CN110442457A - Model training method, device and server based on federation's study - Google Patents
Model training method, device and server based on federation's study Download PDFInfo
- Publication number
- CN110442457A CN110442457A CN201910739086.2A CN201910739086A CN110442457A CN 110442457 A CN110442457 A CN 110442457A CN 201910739086 A CN201910739086 A CN 201910739086A CN 110442457 A CN110442457 A CN 110442457A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- local model
- training pattern
- precision
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- 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]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/94—Hardware or software architectures specially adapted for image or video understanding
- G06V10/95—Hardware or software architectures specially adapted for image or video understanding structured as a network, e.g. client-server architectures
Abstract
The embodiment of the present invention provides a kind of model training method, device and server based on federation's study.This method comprises: multiple working nodes will be sent to training pattern；Receive the local model of multiple working nodes feedback, local model is that each working node treats training pattern according to the data respectively possessed and is trained acquisition；The precision of each local model is determined according to test data set；The weight coefficient of each local model is determined according to the precision of each local model, weight coefficient and precision are positively correlated；According to multiple local models and corresponding weight coefficient, treats training pattern and be updated.The method of the embodiment of the present invention accelerates the convergence rate to training pattern, improves the precision to training pattern by increasing the weight coefficient of the local model of high-precision and reducing the weight coefficient of low precision local model.
Description
Technical field
The present embodiments relate to field of computer technology, and in particular to a kind of model training side based on federation's study
Method, device and server.
Background technique
On the one hand with the continuous reinforcement for constantly improve and monitoring of laws and regulations, the centralized processing of data will face
Huge legal risk；On the other hand it is unwilling between each data owning side for factors such as safety, economic interests shared former
Beginning data.These factors all can cause data to exist in the form of isolated island.In order to break data silo, federation's study meet the tendency of and
It is raw.
Federation's study shares initial data without each data owning side, can be under conditions of guaranteeing safety, fully
Model training is carried out using the initial data of each data owning side, efficiently solves the problems, such as the data silo in artificial intelligence epoch.
Currently based on the model training process of federation's study are as follows: multiple work sections will be issued to training pattern by server node
Point, each working node are trained to what is received to training pattern based on the data respectively possessed, and by trained mould
Shape parameter is back to server node, and server node treats training pattern according to the model parameter of each working node received
It is updated, then repeats the above process until meeting preset performance indicator to training pattern.
During the existing study progress model training based on federation, server node can not effectively identify working node
In low quality node, lead to the introducing of low quality model parameter, to reduce the convergence rate and model of model training
Precision.
Summary of the invention
The embodiment of the present invention provides a kind of model training method, device and server based on federation's study, to solve
The problem that convergence rate is slow and model accuracy is low present in the existing model training method based on federation's study.
In a first aspect, the embodiment of the present invention provides a kind of model training method based on federation's study, comprising:
Multiple working nodes will be sent to training pattern；
The local model of multiple working node feedbacks is received, local model is each working node according to the number respectively possessed
Acquisition is trained according to training pattern is treated；
The precision of each local model is determined according to test data set；
The weight coefficient of each local model, weight coefficient and precision positive are determined according to the precision of each local model
It closes；
According to multiple local models and corresponding weight coefficient, treats training pattern and be updated.
In one possible implementation, the method also includes:
The reward that each working node is determined according to the precision of each local model, so that each working node is according to reward
Participation is adjusted, reward is positively correlated with precision.
In one possible implementation, it will be sent to training pattern before multiple working nodes, the method is also
Include:
It is determined according to training mission to training pattern, and treats training pattern and initialized；
According to the network connection state and training mission between server node and working node, multiple work are determined
Node.
In one possible implementation, it treats after training pattern is updated, the method also includes:
Judge updated whether meet pre-set level to training pattern；
If satisfied, model training process is then terminated, conversely, then continuing repetitive exercise.
In one possible implementation, receiving the local model that multiple working nodes are fed back includes:
Receive the local model that multiple working nodes are fed back within a preset period of time.
In one possible implementation, the precision for determining each local model according to test data set includes:
Test verifying is carried out to each local model according to test data set, determines each local model in test data set
On accuracy rate and/or recall rate；
According to accuracy rate and/or recall rate of each local model in test data set, each local model is determined
Precision.
In one possible implementation, according to multiple local models and corresponding weight coefficient, trained mould is treated
Type is updated, comprising:
Training pattern is treated according to the following formula to be updated:
Wherein, M ' expression is updated to training pattern, and M indicates that, to training pattern, α is the weight system to training pattern
Number, LiIndicate i-th of local model, αiFor the weight coefficient of i-th of local model, N indicates the quantity of local model.
Second aspect, the embodiment of the present invention provide a kind of model training apparatus based on federation's study, comprising:
Sending module, for multiple working nodes will to be sent to training pattern；
Receiving module, for receiving the local model of multiple working nodes feedback, local model be each working node according to
The data respectively possessed treat training pattern and are trained acquisition；
Test module, for determining the precision of each local model according to test data set；
Processing module, for determining the weight coefficient of each local model, weight system according to the precision of each local model
It is several to be positively correlated with precision；
Update module, for treating training pattern and being updated according to multiple local models and corresponding weight coefficient.
The third aspect, the embodiment of the present invention provide a kind of server, comprising:
At least one processor and memory；
Memory stores computer executed instructions；
At least one processor executes the computer executed instructions of memory storage, so that at least one processor executes such as
The described in any item model training methods based on federation's study of first aspect.
Fourth aspect, the embodiment of the present invention provide a kind of computer readable storage medium, the computer-readable storage medium
Computer executed instructions are stored in matter, for realizing any one of such as first aspect when computer executed instructions are executed by processor
The model training method based on federation's study.
Model training method, device and server provided in an embodiment of the present invention based on federation's study, by will be wait instruct
Practice model and be sent to multiple working nodes, and receive the local model of multiple working node feedbacks, then according to test data set
The precision for determining each local model is each local model setting and the positively related weight coefficient of precision, finally according to multiple
Local model and corresponding weight coefficient, treat training pattern and are updated, and realize multiple working nodes for wait train
The coorinated training of model.Due to being not necessarily to shared data between each node, the peace of data be effectively ensured during model training
Full property and data-privacy；By the weight coefficient of increase high quality working node, the weight coefficient of low quality working node is reduced,
The convergence to training pattern is not only accelerated, but also can be improved the precision to training pattern；Further, model training process
Acceleration, also synchronize the workload for reducing each working node, reduce the communication overhead between node.
Detailed description of the invention
Fig. 1 is system architecture schematic diagram of the invention；
Fig. 2 is the flow chart of model training method one embodiment provided by the invention based on federation's study；
Fig. 3 is the flow chart of the model training method another embodiment provided by the invention based on federation's study；
Fig. 4 is the structural schematic diagram of model training apparatus one embodiment provided by the invention based on federation's study；
Fig. 5 is the structural schematic diagram of one embodiment of server provided by the invention.
Specific embodiment
Below by specific embodiment combination attached drawing, invention is further described in detail.Wherein different embodiments
Middle similar component uses associated similar element numbers.In the following embodiments, many datail descriptions be in order to
The application is better understood.However, those skilled in the art can recognize without lifting an eyebrow, part of feature
It is dispensed, or can be substituted by other elements, material, method in varied situations.In some cases, this Shen
Please it is relevant it is some operation there is no in the description show or describe, this is the core in order to avoid the application by mistake
More descriptions are flooded, and to those skilled in the art, these relevant operations, which are described in detail, not to be necessary, they
Relevant operation can be completely understood according to the general technology knowledge of description and this field in specification.
