CN110598870B - Federal learning method and device - Google Patents
Federal learning method and device Download PDFInfo
- Publication number
- CN110598870B CN110598870B CN201910824202.0A CN201910824202A CN110598870B CN 110598870 B CN110598870 B CN 110598870B CN 201910824202 A CN201910824202 A CN 201910824202A CN 110598870 B CN110598870 B CN 110598870B
- Authority
- CN
- China
- Prior art keywords
- federal learning
- participants
- participant
- coordinator
- report
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000012549 training Methods 0.000 claims abstract description 66
- 238000012545 processing Methods 0.000 claims description 12
- 238000003860 storage Methods 0.000 claims description 7
- 238000012360 testing method Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 abstract description 11
- 238000005516 engineering process Methods 0.000 abstract description 7
- 230000005540 biological transmission Effects 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 9
- 238000010801 machine learning Methods 0.000 description 7
- 238000004590 computer program Methods 0.000 description 6
- 238000004891 communication Methods 0.000 description 4
- 230000005611 electricity Effects 0.000 description 4
- 230000006870 function Effects 0.000 description 3
- 230000001788 irregular Effects 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000000737 periodic effect Effects 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000003203 everyday effect Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a federal learning method and a federal learning device, wherein the method comprises the following steps: the coordinator receives reports of a plurality of participants; the coordinator determines the participants meeting preset conditions according to the reports of the participants, and the participants are used as the participants participating in federal learning; wherein the report characterizes the expected available resource situation of the participant; and the coordinator carries out federal learning model training through the participants participating in federal learning. When the method is applied to the financial science and technology (Fintech), the participants which do not meet the expected available resource conditions are removed as much as possible, so that the influence of the transmission efficiency of the participants on the performance of the federal learning model in the federal learning process is reduced in the federal learning process of the coordinator through the participants participating in federal learning.
Description
Technical Field
The invention relates to the field of financial science and technology (Fintech) and the field of federal learning, in particular to a federal learning method and device.
Background
With the development of computer technology, more and more technologies (big data, distributed, blockchain (Blockchain), artificial intelligence, etc.) are applied in the financial field, and the traditional financial industry is gradually changing to financial technology (Fintech). Currently, many financial strategies in the field of financial science and technology are adjusted depending on federal learning results for a large amount of financial transaction data, and the adjustment of the corresponding financial strategy is likely to affect the financial institution's profit and loss. Therefore, the performance of the federal learning model is critical to a financial institution.
When the number of participants in the federal learning is large, particularly when the participants are mobile terminals, the difference between the participants is large, for example, the online time of many participants is irregular and unstable, and disconnection and interruption are easy to occur, so that the training of the federal learning model is affected, and the performance of the obtained federal learning model cannot meet the preset requirement. This is a problem to be solved.
Disclosure of Invention
The embodiment of the application provides a federal learning method and a federal learning device, which solve the problem that the performance of a federal learning model obtained in the prior art cannot meet the preset requirement.
In a first aspect, an embodiment of the present application provides a federal learning method, including: the coordinator receives reports of a plurality of participants; the coordinator determines the participants meeting preset conditions according to the reports of the participants, and the participants are used as the participants participating in federal learning; wherein the report characterizes the expected available resource situation of the participant; and the coordinator carries out federal learning model training through the participants participating in federal learning.
In the method, the coordinator receives the reports of the plurality of participants, and the coordinator screens out the participants meeting the preset conditions from the plurality of participants, and the expected available resource conditions of the participants are represented by the preset conditions, so that the influence of the transmission efficiency of the participants on the performance of the federal learning model in the federal learning process is reduced in the federal learning process of the coordinator through the participants participating in federal learning.
In an alternative embodiment, the report of the participant in federal learning includes an expected idle period of time for the participant in federal learning, the method further comprising: the coordinator takes the public time period in the idle time period expected by the participants participating in the federal learning as the time period trained by the federal learning model.
In the method, the coordinator takes the public time period in the expected idle time period of the participants participating in the federal learning as the time period of the federal learning model training, so that the federal learning training is carried out in the time period in which the participants participating in the federal learning are idle, and the smooth progress of the federal learning model training is ensured.
In an alternative embodiment, the coordinator sends a model update request to the participants participating in federal learning; the model update request instructs the federal learning participants to perform federal learning model training during the common time period.
In the method, the coordinator sends the model update request to the participants participating in the federal learning, and the participants learn the public time period of the federal learning after receiving the model update request so as to transmit information in the federal learning process.
In an alternative embodiment, the report of the participant includes at least one operation index, each operation index corresponding to a preset weight value; the coordinator determines whether the report of the participant satisfies a preset condition in the following manner: the coordinator determines the score of each operation index in at least one operation index according to the at least one operation index and a preset scoring rule; the coordinator determines the report score of the participant according to the scores of each operation index in the at least one operation index and the corresponding preset weight values; the coordinator determines whether the report of the participant meets a preset condition according to the report score of the participant; the preset condition is that the report score of the participant is greater than or equal to a preset score threshold.
Under the mode, according to at least one operation index and a preset scoring rule, the score of each operation index in the at least one operation index is determined, and then according to the score of each operation index in the at least one operation index and the corresponding preset weight value, the report score of the participant is determined; thereby determining whether the participant's report meets a preset condition; the coordinator can comprehensively consider the influence of each operation index according to the weight value of each operation index according to the importance degree of each operation index according to specific situations, thereby improving the accuracy of determining whether the report of the participant meets the preset condition.
