CN114116198A - Asynchronous federal learning method, system, equipment and terminal for mobile vehicle - Google Patents
Asynchronous federal learning method, system, equipment and terminal for mobile vehicle Download PDFInfo
- Publication number
- CN114116198A CN114116198A CN202111229664.1A CN202111229664A CN114116198A CN 114116198 A CN114116198 A CN 114116198A CN 202111229664 A CN202111229664 A CN 202111229664A CN 114116198 A CN114116198 A CN 114116198A
- Authority
- CN
- China
- Prior art keywords
- domain
- aggregation
- edge
- cloud
- vehicle
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 77
- 238000004220 aggregation Methods 0.000 claims abstract description 167
- 230000002776 aggregation Effects 0.000 claims abstract description 158
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 101
- 238000012549 training Methods 0.000 claims abstract description 87
- 238000004891 communication Methods 0.000 claims abstract description 50
- 230000008569 process Effects 0.000 claims abstract description 34
- 238000009826 distribution Methods 0.000 claims abstract description 18
- 230000008901 benefit Effects 0.000 claims abstract description 17
- 238000002474 experimental method Methods 0.000 claims abstract description 12
- 238000010801 machine learning Methods 0.000 claims description 39
- 230000006870 function Effects 0.000 claims description 31
- 230000005540 biological transmission Effects 0.000 claims description 6
- 238000004590 computer program Methods 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000003860 storage Methods 0.000 claims description 5
- 238000013135 deep learning Methods 0.000 claims description 3
- 238000006116 polymerization reaction Methods 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 description 7
- 238000013461 design Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 230000006855 networking Effects 0.000 description 6
- 230000001360 synchronised effect Effects 0.000 description 6
- 230000007547 defect Effects 0.000 description 5
- 238000012935 Averaging Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 4
- 238000011161 development Methods 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000007726 management method Methods 0.000 description 3
- 238000009828 non-uniform distribution Methods 0.000 description 3
- SAZUGELZHZOXHB-UHFFFAOYSA-N acecarbromal Chemical compound CCC(Br)(CC)C(=O)NC(=O)NC(C)=O SAZUGELZHZOXHB-UHFFFAOYSA-N 0.000 description 2
- 230000004069 differentiation Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000002349 favourable effect Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 210000005036 nerve Anatomy 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5072—Grid computing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention belongs to the technical field of vehicle management and discloses an asynchronous federal learning method, an asynchronous federal learning system, equipment and a terminal of a mobile vehicle, wherein the asynchronous federal learning system of the mobile vehicle comprises a user layer, a domain-edge server layer and a data processing center layer; the asynchronous federal learning method for a mobile vehicle comprises the following steps: comprehensively utilizing cloud computing and edge computing to provide a network layered domain-division framework based on a cloud side car; providing an asynchronous federated learning aggregation algorithm aFedV suitable for the cloud-side-vehicle-based network hierarchical domain-division architecture; and aiming at the aggregation algorithm and the hierarchical architecture, experiments are carried out on different data distributions, and the performance of the aFedV algorithm is verified in the aspects of model training accuracy and communication overhead. The invention comprehensively utilizes the advantages of cloud computing and edge computing, adopts an asynchronous mode to update parameters, can reduce the communication times of the whole training process, and solves the problem that the dynamic connection of the mobile federal member can not update the parameters in time in the computing process.
Description
Technical Field
The invention belongs to the technical field of vehicle management, and particularly relates to an asynchronous federal learning method, an asynchronous federal learning system, asynchronous federal learning equipment and an asynchronous federal learning terminal for a mobile vehicle.
Background
At present, the development of the internet of things is promoted by the development of the internet, and the development gradually progresses towards the trend of interconnection of everything, wherein the rising of edge devices is accelerated by the characteristics of 5G large bandwidth, low time delay and the like, and a large amount of information generated by the devices provides original training data for machine learning, so that the edge devices are more intelligent. In an internet of vehicles environment, a vehicle is a typical network edge device, and application data generated by the vehicle is input into a neural network, so that better personalized services can be provided for users.
Under the learning mode of a traditional machine, user data needs to be uploaded to a cloud data processing center (hereinafter referred to as a cloud center) by each vehicle, the result is returned to a local user according to the requirement after the cloud center is trained, but the data transmission process can occupy a large amount of bandwidth, a large amount of flow pressure is caused to a backbone network, a lot of user privacy problems can be brought by transmitting the personal data, and the safety of the user data cannot be guaranteed. In 2017, Google provides a federal learning architecture, the original purpose of the new distributed machine learning is to avoid data leakage, protect privacy and safety of user data, and guarantee that the user data can participate in machine learning training under the condition that the user data is not local. In the federal learning architecture, federal members are defined as a data provider and local training clients, all federal members share one model and cooperate to complete the training of the model, then the local model is uploaded to a cloud center, the cloud center aggregates all the local models to obtain a global model, and finally the result is issued to all the local clients as the initial value of the next model training. Model parameters or gradients are transmitted in the whole interaction process of the cloud-client, and user data are always kept in the local area, so that the privacy of a user is guaranteed.
However, the cloud-client machine learning model has high latency cost and communication overhead, and cannot obtain good performance in many time-sensitive network scenarios. In recent years, edge computing has emerged, because edge nodes are deployed at positions close to the edge of the network, there are more opportunities to contact data generated by network bottom layer equipment, and an edge server generally has powerful computing power, and edge-client communication has the advantages of high bandwidth and low time delay, so that the edge-based machine learning architecture has attracted extensive attention in the field of federal learning.
However, in the machine learning architecture based on the edge or the cloud, when the global model is generated by aggregation, the high level mostly adopts the traditional synchronous aggregation algorithm, such as FedAvg and fedgsd, and the algorithms can effectively realize the model aggregation of the joint learning in the static scene. However, in most mobile scenarios, the federate has the situations that the connection is unstable and asynchronous, and the algorithms cannot solve the problem of how the federate dynamically joins the federate aggregation. In the current internet of vehicles network architecture, data of vehicles are gradually developed to Road Side Unit (RSU) processing at the edge of the network from original cloud center processing, and due to the fact that vehicle nodes have the characteristics of high mobility and flexibility, the vehicles asynchronously participate in federal aggregation to become a normal state. In addition, most of the traditional federated learning aggregation algorithms are based on the idea of a model averaging algorithm, the aggregation time of each round of the aggregation algorithm is usually determined by the client with the longest time, and due to delay, network jitter, computational power conditions of node differentiation and the like, the aggregation time of all rounds can be increased in a certain probability; other aggregation algorithms basically sacrifice local computing resources and increase the number of training sessions for local clients to reduce the number of communications with higher level nodes. However, the method has a certain effect in a specific research field, and the problem that the federal member asynchronously participates in training in a mobile scene is not fundamentally solved.
In conclusion, a network structure tends to be developed towards a cloud side-end layered structure, a large amount of user data are gathered on an end side, a cloud side can contact a wide range of clients, the side has the advantages of rich computing resources, low-cost communication overhead and the like, the development of machine learning, particularly collaborative machine learning is promoted, a new paradigm is provided for machine learning due to the occurrence of federal learning, and in a data sensitive era, the privacy of users is protected while the machine learning efficiency is not reduced. Traditional federal learning is not good in a plurality of mobility scenes such as the internet of vehicles because of only a cloud end or an edge end two-layer structure, and a traditional aggregation algorithm adopts a synchronous parameter updating mode, so that the convergence speed is low. Therefore, a new method and system for asynchronous federal learning of mobile vehicles is needed.
Through the above analysis, the problems and defects of the prior art are as follows:
(1) the existing cloud-client machine learning model has high time delay cost and communication overhead, and cannot obtain good performance in a plurality of time-sensitive network scenes.
(2) The existing federate member adopting the machine learning framework of the traditional synchronous aggregation algorithm has the conditions of unstable connection and asynchronous connection, and the problem of how to dynamically add the federate member into the federate aggregation cannot be solved.
