SVDD (singular value decomposition and direct data decomposition) -based network slice physical node anomaly detection method
Technical Field
The invention belongs to the technical field of mobile communication, and relates to a method for detecting physical node abnormity in a network slicing scene.
Background
With the rapid development of mobile internet technology, next-generation communication systems are being researched to meet more diversified business requirements. One of the major driving forces in the development of mobile communication systems is the need to support various vertical industries, which have distinct business requirements and more diverse business scenarios. The architecture method adopted by the traditional network cannot meet the diversified performance requirements of the vertical market in the aspects of delay, expandability, usability and reliability. How to serve multiple services with different performance requirements on the same underlying physical network arises from the network slicing technology.
The network slice divides a physical network into a plurality of virtual networks, each virtual network can be customized and optimized for specific types of application programs or users, and shared physical network resources are dynamically and efficiently allocated to each logic network slice according to changing user requirements. Technologies such as Software Defined Networking (SDN) and Network Function Virtualization (NFV) can provide operators with the characteristics of programmability, flexibility, and modularity required to create multiple virtual Network slices. The network slices are combined by Virtual Network Functions (VNFs) according to user scenarios or Service models, to form an ordered VNF set, i.e., a Service Function Chain (SFC).
Network slices add complexity to network management, especially where a large number of network slices are deployed, and thus are critical to automated management of network slices. The automatic management of the network slice is realized, and firstly, the abnormity in the network slice can be quickly detected.
Aiming at the prior art, the following defects are found: the existing network slice research mainly aims at the problems of deployment and migration of service function chains, and some researches aim at the problem of abnormal detection of VNFs. Although an anomaly in the underlying physical node in a network slice may cause performance degradation in all network slices deployed on it, few studies have been directed to anomaly detection for the underlying physical node; the conventional network slice anomaly detection research mostly adopts a mode of centralized processing of observation data in a network, and the research introduces extra communication overhead in the network, so that the communication occupies bandwidth resources in the network and brings more time cost; an off-line training mode frequently adopted for the research of machine learning in the network slice anomaly detection conflicts with the dynamic deployment characteristic of the VNF, and the detection model has obvious aging effect, namely the reduction of the detection performance under the condition that the dynamic update cannot be realized.
Disclosure of Invention
In view of this, an object of the present invention is to provide a method for detecting an anomaly of a physical node in distributed online network slices, which can effectively utilize VNF observation information to implement bottom-layer anomaly detection of the physical node, avoid sharing VNF observation data among slices, and dynamically update a detection model.
In order to achieve the purpose, the invention provides the following technical scheme:
a network slice physical node anomaly detection method based on SVDD (singular value decomposition) comprises the steps of generating an observation data model by adopting network slice deployment and a Virtual Network Function (VNF) under a network slice scene, combining unsupervised anomaly detection and VNF observation data network slice sharing, constructing a network slice physical anomaly detection model which is distributed and deployed on each network slice manager, realizing distributed processing of VNF observation data in each slice through a random approximate function, and finally realizing distributed online physical node anomaly detection by adopting a random gradient descent method.
Further, the network slice deployment and VNF generating a model of observation data specifically includes: the bottom layer physical network consists of M physical nodes which are communicated with each other, and N VNFs mapped on the physical nodes M in the network are total; there are three types of network slices deployed onto a physical network: an SFC corresponding to an Enhanced Mobile Broadband (EMBB) network slice consists of 6 VNFs, an SFC corresponding to an Ultra Reliable Low Latency Communication (URLLC) network slice consists of 5 VNFs, and an SFC corresponding to a Mass Machine Type Communication (MMTC) network slice consists of 4 VNFs; each network slice manager collects the observation data of VNF in the slice and at the time tThe observed data for physical node m is s (t) ═ x1(t),x2(t),…,xN(t))TWherein x isnAnd (t), N is 1,2, …, and N is an observation data vector of the nth VNF deployed on the physical node m at the time t, and is composed of a flow rate, a queuing delay, a processing delay and a protocol type.