In an alternative embodiment, the plurality of participant reports are used to instruct the plurality of participants to apply for addition to federal learning.
In the above manner, the plurality of participants inform the coordinator that the plurality of participants desire to participate in federal learning by actively pushing the report, thereby selecting the range of participants participating in federal learning.
In an alternative embodiment, the report of the plurality of participants includes the operating conditions of the plurality of participants.
In the above manner, the report of the plurality of participants includes the operating conditions of the plurality of participants, which facilitates the coordinator to select the appropriate participant.
In a second aspect, the present application provides a federal learning apparatus comprising: a receiving module for receiving reports of a plurality of participants; the processing module is used for determining the participants meeting the preset conditions according to the reports of the participants, and taking the participants as the participants participating in the federal learning; wherein the report characterizes the expected available resource situation of the participant; and for federal learning model training by the participants participating in federal learning.
In an alternative embodiment, the report of the participant in federal learning includes an expected idle period of time of the participant in federal learning, the processing module further configured to: and taking the public time period in the idle time periods expected by the participants participating in the federal learning as the time period trained by the federal learning model.
In an alternative embodiment, the processing module is further configured to: sending a model update request to the participants participating in federal learning; the model update request instructs the federal learning participants to perform federal learning model training during the common time period.
In an alternative embodiment, the report of the participant includes at least one operation index, each operation index corresponding to a preset weight value; the processing module is specifically configured to: determining whether the participant's report meets a preset condition in the following manner: determining the score of at least one operation index according to the at least one operation index and a preset scoring rule; determining a report score of the participant according to the score of the at least one operation index and the corresponding preset weight value; determining whether a report of the participant meets a preset condition according to the report score of the participant; the preset condition is that the report score of the participant is greater than or equal to a preset score threshold.
In an alternative embodiment, the plurality of participant reports are used to instruct the plurality of participants to apply for addition to federal learning.
In an alternative embodiment, the report of the plurality of participants includes the operating conditions of the plurality of participants.
The advantages of the second aspect and the embodiments of the second aspect may be referred to the advantages of the first aspect and the embodiments of the first aspect, and will not be described here again.
In a third aspect, embodiments of the present application provide a computer device comprising a program or instructions which, when executed, are adapted to carry out the methods of the first aspect and the embodiments of the first aspect described above.
In a fourth aspect, embodiments of the present application provide a storage medium including a program or instructions, which when executed, are configured to perform the method of the first aspect and the respective embodiments of the first aspect.
Drawings
FIG. 1 is a schematic diagram of a federal learning architecture provided in an embodiment of the present application;
FIG. 2 is a schematic diagram of a federal learning process according to an embodiment of the present application;
FIG. 3 is a flowchart illustrating steps of a federal learning method according to an embodiment of the present application;
FIG. 4 is a timing diagram of a federal learning method according to an embodiment of the present application;
Fig. 5 is a schematic structural diagram of a federal learning device according to an embodiment of the present application.
Detailed Description
In order to better understand the above technical solutions, the following detailed description will be made with reference to the accompanying drawings and specific embodiments of the present application, and it should be understood that specific features in the embodiments and examples of the present application are detailed descriptions of the technical solutions of the present application, and not limiting the technical solutions of the present application, and the technical features in the embodiments and examples of the present application may be combined with each other without conflict.
Federal learning refers to a method of machine learning by combining different participants. As shown in fig. 1, one joint model parameter update for federal learning is divided into two steps: (a) Each participant (also referred to as a data owner, or client) uses only its own locally owned data to train the machine learning model and obtain model parameter updates, and sends model parameter updates, such as model weights or gradient information, to one coordinator (also referred to as a parameter server (PARAMETER SERVER), or aggregation server (aggregation server); (b) The coordinator fuses (e.g., weight averages) the received model parameter updates from the different participants and redistributes the fused model parameter updates to the individual participants. In federal learning, participants do not need to expose own data to other participants and coordinators, so that federal learning can well protect user privacy and ensure data security.
When the data features of the respective participants overlap more and the user overlap less, the part of the data of the participants, which has the same data features but the users are not identical, is fetched to perform joint machine learning (this joint machine learning mode is hereinafter referred to as a first federal learning mode).
There are two banks in different areas, and their user groups are from the areas where they are located, respectively, and the intersection is small. But their traffic is very similar and the recorded user data characteristics are identical. A first federal learning approach can be used to help two banks build a joint model to predict customer behavior. As another example, a mobile user input prediction (input auto-complement and recommendation) model may be built by combining multiple mobile terminals in a first federal learning manner. For another example, a mobile user search keyword prediction (keyword automatic complement and recommendation) model may be constructed by combining a plurality of mobile terminals in a first federal learning manner. The first federal learning approach may also be used in the field of edge computing and internet of things (Internet of Things, ioT) to solve the problem that internet of things devices (e.g., video sensor nodes and unlimited cameras) do not have sufficient bandwidth to upload large amounts of data.