It is formed respectively in addition, feature described in this description, operation or feature can combine in any suitable way
Kind embodiment.Meanwhile each step in method description or movement can also can be aobvious and easy according to those skilled in the art institute
The mode carry out sequence exchange or adjustment seen.Therefore, the various sequences in the description and the appended drawings are intended merely to clearly describe a certain
A embodiment is not meant to be necessary sequence, and wherein some sequentially must comply with unless otherwise indicated.
It is herein component institute serialization number itself, such as " first ", " second " etc., is only used for distinguishing described object,
Without any sequence or art-recognized meanings.And " connection ", " connection " described in the application, unless otherwise instructed, include directly and
It is indirectly connected with (connection).
Fig. 1 is system architecture schematic diagram of the invention.As shown in Figure 1, server node 101 and working node 102 can be with
It is communicatively coupled by way of wiredly and/or wirelessly.Wherein, server node 101 includes but is not limited to single network clothes
Be engaged in device, multiple network servers composition server group or based on the consisting of a large number of computers or network servers of cloud computing
Cloud；Working node 102 includes but is not limited to computer, smart phone, tablet computer, personal digital assistant messaging device
Deng.
Fig. 1 embodies a kind of completely new distributed artificial intelligence network architecture system.Server node 101 and working node
It is not necessarily to shared initial data between 102 and between each working node, ensure that Information Security and data-privacy.Server
Node 101 is issued to working node 102 to training pattern, and each working node is instructed based on the local data respectively possessed
Practice, and return trained local model, server node 101 merges the local model of multiple working nodes and is polymerized to one
World model continues to issue, and is iterated training until world model restrains.
Below in conjunction with system architecture shown in FIG. 1, the model provided in an embodiment of the present invention based on federation's study is instructed
The some possible application scenarios for practicing method are illustrated.For example, there are first, second, the third three companies for managing video website,
The video recommendations model for thinking acquisition high quality, for recommending video to user, to improve user satisfaction.Each company is possessed
User data it is limited, the video recommendations model of high quality can not be trained based on the user data respectively possessed；In order to protect
The safety of user data is demonstrate,proved, privacy of user is protected, each company can not share the user data respectively possessed.This
When, each company can be by constructing respective working node, and by means of server node coorinated training video recommendations model.
It says, in order to obtain high-precision human face recognition model, can be possessed by server node Issuance model training mission for another example
The terminal device of face image data can be used as working node and participate in model training.Server node can obtain high quality
Human face recognition model, and terminal device can also calculate power by contribution and data obtain corresponding remuneration, realize two-win.
Since the data between server node 101 and working node 102 and between each working node are not shared, information
Asymmetry, and the parameter in the local model returned does not have actual physical meaning yet, therefore server node 101 can not ensure
The quality for the local model that working node 102 returns.It is difficult to avoid that and introduces low-quality local model, low quality local model
The convergence rate and precision that will affect world model are introduced, meanwhile, the reduction of convergence rate also will increase communication overhead, for whole
Very big negative effect can be generated for a system.
It will elaborate how the present invention solves the above problems by specific embodiment below.
Fig. 2 is the flow chart of model training method one embodiment provided by the invention based on federation's study.The present embodiment
The model training method based on federation's study provided can be applied to server node.As shown in Fig. 2, provided in this embodiment
May include: based on the federal model training method learnt
S201, multiple working nodes will be sent to training pattern.
It can include but is not limited to identification model, disaggregated model, detection model, prediction to training pattern in the present embodiment
Model etc., in specific implementation such as can be using artificial nerve network model, support vector machines, convolutional neural networks model,
The present embodiment to this with no restriction.Working node is the node for possessing training data and participating in model interoperability training, the present embodiment
The quantity of the middle working node for participating in model interoperability training is more than or equal to two.
It is understood that server node can be with root before it will be sent to multiple working nodes to training pattern
It determines according to training mission to training pattern.For example, if training mission is to carry out recognition of face, mould to be trained can be determined
Type is identification model；If training mission is to carry out risk profile, it can determine that training pattern be prediction model.It can be preparatory
It establishes training mission and to the mapping relations between training pattern, mould to be trained is determined according to training mission and mapping relations
Type.Server node, can be by building after determining to training pattern in advance between server node and each working node
Vertical network connection will be sent to each working node to training pattern.
S202, the local model for receiving multiple working nodes feedback, local model are each working nodes according to respectively being gathered around
Some data treat training pattern and are trained acquisition.
In the present embodiment when working node receive that server node issues to training pattern after, work can be used
Make the local data stored in node to be trained to what is received to training pattern, generates local model.It is received in local model
It holds back, alternatively, the local model with trained parameter to be passed through to the network pre-established and is connected when the training time reaches preset value
It takes back and reaches server node.The local data stored in each working node in the present embodiment is protected without sharing among the nodes
Safety and the privacy for having data in each working node by oneself are demonstrate,proved.
Optionally, in order to further increase the training speed of model, the local model of multiple working nodes feedbacks is received
A kind of possible implementation are as follows: receive the local model that multiple working nodes are fed back within a preset period of time.
In order to avoid server node waits working node to feed back for a long time in the present embodiment, the training speed of model is reduced
Degree, server node only receive the local model fed back within a preset period of time.Such as server node can issued wait instruct
Opening timing device can close receiving channel refusal and be further continued for receiving local mould after preset time arrival after practicing model
Type, or can directly be discarded in the local model received after preset time reaches.As an example it is assumed that there is A, B and C tri-
A local model, wherein A and B are reached within a preset period of time, and C time-out reaches, then server node is carrying out mould to be trained
When type updates, local model A and B are only considered.Preset time period in the present embodiment for example can be according to answering to training pattern
The factors such as miscellaneous degree, traffic rate between the average calculation power of each working node, node determine.
S203, the precision that each local model is determined according to test data set.
It, can be according to service in the present embodiment after server node receives the local model of working node feedback
Test data set in device node carries out test verifying to the local model received, with the precision of each local model of determination.
The precision of local model can embody the quality of local model, it is to be understood that the corresponding high-precision of the local model of high quality.
Test data set in the present embodiment includes the test data marked in advance, and each working node can not know test data set.
Unified test verifying is carried out by server node, it is possible to prevente effectively from each working node is practised fraud, guarantees test verifying
Accuracy and fairness.Specifically, such as the data of test data concentration can be inputted into local model, determines local model
Reality output, concentrate the desired output marked in advance to be compared the reality output of local model and test data, according to
Comparison result determines the precision of local model.
Optionally, a kind of possible implementation of the precision of each local model is determined according to test data set are as follows: root
Test verifying is carried out to each local model according to test data set, determines accuracy rate of each local model in test data set
And/or recall rate；According to accuracy rate and/or recall rate of each local model in test data set, each local mould is determined
The precision of type.Wherein, accuracy rate can be used for the correctness of measurement model, the correct sample number of accuracy rate=prediction/total sample
Number；Recall rate can be used for the covering surface of measurement model, the recall rate=prediction of positive class is positive the positive class number of the quantity of class/total
Amount.The precision of local model can perhaps the weighted average of recall rate or accuracy rate and recall rate be true according to accuracy rate
It is fixed.
Optionally, the test data set in the present embodiment can be determined according to training pattern.Such as it can pre-establish
It, can be according to the mapping after having determined to training pattern to the mapping relations between training pattern and test data set
Relationship determines corresponding test data set.
S204, the weight coefficient that each local model is determined according to the precision of each local model, weight coefficient and precision
It is positively correlated.
In the present embodiment after the precision for determining each local model, each local mould can be determined according to the precision
The weight coefficient of type.Wherein, the size of weight coefficient embodies the size that local model contributes entire model training process.