Further, the constructing of the network slice physical node anomaly detection model which is distributed and deployed on each network slice manager specifically includes: carrying out distributed parameter updating on the model for generating observation data by network slice deployment and VNF (virtual network function), and obtaining a virtual network element a through multiple iterations*=a1(t)=a2(t)=…=aN(t) is the center of the sphere, R*=R1(t)=R2(t)=…=RN(t) is the radius, ξ*=ξ1(t)=ξ2(t)=…=ξn(t) hypersphere in random feature space for relaxation variables; wherein, an(t)、Rn(t)、ξn(t) respectively updating the parameters in the network slice corresponding to the nth VNF at the time t to obtain the sphere center, the radius and the relaxation variable of the hypersphere in the random feature space; a is*、R*And ξ*Is a distributed updated hypersphere center a in each network slice after multiple iterationsn(t) radius Rn(t) and a slack variable ξn(t) converging the obtained values, in which process the hypersphere defined by the updated parameters in each network slice gradually coincides into the same hypersphere.
Further, the stochastic feature space refers to the observation data x of the original VNF by a stochastic approximation function z (-) in the random spacen(t) high dimensional space to which the original observation data x is mappedn(t) is mapped as z (x)n(t))={z1(xn(t)),…,zD(xn(t))]TWherein D is the dimension of the random feature space;
the hypersphere adopts a support vector data description model to update the parameter a for each network slice manager to VNF observation datan(t),Rn(t),ξn(t), N is a hypersphere defined by 1,2, …, N, eachThe hypersphere parameters on the network slice manager may be different in the initial stage;
relaxation variable ξn(t) is x in order to allow feature spacen(t) mapping z (x)n(t)) and the spherical center of the hypersphere an(t) is greater than the radius R of the hyperspheren(t) is defined in the present invention as ξn(t)>0。
The distributed mode refers to that VNF observation data are distributed and collected and stored by the network slice managers to which the VNFs belong, and are distributed and processed in the network slice managers, and transmission of VNF original observation data among slices does not exist.
Further, the parameter updating specifically includes: parameter a within each network slicen(t)、Rn(t)、ξn(t) updates were all done using a random gradient descent, but with respect to the relaxation variable ξn(t) during updating, the value of the update is required to be ensured to be positive after each updating, so that the result is corrected after each updating; to ensure updated a within each slicen(t)、Rn(t) and ξn(t) final convergence to a common hypersphere parameter a*、R*And ξ*When the physical node is in normal operation state, the parameters a need to be shared among the network slicesn(t),Rn(t),ξn(t) the distributed parameter update in each network slice requires the result of the distributed parameter update at the previous time in all other slice managers.
Further, the random gradient descent method specifically includes: updating the radius R of the hypersphere by a random gradient descent method in each network slice managern(t) and the center of sphere an(t) and a slack variable ξn(t), after the updating is completed, calculating a distributed discriminant function:
wherein sgn is a sign function; if g (x)n(t)). 1, VNFn observation data x at time tn(t) locating within the hypersphere obtained from the updating of the parameters in the random feature space; if g (x)n(t))=-1,X is thenn(t) is located outside the hypersphere.
Further, the distributed online physical node anomaly detection specifically includes: distributed parameter updating is finished on the network slice manager corresponding to each VNF, the real-time decision module judges the running state of the physical node at the current moment according to the distributed discrimination function, the negative influence of abnormal data on model training is considered, and the parameters of observation data updating when the running state of the physical node is abnormal are discarded.
Further, the real-time decision module specifically includes: distributively updating the resulting parameter g (x) according to each network slice manager
n(t)) judgment
Whether or not it is equal to N, if
The running state of the physical node at the time t is normal, if so, the physical node is in a normal running state
The running state of the physical node is abnormal at the moment t; wherein, N is the number of VNFs deployed on a target physical node, and the target physical node is the physical node detected by the method;
if the physical node is in normal operation state, using each distributed parameter a at the current momentn(t),Rn(t),ξn(t) continuing to update the parameters of the random gradient descent method at the next moment; if the current physical node running state is abnormal, discarding the parameter a at the time tn(t),Rn(t),ξn(t) let an(t)=an(t-1),Rn(t)=Rn(t-1),ξn(t)=ξnAnd (t-1) continuing to update the parameters of the random gradient descent method at the next moment.
The invention has the beneficial effects that: according to the invention, the influence of the abnormal operation state of the physical node on the performance of the whole slice network is concerned according to the sharing characteristic of the underlying network of the network slice, and a physical node abnormal detection model based on support vector data description is provided; in order to solve the technical confidentiality and security protection problems faced by the sharing of original observation data slices of a Virtual Network Function (VNF), a distributed anomaly detection model is provided, VNF observation data are processed in a distributed mode in each slice through a random approximation Function, and finally a distributed online physical node anomaly detection method is provided by adopting random gradient descent. Therefore, the invention can realize the physical node abnormity detection by using the virtual network function observation data and simultaneously ensure that the observation data is processed in the slice, thereby solving the problem of information leakage worried by a VNF operator.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objectives and other advantages of the invention may be realized and attained by the means of the instrumentalities and combinations particularly pointed out hereinafter.
Drawings
For the purposes of promoting a better understanding of the objects, aspects and advantages of the invention, reference will now be made to the following detailed description taken in conjunction with the accompanying drawings in which:
FIG. 1 is a schematic diagram of a network slice deployment scenario in the present invention;
FIG. 2 is a schematic view of an observation data collection method according to the present invention;
FIG. 3 is a schematic block diagram of a distributed online anomaly detection method according to the present invention;
FIG. 4 is a flow chart of a distributed online anomaly detection method according to the present invention.
Detailed Description
The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention in a schematic way, and the features in the following embodiments and examples may be combined with each other without conflict.
Referring to fig. 1 to 4, a network slice physical node anomaly detection method based on SVDD is disclosed, which provides a model of network slice deployment and observation data generated by VNF for the problem of anomaly detection in a network slice scene; by combining the consideration of unsupervised anomaly detection and VNF observation data network slice sharing, a network slice physical anomaly detection model which is distributed and deployed on each network slice manager is provided; further, a distributed online physical node anomaly detection method based on random gradient descent is provided.
Fig. 1 is a schematic diagram of a network slice deployment scenario in this embodiment. Referring to fig. 1, after receiving a service request, a network slice management and orchestrator may flexibly deploy a network slice on an underlying physical network according to different service requirements. The infrastructure management module is responsible for managing the underlying physical resources needed for network slice deployment.
Fig. 2 is a schematic diagram of a collection manner of VNF observation data used in this embodiment, and referring to fig. 2, infrastructure-level observation data may be obtained from an infrastructure management module, and includes physical node CPU occupancy, memory occupancy, disk throughput, network throughput, and the like. However, the VNF operation state is closely related to the operation state of the physical node, and data information such as a flow rate, a queuing delay, a processing delay, a protocol type, and the like can be observed through the VNF.
Fig. 3 is a schematic block diagram of a distributed online anomaly detection method in this embodiment, and the method is mainly divided into three parts:
a first part: each network slice manager realizes real-time collection of VNF observation data
A second part: and the abnormal detection parameter updating module which is distributed on each network slice manager realizes local parameter updating in the slice.
And a third part: and a physical node management module in the infrastructure manager realizes real-time abnormal decision according to the parameters transmitted by each distributed parameter updating module, judges the running state of the physical node and selects the mode of updating the next parameter according to the running state of the physical node.
FIG. 4 is a schematic flow chart of a distributed online anomaly detection method according to the present invention, which includes the following steps:
step 401: initializing distributed model parameters at the moment t-0;
step 402: t is t + 1;
step 403: the method comprises the steps of collecting observation data of all VNFs on a target physical node at the moment t in a distributed mode;
step 404: a distributed parameter updating module in each network slice manager calculates a random approximate value of the VNF observation data according to a random approximate function;
step 405: a distributed parameter updating module in each network slice manager calculates the gradient of each parameter at the current moment;
step 406: each network slice manager calculates the current moment model parameters according to the random gradient descent;
step 407: each network slice manager calculates g (x)n(t));
Step 408: the physical node management module updates the obtained g (x) according to each network slice managern(t)) deciding the physical node operation state;
step 409: and if the running state of the physical node is normal, directly returning to the step 401, otherwise, discarding the parameter at the current moment, and then returning to the step 401.
Finally, the above embodiments are only intended to illustrate the technical solutions of the present invention and not to limit the present invention, and although the present invention has been described in detail with reference to the preferred embodiments, it will be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions, and all of them should be covered by the claims of the present invention.