As illustrated in fig. 2, in the primary model parameter update link of the first federal learning mode, a possible procedure is for the coordinator to send a model update request (model update request) message to the participant. The model update request message may have 2 main roles: (a) The model update request message is used for notifying the participants to start model parameter update, namely, is used as a starting signal of local model training of the participants; (b) The model update request message may also carry the latest joint model parameters owned by the coordinator, i.e., the model update request message may be used to distribute the latest joint model parameters to the participants. The joint model parameters may be parameters of a federal learning model, e.g., weight parameter values for connections between nodes of a neural network; alternatively, the joint model parameters may be gradient information of a federal learning model, for example, gradient information in a neural network gradient descent algorithm. The participants may use the joint model parameters as a starting point for local model training in order to continue training the local model.
The coordinator may send the model parameter update message to each participant separately using point-to-point communication. Or the coordinator may also send the model update request message to multiple participants simultaneously using multicast or broadcast.
After the coordinator finishes sending the model update request message, the coordinator waits for receiving the model parameter update sent by the participant even if the coordinator enters a waiting state.
After receiving the model parameter updating message, a participant A can obtain the latest joint model parameters from the message, and perform model training locally or continue to perform model training by using data locally owned by the participant A. The coordinator-distributed joint model parameters received by the participant a may be used as initial values of model parameters trained by the local machine learning model of the participant a, for example, as initial model parameters, or initial gradient information.
After participant a completes the model parameter update locally, participant a may send the locally obtained model parameter update to the coordinator. Participant a may send model parameter updates to the coordinator by means of encryption, for example using homomorphic encryption techniques.
The joint model constructed using the first federal learning mode may be a conventional machine learning model (e.g., linear regression, support vector machine, etc.), or may be various deep learning models (i.e., deep neural network models).
In applying the first federal learning mode technique to jointly construct a machine learning model in conjunction with a plurality of mobile terminals (e.g., smartphones), the participants are mobile terminals and the number of participants will typically be relatively large, e.g., there are thousands or even hundreds of thousands of participants. In the application scenario of the internet of things, the participants of the first federal learning mode may be a large number of wireless sensor nodes, such as wireless cameras.
In such a mobile internet application scenario, the management of the first federal learning mode training process, including selecting participants, model training start time, model training end time, model performance testing, retraining of (periodic) models (periodic model updating), is a complex management effort for the coordinator. In particular, unlike fixed terminals and servers, the online time of a mobile terminal has randomness, i.e. the network connection of the mobile terminal may be interrupted at any time, or even for a longer period of time. If the online time of many participants in a selected first federal learning mode participant is irregular and unstable, e.g., dropped or broken, it can affect the training of the first federal learning mode model, e.g., the training of the model cannot be completed, or the model does not converge, or the performance of the obtained model does not meet the predetermined requirements.
For mobile terminals, participating in federal learning can incur additional overhead, including communication overhead (e.g., mobile network traffic overhead), network bandwidth, power overhead, computing resource overhead, and storage resource overhead. With the popularity of 4G networks and Wi-Fi, traffic overhead has not been a major concern for mobile terminal users. Network speed, battery life (power) and computing resources are always important points of attention of a mobile terminal user and are also key indexes of user experience. If the coordinator selects an inappropriate user to perform the first federal learning mode model training in an inappropriate time, the computing resource overhead of the local model training, the network resource overhead of sending the model parameter update, the power overhead of the participants, etc. are likely to seriously affect the mobile terminal user experience. The mobile terminal user is likely to refuse to participate in the first federal learning mode model training the next time. For example, if the coordinator selects a user to have his cell phone participate in the first federal learning mode model training while playing the online game with the cell phone, the user is likely to discontinue participating in the first federal learning mode model training after the user finds that game play is affected. If there is often a mobile terminal refusing to participate in or discontinue ongoing model training in the first federal learning mode, the coordinator cannot successfully complete the intended model training.
Further, APP management and monitoring software is commonly available on current mobile terminals. If a certain APP based on federal learning is found to occupy a large amount of communication resources, or calculation resources, or consume a relatively large amount of electricity, and influence the user experience, the monitoring software reminds the user, and the user may choose to delete the APP. Thus, APP application popularization based on federal learning is affected, and commercialization of federal learning is affected.
In the operation of a financial institution (e.g., a banking institution, an insurance institution or a securities institution) in a business (e.g., a loan business, a deposit business, etc.), many financial strategies in the field of financial science and technology are adjusted depending on the result of federal learning on a large amount of financial transaction data, and the adjustment of the corresponding financial strategy is likely to affect the profit and loss of the financial institution. When the number of participants in the federal learning is large, particularly when the participants are mobile terminals, the difference between the participants is large, for example, the online time of many participants is irregular and unstable, and disconnection and interruption are easy to occur, so that the training of the federal learning model is affected, and the performance of the obtained federal learning model cannot meet the preset requirement. This situation does not meet the requirements of financial institutions such as banks, and cannot guarantee efficient operation of various businesses of the financial institutions.
To this end, as shown in FIG. 3, an embodiment of the present application provides a federal learning method.
Step 301: the coordinator receives reports of the plurality of participants.
Step 302: and the coordinator determines the participants meeting the preset conditions according to the reports of the participants, and the participants are used as the participants participating in federal learning.
In particular, the report may include operating conditions or available resource conditions. The report characterizes the expected available resource situation of the participant. It should be noted that, the available resource may be a series of information such as "computing resource of the participant", "power of the participant", "network bandwidth of the participant", and the like.
Step 303: and the coordinator carries out federal learning model training through the participants participating in federal learning.