(3) The traditional federal learning is poor in performance in many mobility scenes due to the fact that only a cloud end or an edge end two-layer structure is adopted, and a traditional aggregation algorithm adopts a synchronous parameter updating mode and is low in convergence speed.
(4) Most of the traditional federated learning aggregation algorithm is based on the idea of a model averaging algorithm, aggregation time of all rounds is increased, local computing resources are sacrificed, and training times of local clients are increased.
(5) The prior art schemes all have certain effect in specific research fields, and the problem that federates participate in training asynchronously in a mobile scene is not solved fundamentally.
The difficulty in solving the above problems and defects is: the method needs to design a layered and domain-divided federated learning architecture of the cloud side, wherein asynchronous access modes are needed for parameter updating of edge aggregation and cloud aggregation algorithms; the whole global algorithm updating process should consider the trade-off between the communication overhead of the cloud edge and the local computing resources.
The significance of solving the problems and the defects is as follows: the method is optimized on the basis of traditional federal learning, and a layered domain-divided cloud side federal learning framework completes a machine learning task while protecting user privacy; by utilizing the relation of layering and domain division of the cloud edge, the advantages of cloud computing and edge computing are effectively exerted, and the concept of 'domain' is introduced on the basis of layering, so that the update of model parameters in one edge server is conveniently managed; the method considers the dynamic access characteristics and the resource overhead condition of the mobile vehicle in the design, the dynamic addition and aggregation process of the vehicle is a necessary event, and the optimal condition between the network flow pressure and the machine learning precision can be obtained by balancing the communication overhead and the use condition of the computing resource.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides an asynchronous federal learning method, system, equipment and terminal of a mobile vehicle, and particularly relates to an asynchronous federal learning method, system, equipment and terminal of a mobile vehicle based on a cloud side network layered domain architecture.
The invention is realized by the implementation of an asynchronous federal learning system of a mobile vehicle, which comprises a user layer, a domain-edge server layer and a data processing center layer. The three layers of the whole system jointly complete the generation of a machine learning original data set, model training, parameter aggregation and issuing.
The user layer consists of a plurality of vehicles in the Internet of vehicles environment, and the vehicles are used as end equipment and provide original training data and computing resources required by local training for machine learning; the vehicles are divided into a plurality of domains in geographic space, the vehicles in the signal coverage range of the same road side unit belong to the same domain, and are managed by corresponding domain servers;
a domain-edge server layer consisting of servers at the edge of the network, wherein in the Internet of vehicles environment, the servers are road side units or base stations for providing services for vehicles, and the edge servers frequently communicate with all vehicle users in the domain to update model parameters; and logically abstracting the domain servers, wherein each domain server is a logical client and is only visible to the cloud center.
The cloud data processing center layer is provided with a global view of the whole network and provides an asynchronous aggregation process, each domain-edge server uploads a side model aggregation result to the cloud center, the cloud center performs cloud side global model aggregation, most communication occurs between the cloud center and a logic client, the domain shields bottom vehicle information, and a real client cannot see the cloud and does not need communication.
Another object of the present invention is to provide an asynchronous federal learning method for a mobile vehicle using the asynchronous federal learning system for a mobile vehicle, which includes:
initially, local vehicle equipment acquires machine learning initial parameters or a previous round of aggregation results from a domain server, and after a certain round of local machine learning training, a vehicle uploads a model gradient to a domain; the edge domain server asynchronously receives the model gradient uploaded by the vehicle and then carries out edge aggregation, and the domain edge server issues the aggregation result to all vehicles in the domain to complete local parameter updating in the domain; and performing machine learning training by using local data in the edge aggregation, and acquiring local optimal model parameters by minimizing a local loss function.
The same-time domain server uploads the result to the cloud center for secondary aggregation, the cloud end also adopts an asynchronous mode to receive the model gradient of the edge server, after the cloud side aggregation is completed, the aggregation result is issued to all the edge servers, and the edge servers issue parameters as intermediate media to all vehicles to complete global parameter updating; at the moment, a round of global training process is finished, an asynchronous access mode in the whole system forms a closed loop, and the process is continued until the model training reaches the preset precision or the whole algorithm converges. And the cloud aggregation performs machine learning training by using global data under all domains, and obtains global optimal model parameters by minimizing a global loss function.
Further, the asynchronous federal learning method for a mobile vehicle comprises the following steps:
comprehensively utilizing cloud computing and edge computing, updating parameters in a concept management range of a 'domain', and providing a network layered domain-division architecture based on a cloud side car;
step two, providing an asynchronous federated learning aggregation algorithm aFedV suitable for the cloud-side-vehicle-based network hierarchical domain-division architecture, and updating parameters by adopting an asynchronous mode;
and thirdly, carrying out experiments on different data distributions aiming at the aggregation algorithm and the hierarchical architecture, and verifying the performance of the aFedV algorithm from the aspects of model training accuracy and communication overhead.
Further, the asynchronous federal learning method of the mobile vehicle further comprises the steps of constructing a layered domain division model; wherein the parameters of the layered domain-division model are defined as follows:
there are u edge domain servers under the cloud edge scene, E ═ E1,e2,…,euDenotes a set of u edge domain servers, where euRepresenting the u-th edge domain server; s is the domain where the domain edge server is located, where S ═ S1,s2,…,siDenotes a set of all domains; each edge domain server covers v vehicles in total, and C is { C ═ C }1,c2,…,cvDenotes the set of v vehicles, where cvDenotes the v-th vehicle, cv iIndicating a v-th vehicle in the i-th domain, Di vRepresenting training samples generated by the v-th vehicle in the i-th domain, wherein the number of the samples is D ═ Di vAnd | represents. Wherein, the asynchronous federal learning process parameters of the layering and the domain division further comprise:
l is a local training round index, L is a local training total round, m represents the edge aggregation times, n represents the cloud aggregation times, h is the local machine learning uploading local model gradient to the edge server every time h is executed, and k is the cloud aggregation executed once every k times the edge aggregation is executed.
Assuming that there are | S | domains currently, there is a RSU under each domain, and there is a domain-edge server e in the communication coverage areauThe vehicle driving road section in the region is a target road section; due to transmission of RSUThe radius is much larger than the road width, so the road width is ignored and the underlying road-vehicle model is reduced to a one-dimensional model. In the time period t, the number of vehicles reaching the target road section obeys Poisson distribution with the parameter lambda, namely:
where v represents the number of vehicles arriving within time λ t. Vehicle cv iParticipating in local deep learning training and updating
Finding better parameters by iterating l times
To minimize the local objective loss function
Namely:
after h local training, vehicle cv iUploading the updated gradient to the edge domain server euPerforming edge aggregation, wherein the formula of the edge aggregation is as follows:
wherein, | CiL represents the total number of vehicles v in the ith domain,
representing the number of training samples owned by the v-th vehicle in the i-th domain.
The edge server will update the gradient
The local parameter updating method is issued to all vehicles in the domain, and the vehicles receive the gradient issued by the edge server to update the local parameters while updating the local parameters
Meanwhile, each edge server aggregates the obtained gradient after each k-round edge aggregation
Uploading to a cloud server to execute cloud aggregation, wherein the cloud aggregation formula is as follows:
where | S | represents the total number of fields i. Cloud usage aggregation results
Updating the cloud model, wherein the cloud server minimum global loss function formula is as follows:
optimizing model parameters
After being issued to vehicles in each domain, the local vehicle cv iPerforming global parameters
And (6) updating.
Further, the asynchronous federal learning method for a mobile vehicle further comprises:
the method is characterized in that the problem of asynchronous connection of federal members is provided for a layered domain-divided federal learning framework, in an internet of vehicles environment, dynamic participation of vehicles in federal learning influences an aggregation process and a training result, a model of the problem is established, and the problem is solved by designing an aggregation algorithm.
In a layered domain-division scene, a two-stage model aggregation mode is adopted to respectively minimize a local loss function and a global loss function, and an asynchronous mode can reduce the running time of the whole algorithm and minimize the cloud edge communication times Ms-cNumber of times of communication with edgee-c(ii) a Wherein the optimization function is represented as follows:
s.t.|Ci|＞0(I＝0，1，...|S|)；
ns＞ms＞0(s＝0，1，...|S|)；
wherein M (M; n) represents the total number of times of cloud side communication when the cloud model reaches the preset precision,
a global loss function is represented that is,
representing the local loss function of the v-th vehicle in the i-th domain.
Further, in the second step, the aFedV algorithm is composed of a two-stage aggregation algorithm, and edge side aggregation is performed in the edge domain server and is responsible for updating the intra-domain model parameters and minimizing the local loss function; and the cloud side aggregation is executed in the cloud center and is used for being responsible for updating the global model parameters and minimizing the global loss function, and the two-stage aggregation algorithm adopts an asynchronous mode.
The aFedV algorithm consists of three modules distributed on the cloud side car and is divided into a two-stage polymerization algorithm; when the vehicle is locally trained and the model parameters are updated, the model parameters or gradients issued by the edges are received and updated, and the parameters come from edge aggregation
Or cloud aggregation
The result of (1).
Further, in step two, the aFedV algorithm further includes:
initially, each federate shares a model, and during the training process, each vehicle is trained with its own raw data, so that each client's model
Features that tend to be self-data; after a round of edge aggregation, the current local model
The local model has the characteristics of other vehicle data in the domain, and is subjected to cloud aggregation to obtain the current local model
Will contain characteristics of global vehicle data, the cloud model at that time
The vehicle training system has the characteristics of all the data of the vehicles participating in training uniformly; based on the parameter issuing function in the asynchronous algorithm, the newly added vehicle participating in training always obtains the current latest model, and the characteristic has great advantage in the time sensitivity scene of the Internet of vehicles.
It is a further object of the invention to provide a computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of:
initially, local vehicle equipment acquires machine learning initial parameters or a previous round of aggregation results from a domain server, and after a certain round of local machine learning training, a vehicle uploads a model gradient to a domain; the edge domain server asynchronously receives the model gradient uploaded by the vehicle and then carries out edge aggregation, and the domain edge server issues the aggregation result to all vehicles in the domain to complete local parameter updating in the domain;
the same-time domain server uploads the result to the cloud center for secondary aggregation, the cloud end also adopts an asynchronous mode to receive the model gradient of the edge server, after the cloud side aggregation is completed, the aggregation result is issued to all the edge servers, and the edge servers issue parameters as intermediate media to all vehicles to complete global parameter updating; at the moment, a round of global training process is finished, an asynchronous access mode in the whole system forms a closed loop, and the process is continued until the model training reaches the preset precision or the whole algorithm converges.
It is another object of the present invention to provide a computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of:
initially, local vehicle equipment acquires machine learning initial parameters or a previous round of aggregation results from a domain server, and after a certain round of local machine learning training, a vehicle uploads a model gradient to a domain; the edge domain server asynchronously receives the model gradient uploaded by the vehicle and then carries out edge aggregation, and the domain edge server issues the aggregation result to all vehicles in the domain to complete local parameter updating in the domain;
the same-time domain server uploads the result to the cloud center for secondary aggregation, the cloud end also adopts an asynchronous mode to receive the model gradient of the edge server, after the cloud side aggregation is completed, the aggregation result is issued to all the edge servers, and the edge servers issue parameters as intermediate media to all vehicles to complete global parameter updating; at the moment, a round of global training process is finished, an asynchronous access mode in the whole system forms a closed loop, and the process is continued until the model training reaches the preset precision or the whole algorithm converges.
Another object of the present invention is to provide an information data processing terminal, which is used for implementing the asynchronous federal learning system of a mobile vehicle.
By combining all the technical schemes, the invention has the advantages and positive effects that: the invention provides an asynchronous federal learning method of a mobile vehicle, which provides a network layered domain-division framework based on a cloud side vehicle, wherein the framework comprehensively utilizes the advantages of cloud computing and edge computing and manages parameter updating in a range by introducing a domain concept; an Asynchronous Federated Learning aggregation algorithm (Asynchronous fed Learning of Internet of Vehicles, aFedV) suitable for the framework is provided, and the algorithm adopts an Asynchronous mode to update parameters, so that the problem that the parameters cannot be updated timely due to dynamic connection of mobile federal members in the calculation process can be solved. Aiming at different data characteristic distribution scenes, the method is experimentally verified and analyzed, and the model training accuracy and the communication overhead prove that the aFedV algorithm has good performance under the cloud side car-based network layered domain-division architecture.
The aFedV asynchronous aggregation algorithm provided by the invention is a federated optimization algorithm, is optimized on the existing model averaging algorithm, and can effectively solve the problem of uneven distribution of data sets in a mobile internet scene. The invention provides an asynchronous federated learning framework based on cloud side network layering and domain division in a vehicle networking scene, and provides an asynchronous aggregation algorithm aFedV suitable for the framework. Aiming at unbalanced data distribution, the method disclosed by the invention is used for developing an experiment on an MNIST data set, and from the view of a cloud model result, the aFedV can be quickly and effectively converged under the condition of ensuring certain precision. In the future, the invention combines the block chain technology with a layered and domain-divided asynchronous federal learning framework and focuses on a privacy protection mechanism based on mobility in a system.
The total time of traditional federal learning training is limited by the bottom client and the aggregation algorithm, which includes uncertain training delay, queuing delay, transmission delay, etc., and the aggregation time of each round of the synchronous aggregation mode depends on the client which takes the longest time. The asynchronous aggregation algorithm provided by the invention adopts an asynchronous access mode, and is particularly suitable for mobile terminalsThe algorithm makes a balance on communication turns and calculation overhead by utilizing the communication advantages of cloud edges and edge-ends.
And
the method makes up the defects of the traditional algorithm, and the running time of each round of the algorithm is determined by the vehicle trained firstly, so that the running time of the whole training can be effectively reduced by the asynchronous mode, and the method is favorable for the quick convergence of the whole algorithm.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments of the present invention will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
Fig. 1 is a flowchart of an asynchronous federal learning method for a mobile vehicle according to an embodiment of the present invention.
Fig. 2 is an asynchronous federal learning method for a mobile vehicle according to an embodiment of the present invention.
Fig. 3 is an abstract diagram of a cloud edge network layered domain system architecture provided in an embodiment of the present invention.
Fig. 4 is a flowchart of asynchronous federal learning based on hierarchical domain division of a cloud edge network according to an embodiment of the present invention.
Fig. 5 is a schematic diagram of accuracy of a cloud server CNN model under an unbalanced MNIST data set according to an embodiment of the present invention.
Fig. 6 is a schematic representation of poisson distribution of arriving vehicles in different fields provided by embodiments of the present invention.
Fig. 7 is a schematic precision diagram of a local vehicle CNN model under a balanced MNIST data set according to an embodiment of the present invention.
Fig. 8 is a schematic diagram of accuracy of a cloud server CNN model based on a balanced MNIST data set according to an embodiment of the present invention.
Fig. 9 is a schematic diagram of the size and scale of a vehicle sample set in an unbalanced MNIST data set according to an embodiment of the present invention.
Fig. 10 is a schematic loss value diagram of a local vehicle CNN model under an unbalanced MNIST data set according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
In view of the problems in the prior art, the invention provides an asynchronous federal learning method, system, device and terminal for a mobile vehicle, and the invention is described in detail below with reference to the accompanying drawings.
As shown in fig. 1, an asynchronous federal learning method for a mobile vehicle according to an embodiment of the present invention includes the following steps:
s101, comprehensively utilizing cloud computing and edge computing, updating parameters in a concept management range of a 'domain', and providing a network layered domain-division framework based on a cloud side car;
s102, providing an asynchronous federated learning aggregation algorithm aFedV suitable for the cloud-side-vehicle-based network hierarchical domain-division architecture, and updating parameters by adopting an asynchronous mode;
s103, aiming at the aggregation algorithm and the hierarchical architecture, experiments are carried out on different data distributions, and the performance of the aFedV algorithm is verified in the aspects of model training accuracy and communication overhead.
The technical solution of the present invention is further described below with reference to specific examples.
1. The invention provides a network layered domain-division architecture based on a cloud side car, which comprehensively utilizes the advantages of cloud computing and edge computing and manages parameter updating in a range by introducing a domain concept; an Asynchronous Federated Learning aggregation algorithm (Asynchronous fed Learning of Internet of Vehicles, aFedV) suitable for the framework is provided, and the algorithm adopts an Asynchronous mode to update parameters, so that the problem that the parameters cannot be updated timely due to dynamic connection of mobile federal members in the calculation process can be solved. Aiming at different data characteristic distribution scenes, the method is experimentally verified and analyzed, and the model training accuracy and the communication overhead prove that the aFedV algorithm has good performance under the cloud side car-based network layered domain-division architecture.
The invention provides a layered and domain-divided asynchronous federated learning framework, such as a graph 2 and an asynchronous aggregation algorithm aFedV suitable for the framework, and solves the problem of asynchronous connection of federated members in a mobile scene by comprehensively utilizing the advantages of a cloud-edge-end network framework. The main contributions of the present invention are as follows:
(1) the asynchronous federated learning framework based on cloud side end network layering and domain division in the Internet of vehicles environment is provided, and under the framework, a vehicle serves as end equipment to provide training data and local computing resources for machine learning; an edge server is arranged under each domain, and a plurality of sub-domain edge servers aggregate client models belonging to the domain; and aggregating the models uploaded by the edge servers by the cloud center.
(2) Aiming at the problem that a hierarchical domain-divided federal learning framework provides asynchronous connection of federal members, in an internet of vehicles environment, dynamic participation of vehicles in federal learning influences an aggregation process and a training result, the invention establishes a model of the problem and solves the problem by designing an aggregation algorithm.
(3) The proposed aFedV algorithm is a two-stage aggregation algorithm, aggregation is respectively carried out on the edge servers and the cloud servers, the aFedV algorithm adopts an asynchronous aggregation mode on each layer, the aggregation mode not only utilizes a plurality of widely distributed domain edge servers to obtain a large amount of data and cheap edge-end communication overhead, but also utilizes the strong computing power of the cloud servers to exchange machine learning model parameters with the plurality of edge servers, and global parameters are integrated.
The rest part of the invention is arranged as follows, the second part provides a cloud edge terminal layered and domain-divided asynchronous federated learning framework and defines a system model, in the third part, the invention designs and analyzes an asynchronous aggregation algorithm aFedV, the fourth part of the invention develops experiments on different data distributions aiming at the proposed aggregation algorithm and layered and hierarchical framework, and the fifth part obtains a conclusion.
2. Asynchronous federated learning system
2.1 System architecture
The asynchronous federal learning architecture based on cloud side end network layering and domain division in the internet of vehicles environment is shown in fig. 2, and the whole architecture is divided into three layers, namely a user layer, an edge server layer and a cloud center layer. On the basis of a three-layer architecture, the invention introduces the concept of 'domain' to manage the vehicle users under the whole edge server, and each domain is regarded as a single logical node user through logical abstraction, so that the whole logical view still keeps the two-layer architecture unchanged, and the logical view after abstraction is shown in fig. 3.
(1) The user layer consists of a number of vehicles in a car networking environment that generate the raw training data needed for machine learning and provide the computing resources needed for local training. The vehicles are divided into a plurality of domains in geographic space, the vehicles in the signal coverage range of the same road side unit belong to the same domain, and are managed by the corresponding domain server;
(2) the domain-edge server layer consists of servers at the edge of the network, which in a car networking environment may be road side units, base stations, etc. that serve the vehicles. Due to the large bandwidth and low latency of edge-to-end communications, the edge server may communicate frequently with all vehicle users within its domain to update the model parameters. In addition, the invention carries out logic abstraction on the domain servers, each domain server is a logic client and is only visible to the cloud center.
(3) The cloud data processing center has a global view of the whole network, and generally, the cloud-end communication bandwidth is small, the communication overhead is large, and the real-time performance is poor. Therefore, in the asynchronous aggregation process provided by the invention, the cloud-end communication frequency is very low, each domain-edge server uploads the side model aggregation result to the cloud center, the cloud center performs cloud-side global model aggregation, in fact, most communication occurs between the cloud center and the logic client, the domain shields the information of the bottom layer vehicle, the real client is invisible to the cloud and does not need communication, and therefore, the complexity of the whole network communication is reduced.
An asynchronous federal learning flow chart of cloud edge network layering and domain division is shown in fig. 4, initially, local vehicle equipment obtains machine learning initial parameters or a previous round of aggregation result from a belonging domain server, and after a certain round of local machine learning training, a vehicle uploads a model gradient to a belonging domain; the edge domain server asynchronously receives the model gradient uploaded by the vehicle, then carries out edge aggregation, and issues the aggregation result to all vehicles in the domain to complete local parameter updating in the domain; and the time domain server uploads the result to the cloud center for secondary (fuser and advanced) aggregation, the cloud end also adopts an asynchronous mode to receive the model gradient of the edge server, after the cloud side aggregation is completed, the aggregation result is issued to all the edge servers, and the edge servers serve as intermediate issuing parameters to all vehicles to complete global parameter updating. At the moment, a round of global training process is finished, an asynchronous access mode in the whole system forms a closed loop, and the process is continued until the model training reaches the preset precision or the whole algorithm converges.
2.2 model definition
For convenience of scene description, in the system proposed by the present invention, the parameters of the hierarchical domain model are defined as follows:
there are u edge domain servers under the cloud edge scene, E ═ E1,e2,…,euDenotes a set of u edge domain servers, where euRepresenting the u-th edge domain server; s is the domain where the domain edge server is located, where S ═ S1,s2,…,siDenotes a set of all domains; each edge domain server covers v vehicles in total, and C is { C ═ C }1,c2,…,cvDenotes the set of v vehicles, where cvDenotes the v-th vehicle, cv iIndicating a v-th vehicle in the i-th domain, Di vRepresenting training samples generated by the v-th vehicle in the i-th domain, wherein the number of the samples is D ═ Di vI represents. Other parameters are illustrated in table 1.
TABLE 1 hierarchical and domain-based asynchronous federated learning process parameter table
Suppose there are | S | domains currently, one roadside unit under each domain, and one domain-edge server e exists in the communication coverage areauAnd the vehicle driving road section in the domain is the target road section. Since the transmission radius of the RSU is much larger than the road width, the invention omits the road width and simplifies the road-vehicle (RSU-vehicle) model into a one-dimensional model. In the time period t, the number of vehicles reaching the target road section obeys Poisson distribution with the parameter lambda, namely:
where v represents the number of vehicles arriving within time λ t. Vehicle cv iParticipating in local deep learning training and updating
It finds the better parameter by iterating l times
To minimize the local objective loss function
Namely:
after h local training, vehicle cv iUploading the updated gradient to the edge domain server euPerforming edge aggregation, wherein the formula of the edge aggregation is as follows:
|Cil represents the total number of vehicles v in the ith domain,
representing the number of training samples owned by the v-th vehicle in the i-th domain. The edge server then compares the updated gradient
The local parameter updating method is issued to all vehicles in the domain, and the vehicles receive the gradient issued by the edge server to update the local parameters while updating the local parameters
Meanwhile, each edge server aggregates the obtained gradient after each k-round edge aggregation
Uploading to a cloud server to execute cloud aggregation, wherein the cloud aggregation formula is as follows:
where | s | represents the total number of fields i. Cloud usage aggregation results
Updating the cloud model, wherein the cloud server minimum global loss function formula is as follows:
optimizing model parameters
Down to vehicles in each domain, and local vehicles cv iPerforming global parameters
And (6) updating.
2.3 problem formulation
In the layered domain-division scene provided by the invention, model parameters or gradient information need to be exchanged between the cloud-side-vehicle three-layer architectures, which causes a large amount of communication overhead, and the time when the vehicle is accessed and leaves the RSU is a probability problem, so that the participating vehicles can not guarantee simultaneous online training. By optimizing the algorithm, the invention respectively minimizes the local loss function and the global loss function by adopting a two-stage model aggregation mode, and the asynchronous mode can reduce the running time of the whole algorithm and minimize the cloud edge communication times Ms-cNumber of times of communication with edgee-cUnder the condition of ensuring certain accuracy, the communication overhead is reduced, and the asynchronism of vehicle access and the real-time performance of available models are improved. The optimization function is represented as follows:
s.t.|Ci|＞0(i＝0，1，...|S|) (8)
ns＞ms＞0(s＝0，1，...|S|) (9)
wherein M (M; n) represents the total number of times of cloud side communication when the cloud model reaches the preset precision,
a global loss function is represented that is,
representing the local loss function of the v-th vehicle in the i-th domain.
3. Design and analysis of aFedV algorithm
The part of the invention designs an asynchronous federated learning aggregation algorithm based on cloud side terminal hierarchical domains, and the algorithm considers a vehicle moving scene and integrates the advantages of cloud computing and edge computing.
3.1 design of aggregation Algorithm
The aFedV algorithm consists of a two-stage aggregation algorithm, and edge side aggregation is executed in an edge domain server and is responsible for updating the model parameters in the domain and minimizing a local loss function; and the cloud side aggregation is executed in the cloud center and is responsible for updating the global model parameters and minimizing the global loss function, the two-stage aggregation algorithm adopts an asynchronous mode, and pseudo codes of the algorithm are shown in a table 2.
TABLE 2 pseudo code for aFedV Algorithm
The aFedV algorithm consists of three modules distributed on the cloud side car and is divided into a two-stage polymerization algorithm. When the vehicle is locally trained and the model parameters are updated, the model parameters or gradients issued by the edges are received and updated, and the parameters come from edge aggregation
Or cloud aggregation
The result of (1).
3.2 aFedV Algorithm analysis
The total time of traditional federal learning training is limited by the bottom client and the aggregation algorithm, which includes uncertain training delay, queuing delay, transmission delay, etc., and the aggregation time of each round of the synchronous aggregation mode depends on the client which takes the longest time. The asynchronous aggregation algorithm provided by the invention adopts an asynchronous access mode, is particularly suitable for mobile terminals, and makes a balance on communication turns and calculation overhead by utilizing the communication advantages of cloud edges and edge-end.
And
the method makes up the defects of the traditional algorithm, and the running time of each round of the algorithm is determined by the vehicle trained firstly, so that the running time of the whole training can be effectively reduced by the asynchronous mode, and the method is favorable for the quick convergence of the whole algorithm.
The federated learning is a master-slave machine learning training architecture which is independent of a model, and the cloud edge-side hierarchical domain system architecture provided by the invention inherits the advantages and has the characteristics of strong robustness and generalization capability. Initially, each federate shares a model, and during the training process, each vehicle is trained with its own raw data, so that each client's model
Features that favor self data, however after a round of edge aggregation, the current local model
The local model has the characteristics of other vehicle data in the domain, and is subjected to cloud aggregation to obtain the current local model
Will contain characteristics of global vehicle data, the cloud model at that time
Homogeneous (homogeneous) possesses all the features that participate in training vehicle data.
In addition, due to asynchronous access of vehicles, uncertain network delay in the car networking environment, node differentiation computational force conditions and the like, the participation time and the ending time of the federates are uncertain in the time dimension, so that a part of vehicles iterates for a plurality of times when the part of vehicles are ready to participate in training. Due to the fact that the parameter issuing function in the asynchronous algorithm is benefited, the vehicles which are newly added into the training always obtain the current latest model, and the characteristic has great advantages in the time sensitivity scene of the Internet of vehicles.
The aFedV asynchronous aggregation algorithm provided by the invention is a federated optimization algorithm, is optimized on the existing model averaging algorithm, and can effectively solve the problem of uneven distribution of data sets in a mobile internet scene, as shown in figure 5.
4. Results of the experiment
The part aims at the layered domain-division framework and the asynchronous aggregation algorithm provided by the invention to develop simulation experiments so as to prove the advantages of the invention.
The invention develops experiments on MNIST data sets and convolutional neural networks and a plurality of computing units. The MNIST data set is a ten-class hand-written digit recognition data set and consists of 60000 training samples and 10000 testing samples, and each sample is a 28-by-28-pixel gray-scale hand-written digit picture. The invention uses CNN to classify MNIST data set, wherein the convolution nerve network structure is two convolution layers of 5 × 5, each convolution layer has 10 and 20 channels, each convolution layer is followed by a 2 × 2 pooling layer and a ReLu activation function, and finally output through the linear layers of 320 nerve units, and the whole network model has 8490 parameters. Regarding the computing unit, the invention adopts a vehicle arrival model based on Poisson distribution to dynamically generate vehicles and set the life cycle of the vehicles, and simulates the vehicle offline condition in the computing process in a vehicle networking scene, wherein the vehicles are distributed in 4 domain servers, and the domain-edge servers are connected to the cloud center.
Assuming that the vehicle-to-RSU signal range follows a poisson distribution with λ ═ 0.7, the number of arriving vehicles per second is shown in fig. 6.
According to fig. 6, the vehicles are asynchronous in the time dimension when they start participating in hierarchical domain federated learning, with edge aggregation and cloud aggregation being accessed by the aFedV algorithm.
The accuracy of the balanced MNIST dataset local vehicle CNN model is shown in fig. 7.
The invention sets that when a local vehicle trains a complete epoch, the test set is used for testing, and in a certain experiment, the test precision of all vehicle local models is shown in figure 8. As can be seen from the accuracy curve of the cloud model, the accuracy of 85% preset by the invention is achieved through 39 iterations on average.
The invention finds that the acc curve lengths of each vehicle are different in reaching the preset accuracy, i.e. they have different iteration times, because the vehicles arrive in the RSU signal range based on poisson distribution and join federal learning at different time points. It can also be seen from fig. 8 that although the number of iterations varies from vehicle to vehicle, the cloud model tends to converge eventually.
In an actual internet of vehicles scene, non-uniform distribution of data on vehicles often occurs, and due to different use conditions of vehicle users, the sizes of data generated by the vehicle users are different, and fig. 9 shows the number and proportion of data samples of different vehicles in a data set non-uniform distribution experiment.
Fig. 10 shows a local training loss curve of a vehicle in a sample non-uniform distribution scene, and since the latest model issued by the edge is obtained when the vehicle added later participates in the federal learning initially, the loss curve has a very fast descending speed, and the latest minimum loss value is reached after 2 iterations.
TABLE 3 hierarchical and domain-divided asynchronous federal learning cloud frontier communication times
In the above experiment, the cloud-edge and edge-end communication times are counted, see table 3, and it can be seen that the cloud-edge communication times are 6 times less than the edge-end communication times under the cloud-edge hierarchical domain division system architecture, because the cloud-edge and edge-end communication methods and the cloud-edge and edge-end server servers are used as intermediate media to reduce the communication times by designing an asynchronous aggregation algorithm.
5. Summary of the invention
The invention provides an asynchronous federated learning framework based on cloud side network layering and domain division in a vehicle networking scene, and provides an asynchronous aggregation algorithm aFedV suitable for the framework. Aiming at unbalanced data distribution, the method disclosed by the invention is used for developing an experiment on an MNIST data set, and from the view of a cloud model result, the aFedV can be quickly and effectively converged under the condition of ensuring certain precision. In the future, the invention combines the block chain technology with a layered and domain-divided asynchronous federal learning framework and focuses on a privacy protection mechanism based on mobility in a system.
In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When used in whole or in part, can be implemented in a computer program product that includes one or more computer instructions. When loaded or executed on a computer, cause the flow or functions according to embodiments of the invention to occur, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, the computer instructions may be transmitted from one website site, computer, server, or data center to another website site, computer, server, or data center via wire (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL), or wireless (e.g., infrared, wireless, microwave, etc.)). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that includes one or more of the available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.
The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.
Claims (10)
1. An asynchronous federal learning method for a mobile vehicle, comprising:
initially, local vehicle equipment acquires machine learning initial parameters or a previous round of aggregation results from a domain server, and after a certain round of local machine learning training, a vehicle uploads a model gradient to a domain; the edge domain server asynchronously receives the model gradient uploaded by the vehicle and then carries out edge aggregation, and the domain edge server issues the aggregation result to all vehicles in the domain to complete local parameter updating in the domain;
the same-time domain server uploads the result to the cloud center for secondary aggregation, the cloud end also adopts an asynchronous mode to receive the model gradient of the edge server, after the cloud side aggregation is completed, the aggregation result is issued to all the edge servers, and the edge servers issue parameters as intermediate media to all vehicles to complete global parameter updating; at the moment, a round of global training process is finished, an asynchronous access mode in the whole system forms a closed loop, and the process is continued until the model training reaches the preset precision or the whole algorithm converges.
2. The asynchronous federal learning method for a mobile vehicle as claimed in claim 1, wherein the asynchronous federal learning method for a mobile vehicle comprises the steps of:
comprehensively utilizing cloud computing and edge computing, updating parameters in a concept management range of a 'domain', and providing a network layered domain-division architecture based on a cloud side car;
step two, providing an asynchronous federated learning aggregation algorithm aFedV suitable for the cloud-side-vehicle-based network hierarchical domain-division architecture, and updating parameters by adopting an asynchronous mode;
and thirdly, carrying out experiments on different data distributions aiming at the aggregation algorithm and the hierarchical architecture, and verifying the performance of the aFedV algorithm from the aspects of model training accuracy and communication overhead.
3. The asynchronous federal learning method for a mobile vehicle as claimed in claim 1, further comprising the steps of constructing a hierarchical domain model; wherein the parameters of the layered domain-division model are defined as follows:
there are u edge domain servers under the cloud edge scene, E ═ E1,e2,…,euDenotes a set of u edge domain servers, where euRepresenting the u-th edge domain server; s is the domain where the domain edge server is located, where S ═ S1,s2,…,siDenotes a set of all domains; each edge domain server covers v vehicles in total, and C is { C ═ C }1,c2,…,cvDenotes the set of v vehicles, where cvDenotes the v-th vehicle, cv iIndicating a v-th vehicle in the i-th domain, Di vRepresenting training samples generated by the v-th vehicle in the i-th domain, wherein the number of the samples is D ═ Di vI represents; wherein, the asynchronous federal learning process parameters of the layering and the domain division further comprise:
l is a local training round index, L is a local training total round, m represents the edge aggregation times, n represents the cloud aggregation times, h is the local machine learning uploading local model gradient to the edge server every time h is executed, and k is the cloud aggregation executed once every k times the edge aggregation is executed;
assuming that there are | S | domains currently, there is a RSU under each domain, and there is a domain-edge server e in the communication coverage areauThe vehicle driving road section in the region is a target road section; the transmission radius of the RSU is far larger than the width of the road, so that the width of the road is ignored, and the bottom road vehicle model is simplified into a one-dimensional model; in the time period t, the number of vehicles reaching the target road section obeys Poisson distribution with the parameter lambda, namely:
where v represents the number of vehicles arriving within time λ t; vehicle cv iParticipate in local deep learning trainingExercise and renewal
Finding better parameters by iterating l times
To minimize the local objective loss function
Namely:
after h local training, vehicle cv iUploading the updated gradient to the edge domain server euPerforming edge aggregation, wherein the formula of the edge aggregation is as follows:
wherein, | CiL represents the total number of vehicles v in the ith domain,
representing the number of training samples owned by the v vehicle in the i domain;
the edge server will update the gradient
The local parameter updating method is issued to all vehicles in the domain, and the vehicles receive the gradient issued by the edge server to update the local parameters while updating the local parameters
Meanwhile, each edge server aggregates the obtained gradient after each k-round edge aggregation
Uploading to a cloud server to execute cloud aggregation, wherein the cloud aggregation formula is as follows:
wherein, | s | represents the total number of fields i; cloud usage aggregation results
Updating the cloud model, wherein the cloud server minimum global loss function formula is as follows:
optimizing model parameters
After being issued to vehicles in each domain, the local vehicles
Performing global parameters
And (6) updating.
4. The asynchronous federal learning method for a mobile vehicle as claimed in claim 1, further comprising:
the method comprises the steps that a problem of asynchronous connection of federal members is provided for a layered domain-divided federal learning framework, in an internet of vehicles environment, dynamic participation of vehicles in federal learning influences an aggregation process and a training result, a model of the problem is established, and the problem is solved by designing an aggregation algorithm;
in a layered domain-division scene, two-stage model aggregation is adoptedThe mode respectively minimizes a local loss function and a global loss function, and the asynchronous mode can reduce the running time of the whole algorithm and minimize the cloud edge communication times Ms-cNumber of times of communication with edgee-c(ii) a Wherein the optimization function is represented as follows:
s.t.|Ci|＞0(i＝0，1，…|S|)；
ns＞ms＞0(s＝0，1，...|S|)；
wherein M (M; n) represents the total number of times of cloud side communication when the cloud model reaches the preset precision,
a global loss function is represented that is,
representing the local loss function of the v-th vehicle in the i-th domain.
5. The asynchronous federal learning method for a mobile vehicle as claimed in claim 2, wherein in step two, the aFedV algorithm is composed of a two-stage aggregation algorithm, and the edge-side aggregation is performed at the edge domain server for taking charge of the update of the intra-domain model parameters and minimizing the local loss function; the cloud side aggregation is executed in the cloud center and is used for being responsible for updating global model parameters and minimizing a global loss function, and both two-stage aggregation algorithms adopt asynchronous modes;
the aFedV algorithm consists of three modules distributed on the cloud side car and is divided into a two-stage polymerization algorithm; when the vehicle is locally trained and the model parameters are updated, the model parameters or gradients issued by the edges are received at the same timeUpdating, these parameters come from edge aggregation
Or cloud aggregation
The result of (1).
6. The asynchronous federal learning method for a mobile vehicle as claimed in claim 2, wherein in step two, the aFedV algorithm further comprises:
initially, each federate shares a model, and during the training process, each vehicle is trained with its own raw data, so that each client's model
Features that tend to be self-data; after a round of edge aggregation, the current local model
The local model has the characteristics of other vehicle data in the domain, and is subjected to cloud aggregation to obtain the current local model
Will contain characteristics of global vehicle data, the cloud model at that time
The vehicle training system has the characteristics of all the data of the vehicles participating in training uniformly; based on the parameter issuing function in the asynchronous algorithm, the newly added vehicle participating in training always obtains the current latest model, and the characteristic has great advantage in the time sensitivity scene of the Internet of vehicles.
7. A computer device, characterized in that the computer device comprises a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to carry out the steps of:
initially, local vehicle equipment acquires machine learning initial parameters or a previous round of aggregation results from a domain server, and after a certain round of local machine learning training, a vehicle uploads a model gradient to a domain; the edge domain server asynchronously receives the model gradient uploaded by the vehicle and then carries out edge aggregation, and the domain edge server issues the aggregation result to all vehicles in the domain to complete local parameter updating in the domain;
the same-time domain server uploads the result to the cloud center for secondary aggregation, the cloud end also adopts an asynchronous mode to receive the model gradient of the edge server, after the cloud side aggregation is completed, the aggregation result is issued to all the edge servers, and the edge servers issue parameters as intermediate media to all vehicles to complete global parameter updating; at the moment, a round of global training process is finished, an asynchronous access mode in the whole system forms a closed loop, and the process is continued until the model training reaches the preset precision or the whole algorithm converges.
8. A computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of:
initially, local vehicle equipment acquires machine learning initial parameters or a previous round of aggregation results from a domain server, and after a certain round of local machine learning training, a vehicle uploads a model gradient to a domain; the edge domain server asynchronously receives the model gradient uploaded by the vehicle and then carries out edge aggregation, and the domain edge server issues the aggregation result to all vehicles in the domain to complete local parameter updating in the domain;
the same-time domain server uploads the result to the cloud center for secondary aggregation, the cloud end also adopts an asynchronous mode to receive the model gradient of the edge server, after the cloud side aggregation is completed, the aggregation result is issued to all the edge servers, and the edge servers issue parameters as intermediate media to all vehicles to complete global parameter updating; at the moment, a round of global training process is finished, an asynchronous access mode in the whole system forms a closed loop, and the process is continued until the model training reaches the preset precision or the whole algorithm converges.
9. An asynchronous federal learning system of a mobile vehicle for implementing the asynchronous federal learning method of the mobile vehicle as claimed in any one of claims 1 to 6, wherein the asynchronous federal learning system of the mobile vehicle comprises a user layer, a domain-edge server layer and a data processing center layer;
the user layer consists of a plurality of vehicles in the Internet of vehicles environment, and the vehicles are used as end equipment and provide original training data and computing resources required by local training for machine learning; the vehicles are divided into a plurality of domains in geographic space, the vehicles in the signal coverage range of the same road side unit belong to the same domain, and are managed by corresponding domain servers;
a domain-edge server layer consisting of servers at the edge of the network, wherein in the Internet of vehicles environment, the servers are road side units or base stations for providing services for vehicles, and the edge servers frequently communicate with all vehicle users in the domain to update model parameters; performing logic abstraction on the domain servers, wherein each domain server is a logic client and is only visible to the cloud center;
the cloud data processing center layer is provided with a global view of the whole network and provides an asynchronous aggregation process, each domain-edge server uploads a side model aggregation result to the cloud center, the cloud center performs cloud side global model aggregation, most communication occurs between the cloud center and a logic client, the domain shields bottom vehicle information, and a real client cannot see the cloud and does not need communication.
10. An information data processing terminal, characterized in that the information data processing terminal is configured to implement the asynchronous federal learning system for mobile vehicles as claimed in claim 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111229664.1A CN114116198A (en) | 2021-10-21 | 2021-10-21 | Asynchronous federal learning method, system, equipment and terminal for mobile vehicle |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111229664.1A CN114116198A (en) | 2021-10-21 | 2021-10-21 | Asynchronous federal learning method, system, equipment and terminal for mobile vehicle |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114116198A true CN114116198A (en) | 2022-03-01 |
Family
ID=80376499
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111229664.1A Pending CN114116198A (en) | 2021-10-21 | 2021-10-21 | Asynchronous federal learning method, system, equipment and terminal for mobile vehicle |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114116198A (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114745383A (en) * | 2022-04-08 | 2022-07-12 | 浙江金乙昌科技股份有限公司 | Mobile edge calculation assisted multilayer federal learning method |
| CN114827198A (en) * | 2022-03-31 | 2022-07-29 | 电子科技大学 | Multilayer center asynchronous federal learning method applied to Internet of vehicles |
| CN114945044A (en) * | 2022-07-25 | 2022-08-26 | 北京智芯微电子科技有限公司 | Method, device and equipment for constructing digital twin platform based on federal learning |
| CN115037618A (en) * | 2022-06-06 | 2022-09-09 | 电子科技大学 | Lightweight edge intelligent collaborative federal learning platform based on KubeEdge |
| CN115277689A (en) * | 2022-04-29 | 2022-11-01 | 国网天津市电力公司 | Yun Bianwang network communication optimization method and system based on distributed federal learning |
| CN115277446A (en) * | 2022-07-12 | 2022-11-01 | 中国信息通信研究院 | Energy-saving online internet connection learning network and method |
| CN115376031A (en) * | 2022-10-24 | 2022-11-22 | 江西省科学院能源研究所 | Road unmanned aerial vehicle routing inspection data processing method based on federal adaptive learning |
| CN115422820A (en) * | 2022-06-02 | 2022-12-02 | 国汽智控(北京)科技有限公司 | Federal learning model training method and road condition prediction method applied to road condition prediction |
| CN115481752A (en) * | 2022-09-23 | 2022-12-16 | 中国电信股份有限公司 | Model training method and device, electronic equipment and storage medium |
| CN115987694A (en) * | 2023-03-20 | 2023-04-18 | 杭州海康威视数字技术股份有限公司 | Equipment privacy protection method, system and device based on multi-domain federation |
| CN116094993A (en) * | 2022-12-22 | 2023-05-09 | 电子科技大学 | Federal learning security aggregation method suitable for edge computing scene |
| CN116546429A (en) * | 2023-06-06 | 2023-08-04 | 江南大学 | Vehicle selection method and system in federal learning of Internet of vehicles |
| WO2023208043A1 (en) * | 2022-04-29 | 2023-11-02 | 索尼集团公司 | Electronic device and method for wireless communication system, and storage medium |
| CN117076132A (en) * | 2023-10-12 | 2023-11-17 | 北京邮电大学 | Resource allocation and aggregation optimization method and device for hierarchical federal learning system |
| CN117812564A (en) * | 2024-02-29 | 2024-04-02 | 湘江实验室 | Federal learning method, device, equipment and medium applied to Internet of vehicles |
| CN116094993B (en) * | 2022-12-22 | 2024-05-31 | 电子科技大学 | Federal learning security aggregation method suitable for edge computing scene |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021016596A1 (en) * | 2019-07-25 | 2021-01-28 | Nvidia Corporation | Deep neural network for segmentation of road scenes and animate object instances for autonomous driving applications |
| CN112817653A (en) * | 2021-01-22 | 2021-05-18 | 西安交通大学 | Cloud-side-based federated learning calculation unloading computing system and method |
| US11017322B1 (en) * | 2021-01-28 | 2021-05-25 | Alipay Labs (singapore) Pte. Ltd. | Method and system for federated learning |
| CN113204443A (en) * | 2021-06-03 | 2021-08-03 | 京东科技控股股份有限公司 | Data processing method, equipment, medium and product based on federal learning framework |
| CN113283177A (en) * | 2021-06-16 | 2021-08-20 | 江南大学 | Mobile perception caching method based on asynchronous federated learning |
| CN113419857A (en) * | 2021-06-24 | 2021-09-21 | 广东工业大学 | Federal learning method and system based on edge digital twin association |
| CN113435472A (en) * | 2021-05-24 | 2021-09-24 | 西安电子科技大学 | Vehicle-mounted computing power network user demand prediction method, system, device and medium |
| CN113467927A (en) * | 2021-05-20 | 2021-10-01 | 杭州趣链科技有限公司 | Block chain based trusted participant federated learning method and device |
-
2021
- 2021-10-21 CN CN202111229664.1A patent/CN114116198A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021016596A1 (en) * | 2019-07-25 | 2021-01-28 | Nvidia Corporation | Deep neural network for segmentation of road scenes and animate object instances for autonomous driving applications |
| CN112817653A (en) * | 2021-01-22 | 2021-05-18 | 西安交通大学 | Cloud-side-based federated learning calculation unloading computing system and method |
| US11017322B1 (en) * | 2021-01-28 | 2021-05-25 | Alipay Labs (singapore) Pte. Ltd. | Method and system for federated learning |
| CN113467927A (en) * | 2021-05-20 | 2021-10-01 | 杭州趣链科技有限公司 | Block chain based trusted participant federated learning method and device |
| CN113435472A (en) * | 2021-05-24 | 2021-09-24 | 西安电子科技大学 | Vehicle-mounted computing power network user demand prediction method, system, device and medium |
| CN113204443A (en) * | 2021-06-03 | 2021-08-03 | 京东科技控股股份有限公司 | Data processing method, equipment, medium and product based on federal learning framework |
| CN113283177A (en) * | 2021-06-16 | 2021-08-20 | 江南大学 | Mobile perception caching method based on asynchronous federated learning |
| CN113419857A (en) * | 2021-06-24 | 2021-09-21 | 广东工业大学 | Federal learning method and system based on edge digital twin association |
Non-Patent Citations (3)
| Title |
|---|
| YUNLONG LU: ""Differentially Private Asynchronous Federated Learning for Mobile Edge Computing in Urban Informatics"", 《IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS》, vol. 16, no. 3, 31 March 2020 (2020-03-31), pages 2134 - 2143, XP011768092, DOI: 10.1109/TII.2019.2942179 * |
| 夕月一弯: ""联邦学习综述"", Retrieved from the Internet <URL:《https://www.cnblogs.com/wt869054461/p/12375011.html》> * |
| 廖钰盈: ""面向异构边缘节点的融合联邦学习机制研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 2021, 15 May 2021 (2021-05-15), pages 138 - 630 * |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114827198A (en) * | 2022-03-31 | 2022-07-29 | 电子科技大学 | Multilayer center asynchronous federal learning method applied to Internet of vehicles |
| CN114827198B (en) * | 2022-03-31 | 2023-04-28 | 电子科技大学 | Multi-layer center asynchronous federal learning method applied to Internet of vehicles |
| CN114745383A (en) * | 2022-04-08 | 2022-07-12 | 浙江金乙昌科技股份有限公司 | Mobile edge calculation assisted multilayer federal learning method |
| WO2023208043A1 (en) * | 2022-04-29 | 2023-11-02 | 索尼集团公司 | Electronic device and method for wireless communication system, and storage medium |
| CN115277689A (en) * | 2022-04-29 | 2022-11-01 | 国网天津市电力公司 | Yun Bianwang network communication optimization method and system based on distributed federal learning |
| CN115277689B (en) * | 2022-04-29 | 2023-09-22 | 国网天津市电力公司 | Cloud edge network communication optimization method and system based on distributed federal learning |
| CN115422820A (en) * | 2022-06-02 | 2022-12-02 | 国汽智控(北京)科技有限公司 | Federal learning model training method and road condition prediction method applied to road condition prediction |
| CN115422820B (en) * | 2022-06-02 | 2023-09-19 | 国汽智控(北京)科技有限公司 | Federal learning model training method applied to road condition prediction and road condition prediction method |
| CN115037618B (en) * | 2022-06-06 | 2023-11-07 | 电子科技大学 | Lightweight edge intelligent collaborative federal learning platform based on KubeEdge |
| CN115037618A (en) * | 2022-06-06 | 2022-09-09 | 电子科技大学 | Lightweight edge intelligent collaborative federal learning platform based on KubeEdge |
| CN115277446A (en) * | 2022-07-12 | 2022-11-01 | 中国信息通信研究院 | Energy-saving online internet connection learning network and method |
| CN114945044A (en) * | 2022-07-25 | 2022-08-26 | 北京智芯微电子科技有限公司 | Method, device and equipment for constructing digital twin platform based on federal learning |
| CN114945044B (en) * | 2022-07-25 | 2022-11-08 | 北京智芯微电子科技有限公司 | Method, device and equipment for constructing digital twin platform based on federal learning |
| CN115481752B (en) * | 2022-09-23 | 2024-03-19 | 中国电信股份有限公司 | Model training method, device, electronic equipment and storage medium |
| CN115481752A (en) * | 2022-09-23 | 2022-12-16 | 中国电信股份有限公司 | Model training method and device, electronic equipment and storage medium |
| CN115376031B (en) * | 2022-10-24 | 2023-03-24 | 江西省科学院能源研究所 | Road unmanned aerial vehicle routing inspection data processing method based on federal adaptive learning |
| CN115376031A (en) * | 2022-10-24 | 2022-11-22 | 江西省科学院能源研究所 | Road unmanned aerial vehicle routing inspection data processing method based on federal adaptive learning |
| CN116094993A (en) * | 2022-12-22 | 2023-05-09 | 电子科技大学 | Federal learning security aggregation method suitable for edge computing scene |
| CN116094993B (en) * | 2022-12-22 | 2024-05-31 | 电子科技大学 | Federal learning security aggregation method suitable for edge computing scene |
| CN115987694A (en) * | 2023-03-20 | 2023-04-18 | 杭州海康威视数字技术股份有限公司 | Equipment privacy protection method, system and device based on multi-domain federation |
| CN116546429A (en) * | 2023-06-06 | 2023-08-04 | 江南大学 | Vehicle selection method and system in federal learning of Internet of vehicles |
| CN116546429B (en) * | 2023-06-06 | 2024-01-16 | 杭州一诺科创信息技术有限公司 | Vehicle selection method and system in federal learning of Internet of vehicles |
| CN117076132A (en) * | 2023-10-12 | 2023-11-17 | 北京邮电大学 | Resource allocation and aggregation optimization method and device for hierarchical federal learning system |
| CN117076132B (en) * | 2023-10-12 | 2024-01-05 | 北京邮电大学 | Resource allocation and aggregation optimization method and device for hierarchical federal learning system |
| CN117812564A (en) * | 2024-02-29 | 2024-04-02 | 湘江实验室 | Federal learning method, device, equipment and medium applied to Internet of vehicles |
| CN117812564B (en) * | 2024-02-29 | 2024-05-31 | 湘江实验室 | Federal learning method, device, equipment and medium applied to Internet of vehicles |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114116198A (en) | Asynchronous federal learning method, system, equipment and terminal for mobile vehicle | |
| Elgendy et al. | Joint computation offloading and task caching for multi-user and multi-task MEC systems: reinforcement learning-based algorithms | |
| Lu et al. | Edge QoE: Computation offloading with deep reinforcement learning for Internet of Things | |
| Zhou et al. | Machine learning-based offloading strategy for lightweight user mobile edge computing tasks | |
| CN113435472A (en) | Vehicle-mounted computing power network user demand prediction method, system, device and medium | |
| Zhai et al. | Toward reinforcement-learning-based service deployment of 5G mobile edge computing with request-aware scheduling | |
| CN112395090B (en) | Intelligent hybrid optimization method for service placement in mobile edge calculation | |
| CN104995870A (en) | Multi-objective server placement determination | |
| Aral et al. | Staleness control for edge data analytics | |
| CN110968426A (en) | Edge cloud collaborative k-means clustering model optimization method based on online learning | |
| Xiao et al. | Collaborative cloud-edge service cognition framework for DNN configuration toward smart IIoT | |
| CN111488528A (en) | Content cache management method and device and electronic equipment | |
| Xiong et al. | A self-adaptive approach to service deployment under mobile edge computing for autonomous driving | |
| Qayyum et al. | Mobility-aware hierarchical fog computing framework for Industrial Internet of Things (IIoT) | |
| Qi et al. | An efficient pruning scheme of deep neural networks for Internet of Things applications | |
| Jin et al. | A real-time multimedia streaming transmission control mechanism based on edge cloud computing and opportunistic approximation optimization | |
| Qu et al. | Resource allocation for MEC system with multi-users resource competition based on deep reinforcement learning approach | |
| Li et al. | High-Precision Cluster Federated Learning for Smart Home: An Edge-Cloud Collaboration Approach | |
| EP4158556A1 (en) | Collaborative machine learning | |
| CN111770152A (en) | Edge data management method, medium, edge server and system | |
| Li et al. | Routing algorithm based on triangular fuzzy layer model and multi‐layer clustering for opportunistic network | |
| CN113992520B (en) | Virtual network resource deployment method and system | |
| Chen et al. | A vehicle-assisted computation offloading algorithm based on proximal policy optimization in vehicle edge networks | |
| WO2022253454A2 (en) | Dimensioning of telecommunication infrastructure | |
| Guo et al. | Edge Learning for Distributed Big Data Analytics: Theory, Algorithms, and System Design |
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 |