It is understood that introducing the local model of high quality, the convergence to training pattern can speed up, improve the essence to training pattern
Degree, is beneficial to entire model training process；And low-quality local model is introduced, have no not only for entire model training process
It helps, will increase the workload of other working nodes instead, increase the communication overhead of whole system.Therefore, each in the present embodiment
The weight coefficient of local model and its precision are positively correlated, i.e. the weight coefficient of the higher local model acquisition of precision is bigger.It is optional
, the local model of default precision threshold is lower than for precision, its weight coefficient 0 can be set as, may be brought to avoid it
Adverse effect.
By introducing attention mechanism in the present embodiment, increases the weight coefficient of high quality working node, reduce low quality
The weight coefficient of working node is conducive to acceleration model convergence, improves training speed.
S205, according to multiple local models and corresponding weight coefficient, treat training pattern and be updated.
Server node can treat training after the weight coefficient for determining each local model accordingly in the present embodiment
Model is updated.For example, can treat training pattern using the weighted average of multiple local models and be updated.
Model training method provided in this embodiment based on federation's study, by the way that multiple works will be sent to training pattern
Make node, and receive the local model of multiple working node feedbacks, each local model is then determined according to test data set
Precision, for each local model setting and the positively related weight coefficient of precision, finally according to multiple local models and corresponding
Weight coefficient is treated training pattern and is updated, and realizes multiple working nodes for the coorinated training to training pattern.Due to
During model training, it is not necessarily to shared data between each node, safety and the data-privacy of data has been effectively ensured；Pass through
Increase the weight coefficient of high quality working node, reduces the weight coefficient of low quality working node, not only accelerate mould to be trained
The convergence of type, and can be improved the precision to training pattern；Further, the acceleration of model training process, it is also synchronous to reduce
The workload of each working node, reduces the communication overhead between node.
In order to transfer the enthusiasm of each working node, so that it can play an active part in model training, it will usually to each work
Node gives excitation appropriate.Currently, being broadly divided into two classes to the incentive mechanism of working node: the incentive mechanism based on game theory
With the incentive mechanism based on contract theory.Prize is distributed based on the contribution of the incentive mechanism Equilibrium Game working node of game theory
It encourages, is mostly that total system is contributed to stimulate working node.In the mechanism, the work quality or work of evaluation work node
The contribution that node is done by system has following two mode: being overall model institute according to after a working node participation model training
The decline of overall model income, is commented caused by the promotion of bring income or a working node exit model training
Estimate.Specifically, can be according to indexs such as operation time, communication overhead and model accuracies come comprehensive descision, and then determine work
The work quality of node.Incentive mechanism based on contract theory be substantially the incentive mechanism based on game theory a kind of derivative and
Variation can preset the specification that a plurality of types of contracts sign reward, and the distributed of information asymmetry that be more adaptive to motivates
Situation.Server node can set up a revenue function, with lift scheme precision and receipts as far as possible under the paying that contract limits
Speed is held back to reduce expense, strives for the maximum return of oneself；Working node can set up utility function, in data quality and
Strive for high-grade contract as far as possible under conditions of computing capability to improve the income of oneself.Both sides are in the state of a game
Under, mutual game is to reach balance.
Above-mentioned incentive mechanism is difficult to avoid that malicious node gains reward by cheating, and the node that can not effectively prevent doing evil is to model training mistake
The influence of journey.By taking the incentive mechanism based on contract theory as an example, it is assumed that server node has been signed high-grade with working node
Contract, server node not can guarantee the data of working node and calculate power and do not play tricks.For working node, if intentionally
It does evil or in order to gain excitation by cheating, model can be returned without model training and directly completely, or do not have in local model
Model is returned under convergent state, or directly makes up parameter passback model.Such passback model is low-quality model, right
What model convergence not only help without, will increase the workload of other working nodes instead.
In order to avoid there is the phenomenon that node gains reward by cheating, while in order to transfer the enthusiasm of each node, in above-mentioned implementation
On the basis of example, method provided in this embodiment can also include: to determine each work section according to the precision of each local model
The reward of point, so that each working node adjusts participation according to reward, reward is positively correlated with precision.
Reward in the present embodiment for example can be Permission Levels, calculating money of each working node using trained model
Source, storage resource etc..Can according to specific application scenarios determine, the present embodiment for reward concrete form with no restriction.
Server node, can be according to the precision using clothes after the precision for determining each local model in the present embodiment
The income of the utility function evaluating server node of business device node, is distributed according to the income of server node to each working node
Reward.Each working node is assessed the income of each working node according to respective utility function, is adjusted accordingly after being rewarded
Participation, specifically, data volume, computing resource, storage resource, communication used in adjustable participation model training process
Resource etc..
It is understood that high-precision local model can make server node obtain higher income, therefore in order to
The working node of high quality is attracted to participate in model training process, it can be to the work of the local model of offer high-precision in the present embodiment
The more rewards of node distribution, i.e., precision is higher, and the reward that can be obtained is also more.The reward of great number can promote working node product
The training of the participation model of pole, contributes higher calculation power and more data, provides high-precision local.Optionally, for essence
Lower than the local model for presetting precision threshold, server node can be refused to distribute to the working node for providing the local model degree
Reward.Therefore, it for the working node for making up parameter passback model, possibly can not be rewarded, or be only capable of obtaining very
Small reward can effectively reduce the phenomenon that working node gains reward by cheating.
It is provided in this embodiment based on federation study model training method, on the basis of the above embodiments, pass through to
The positively related reward of precision for the local model that each working node distribution is provided with it, the working node for promoting high quality are positive
The training process for participating in model, alleviates the phenomenon that malicious node gains reward by cheating, has further speeded up the convergence to training pattern
And improve precision to training pattern.
Fig. 3 is the flow chart of the model training method another embodiment provided by the invention based on federation's study.Such as Fig. 3 institute
Show, method provided in this embodiment may include:
S301, it is determined according to training mission to training pattern, and treats training pattern and initialized.
Determine that the specific implementation process to training pattern can refer to above-described embodiment according to training mission, it is no longer superfluous herein
It states.After determining to training pattern, server node can treat training pattern and be initialized.Such as training can be treated
Model carries out random initializtion, is random value by the parameter initialization to training pattern；Alternatively, can will be initial to training pattern
Turn to pre-stored model in server node.
It should be noted that server node is only treated training pattern and is carried out once during being iterated training
Initialization.
S302, according to the network connection state and training mission between server node and working node, determine multiple
Working node.
It, can be according to server node and working node in order to improve the quality for the working node for participating in model training process
Between network connection state, training mission, the precision of local model that provides in history iterative process etc. to working node into
Row screening.It for example, can be from if during being determined according to training mission 3 working nodes being needed to participate in model trainings
3 working nodes are randomly selected in good 4 working nodes of network connection state and participate in model training, or can be according to each
The history of working node shows, and higher 3 working nodes of quality are chosen from good 4 working nodes of network connection state
Participate in model training.
It should be noted that can be chosen to working node during every wheel repetitive exercise.Namely
It says, the node that model training is participated in during repetitive exercise can not be identical.For example, in first round training process, work
The network connection state made between node W and server node is good, the selected participation model training of working node W, however
During second wheel training, the network connection state between working node W and server node deteriorates, and can not efficiently be led to
Letter, working node W will be removed.It is possible to prevente effectively from the introducing of low quality working node.
S303, multiple working nodes will be sent to training pattern.
Specific implementation process can be with reference to the step S201 in above-described embodiment, and details are not described herein again.
S304, the local model that multiple working nodes are fed back within a preset period of time is received.
Specific implementation process can refer to above-described embodiment, and details are not described herein again.
S305, the precision that each local model is determined according to test data set.
Specific implementation process can be with reference to the step S203 in above-described embodiment, and details are not described herein again.
It should be noted that improving the generalization ability to training pattern in order to avoid there is over-fitting, changing
During generation training, test data set can be some or all of not identical.For example, if being stored in server node
10000 test datas marked in advance can be selected at random during every wheel repetitive exercise from 10000 test datas
It takes, or chooses 100 formation test data sets according to preset rules, for being verified to local model.
S306, the weight coefficient that each local model is determined according to the precision of each local model, weight coefficient and precision
It is positively correlated.
Specific implementation process can be with reference to the step S204 in above-described embodiment, and details are not described herein again.
S307, according to multiple local models and corresponding weight coefficient, treat training pattern and be updated.
Optionally, according to multiple local models and corresponding weight coefficient, one kind that training pattern is updated is treated
Possible implementation are as follows:
Training pattern is treated according to the following formula to be updated:
Wherein, M ' expression is updated to training pattern, and M indicates that, to training pattern, α is the weight system to training pattern
Number, α ∈ [0,1], LiIndicate i-th of local model, αiFor the weight coefficient of i-th of local model, αi∈ [0,1], N indicate this
The quantity of ground model,
It should be noted that at repetitive exercise initial stage, due to unstable to training pattern M, for the training of acceleration model
Process can reduce α, increase αi, such as using random fashion initialized to training pattern, α can be set as
0；And in the repetitive exercise later period, since M has tended to convergence state, in order to improve the stability of training process, α can be increased,
Reduce αi。
S308, judge it is updated whether meet pre-set level to training pattern, if satisfied, then terminating model training mistake
Journey continues repetitive exercise if not satisfied, thening follow the steps S302.
Whether whether the pre-set level in the present embodiment can be for one or more of following index: restraining, reach
Whether maximum number of iterations reaches the maximum training time etc..Various indexs can be combined and be judged, such as can sentenced first
Break it is updated whether restrained to training pattern, if convergence if terminate model training process, otherwise continue to determine whether to have reached
To maximum number of iterations, model training process is terminated if reaching, and otherwise continues to determine whether to have reached the maximum the training time,
Model training process is terminated if reaching, no person executes step S302, continues repetitive exercise.
Model training method provided in this embodiment based on federation's study, on the basis of the above embodiments, further
Improve training speed and model accuracy.
The embodiment of the present invention also provides a kind of model training apparatus based on federation's study, shown in Figure 4, the present invention
Embodiment is only illustrated by taking Fig. 4 as an example, is not offered as that present invention is limited only to this.Fig. 4 is provided by the invention based on federal
The structural schematic diagram of one embodiment of model training apparatus of habit.As shown in figure 4, the mould provided in this embodiment based on federation's study
Type training device 40 may include: sending module 401, receiving module 402, test module 403, processing module 404 and update mould
Block 405.
Sending module 401, for multiple working nodes will to be sent to training pattern；
Receiving module 402, for receiving the local model of multiple working node feedbacks, local model is each working node root
Training pattern, which is treated, according to the data respectively possessed is trained acquisition；
Test module 403, for determining the precision of each local model according to test data set；
Processing module 404, for determining the weight coefficient of each local model, weight according to the precision of each local model
Coefficient and precision are positively correlated；
Update module 405, for treating training pattern and carrying out more according to multiple local models and corresponding weight coefficient
Newly.
The device of the present embodiment can be used for executing the technical solution of embodiment of the method shown in Fig. 2, realization principle and skill
Art effect is similar, and details are not described herein again.
Optionally, the model training apparatus 40 based on federation's study can also include distribution module (not shown), tool
Body is used to determine the reward of each working node according to the precision of each local model, so that each working node is adjusted according to reward
Whole participation, reward are positively correlated with precision.
Optionally, the model training apparatus 40 based on federation's study can also include initialization module (not shown),
Specifically for before it will be sent to multiple working nodes to training pattern, determined according to training mission to training pattern, and right
It is initialized to training pattern；According to the network connection state and training mission between server node and working node,
Determine multiple working nodes.
Optionally, the model training apparatus 40 based on federation's study can also include judgment module (not shown), tool
Body is used for after treating training pattern and being updated, and judges updated whether meet pre-set level to training pattern；If full
Foot, then terminate model training process, conversely, then continuing repetitive exercise.
Optionally, receiving module 402 specifically can be used for, and receive the sheet that multiple working nodes are fed back within a preset period of time
Ground model.
Optionally, test module 403 specifically can be used for, and carries out test to each local model according to test data set and tests
Card determines accuracy rate and/or recall rate of each local model in test data set；According to each local model in test number
According to the accuracy rate and/or recall rate on collection, the precision of each local model is determined.
Optionally, update module 405 specifically can be used for, and treats training pattern according to the following formula and is updated:
Wherein, M ' expression is updated to training pattern, and M indicates that, to training pattern, α is the weight system to training pattern
Number, LiIndicate i-th of local model, αiFor the weight coefficient of i-th of local model, N indicates the quantity of local model.
The embodiment of the present invention also provides a kind of server, shown in Figure 5, the embodiment of the present invention only by taking Fig. 5 as an example into
Row explanation, is not offered as that present invention is limited only to this.Fig. 5 is the structural schematic diagram of one embodiment of server provided by the invention.Such as
Shown in Fig. 5, server 50 provided in this embodiment may include: memory 501, processor 502 and bus 503.Wherein, bus
503 for realizing the connection between each element.
Computer program is stored in memory 501, computer program may be implemented above-mentioned when being executed by processor 502
The technical solution that one embodiment of the method provides.
Wherein, be directly or indirectly electrically connected between memory 501 and processor 502, with realize data transmission or
Interaction.It is electrically connected for example, these elements can be realized between each other by one or more of communication bus or signal wire, such as
It can be connected by bus 503.The computer journey for realizing the model training method based on federation's study is stored in memory 501
Sequence, the software function module that can be stored in the form of software or firmware including at least one in memory 501, processor 502
By running the software program and module that are stored in memory 501, thereby executing various function application and data processing.
Memory 501 may be, but not limited to, random access memory (Random Access Memory, referred to as:
RAM), read-only memory (Read Only Memory, referred to as: ROM), programmable read only memory (Programmable
Read-Only Memory, referred to as: PROM), erasable read-only memory (Erasable Programmable Read-Only
Memory, referred to as: EPROM), electricallyerasable ROM (EEROM) (Electric Erasable Programmable Read-
Only Memory, referred to as: EEPROM) etc..Wherein, memory 501 is for storing program, and processor 502 refers to receiving execution
After order, program is executed.Further, the software program in above-mentioned memory 501 and module may also include operating system, can
Including the various component softwares for management system task (such as memory management, storage equipment control, power management etc.) and/or
Driving, and can be in communication with each other with various hardware or component software, to provide the running environment of other software component.
Processor 502 can be a kind of IC chip, the processing capacity with signal.Above-mentioned processor 502 can
To be general processor, including central processing unit (Central Processing Unit, referred to as: CPU), network processing unit
(Network Processor, referred to as: NP) etc..It may be implemented or execute disclosed each method, the step in the embodiment of the present invention
Rapid and logic diagram.General processor can be microprocessor or the processor is also possible to any conventional processor etc..
It is appreciated that Fig. 5 structure be only illustrate, can also include than shown in Fig. 5 more perhaps less component or have with
Different configuration shown in Fig. 5.Each component shown in Fig. 5 can use hardware and/or software realization.
It should be noted that server provided in this embodiment includes but is not limited to single network server, multiple networks
The server group that server forms or the cloud consisting of a large number of computers or network servers based on cloud computing, wherein cloud meter
It is one kind of distributed computing, a super virtual computer being made of the computer of a group loose couplings.
The embodiment of the present invention also provides a kind of computer readable storage medium, is stored thereon with computer program, computer
The model training method based on federation's study that any of the above-described embodiment of the method provides may be implemented when program is executed by processor.
Computer readable storage medium in the present embodiment can be any usable medium that computer can access, or include one
Data storage devices, the usable mediums such as a or multiple usable mediums integrated server, data center can be magnetic medium,
(for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or semiconductor medium (such as SSD) etc..
It will be understood by those skilled in the art that all or part of function of various methods can pass through in above embodiment
The mode of hardware is realized, can also be realized by way of computer program.When function all or part of in above embodiment
When being realized by way of computer program, which be can be stored in a computer readable storage medium, and storage medium can
To include: read-only memory, random access memory, disk, CD, hard disk etc., it is above-mentioned to realize which is executed by computer
Function.For example, program is stored in the memory of equipment, when executing program in memory by processor, can be realized
State all or part of function.In addition, when function all or part of in above embodiment is realized by way of computer program
When, which also can store in storage mediums such as server, another computer, disk, CD, flash disk or mobile hard disks
In, through downloading or copying and saving into the memory of local device, or version updating is carried out to the system of local device, when logical
When crossing the program in processor execution memory, all or part of function in above embodiment can be realized.
Use above specific case is illustrated the present invention, is merely used to help understand the present invention, not to limit
The system present invention.For those skilled in the art, according to the thought of the present invention, can also make several simple
It deduces, deform or replaces.
Claims (10)
1. a kind of model training method based on federation's study, is applied to server node characterized by comprising
Multiple working nodes will be sent to training pattern；
Receive the local model of multiple working nodes feedbacks, the local model is each working node according to respectively being possessed
Data be trained acquisition to training pattern to described；
The precision of each local model is determined according to test data set；
The weight coefficient of each local model, the weight coefficient and institute are determined according to the precision of each local model
State precision positive correlation；
According to multiple local models and corresponding weight coefficient, it is updated to described to training pattern.
2. the method as described in claim 1, which is characterized in that the method also includes:
The reward that each working node is determined according to the precision of each local model, so that each working node
Participation is adjusted according to the reward, the reward is positively correlated with the precision.
3. the method as described in claim 1, which is characterized in that it is described will be sent to training pattern multiple working nodes it
Before, the method also includes:
It is initialized to training pattern, and to described to training pattern according to training mission determination is described；
According to the network connection state and the training mission between the server node and working node, determine described in
Multiple working nodes.
4. the method as described in claim 1, which is characterized in that it is described to it is described be updated to training pattern after, it is described
Method further include:
Judge updated whether meet pre-set level to training pattern；
If satisfied, model training process is then terminated, conversely, then continuing repetitive exercise.
5. the method as described in claim 1, which is characterized in that the local model for receiving multiple working node feedbacks
Include:
Receive the local model that multiple working nodes are fed back within a preset period of time.
6. the method according to claim 1 to 5, which is characterized in that it is described determined according to test data set it is each described
The precision of local model includes:
Test verifying is carried out to each local model according to test data set, determines each local model in the survey
Try the accuracy rate and/or recall rate on data set；
According to accuracy rate and/or recall rate of each local model in the test data set, each described is determined
The precision of ground model.
7. the method according to claim 1 to 5, which is characterized in that described according to multiple local models and right
The weight coefficient answered is updated to described to training pattern, comprising:
It is updated according to the following formula to described to training pattern:
Wherein, M ' expression is updated to training pattern, and M indicates described to training pattern, and α is the weight to training pattern
Coefficient, LiIndicate i-th of local model, αiFor the weight coefficient of i-th of local model, N indicates the quantity of local model.
8. a kind of model training apparatus based on federation's study characterized by comprising
Sending module, for multiple working nodes will to be sent to training pattern；
Receiving module, for receiving the local model of multiple working node feedbacks, the local model is each working node
Acquisition is trained to training pattern to described according to the data respectively possessed；
Test module, for determining the precision of each local model according to test data set；
Processing module, for determining the weight coefficient of each local model, institute according to the precision of each local model
It states weight coefficient and the precision is positively correlated；
Update module, for being carried out to training pattern to described according to multiple local models and corresponding weight coefficient
It updates.
9. a kind of server characterized by comprising at least one processor and memory；
The memory stores computer executed instructions；
At least one described processor executes the computer executed instructions of the memory storage, so that at least one described processing
Device is executed such as the described in any item model training methods based on federation's study of claim 1-7.
10. a kind of computer readable storage medium, which is characterized in that be stored with computer in the computer readable storage medium
It executes instruction, for realizing such as described in any item bases of claim 1-7 when the computer executed instructions are executed by processor
In the model training method of federation's study.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910739086.2A CN110442457A (en) | 2019-08-12 | 2019-08-12 | Model training method, device and server based on federation's study |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910739086.2A CN110442457A (en) | 2019-08-12 | 2019-08-12 | Model training method, device and server based on federation's study |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110442457A true CN110442457A (en) | 2019-11-12 |
Family
ID=68434570
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910739086.2A Pending CN110442457A (en) | 2019-08-12 | 2019-08-12 | Model training method, device and server based on federation's study |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110442457A (en) |
Cited By (54)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110956275A (en) * | 2019-11-27 | 2020-04-03 | 支付宝(杭州)信息技术有限公司 | Risk prediction and risk prediction model training method and device and electronic equipment |
| CN110990870A (en) * | 2019-11-29 | 2020-04-10 | 上海能塔智能科技有限公司 | Operation and maintenance, processing method, device, equipment and medium using model library |
| CN111027715A (en) * | 2019-12-11 | 2020-04-17 | 支付宝(杭州)信息技术有限公司 | Monte Carlo-based federated learning model training method and device |
| CN111046433A (en) * | 2019-12-13 | 2020-04-21 | 支付宝(杭州)信息技术有限公司 | Model training method based on federal learning |
| CN111091200A (en) * | 2019-12-20 | 2020-05-01 | 深圳前海微众银行股份有限公司 | Updating method, system, agent, server and storage medium of training model |
| CN111260061A (en) * | 2020-03-09 | 2020-06-09 | 厦门大学 | Differential noise adding method and system in federated learning gradient exchange |
| CN111369042A (en) * | 2020-02-27 | 2020-07-03 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
| CN111383096A (en) * | 2020-03-23 | 2020-07-07 | 中国建设银行股份有限公司 | Fraud detection and model training method and device thereof, electronic equipment and storage medium |
| CN111460528A (en) * | 2020-04-01 | 2020-07-28 | 支付宝(杭州)信息技术有限公司 | Multi-party combined training method and system based on Adam optimization algorithm |
| CN111488995A (en) * | 2020-04-08 | 2020-08-04 | 北京字节跳动网络技术有限公司 | Method and apparatus for evaluating a joint training model |
| CN111678696A (en) * | 2020-06-17 | 2020-09-18 | 南昌航空大学 | Intelligent mechanical fault diagnosis method based on federal learning |
| CN111695701A (en) * | 2020-06-12 | 2020-09-22 | 上海富数科技有限公司 | System for realizing data set construction processing based on federal learning and construction generation method thereof |
| CN111814985A (en) * | 2020-06-30 | 2020-10-23 | 平安科技（深圳）有限公司 | Model training method under federated learning network and related equipment thereof |
| CN111882133A (en) * | 2020-08-03 | 2020-11-03 | 重庆大学 | Prediction-based federated learning communication optimization method and system |
| CN111881966A (en) * | 2020-07-20 | 2020-11-03 | 北京市商汤科技开发有限公司 | Neural network training method, device, equipment and storage medium |
| CN111898764A (en) * | 2020-06-23 | 2020-11-06 | 华为技术有限公司 | Method, device and chip for federal learning |
| CN111931876A (en) * | 2020-10-12 | 2020-11-13 | 支付宝(杭州)信息技术有限公司 | Target data side screening method and system for distributed model training |
| CN111931242A (en) * | 2020-09-30 | 2020-11-13 | 国网浙江省电力有限公司电力科学研究院 | Data sharing method, computer equipment applying same and readable storage medium |
| CN112001500A (en) * | 2020-08-13 | 2020-11-27 | 星环信息科技（上海）有限公司 | Model training method, device and storage medium based on longitudinal federated learning system |
| CN112101528A (en) * | 2020-09-17 | 2020-12-18 | 上海交通大学 | Terminal contribution measurement method based on back propagation |
| CN112232519A (en) * | 2020-10-15 | 2021-01-15 | 成都数融科技有限公司 | Joint modeling method based on federal learning |
| KR20210065069A (en) * | 2019-11-22 | 2021-06-03 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | How to create a shared encoder, device and electronics |
| CN112949837A (en) * | 2021-04-13 | 2021-06-11 | 中国人民武装警察部队警官学院 | Target recognition federal deep learning method based on trusted network |
| WO2021120677A1 (en) * | 2020-07-07 | 2021-06-24 | 平安科技（深圳）有限公司 | Warehousing model training method and device, computer device and storage medium |
| WO2021138877A1 (en) * | 2020-01-09 | 2021-07-15 | 深圳前海微众银行股份有限公司 | Vertical federated learning model training optimization method and apparatus, device, and medium |
| CN113128528A (en) * | 2019-12-27 | 2021-07-16 | 无锡祥生医疗科技股份有限公司 | Ultrasonic image deep learning distributed training system and training method |
| CN113158550A (en) * | 2021-03-24 | 2021-07-23 | 北京邮电大学 | Method and device for federated learning, electronic equipment and storage medium |
| CN113191479A (en) * | 2020-01-14 | 2021-07-30 | 华为技术有限公司 | Method, system, node and storage medium for joint learning |
| CN113222211A (en) * | 2021-03-31 | 2021-08-06 | 中国科学技术大学先进技术研究院 | Multi-region diesel vehicle pollutant emission factor prediction method and system |
| CN113239351A (en) * | 2020-12-08 | 2021-08-10 | 武汉大学 | Novel data pollution attack defense method for Internet of things system |
| WO2021164404A1 (en) * | 2020-02-20 | 2021-08-26 | 中国银联股份有限公司 | Inspection method and apparatus |
| CN113312169A (en) * | 2020-02-27 | 2021-08-27 | 香港理工大学深圳研究院 | Method and device for distributing computing resources |
| CN113377797A (en) * | 2021-07-02 | 2021-09-10 | 支付宝(杭州)信息技术有限公司 | Method, device and system for jointly updating model |
| WO2021179196A1 (en) * | 2020-03-11 | 2021-09-16 | Oppo广东移动通信有限公司 | Federated learning-based model training method, electronic device, and storage medium |
| CN113516250A (en) * | 2021-07-13 | 2021-10-19 | 北京百度网讯科技有限公司 | Method, device and equipment for federated learning and storage medium |
| CN113537633A (en) * | 2021-08-09 | 2021-10-22 | 中国电信股份有限公司 | Prediction method, device, equipment, medium and system based on longitudinal federal learning |
| CN113688855A (en) * | 2020-05-19 | 2021-11-23 | 华为技术有限公司 | Data processing method, federal learning training method, related device and equipment |
| CN113705823A (en) * | 2020-05-22 | 2021-11-26 | 华为技术有限公司 | Model training method based on federal learning and electronic equipment |
| CN113792632A (en) * | 2021-09-02 | 2021-12-14 | 广州广电运通金融电子股份有限公司 | Finger vein identification method, system and storage medium based on multi-party cooperation |
| CN113807157A (en) * | 2020-11-27 | 2021-12-17 | 京东科技控股股份有限公司 | Method, device and system for training neural network model based on federal learning |
| CN113837766A (en) * | 2021-10-08 | 2021-12-24 | 支付宝(杭州)信息技术有限公司 | Risk identification method and device and electronic equipment |
| WO2022057433A1 (en) * | 2020-09-18 | 2022-03-24 | 华为技术有限公司 | Machine learning model training method and related device |
| WO2022096919A1 (en) * | 2020-11-05 | 2022-05-12 | Telefonaktiebolaget Lm Ericsson (Publ) | Managing training of a machine learning model |
| CN114548426A (en) * | 2022-02-17 | 2022-05-27 | 北京百度网讯科技有限公司 | Asynchronous federal learning method, business service prediction method, device and system |
| WO2022110248A1 (en) * | 2020-11-30 | 2022-06-02 | 华为技术有限公司 | Federated learning method, device and system |
| WO2022121032A1 (en) * | 2020-12-10 | 2022-06-16 | 广州广电运通金融电子股份有限公司 | Data set division method and system in federated learning scene |
| CN114662340A (en) * | 2022-04-29 | 2022-06-24 | 烟台创迹软件有限公司 | Weighing model scheme determination method and device, computer equipment and storage medium |
| CN114679332A (en) * | 2022-04-14 | 2022-06-28 | 浙江工业大学 | APT detection method of distributed system |
| CN114741611A (en) * | 2022-06-08 | 2022-07-12 | 杭州金智塔科技有限公司 | Federal recommendation model training method and system |
| CN114844653A (en) * | 2022-07-04 | 2022-08-02 | 湖南密码工程研究中心有限公司 | Credible federal learning method based on alliance chain |
| CN115196730A (en) * | 2022-07-19 | 2022-10-18 | 南通派菲克水务技术有限公司 | Intelligent sodium hypochlorite adding system for water plant |
| WO2023030221A1 (en) * | 2021-08-30 | 2023-03-09 | 华为技术有限公司 | Communication method and apparatus |
| CN116127417A (en) * | 2023-04-04 | 2023-05-16 | 山东浪潮科学研究院有限公司 | Code defect detection model construction method, device, equipment and storage medium |
| CN116245172A (en) * | 2023-03-14 | 2023-06-09 | 南京航空航天大学 | Coalition building game method facing individual model performance optimization in cross-island federal learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108898219A (en) * | 2018-06-07 | 2018-11-27 | 广东工业大学 | A kind of neural network training method based on block chain, device and medium |
| CN109189825A (en) * | 2018-08-10 | 2019-01-11 | 深圳前海微众银行股份有限公司 | Lateral data cutting federation learning model building method, server and medium |
| CN109635462A (en) * | 2018-12-17 | 2019-04-16 | 深圳前海微众银行股份有限公司 | Model parameter training method, device, equipment and medium based on federation's study |
| CN109871702A (en) * | 2019-02-18 | 2019-06-11 | 深圳前海微众银行股份有限公司 | Federal model training method, system, equipment and computer readable storage medium |
| US20190227980A1 (en) * | 2018-01-22 | 2019-07-25 | Google Llc | Training User-Level Differentially Private Machine-Learned Models |
-
2019
- 2019-08-12 CN CN201910739086.2A patent/CN110442457A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190227980A1 (en) * | 2018-01-22 | 2019-07-25 | Google Llc | Training User-Level Differentially Private Machine-Learned Models |
| CN108898219A (en) * | 2018-06-07 | 2018-11-27 | 广东工业大学 | A kind of neural network training method based on block chain, device and medium |
| CN109189825A (en) * | 2018-08-10 | 2019-01-11 | 深圳前海微众银行股份有限公司 | Lateral data cutting federation learning model building method, server and medium |
| CN109635462A (en) * | 2018-12-17 | 2019-04-16 | 深圳前海微众银行股份有限公司 | Model parameter training method, device, equipment and medium based on federation's study |
| CN109871702A (en) * | 2019-02-18 | 2019-06-11 | 深圳前海微众银行股份有限公司 | Federal model training method, system, equipment and computer readable storage medium |
Cited By (84)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2022512053A (en) * | 2019-11-22 | 2022-02-02 | ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド | Shared encoder generation methods, devices and electronic devices |
| JP7159460B2 (en) | 2019-11-22 | 2022-10-24 | ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド | Shared encoder generation method, apparatus and electronic equipment |
| KR102532368B1 (en) * | 2019-11-22 | 2023-05-15 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | Shared encoder generation method, device and electronic device |
| KR20210065069A (en) * | 2019-11-22 | 2021-06-03 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | How to create a shared encoder, device and electronics |
| CN110956275B (en) * | 2019-11-27 | 2021-04-02 | 支付宝(杭州)信息技术有限公司 | Risk prediction and risk prediction model training method and device and electronic equipment |
| CN110956275A (en) * | 2019-11-27 | 2020-04-03 | 支付宝(杭州)信息技术有限公司 | Risk prediction and risk prediction model training method and device and electronic equipment |
| CN110990870A (en) * | 2019-11-29 | 2020-04-10 | 上海能塔智能科技有限公司 | Operation and maintenance, processing method, device, equipment and medium using model library |
| CN111027715B (en) * | 2019-12-11 | 2021-04-02 | 支付宝(杭州)信息技术有限公司 | Monte Carlo-based federated learning model training method and device |
| CN111027715A (en) * | 2019-12-11 | 2020-04-17 | 支付宝(杭州)信息技术有限公司 | Monte Carlo-based federated learning model training method and device |
| CN111046433A (en) * | 2019-12-13 | 2020-04-21 | 支付宝(杭州)信息技术有限公司 | Model training method based on federal learning |
| WO2021121029A1 (en) * | 2019-12-20 | 2021-06-24 | 深圳前海微众银行股份有限公司 | Training model updating method and system, and agent, server and computer-readable storage medium |
| CN111091200A (en) * | 2019-12-20 | 2020-05-01 | 深圳前海微众银行股份有限公司 | Updating method, system, agent, server and storage medium of training model |
| CN113128528A (en) * | 2019-12-27 | 2021-07-16 | 无锡祥生医疗科技股份有限公司 | Ultrasonic image deep learning distributed training system and training method |
| WO2021138877A1 (en) * | 2020-01-09 | 2021-07-15 | 深圳前海微众银行股份有限公司 | Vertical federated learning model training optimization method and apparatus, device, and medium |
| CN113191479A (en) * | 2020-01-14 | 2021-07-30 | 华为技术有限公司 | Method, system, node and storage medium for joint learning |
| TWI770749B (en) * | 2020-02-20 | 2022-07-11 | 大陸商中國銀聯股份有限公司 | Inspection method and device |
| WO2021164404A1 (en) * | 2020-02-20 | 2021-08-26 | 中国银联股份有限公司 | Inspection method and apparatus |
| CN111369042B (en) * | 2020-02-27 | 2021-09-24 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
| CN113312169B (en) * | 2020-02-27 | 2023-12-19 | 香港理工大学深圳研究院 | Computing resource allocation method and device |
| WO2021169577A1 (en) * | 2020-02-27 | 2021-09-02 | 山东大学 | Wireless service traffic prediction method based on weighted federated learning |
| CN113312169A (en) * | 2020-02-27 | 2021-08-27 | 香港理工大学深圳研究院 | Method and device for distributing computing resources |
| CN111369042A (en) * | 2020-02-27 | 2020-07-03 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
| CN111260061A (en) * | 2020-03-09 | 2020-06-09 | 厦门大学 | Differential noise adding method and system in federated learning gradient exchange |
| CN111260061B (en) * | 2020-03-09 | 2022-07-19 | 厦门大学 | Differential noise adding method and system in federated learning gradient exchange |
| WO2021179196A1 (en) * | 2020-03-11 | 2021-09-16 | Oppo广东移动通信有限公司 | Federated learning-based model training method, electronic device, and storage medium |
| CN111383096A (en) * | 2020-03-23 | 2020-07-07 | 中国建设银行股份有限公司 | Fraud detection and model training method and device thereof, electronic equipment and storage medium |
| CN111460528B (en) * | 2020-04-01 | 2022-06-14 | 支付宝(杭州)信息技术有限公司 | Multi-party combined training method and system based on Adam optimization algorithm |
| CN111460528A (en) * | 2020-04-01 | 2020-07-28 | 支付宝(杭州)信息技术有限公司 | Multi-party combined training method and system based on Adam optimization algorithm |
| WO2021203919A1 (en) * | 2020-04-08 | 2021-10-14 | 北京字节跳动网络技术有限公司 | Method and apparatus for evaluating joint training model |
| CN111488995A (en) * | 2020-04-08 | 2020-08-04 | 北京字节跳动网络技术有限公司 | Method and apparatus for evaluating a joint training model |
| CN113688855A (en) * | 2020-05-19 | 2021-11-23 | 华为技术有限公司 | Data processing method, federal learning training method, related device and equipment |
| CN113688855B (en) * | 2020-05-19 | 2023-07-28 | 华为技术有限公司 | Data processing method, federal learning training method, related device and equipment |
| CN113705823A (en) * | 2020-05-22 | 2021-11-26 | 华为技术有限公司 | Model training method based on federal learning and electronic equipment |
| CN111695701B (en) * | 2020-06-12 | 2021-08-13 | 上海富数科技有限公司 | System for realizing data set construction processing based on federal learning and construction generation method thereof |
| CN111695701A (en) * | 2020-06-12 | 2020-09-22 | 上海富数科技有限公司 | System for realizing data set construction processing based on federal learning and construction generation method thereof |
| CN111678696A (en) * | 2020-06-17 | 2020-09-18 | 南昌航空大学 | Intelligent mechanical fault diagnosis method based on federal learning |
| CN111898764A (en) * | 2020-06-23 | 2020-11-06 | 华为技术有限公司 | Method, device and chip for federal learning |
| CN111814985A (en) * | 2020-06-30 | 2020-10-23 | 平安科技（深圳）有限公司 | Model training method under federated learning network and related equipment thereof |
| WO2021120676A1 (en) * | 2020-06-30 | 2021-06-24 | 平安科技（深圳）有限公司 | Model training method for federated learning network, and related device |
| CN111814985B (en) * | 2020-06-30 | 2023-08-29 | 平安科技（深圳）有限公司 | Model training method under federal learning network and related equipment thereof |
| WO2021120677A1 (en) * | 2020-07-07 | 2021-06-24 | 平安科技（深圳）有限公司 | Warehousing model training method and device, computer device and storage medium |
| CN111881966A (en) * | 2020-07-20 | 2020-11-03 | 北京市商汤科技开发有限公司 | Neural network training method, device, equipment and storage medium |
| CN111882133A (en) * | 2020-08-03 | 2020-11-03 | 重庆大学 | Prediction-based federated learning communication optimization method and system |
| CN112001500A (en) * | 2020-08-13 | 2020-11-27 | 星环信息科技（上海）有限公司 | Model training method, device and storage medium based on longitudinal federated learning system |
| CN112101528B (en) * | 2020-09-17 | 2023-10-24 | 上海交通大学 | Terminal contribution measurement method based on back propagation |
| CN112101528A (en) * | 2020-09-17 | 2020-12-18 | 上海交通大学 | Terminal contribution measurement method based on back propagation |
| WO2022057433A1 (en) * | 2020-09-18 | 2022-03-24 | 华为技术有限公司 | Machine learning model training method and related device |
| CN111931242B (en) * | 2020-09-30 | 2021-02-19 | 国网浙江省电力有限公司电力科学研究院 | Data sharing method, computer equipment applying same and readable storage medium |
| CN111931242A (en) * | 2020-09-30 | 2020-11-13 | 国网浙江省电力有限公司电力科学研究院 | Data sharing method, computer equipment applying same and readable storage medium |
| CN111931876A (en) * | 2020-10-12 | 2020-11-13 | 支付宝(杭州)信息技术有限公司 | Target data side screening method and system for distributed model training |
| US11449805B2 (en) | 2020-10-12 | 2022-09-20 | Alipay (Hangzhou) Information Technology Co., Ltd. | Target data party selection methods and systems for distributed model training |
| CN112232519B (en) * | 2020-10-15 | 2024-01-09 | 成都数融科技有限公司 | Joint modeling method based on federal learning |
| CN112232519A (en) * | 2020-10-15 | 2021-01-15 | 成都数融科技有限公司 | Joint modeling method based on federal learning |
| WO2022096919A1 (en) * | 2020-11-05 | 2022-05-12 | Telefonaktiebolaget Lm Ericsson (Publ) | Managing training of a machine learning model |
| CN113807157A (en) * | 2020-11-27 | 2021-12-17 | 京东科技控股股份有限公司 | Method, device and system for training neural network model based on federal learning |
| WO2022110248A1 (en) * | 2020-11-30 | 2022-06-02 | 华为技术有限公司 | Federated learning method, device and system |
| CN113239351B (en) * | 2020-12-08 | 2022-05-13 | 武汉大学 | Novel data pollution attack defense method for Internet of things system |
| CN113239351A (en) * | 2020-12-08 | 2021-08-10 | 武汉大学 | Novel data pollution attack defense method for Internet of things system |
| WO2022121032A1 (en) * | 2020-12-10 | 2022-06-16 | 广州广电运通金融电子股份有限公司 | Data set division method and system in federated learning scene |
| CN113158550A (en) * | 2021-03-24 | 2021-07-23 | 北京邮电大学 | Method and device for federated learning, electronic equipment and storage medium |
| CN113222211A (en) * | 2021-03-31 | 2021-08-06 | 中国科学技术大学先进技术研究院 | Multi-region diesel vehicle pollutant emission factor prediction method and system |
| CN113222211B (en) * | 2021-03-31 | 2023-12-12 | 中国科学技术大学先进技术研究院 | Method and system for predicting pollutant emission factors of multi-region diesel vehicle |
| CN112949837A (en) * | 2021-04-13 | 2021-06-11 | 中国人民武装警察部队警官学院 | Target recognition federal deep learning method based on trusted network |
| CN112949837B (en) * | 2021-04-13 | 2022-11-11 | 中国人民武装警察部队警官学院 | Target recognition federal deep learning method based on trusted network |
| CN113377797A (en) * | 2021-07-02 | 2021-09-10 | 支付宝(杭州)信息技术有限公司 | Method, device and system for jointly updating model |
| CN113516250A (en) * | 2021-07-13 | 2021-10-19 | 北京百度网讯科技有限公司 | Method, device and equipment for federated learning and storage medium |
| CN113516250B (en) * | 2021-07-13 | 2023-11-03 | 北京百度网讯科技有限公司 | Federal learning method, device, equipment and storage medium |
| CN113537633A (en) * | 2021-08-09 | 2021-10-22 | 中国电信股份有限公司 | Prediction method, device, equipment, medium and system based on longitudinal federal learning |
| CN113537633B (en) * | 2021-08-09 | 2023-04-18 | 中国电信股份有限公司 | Prediction method, device, equipment, medium and system based on longitudinal federal learning |
| WO2023030221A1 (en) * | 2021-08-30 | 2023-03-09 | 华为技术有限公司 | Communication method and apparatus |
| CN113792632A (en) * | 2021-09-02 | 2021-12-14 | 广州广电运通金融电子股份有限公司 | Finger vein identification method, system and storage medium based on multi-party cooperation |
| CN113837766A (en) * | 2021-10-08 | 2021-12-24 | 支付宝(杭州)信息技术有限公司 | Risk identification method and device and electronic equipment |
| CN114548426A (en) * | 2022-02-17 | 2022-05-27 | 北京百度网讯科技有限公司 | Asynchronous federal learning method, business service prediction method, device and system |
| CN114548426B (en) * | 2022-02-17 | 2023-11-24 | 北京百度网讯科技有限公司 | Asynchronous federal learning method, business service prediction method, device and system |
| CN114679332A (en) * | 2022-04-14 | 2022-06-28 | 浙江工业大学 | APT detection method of distributed system |
| CN114662340A (en) * | 2022-04-29 | 2022-06-24 | 烟台创迹软件有限公司 | Weighing model scheme determination method and device, computer equipment and storage medium |
| CN114662340B (en) * | 2022-04-29 | 2023-02-28 | 烟台创迹软件有限公司 | Weighing model scheme determination method and device, computer equipment and storage medium |
| CN114741611B (en) * | 2022-06-08 | 2022-10-14 | 杭州金智塔科技有限公司 | Federal recommendation model training method and system |
| CN114741611A (en) * | 2022-06-08 | 2022-07-12 | 杭州金智塔科技有限公司 | Federal recommendation model training method and system |
| CN114844653A (en) * | 2022-07-04 | 2022-08-02 | 湖南密码工程研究中心有限公司 | Credible federal learning method based on alliance chain |
| CN115196730A (en) * | 2022-07-19 | 2022-10-18 | 南通派菲克水务技术有限公司 | Intelligent sodium hypochlorite adding system for water plant |
| CN116245172A (en) * | 2023-03-14 | 2023-06-09 | 南京航空航天大学 | Coalition building game method facing individual model performance optimization in cross-island federal learning |
| CN116245172B (en) * | 2023-03-14 | 2023-10-17 | 南京航空航天大学 | Federation construction method for optimizing individual model performance in cross-island federation learning |
| CN116127417A (en) * | 2023-04-04 | 2023-05-16 | 山东浪潮科学研究院有限公司 | Code defect detection model construction method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110442457A (en) | Model training method, device and server based on federation's study | |
| JP7296978B2 (en) | Providing incentives to players to engage in competitive gameplay | |
| US7537523B2 (en) | Dynamic player groups for interest management in multi-character virtual environments | |
| CN110189174A (en) | A kind of mobile intelligent perception motivational techniques based on quality of data perception | |
| CN110610242A (en) | Method and device for setting participant weight in federated learning | |
| CN111282267B (en) | Information processing method, information processing apparatus, information processing medium, and electronic device | |
| CN107483986A (en) | A kind of method and system of gifts | |
| Zhan et al. | Incentive mechanism design for federated learning: Challenges and opportunities | |
| EP3852014A1 (en) | Method and apparatus for training learning model, and computing device | |
| Ho et al. | Towards social norm design for crowdsourcing markets | |
| CN107229966A (en) | A kind of model data update method, apparatus and system | |
| CN106529189A (en) | User classifying method, application server and application client-side | |
| CN112990987B (en) | Information popularization method and device, electronic equipment and storage medium | |
| CN107944460A (en) | One kind is applied to class imbalance sorting technique in bioinformatics | |
| CN111957047A (en) | Checkpoint configuration data adjusting method, computer equipment and storage medium | |
| Xiao et al. | Incentive mechanism design for federated learning: A two-stage stackelberg game approach | |
| CN110533178A (en) | A kind of neural network model training method, apparatus and system | |
| CN114372589A (en) | Federated learning method and related device | |
| CN114021735A (en) | Method and device for processing data in federated learning | |
| CN104079472A (en) | Method for mobile game products to achieve real-time social contact functions | |
| CN116866345A (en) | Computer enhancement to increase service growth rate | |
| CN116339973A (en) | Digital twin cloud platform computing resource scheduling method based on particle swarm optimization algorithm | |
| CN109543725A (en) | A kind of method and device obtaining model parameter | |
| CN114530073A (en) | Training method and device based on virtual reality | |
| CN108984479B (en) | Method for improving operating efficiency of crowdsourcing platform |
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 |