In the method from step 301 to step 303, the coordinator receives the reports of the plurality of participants, and the coordinator screens out the participants meeting the preset conditions among the plurality of participants, and as the preset conditions characterize the expected available resource conditions of the participants, the influence of the transmission efficiency of the participants on the performance of the federal learning model in the federal learning process is reduced in the federal learning process of the coordinator through the participants participating in federal learning.
In step 301, the reporting of each participant in the plurality of participant reports includes, but is not limited to, at least one of: the difference between the participant's current performance index and the expected performance index; the amount of data owned by the participant; the amount of electricity of the participant; computing resources of the participant; the network bandwidth of the participant. The report includes content only by way of example.
In an alternative embodiment of step 301, the plurality of participant reports are used to indicate that the plurality of participants apply to join federal learning.
In another alternative embodiment of step 301, the report of the plurality of participants includes the operating conditions of the plurality of participants.
In an alternative embodiment of step 302, the report of the participant includes at least one operation index, each operation index corresponding to a preset weight value; the coordinator determines whether the report of the participant satisfies a preset condition in the following manner: the coordinator determines the score of at least one operation index according to the at least one operation index and a preset scoring rule; the coordinator determines the report score of the participant according to the score of the at least one operation index and the corresponding preset weight value; the coordinator determines whether the report of the participant meets a preset condition according to the report score of the participant; the preset condition is that the report score of the participant is greater than or equal to a preset score threshold.
The score value of the at least one operation index has the following meaning: (1) When the at least one operation index only contains one index, the score of the at least one operation index is the score of the only one index; (2) When the at least one operation index contains two or more indexes, the score of the at least one operation index is a combination of the scores of each operation index of the at least one operation index.
The score of the at least one operation index and the corresponding preset weight value have the following meanings: (1) When at least one running index only contains one index, the score of at least one running index and the corresponding preset weight value are the score of only one index and the preset weight value of only one index; (2) When at least one running index contains two or more indexes, the score of at least one running index and the corresponding preset weight value are the combination of the score of each running index of at least one running index and the corresponding preset weight value.
Under the mode, according to at least one operation index and a preset scoring rule, the score of each operation index in the at least one operation index is determined, and then according to the score of each operation index in the at least one operation index and the corresponding preset weight value, the report score of the participant is determined; thereby determining whether the participant's report meets a preset condition; the coordinator can comprehensively consider the influence of each operation index according to the weight value of each operation index according to the importance degree of each operation index according to specific situations, thereby improving the accuracy of determining whether the report of the participant meets the preset condition.
For example, the coordinator considers two factors: network bandwidth of the participant, power of the participant. The coordinator considers the network bandwidth of the participant to be more important, the network bandwidth weight value of the participant is set to 10, and the electric quantity weight value of the participant is set to 8. Participant 1 had a network bandwidth of 6MB/s (megabytes/second) and participant 2 had a network bandwidth of 8MB/s; the electricity quantity of the participant 1 is 1600mAh, and the electricity quantity of the participant 2 is 1000mAh. The preset scoring rule of the network bandwidth is that the corresponding score of the physical bandwidth is 10 (network bandwidth/10 MB/s) in the range of 0-10 MB/s; the preset scoring rule of the electric quantity is that the corresponding score of the electric quantity is 10 x (electric quantity/2000 mAh) in 0-2000 mAh. Thus participant 1 reported a score of 10 x 6+8 x 8 = 124; participant 2 reports with a score of 10 x 8+8 x 5 = 120, assuming a preset score threshold of 110, and participant 1 and participant 2 reports meet preset conditions.
In addition, the participant can be scored by combining the difference between the current performance index and the expected performance index of the participant, the data amount owned by the participant, the computing resource of the participant and other operation indexes. For example, the preset scoring rule of the difference between the current performance index and the expected performance index of the participant is to score according to the ratio of the difference to the expected performance index, and establish a mapping relationship between the ratio and the score; the preset scoring rule of the data volume owned by the participant is that scoring is carried out according to the proportion of the data volume of the participant in the data volume of all participants participating in the training of the federal learning model, and the score is higher when the proportion is higher; the preset scoring rule of the computing resource of the participant is that, since the computing resource is usually a limited value and has an upper limit, the mapping relationship between the computing resource of the participant and the score can be established.
In steps 301-303, the period of federal learning may be determined according to the following alternative embodiments:
The report of the participant in federal learning includes an expected idle period of time for the participant in federal learning, the method further comprising: the coordinator takes the public time period in the idle time period expected by the participants participating in the federal learning as the time period trained by the federal learning model.
In the method, the coordinator takes the public time period in the expected idle time period of the participants participating in the federal learning as the time period of the federal learning, so that the federal learning training is carried out in the time period in which the participants participating in the federal learning are idle, and the available resource condition of the participants in the federal learning process is ensured.
In an alternative embodiment of steps 301 to 303, the coordinator sends a model update request to the participants participating in federal learning; the model update request instructs the federal learning participants to perform federal learning model training during the common time period.
In the method, the coordinator sends the model update request to the participants participating in the federal learning, and the participants learn the public time period of the federal learning after receiving the model update request so as to transmit information in the federal learning process.
A federal learning method according to an embodiment of the present application is described in detail below with reference to fig. 4.
The core idea of the technical scheme provided by the application is that a mobile terminal (or an internet traffic (IoT) device) actively selects to apply to a coordinator to participate in the training of a horizontal federal learning model, and actively reports the running state and the network connection condition of the mobile terminal (or the IoT device). After the coordinator receives the applications and reports from the multiple mobile terminals (or IoT devices), the coordinator selects participants from the candidate mobile terminals (or IoT devices), and determines the starting time of the federal learning model training according to the report content sent by the devices. In other words, it is the coordinator that gathers candidate participants and candidate participant information for lateral federal learning through active applications and reports of the mobile terminal (or IoT device).
As illustrated in fig. 4, after the coordinator (also referred to as a parameter server, or fusion server) completes the startup online, the coordinator sends a message to the mobile terminal or IoT device that the coordinator is "ready" or "coordinator ready" (coordinator _ready). The "coordinator _ready" message is mainly used to inform the mobile terminal or IoT device that the device report can start to send to the coordinator, and the "coordinator _ready" message can carry key information at the same time. The coordinator may send the "coordinator _ready" message to a mobile terminal or IoT device by means of point-to-point communication, or the coordinator may send the "coordinator _ready" message to multiple mobile terminals or IoT devices simultaneously by means of multicast or broadcast. For ease of description, the mobile terminals or IoT devices, hereinafter abbreviated as possible applications, are devices, including other possible devices, such as fixed wireless terminals. As shown in fig. 4, device D does not send a device report to the coordinator; the equipment C is not selected to participate in the training of the transverse federal learning model; only devices a and B are selected to participate in lateral federal learning model training
After receiving the "coordinator _ready" message sent by the coordinator, a device may choose to send a "device report" message to the coordinator. The mobile terminal or IoT device actively sends a "device report" to the coordinator, applies to the coordinator to participate in the lateral federal learning model training, and reports the report of the device to the coordinator. The coordinator collects candidate participant information by receiving a "device report" sent by the plurality of devices, and selects a participant of the lateral federal learning model training from the candidate participants according to the "device report" content and determines a start time of the model training. The coordinator sends a "model update request" message to the selected participants.
In particular, the device report message may have two roles.
The first effect is that a device applies to the coordinator via a "device_report" message that it wishes to join the horizontal federal learning model training, e.g., the coordinator may consider that a device is sending to the coordinator. The device report message indicates that the device wishes to join in the lateral federal learning model training.
The second effect is that a device reports some of its operating conditions to the coordinator via a "device_report" message. The device operational status, i.e., the "device_report" message, may include one or more of the following information: the degree of urgency that the device needs to update the model (e.g., the greater the gap between the performance of the existing model and the predetermined performance index, the more urgent the gap is to account for federal learning, the more urgent the degree of updating the model) the device has the amount of data that can be used for federal learning model training (i.e., the training set size), the amount of data that the device has that can be used to test the performance of the federal learning model (i.e., the testing set size), the power of the device (e.g., the device is in an active power state), the computing resources of the device (e.g., the device is idle), the network connectivity and bandwidth of the device (e.g., in Wi-Fi high speed connection), and the time that the current operating condition of the device can be maintained (e.g., the device will remain in the current state for the next 2 hours). The time that the current running condition of the device can be maintained can be understood as the time that the device can stably participate in the training of the transverse federal learning model.
The device may calculate the historical usage/operation condition of the device, so as to obtain an operation condition estimation of the device, for example, a smart phone may calculate that the smart phone is in an idle state after 12 nights every day, keep Wi-Fi connection, and plug in a power supply or have sufficient electric quantity.
After receiving the "device_report" sent by the plurality of devices, i.e., the coordinator collects the candidate device information by receiving the "device report" sent by the plurality of devices, the coordinator may select a part or all of the devices to participate in the horizontal federal learning model training, e.g., the coordinator may rank according to the data amount and the network bandwidth, and further select the devices.
After selecting the participants, the coordinator may determine a start time for federal learning model training based on the time that the participants may next stably participate in the federal learning model training. The coordinator can start to idle according to the latest time of the selected participants, for example, 3 participants A, B and C can enter idle states at 1, 2 and 3 early morning respectively and can participate in the training of the horizontal federal learning model, and then the coordinator can choose to start to train the horizontal federal learning model at 3 early morning.
Further, the coordinator may determine the total time of model training and the end time of training, e.g., the total duration of model training is 2 hours.
After the coordinator determines the start time of the federal learning model training, the coordinator may send a "model update request" message to the selected device, instructing the participant to start the participant local model training or to continue the model training. The coordinator can send the model_update_request message to the selected device in a point-to-point unicast mode; or the coordinator can send the 'model_update_request' message to a plurality of selected devices simultaneously in a multicast or multicast mode. Or the coordinator may send the "model_update_request" message to a plurality of selected devices simultaneously by broadcasting, and in this implementation, the broadcasting "model_update_request" message needs to carry the selected device ID information so as to distinguish the selected devices from the non-selected devices.
In the technical scheme provided by the application, the mobile terminal or the IoT device actively transmits an application for participating in horizontal federal learning and a report of the mobile terminal or the IoT device ("device report") to the coordinator. This facilitates the coordinator to collect information of the alternative devices by receiving the "device report". Therefore, the coordinator can be effectively prevented from actively inquiring the report of each participant, and the operation and maintenance burden of the coordinator can be effectively reduced. Especially in the scenario that requires the coordinator to query reports of tens of thousands of mobile terminals, even hundreds of thousands of mobile terminals, the operation and maintenance costs of the coordinator can be significantly reduced.
The equipment actively reports the report of the equipment ("equipment report"), a coordinator can fully know the running state and the network connection condition of each equipment, the coordinator can determine the starting time of the federal learning model training according to the idle time and the network connection condition of the selected federal learning participants, enough time can be ensured to complete the federal learning model training, and the condition that partial participants have network connection interruption can be effectively avoided.
The device actively reports the report of the device ("device report") and the application participates in the training of the transverse federal learning model, thereby conforming to the own wish of the mobile terminal and reducing the influence on the user experience of the mobile terminal as much as possible.
As shown in fig. 5, the present application provides a federal learning apparatus including: a receiving module for receiving reports of a plurality of participants; the processing module is used for determining the participants meeting the preset conditions according to the reports of the participants, and taking the participants as the participants participating in the federal learning; wherein the report characterizes the expected available resource situation of the participant; and for federal learning model training by the participants participating in federal learning.
In an alternative embodiment, the report of the participant in federal learning includes an expected idle period of time of the participant in federal learning, and the processing module 502 is further configured to: and taking the public time period in the idle time periods expected by the participants participating in the federal learning as the time period trained by the federal learning model.
In an alternative embodiment, the processing module 502 is further configured to: sending a model update request to the participants participating in federal learning; the model update request instructs the federal learning participants to perform federal learning model training during the common time period.
In an alternative embodiment, the report of the participant includes at least one operation index, each operation index corresponding to a preset weight value; the processing module 502 is specifically configured to: determining whether the participant's report meets a preset condition in the following manner: determining the score of each operation index in at least one operation index according to the at least one operation index and a preset scoring rule; determining the report score of the participant according to the scores of each operation index in the at least one operation index and the corresponding preset weight values; determining whether a report of the participant meets a preset condition according to the report score of the participant; the preset condition is that the report score of the participant is greater than or equal to a preset score threshold.
In an alternative embodiment, the plurality of participant reports are used to instruct the plurality of participants to apply for addition to federal learning.
In an alternative embodiment, the report of the plurality of participants includes the operating conditions of the plurality of participants.
The embodiment of the application provides a computer device, which comprises a program or an instruction, and the program or the instruction are used for executing the federal learning method and any optional method provided by the embodiment of the application when being executed.
The embodiment of the application provides a storage medium, which comprises a program or an instruction, and the program or the instruction are used for executing the federal learning method and any optional method provided by the embodiment of the application when being executed.
Finally, it should be noted that: it will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the scope of the application. Thus, it is intended that the present application also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.
Claims (7)
1. A federal learning method, comprising:
the coordinator receives device reports actively sent by a plurality of participants; the device report of the plurality of participants is used for indicating the plurality of participants to apply for joining federal learning;
The coordinator determines the participants meeting preset conditions according to the equipment reports of the participants, and the participants are used as the participants participating in federal learning; wherein the device report characterizes expected available resource conditions of the participant, the device report further comprises device operation conditions of the participant, the device operation conditions comprise emergency degree of the device needing to update a model, data amount which can be used for training a federal learning model, data amount which can be used for testing performance of the federal learning model, electric quantity of the device, network connection condition and bandwidth of the device and time for keeping current operation condition of the device, and the device report further comprises expected idle time period of the participant participating in federal learning;
And the coordinator carries out federal learning model training through the participants participating in federal learning, and takes a public time period in the idle time period expected by the participants participating in federal learning as a time period of federal learning model training.
2. The method of claim 1, wherein the coordinator determines, based on the device reports of the plurality of participants, a participant that satisfies a preset condition, and after the participant is a participant in federal learning, the coordinator further comprising, before performing federal learning model training by the participant in federal learning:
The coordinator sends a model update request to the participants participating in federal learning; the model update request instructs the federal learning participants to perform federal learning model training during the common time period.
3. The method of any of claims 1 to 2, wherein the participant's device report includes at least one operational indicator, each operational indicator corresponding to a preset weight value; the coordinator determines whether the device report of the participant satisfies a preset condition in the following manner:
The coordinator determines the score of at least one operation index according to the at least one operation index and a preset scoring rule;
The coordinator determines the equipment report score of the participant according to the score of the at least one operation index and the corresponding preset weight value;
The coordinator determines whether the device report of the participant meets a preset condition according to the device report score of the participant; the preset condition is that the equipment report score of the participant is greater than or equal to a preset score threshold.
4. The method of any of claims 1-2, wherein the device report of the plurality of participants includes operating conditions of the plurality of participants.
5. A federal learning apparatus, comprising:
The receiving module is used for receiving the equipment reports actively sent by the participants; the device report of the plurality of participants is used for indicating the plurality of participants to apply for joining federal learning;
The processing module is used for determining the participants meeting the preset conditions according to the equipment reports of the participants, and taking the participants as the participants participating in federal learning; wherein the device report characterizes expected available resource conditions of the participant, the device report further comprises device operation conditions of the participant, the device operation conditions comprise emergency degree of the device needing to update a model, data amount which can be used for training of a federal learning model, data amount which can be used for testing performance of the federal learning model, electric quantity of the device, network connection condition and bandwidth of the device and time for maintaining current operation conditions of the device, and the device report further comprises expected idle time period of the participant participating in federal learning; and the method is used for training the federal learning model through the participants participating in federal learning, and taking the public time period in the idle time period expected by the participants participating in federal learning as the time period for the federal learning model training.
6. A computer device comprising a program or instructions which, when executed, performs the method of any of claims 1 to 4.
7. A storage medium comprising a program or instructions which, when executed, perform the method of any one of claims 1 to 4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910824202.0A CN110598870B (en) | 2019-09-02 | 2019-09-02 | Federal learning method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910824202.0A CN110598870B (en) | 2019-09-02 | 2019-09-02 | Federal learning method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110598870A CN110598870A (en) | 2019-12-20 |
| CN110598870B true CN110598870B (en) | 2024-04-30 |
Family
ID=68857106
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910824202.0A Active CN110598870B (en) | 2019-09-02 | 2019-09-02 | Federal learning method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110598870B (en) |
Families Citing this family (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111325572B (en) * | 2020-01-21 | 2024-05-03 | 深圳前海微众银行股份有限公司 | Data processing method and device |
| CN111275491B (en) * | 2020-01-21 | 2023-12-26 | 深圳前海微众银行股份有限公司 | Data processing method and device |
| CN113159332B (en) * | 2020-01-23 | 2024-01-30 | 华为技术有限公司 | Method and equipment for realizing model update |
| CN113496291A (en) * | 2020-03-18 | 2021-10-12 | 索尼公司 | Apparatus, method and storage medium for federal learning |
| KR102567565B1 (en) * | 2020-03-31 | 2023-08-21 | 한국전자기술연구원 | Apparatus and system for managing federated learning resource, and resource efficiency method thereof |
| CN111428885B (en) * | 2020-03-31 | 2021-06-04 | 深圳前海微众银行股份有限公司 | User indexing method in federated learning and federated learning device |
| CN111510327B (en) | 2020-04-08 | 2022-01-21 | 北京字节跳动网络技术有限公司 | Network connection method, device, system and server for training participants of co-training model |
| CN111538598A (en) * | 2020-04-29 | 2020-08-14 | 深圳前海微众银行股份有限公司 | Federal learning modeling method, device, equipment and readable storage medium |
| GB2595849A (en) * | 2020-06-02 | 2021-12-15 | Nokia Technologies Oy | Collaborative machine learning |
| CN111539731A (en) * | 2020-06-19 | 2020-08-14 | 支付宝(杭州)信息技术有限公司 | Block chain-based federal learning method and device and electronic equipment |
| US11429903B2 (en) | 2020-06-24 | 2022-08-30 | Jingdong Digits Technology Holding Co., Ltd. | Privacy-preserving asynchronous federated learning for vertical partitioned data |
| WO2022014731A1 (en) * | 2020-07-14 | 2022-01-20 | 엘지전자 주식회사 | Scheduling method and device for aircomp-based federated learning |
| CN114039864A (en) * | 2020-07-21 | 2022-02-11 | 中国移动通信有限公司研究院 | Multi-device cooperation model generation method, device and equipment |
| CN111880568A (en) * | 2020-07-31 | 2020-11-03 | 深圳前海微众银行股份有限公司 | Optimization training method, device and equipment for automatic control of unmanned aerial vehicle and storage medium |
| CN112037800A (en) * | 2020-09-22 | 2020-12-04 | 平安科技（深圳）有限公司 | Voiceprint nuclear model training method and device, medium and electronic equipment |
| EP4231207A1 (en) * | 2020-10-13 | 2023-08-23 | Beijing Xiaomi Mobile Software Co., Ltd. | Wireless communication method and apparatus, communication device, and storage medium |
| CN112464278B (en) * | 2020-11-24 | 2023-07-21 | 平安科技（深圳）有限公司 | Federal modeling method based on non-uniformly distributed data and related equipment |
| CN113255924B (en) * | 2020-11-25 | 2023-03-24 | 中兴通讯股份有限公司 | Federal learning participant selection method, device, equipment and storage medium |
| CN112418439B (en) * | 2020-11-25 | 2023-09-26 | 脸萌有限公司 | Model using method, device, storage medium and equipment |
| EP4231209A4 (en) * | 2020-11-30 | 2023-12-20 | Huawei Technologies Co., Ltd. | Federated learning method, device and system |
| CN112508205A (en) * | 2020-12-04 | 2021-03-16 | 中国科学院深圳先进技术研究院 | Method, device and system for scheduling federated learning |
| CN112671613B (en) * | 2020-12-28 | 2022-08-23 | 深圳市彬讯科技有限公司 | Federal learning cluster monitoring method, device, equipment and medium |
| CN112766514B (en) * | 2021-01-22 | 2021-12-24 | 支付宝(杭州)信息技术有限公司 | Method, system and device for joint training of machine learning model |
| US20240028961A1 (en) * | 2021-01-25 | 2024-01-25 | Nokia Technologies Oy | Enablement of federated machine learning for terminals to improve their machine learning capabilities |
| US11711348B2 (en) | 2021-02-22 | 2023-07-25 | Begin Ai Inc. | Method for maintaining trust and credibility in a federated learning environment |
| US11755954B2 (en) | 2021-03-11 | 2023-09-12 | International Business Machines Corporation | Scheduled federated learning for enhanced search |
| CN113537518B (en) * | 2021-07-19 | 2022-09-30 | 哈尔滨工业大学 | Model training method and device based on federal learning, equipment and storage medium |
| CN115842851A (en) * | 2021-09-18 | 2023-03-24 | 大唐移动通信设备有限公司 | Method and device for processing federated learning group and functional entity |
| CN113836809B (en) * | 2021-09-26 | 2023-12-01 | 上海万向区块链股份公司 | Cross-industry data joint modeling method and system based on block chain and federal learning |
| CN116032917A (en) * | 2021-10-26 | 2023-04-28 | 展讯通信（上海）有限公司 | Federal learning method and device, and communication method and device |
| CN116567702A (en) * | 2022-01-26 | 2023-08-08 | 展讯通信（上海）有限公司 | User equipment selection method, device, chip and module equipment |
| CN115021883B (en) * | 2022-07-13 | 2022-12-27 | 北京物资学院 | Signaling mechanism for application of federal learning in wireless cellular systems |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104618381A (en) * | 2015-02-06 | 2015-05-13 | 百度在线网络技术（北京）有限公司 | Information interaction method and device |
| CN105303365A (en) * | 2015-10-16 | 2016-02-03 | 东华大学 | Office schedule planning method based on intelligent terminals |
| CN108460567A (en) * | 2017-02-20 | 2018-08-28 | 百度在线网络技术（北京）有限公司 | Determination method, apparatus, equipment and the storage medium of activity time |
| CN108683743A (en) * | 2018-05-21 | 2018-10-19 | 陕西师范大学 | A kind of method for allocating tasks by multiple mobile terminal gathered datas |
| 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 |
| CN109299728A (en) * | 2018-08-10 | 2019-02-01 | 深圳前海微众银行股份有限公司 | Federal learning method, system and readable storage medium storing program for executing |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110047230A1 (en) * | 2006-11-17 | 2011-02-24 | Mcgee Steven J | Method / process / procedure to enable: The Heart Beacon Rainbow Force Tracking |
-
2019
- 2019-09-02 CN CN201910824202.0A patent/CN110598870B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104618381A (en) * | 2015-02-06 | 2015-05-13 | 百度在线网络技术（北京）有限公司 | Information interaction method and device |
| CN105303365A (en) * | 2015-10-16 | 2016-02-03 | 东华大学 | Office schedule planning method based on intelligent terminals |
| CN108460567A (en) * | 2017-02-20 | 2018-08-28 | 百度在线网络技术（北京）有限公司 | Determination method, apparatus, equipment and the storage medium of activity time |
| CN108683743A (en) * | 2018-05-21 | 2018-10-19 | 陕西师范大学 | A kind of method for allocating tasks by multiple mobile terminal gathered datas |
| 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 |
| CN109299728A (en) * | 2018-08-10 | 2019-02-01 | 深圳前海微众银行股份有限公司 | Federal learning method, system and readable storage medium storing program for executing |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110598870A (en) | 2019-12-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110598870B (en) | Federal learning method and device | |
| CN112181666B (en) | Equipment assessment and federal learning importance aggregation method based on edge intelligence | |
| CN110443375B (en) | Method and device for federated learning | |
| Zhang et al. | Deep reinforcement learning (DRL)-based device-to-device (D2D) caching with blockchain and mobile edge computing | |
| Yuan et al. | CSEdge: Enabling collaborative edge storage for multi-access edge computing based on blockchain | |
| Wang et al. | A novel reputation-aware client selection scheme for federated learning within mobile environments | |
| Liwang et al. | Game theory based opportunistic computation offloading in cloud-enabled IoV | |
| CN107889082B (en) | D2D device discovery method utilizing social relationship among users | |
| US20220294706A1 (en) | Methods, apparatus and machine-readable media relating to machine-learning in a communication network | |
| CN114301935B (en) | Reputation-based internet of things edge cloud collaborative federal learning node selection method | |
| Aghazadeh et al. | Proactive content caching in edge computing environment: A review | |
| CN110535770A (en) | A kind of video flowing method for intelligently routing based on QoS perception under SDN environment | |
| CN103430511B (en) | The method and apparatus of replicate data in the peer-to-peer network including multiple equipment | |
| Banerjee et al. | On the implementation of cognitive autonomous networks | |
| Huang et al. | A reliable and fair federated learning mechanism for mobile edge computing | |
| CN102546230A (en) | Overlay-network topological optimization method in P2P (Peer-To-Peer) streaming media system | |
| Shen et al. | CCLBR: Congestion control-based load balanced routing in unstructured P2P systems | |
| CN109040075B (en) | Management method, server and system for nodes in wireless mobile sensor network | |
| CN111798223A (en) | Data collection method and device, storage medium and electronic equipment | |
| Brik et al. | Gss-vf: A game-theoretic approach for service discovery in fog network of vehicles | |
| CN115186833A (en) | Federal learning method, system, device and storage medium | |
| CN115001692A (en) | Model updating method and device, computer readable storage medium and electronic device | |
| Yin et al. | Research on the content sharing system for mobile edge caching networks: a hierarchical architecture | |
| Pouryazdan et al. | On coalitional and non-coalitional games in the design of user incentives for dependable mobile crowdsensing services | |
| Liu et al. | Deep reinforcement learning (DRL)-based transcoder selection for blockchain-enabled video streaming |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |