data;
step 302: the method based on consistency regularization adopts a consistency loss function between clients to process the label-free data, and is defined as follows:
Figure FDA0004052259310000021
wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure FDA0004052259310000022
global parameters selected for the server based on similarity of untrained local models in the network slices, representing consensus models with output as labels, KL representing the parameters being fixed, KL representing the relative entropy, quantized probability distribution differences, u representing all unlabeled data, y the predicted output>
Figure FDA0004052259310000023
For the output of the local model to the input u, representing the prediction result of the local model, and the consistency loss function between clients represents the difference between the prediction result of the local model and the labels provided by the consensus models;
step 303: the server selects H global parameters in each communication and broadcasts, and trains the unlabeled data by adopting a final consistency regularization loss function guiding model, wherein the expression is as follows:
Figure FDA0004052259310000024
Figure FDA0004052259310000025
where Φ (·) is the final consistency regularized loss function, Θ (·) represents generating a Shan Re tag with a given maximum normalized exponential function value, max (·) represents outputting a tag in a single thermal form on the class with the largest protocol to satisfy the computation of the cross entropy loss function,
Figure FDA0004052259310000028
representing the resulting class probability, pi (·) is a random transformation function, < >>
Figure FDA0004052259310000026
For protocol-based pseudo tags, i.e. tags of the true class in the form of single heat, the low confidence prediction data below the confidence threshold τ are discarded when the pseudo tag is generated, and the pseudo tag is reused->
Figure FDA0004052259310000027
Performing standard cross entropy minimization, cross entropy being a cross entropy loss function, representing a scale of single thermal form for class probability and true classPerforming cross entropy calculation by the sign;
step 304: model neural network p θ The weights θ of (y|x) are divided into supervised learning parameters σ and unsupervised learning parameters ψ, which are used for supervised learning and unsupervised learning, respectively, to reduce the interaction between supervised learning and unsupervised learning, and during supervised learning with a labeled dataset, the unsupervised learning parameters ψ are fixed, i.e. no back propagation is performed, to minimize the loss term, and the corresponding training objectives are as follows:
Figure FDA0004052259310000029
wherein x and y are elements in the labeled data set S, x is all labeled data, y is labels respectively corresponding to all labeled data, and λs is a super parameter for controlling learning proportion among items;
step 305: unsupervised learning and training are performed on the unlabeled dataset by using a consistency regularization method, and a supervised learning parameter sigma is kept unchanged in the learning process so as to minimize loss, wherein corresponding training targets are as follows:
Figure FDA0004052259310000031
wherein lambda is I Is the parameter of the ultrasonic wave to be used as the ultrasonic wave,
Figure FDA0004052259310000032
for the L2 regular term, the L2 paradigm is represented to preserve knowledge learned from the supervised learning parameters sigma such that the smaller the difference between the supervised learning parameters sigma and the unsupervised learning parameters ψ, the ++>
Figure FDA0004052259310000033
For the L1 regularization item, the unsupervised parameter psi set contains a plurality of 0 items so as to improve the communication efficiency of federal learning and sparse under the condition of not affecting the previous learning effect;
step 306: for each round of model training, training and updating the supervised learning parameters and the unsupervised learning parameters in the local model, after the updating is completed, transmitting the parameters of the local model to a server by each client, aggregating the local model by using a model aggregation method by the server, selecting a designated number of global variables, and if the performance of the aggregated global model does not reach the expectations, transmitting the global variables to each client again by the server, and repeatedly training and aggregating the model until the performance of the global model reaches the expectations;
in the step 304, the relationship between the supervised learning parameter σ and the unsupervised learning parameter ψ is as follows:
θ＝σ+ψ。
2. the intelligent optimization method for the blockchain network slice security of the 5G intelligent network system according to claim 1, wherein in the step 1, the core architecture of the mobile blockchain network includes a 5G network access layer, a 5G network slice layer and a mobile blockchain network application layer which are arranged from bottom to top.
3. The blockchain network slicing security intelligent optimization method for the 5G intelligent network system according to claim 2, wherein the network architecture of the 5G network access layer is a centralized wireless access network so as to greatly reduce the cost and energy consumption of communication equipment, and the 5G network access layer is used for displaying service users of the 5G access network, wherein the service users comprise personal users, cities, network operators and enterprises;
the 5G network slice layer is a combination of a technical premise of providing differentiated service requirements for different users by 5G and arrangement of 5G network functions, and is used for displaying topology formed by block chain network slices and block transmission among the network slices;
the mobile blockchain network application layer is used for showing the whole process from transaction creation to blockbroadcasting of the blockchain nodes, and the expansibility and the blocksynchronization speed of the mobile blockchain network application layer are greatly improved based on the low-delay and high-expansibility characteristics of the 5G network.
4. A blockchain network slice security intelligent optimization method for a 5G intelligent network system according to claim 3, wherein the blockchain network slice includes blockchain links and blockchain nodes, the blockchain nodes implement specific network functions through network function virtualization running in a data center, the blockchain nodes include key characteristics of the blockchain, the key characteristics of the blockchain include anonymity, security, randomly generated addresses, blockchain data storage and workload certification mechanisms, links between the blockchain nodes are established through an SDN controller based on global information of a network, and an administrator is allowed to remotely configure a physical network to reserve resources for the network slice with network resource requirements, and the network slices communicate with each other for block transmission through a transmission link.
5. The intelligent optimization method for blocking chain network slicing security of a 5G intelligent network system according to claim 4, wherein in step 2, when a transmission link failure occurs, part of the data which is affected is mixed in the normal data traffic, the data which is retracted is collected to obtain a retracted data packet, the retracted data packet is marked and used as a labeled data set, and the rest of the data is used as an unlabeled data set.
6. The method for intelligent optimization of blockchain network slicing security for a 5G intelligent network system according to claim 1, wherein in step 3, the link state inference algorithm establishes a federal semi-supervised learning model by using a federal semi-supervised learning algorithm, learns the characteristics of the retracted data packets, predicts the range of the failed transmission link, and thereby reroutes the affected data packets between blockchain nodes, so that the connection between blockchain nodes is quickly restored, and the consensus convergence speed of the mobile blockchain network is improved.
CN202111203106.8A 2021-10-15 2021-10-15 Block chain network slicing security intelligent optimization method for 5G intelligent networking system Active CN114070775B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111203106.8A CN114070775B (en) 2021-10-15 2021-10-15 Block chain network slicing security intelligent optimization method for 5G intelligent networking system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111203106.8A CN114070775B (en) 2021-10-15 2021-10-15 Block chain network slicing security intelligent optimization method for 5G intelligent networking system

Publications (2)

Publication Number Publication Date
CN114070775A CN114070775A (en) 2022-02-18
CN114070775B true CN114070775B (en) 2023-07-07

Family

ID=80234682

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111203106.8A Active CN114070775B (en) 2021-10-15 2021-10-15 Block chain network slicing security intelligent optimization method for 5G intelligent networking system

Country Status (1)

Country Link
CN (1) CN114070775B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114726743A (en) * 2022-03-04 2022-07-08 重庆邮电大学 Service function chain deployment method based on federal reinforcement learning
CN116701939B (en) * 2023-06-09 2023-12-15 浙江大学 Classifier training method and device based on machine learning

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109544334A (en) * 2018-10-22 2019-03-29 绿州蔚来(深圳)控股有限公司 A kind of network scalability block chain implementation method
CN112348204A (en) * 2020-11-05 2021-02-09 大连理工大学 Safe sharing method for marine Internet of things data under edge computing framework based on federal learning and block chain technology
CN112615730A (en) * 2020-11-23 2021-04-06 北京邮电大学 Resource allocation method and device based on block chain network slice proxy
CN112887145A (en) * 2021-01-27 2021-06-01 重庆邮电大学 Distributed network slice fault detection method
CN113379066A (en) * 2021-06-10 2021-09-10 重庆邮电大学 Federal learning method based on fog calculation

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111213134A (en) * 2018-09-19 2020-05-29 因特比有限公司 Method and system for performing hyper-convergence using blockchains
CN112001502B (en) * 2020-08-24 2022-06-21 平安科技（深圳）有限公司 Federal learning training method and device for high-delay network environment robustness
CN112906859B (en) * 2021-01-27 2022-07-01 重庆邮电大学 Federal learning method for bearing fault diagnosis
CN112861152B (en) * 2021-02-08 2024-02-20 北京航空航天大学 Federal learning incentive method and system based on license chain
CN113159333B (en) * 2021-03-27 2022-10-11 北京邮电大学 Federal learning method, system and device based on hierarchical fragment block chain
CN113194126B (en) * 2021-04-21 2023-10-27 泉州华中科技大学智能制造研究院 Transverse federal learning model construction method based on blockchain

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109544334A (en) * 2018-10-22 2019-03-29 绿州蔚来(深圳)控股有限公司 A kind of network scalability block chain implementation method
CN112348204A (en) * 2020-11-05 2021-02-09 大连理工大学 Safe sharing method for marine Internet of things data under edge computing framework based on federal learning and block chain technology
CN112615730A (en) * 2020-11-23 2021-04-06 北京邮电大学 Resource allocation method and device based on block chain network slice proxy
CN112887145A (en) * 2021-01-27 2021-06-01 重庆邮电大学 Distributed network slice fault detection method
CN113379066A (en) * 2021-06-10 2021-09-10 重庆邮电大学 Federal learning method based on fog calculation

Also Published As

Publication number Publication date
CN114070775A (en) 2022-02-18

Similar Documents

Publication Publication Date Title
Jiang Graph-based deep learning for communication networks: A survey
CN114070775B (en) Block chain network slicing security intelligent optimization method for 5G intelligent networking system
Mohammed et al. Budgeted online selection of candidate IoT clients to participate in federated learning
US20220030031A1 (en) 3s-chain: smart, secure, and software-defined networking (sdn)-powered blockchain-powered networking and monitoring system
CN106603293A (en) Network fault diagnosis method based on deep learning in virtual network environment
Alnawayseh et al. Smart congestion control in 5g/6g networks using hybrid deep learning techniques
Xu et al. Communication quality prediction for Internet of vehicle (IoV) networks: an elman approach
Aouedi et al. Intrusion detection for softwarized networks with semi-supervised federated learning
CN114710437B (en) Internet of things edge network routing system combining block chains
Naeem et al. Federated-learning-empowered semi-supervised active learning framework for intrusion detection in zsm
Yao et al. Developing networks using artificial intelligence
Chen et al. ALBLP: adaptive load-balancing architecture based on link-state prediction in software-defined networking
Li et al. Research and application of AI in 5G network operation and maintenance
Cui et al. Collaborative intrusion detection system for sdvn: A fairness federated deep learning approach
Geepthi et al. RETRACTED ARTICLE: Network traffic detection for peer-to-peer traffic matrices on bayesian network in WSN
Huang et al. Collective reinforcement learning based resource allocation for digital twin service in 6G networks
Guo et al. Capturing spatial–temporal correlations with Attention based Graph Convolutional Network for network traffic prediction
CN116319583A (en) Encryption network traffic classification method based on GCNN and MoE
Shi et al. PFCC: Predictive fast consensus convergence for mobile blockchain over 5G slicing-enabled IoT
Kurniawan et al. Mobile computing and communications-driven fog-assisted disaster evacuation techniques for context-aware guidance support: A survey
Guo et al. ML-SDNIDS: an attack detection mechanism for SDN based on machine learning
Hong et al. Retracted: Artificial intelligence point‐to‐point signal communication network optimization based on ubiquitous clouds
Fang Security evaluation method of distance education network nodes based on machine learning
Al-Yassari et al. Automatic Spike Neural Technique for Slicing Bandwidth Estimated Virtual Buffer-Size in Network Environment
Popoola Federated deep learning for botnet attack detection in IoT networks

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant