WO2021174934A1

WO2021174934A1 - Federated model parameter determination method, apparatus and device, and computer storage medium

Info

Publication number: WO2021174934A1
Application number: PCT/CN2020/134668
Authority: WO
Inventors: 鞠策; 高大山; 魏锡光
Original assignee: 深圳前海微众银行股份有限公司
Priority date: 2020-03-06
Filing date: 2020-12-08
Publication date: 2021-09-10
Also published as: CN111382875A

Abstract

A federated model parameter determination method, apparatus and device, and a computer storage medium. Said method comprises: acquiring multiple parameters of a target model, and performing random encoding and combination processing on the multiple parameters to obtain multiple groups of first parameter codes; respectively performing a federated learning process on the target model to obtain multiple federated models, and determining the accuracy rates of the multiple federated models; if among the multiple federated models, a federated model of which the accuracy rate meets a preset condition does not converge, on the basis of the accuracy rates of the multiple federated models, selecting multiple groups of second parameter codes from the multiple groups of first parameter codes for crossover and mutation processing to obtain multiple groups of third parameter codes, and then performing a federated learning process on the target model; and if among the multiple federated models, a federated model of which the accuracy rate meets the preset condition converges, determining a first target parameter encoding group corresponding to the converged federated model, and determining first target parameters corresponding to the first target parameter encoding group as federated model parameters.

Description

Method, device, equipment and computer storage medium for determining parameters of federation model

This application requires: the priority of the Chinese patent application filed on March 6, 2020, with the application number 202010155526.2, and the name "Federal model parameter determination method, device, equipment and storage medium", which is hereby incorporated by reference.

Technical field

This application relates to the technical field of financial technology (Fintech), and in particular to a method, device, equipment, and computer storage medium for determining federal model parameters.

Background technique

With the continuous development of financial technology (Fintech), especially Internet technology finance, more and more technologies (such as artificial intelligence, big data analysis, cloud storage, etc.) are applied in the financial field, but the financial field also proposes various technologies Higher requirements, such as requiring more accurate generation of the parameters of the federated model, in order to improve the accuracy of the federated model's prediction.

The parameters of the current federation model are usually obtained by training data collected by various sensors, but the built-in parameters of different sensors are different, resulting in different batches of data used for federation model training; after training, the model parameters of the federation model are obtained. Later, the use of the trained federated model for prediction has poor adaptability; in some scenarios, the prediction accuracy rate is high, while in other scenarios, the prediction accuracy rate is low, and the expected prediction effect cannot be achieved.

The above content is only used to assist the understanding of the technical solution of the application, and does not mean that the above content is recognized as prior art.

Technical solutions

The main purpose of this application is to provide a method, device, equipment, and computer storage medium for determining the parameters of a federated model. Technical problem of poor compatibility.

In order to achieve the above-mentioned purpose, the present application provides a method for determining parameters of a federation model. The method for determining parameters of a federation model includes the following steps:

Acquire multiple parameters of the target model, and perform random coding and combination processing on the multiple parameters to obtain multiple sets of first parameter codes;

Performing a federated learning process on the target model based on multiple sets of the first parameter codes to obtain multiple federated models, and determine the accuracy of the multiple federated models;

If the federated model whose accuracy meets the preset condition among the multiple federated models does not converge, then based on the accuracy of the multiple federated models, select multiple sets of second parameter encoding from multiple sets of the first parameter encoding. Cross mutation processing to obtain multiple sets of third parameter codes, and based on multiple sets of third parameter codes, perform a federated learning process on the target model; and

If the federated model whose accuracy meets the preset condition among the plurality of federated models converges, the first target parameter encoding group corresponding to the convergent federated model is determined, and the first target parameter corresponding to the first target parameter encoding group is determined , Determined as the federated model parameter.

In one embodiment, the step of selecting multiple sets of second parameter codes from multiple sets of the first parameter codes to perform cross-mutation processing based on the accuracy rates of the multiple federated models to obtain multiple sets of third parameter codes include:

Comparing the accuracy rates of the multiple federated models with a preset threshold to determine a target accuracy rate of the accuracy rates of the multiple federated models that is greater than the preset threshold;

Screening multiple sets of the first parameter codes according to each of the target accuracy rates to obtain multiple sets of second parameter codes; and

Cross-mutation processing is performed on multiple sets of the second parameter codes to obtain multiple sets of third parameter codes.

In an embodiment, the step of performing cross-mutation processing on multiple sets of the second parameter codes to obtain multiple sets of third parameter codes includes:

Randomly divide the multiple groups of the second parameter codes into multiple data group categories according to the preset number of cross items;

According to the preset number of crossover bits, cross each of the second parameter encoding groups in each of the data group classes to generate multiple sets of to-be-mutated second parameter codes; and

Perform mutation processing on multiple sets of the second parameter codes to be mutated to generate multiple sets of third parameter codes.

In an embodiment, the step of performing mutation processing on multiple sets of the second parameter codes to be mutated, and generating multiple sets of third parameter codes includes:

According to the preset mutation ratio, randomly select the parameter codes to be mutated in each of the second parameter codes to be mutated; and

Each of the multiple sets of the second parameter codes to be mutated is respectively mutated to generate multiple sets of third parameter codes.

In an embodiment, before the step of if the federated model whose accuracy rate meets the preset condition among the plurality of federated models does not converge, the method further includes:

Comparing the accuracy rates of the multiple federated models to determine the target accuracy rate with the largest value among the accuracy rates of the multiple federated models; and

The federated model with the target accuracy rate is determined as the federated model whose accuracy rate meets the preset condition among the plurality of federated models.

In an embodiment, after the step of determining the federated model with the target accuracy rate as the federated model whose accuracy rate meets a preset condition among the plurality of federated models, the method further includes:

It is determined whether the federated model whose accuracy rate meets the preset condition among the plurality of federated models has converged.

In an embodiment, the step of determining the accuracy of multiple federated models includes:

Acquiring test results generated by processing preset test data in a plurality of federated models in the federated learning process; and

The reference results corresponding to the preset test data are respectively compared with each of the test results to generate accuracy rates of multiple federated models.

In an embodiment, after the step of judging whether the federated model whose accuracy rate meets the preset condition among the plurality of federated models has converged, the method further includes: if it is determined that the accuracy rate of the federated models meets the preset condition If the conditional federated model does not converge, the federated learning is re-executed on the target model to generate new multiple federated models whose accuracy rate meets the preset condition to determine whether to converge.

In an embodiment, the crossover mutation processing is implemented by a genetic algorithm, where the genetic algorithm includes selection, crossover, and mutation.

In one embodiment, the preset number of crossing items includes: crossing two sets of second parameter codes and crossing three sets of third parameter codes; the preset number of crossing bits includes: setting the number of crossing bits as Five-five crosses, four-six crosses.

In an embodiment, the step of mutating each of the multiple sets of the second parameter codes to be mutated to be mutated respectively, and the step of generating multiple sets of third parameter codes includes: preset mutation positions for mutation, wherein The mutation position represents the position of the mutation code, and the position includes multiple positions and single-out positions.

In an embodiment, in order to achieve the foregoing objective, the present application further provides a federated model parameter determining device, the federated model parameter determining device including:

The acquiring module is used to acquire multiple parameters of the target model, and perform random coding and combination processing on the multiple parameters to obtain multiple sets of first parameter codes;

An execution module, configured to execute a federated learning process on the target model based on multiple sets of the first parameter codes to obtain multiple federated models, and determine the accuracy of the multiple federated models;

The selection module is configured to select multiple groups from multiple sets of the first parameter codes based on the accuracy rates of the multiple federation models if the federation models whose accuracy rates meet the preset conditions among the multiple federation models do not converge Perform cross-mutation processing on the second parameter encoding to obtain multiple sets of third parameter encodings, and then perform a federated learning process on the target model based on the multiple sets of third parameter encodings; and

The determining module is used for determining the first target parameter encoding group corresponding to the convergent federated model if the federal model whose accuracy rate meets the preset condition among the plurality of federal models converges, and corresponding the first target parameter encoding group The first target parameter of is determined as the federated model parameter.

In an embodiment, the selection module further includes:

A comparison unit, configured to compare the accuracy rates of the multiple federation models with a preset threshold, and determine a target accuracy rate of the accuracy rates of the multiple federation models that is greater than the preset threshold;

A screening unit, configured to screen multiple sets of the first parameter codes according to the target accuracy rates to obtain multiple sets of second parameter codes; and

The processing unit is configured to perform cross-mutation processing on multiple sets of the second parameter codes to obtain multiple sets of third parameter codes.

In an embodiment, the processing unit is further configured to:

In an embodiment, the federal model parameter determination device further includes:

The comparison module is used to compare the accuracy rates of multiple federated models and determine the target accuracy rate with the largest value among the accuracy rates of the multiple federated models; and

The determining module is further configured to determine the federation model with the target accuracy rate as a federation model whose accuracy rate meets a preset condition among the plurality of federation models.

The judging module is used for judging whether the federated model whose accuracy rate meets the preset condition among the multiple federated models has converged.

In an embodiment, the execution module further includes:

An obtaining unit, configured to obtain test results generated by processing preset test data on a plurality of federated models in the federated learning process; and

The generating unit is configured to compare the reference results corresponding to the preset test data with each of the test results to generate the accuracy rates of multiple federated models.

In an embodiment, in order to achieve the above object, the present application also provides a federated model parameter determining device. The federated model parameter determining device includes a memory, a processor, and is stored on the memory and can be stored on the processor. A running federation model parameter determination program, which implements the steps of the federation model parameter determination method described above when the federation model parameter determination program is executed by the processor.

In an embodiment, in order to achieve the above object, the present application also provides a computer storage medium, the computer storage medium stores a federated model parameter determination program, and the federated model parameter determination program is executed by a processor to achieve the above The steps of the method for determining the parameters of the federated model described above.

The method for determining the parameters of the federated model of the present application firstly performs random coding combinations on the acquired multiple parameters of the target model to generate multiple sets of first parameter codes, and respectively execute the federated learning process on the target model based on the multiple sets of first parameter codes. Obtain multiple federated models, and determine the accuracy of multiple federated models; after that, if the federated model whose accuracy is determined to meet the preset conditions among multiple federated models does not converge, then based on the accuracy of multiple federated models, the accuracy of multiple federated models is determined. In the first parameter encoding, multiple sets of second parameter encodings are selected for cross-mutation processing to obtain multiple sets of third parameter encodings, and based on the multiple sets of third parameter encodings, continue the federated learning process for the target model; if multiple federated models are determined If the federated model whose medium accuracy meets the preset condition converges, the first target parameter encoding group corresponding to the converged federated model is determined, and the first target parameter corresponding to the first target parameter encoding group is determined as the federated model parameter. The accuracy of multiple federated models characterizes the accuracy of the data processing of the federated model with the corresponding parameters of each group of first parameter encoding; the higher the accuracy, the higher the degree of applicability of the parameters; thus, the accuracy of each parameter Combine the convergent federated model with the rate to determine the federated model parameters. In the process of determining the parameters of the federation model, multiple sets of second parameter codes are selected from multiple sets of first parameter codes to perform cross-mutation processing to obtain the third parameter code for continuing training, which avoids the difference between each batch The model parameters are trained using flexible training data, which leads to the problem of poor adaptability of the federated model, which improves the accuracy of the federated model's prediction for different scenarios.

Description of the drawings

FIG. 1 is a schematic diagram of the structure of the device hardware operating environment involved in the embodiment of the device for determining the parameters of the federation model of the application;

FIG. 2 is a schematic flowchart of the first embodiment of the method for determining the parameters of the federation model of this application;

Fig. 3 is a schematic diagram of functional modules of a preferred embodiment of a device for determining parameters of a federated model of this application.

The realization, functional characteristics, and advantages of the purpose of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

Embodiments of the present invention

It should be understood that the specific embodiments described here are only used to explain the application, and are not used to limit the application.

This application provides a device for determining parameters of a federated model. Referring to FIG. 1, FIG. 1 is a schematic structural diagram of a device hardware operating environment involved in an embodiment of a device for determining a device for determining parameters of a federated model of this application.

As shown in FIG. 1, the federation model parameter determination device may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a magnetic disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

Those skilled in the art can understand that the hardware structure of the device for determining the parameters of the federation model shown in FIG. 1 does not constitute a limitation on the device for determining the parameters of the federation model. Components, or different component arrangements.

As shown in Fig. 1, a memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a federation model parameter determination program. Among them, the operating system is a program that manages and controls the federation model parameter determination equipment and software resources, and supports the operation of the network communication module, user interface module, federation model parameter determination program, and other programs or software; the network communication module is used to manage and control the network Interface 1004: The user interface module is used to manage and control the user interface 1003.

In the hardware structure of the device for determining the parameters of the federation model shown in Figure 1, the network interface 1004 is mainly used to connect to the back-end server and communicate with the back-end server; the user interface 1003 is mainly used to connect to the client (user side) and communicate with the client. Data communication: The processor 1001 can call the federation model parameter determination program stored in the memory 1005, and perform the following operations:

If the federated model whose accuracy meets the preset condition among the multiple federated models does not converge, then based on the accuracy of the multiple federated models, select multiple sets of second parameter encoding from multiple sets of the first parameter encoding. Cross mutation processing to obtain multiple sets of third parameter codes, and based on multiple sets of third parameter codes, perform a federated learning process on the target model;

Screening multiple sets of the first parameter codes according to each target accuracy rate to obtain multiple sets of second parameter codes;

According to the preset number of crossover bits, cross each of the second parameter encoding groups in each of the data group classes to generate multiple sets of to-be-mutated second parameter codes;

According to the preset mutation ratio, randomly select the to-be-mutated parameter codes in each of the to-be-mutated second parameter coding groups;

In an embodiment, before the step of if the federated model whose accuracy rate meets the preset condition in the plurality of federated models does not converge, the processor 1001 may call the federated model parameter determination program stored in the memory 1005, and execute the following operate:

Comparing the accuracy rates of the multiple federated models to determine the target accuracy rate with the largest value among the accuracy rates of the multiple federated models;

In one embodiment, after the step of determining the federation model with the target accuracy rate as the federation model with the accuracy rate meeting preset conditions among the plurality of federation models, the processor 1001 may call the federation model stored in the memory 1005 The federation model parameter determination procedure, and performs the following operations:

Acquiring test results generated by processing preset test data for multiple federated models in the federated learning process;

The specific implementation of the device for determining the parameters of the federation model of the present application is basically the same as the following embodiments of the method for determining the federation model parameters, and will not be repeated here.

This application also provides a method for determining the parameters of the federation model.

Referring to FIG. 2, FIG. 2 is a schematic flowchart of a first embodiment of a method for determining a federal model parameter of this application.

The embodiment of the application provides an embodiment of the method for determining the parameters of the federation model. It should be noted that although the logical sequence is shown in the flowchart, in some cases, the sequence shown may be executed in a different order than here. Or the steps described. Specifically, the method for determining federation model parameters in this embodiment includes:

Step S10: Obtain multiple parameters of the target model, and perform random coding combination processing on the multiple parameters to obtain multiple sets of first parameter codes.

The method for determining the parameters of the federation model in this embodiment is applied to the server, and is suitable for determining the optimal model parameters of the federation model through the server. The federation model can be a horizontal federation model or a vertical federation model. This embodiment is preferably described as a horizontal federation model. The horizontal federation model is a joint model constructed based on horizontal federated learning. The horizontal federated learning is based on two data sets. When the user features overlap more and the user overlaps less, the data set is divided into horizontal (user dimension), and the part of the data with the same user characteristics but not the same user is taken out for training. The server generates multiple sets of parameters as the model parameters of the horizontal federation model, and the horizontal federation model is trained on the basis of the multiple sets of model parameters in the way of horizontal federation learning, and multiple training results are obtained. The target training result with the best convergence characteristics is found from the multiple training results, and the model parameter that generates the target training result is the optimal model parameter suitable for the horizontal federated model.

In one embodiment, the untrained federated model is used as the target model. Before the server generates multiple sets of model parameters for training the target model, it first sets multiple adjustable parameters for the target model. The parameters include, but are not limited to, the number of neurons in the neural network algorithm, the number of layers of the neural network, the window size, and the number of features. After that, the length l of each parameter is given, so that the adjustable range of each parameter is within [0,], and each parameter is acquired and encoded, and the binary code of each parameter is generated. Furthermore, according to the parameters, the binary codes of the parameters are randomly combined into multiple groups of first parameter codes with the same length and the same meaning for the target model to perform the federated learning process. It should be noted that each group of first parameter encoding includes the encoding of each parameter, but the encoding of each parameter takes a different value within the adjustable range, so that the value of multiple sets of first parameter encoding is different; The first parameter codes with the same group length, the same meaning, and different values.

Step S20: Perform a federated learning process on the target model respectively based on the multiple sets of the first parameter codes to obtain multiple federated models, and determine the accuracy of the multiple federated models;

In another embodiment, the target model executes a federated learning process based on each group of first parameter codes, and processes the training samples used for training to obtain multiple federated models, the number of federated models and the number of groups of first parameter codes Unanimous. In addition, preset test data for testing are set in advance, and the preset test data are respectively transmitted to multiple federated models, processed by multiple federated models, and the accuracy of multiple federated models is determined. The accuracy of a federated model for the processing of preset test data. Specifically, the step of determining the accuracy of multiple federated models includes:

Step S21: Obtain test results generated by processing preset test data for multiple federated models in the federated learning process;

In step S22, the reference results corresponding to the preset test data are respectively compared with each of the test results to generate accuracy rates of multiple federated models.

In one embodiment, each federated model uses the parameter corresponding to the first parameter code as its own network parameter during the federated learning process, runs with its own network parameter, and performs test processing on the preset test data to generate a test result. The server obtains the test results of each federation model, retrieves the reference results corresponding to the preset test data, compares the test results with each test result, and generates a characterization that each test result is consistent with the reference result. The numerical value of the degree of sexuality. The various values are the accuracy of each federation model running with the network parameters corresponding to each first parameter code, and the processing results generated by processing the preset test data, that is, the accuracy of multiple federation models. The higher the accuracy rate, the higher the accuracy of the federated model's processing of the preset test data. On the contrary, the lower the accuracy of the federated model's processing of the preset test data.

Step S30, if the federation models whose accuracy rates meet the preset conditions among the multiple federation models do not converge, then based on the accuracy rates of the multiple federation models, select multiple sets of second parameter codes from multiple sets of the first parameter codes. Perform cross-mutation processing on the parameter encoding to obtain multiple sets of third parameter encodings, and then perform a federated learning process on the target model based on the multiple sets of third parameter encodings;

In one embodiment, the accuracy rate is one of the basis for determining the optimal federated model parameters of the federation model, and the processing accuracy of the federated model represented by different accuracy rates is different. In order to determine the optimal federated model parameters, the accuracy of the federated model can be determined. The low federation model is eliminated. Specifically, if the federated model whose accuracy rate meets the preset condition among multiple federated models does not converge before the step, it also includes:

Step a: Comparing the accuracy rates of the multiple federation models to determine the target accuracy rate with the largest value among the accuracy rates of the multiple federation models;

Step b: Determine the federated model with the target accuracy rate as a federated model whose accuracy rate meets a preset condition among the plurality of federated models.

Step c: Judge whether the federated model whose accuracy rate meets the preset condition among the plurality of federated models has converged.

In one embodiment, a preset condition for removal is set, such as a numerical value condition or a maximum value condition; when the accuracy of the federation model satisfies the numerical value condition or the maximum value condition, the federated model is used to determine Federated model parameters, otherwise the federated model will not be used to determine the federated model parameters. In this embodiment, the preset condition for excluding the federated model with low accuracy is set as the maximum value condition. After the target model executes the federation learning process to obtain multiple federated models, and the accuracy of the federated models is determined, compare the accuracy of each federated model to determine the target accuracy with the largest value among the accuracy of multiple federated models . Then find the federated model with the target accuracy rate, and determine it as the federated model whose accuracy rate meets the preset conditions among multiple federated models, so as to realize the elimination of other federated models.

In one embodiment, it is determined whether the federal model remaining after the elimination operation, that is, the federal model whose accuracy meets the preset condition among the multiple federal models, is converged. Among them, the convergence characterization federation model processes the preset test data with the parameters corresponding to the first parameter encoding, and the test results obtained have the highest accuracy rate; or even if it does not have the highest accuracy rate every time, most of the times It is the highest accuracy rate, and there is little difference between the accuracy rate and the highest accuracy rate of other times. Therefore, the accuracy of the test results of the previous federated model processing preset test data can be used to determine whether the federated model converges.

In one embodiment, if it is determined that the federated model whose accuracy rate meets the preset condition among the multiple federated models does not converge, then the federated learning is performed on the target model again to generate a new multiple federated model judgement whose accuracy rate meets the preset condition Whether to converge. Specifically, because the federated model for convergence judgment is based on the premise that the accuracy rate meets the preset conditions, in order to ensure that the accuracy of each federated model meets the preset conditions, it is set to encode multiple sets of first parameters based on the accuracy of each federated model The screening mechanism is to screen out the second parameter code with relatively high characterization accuracy.

In an embodiment, the server performs cross-mutation processing on the second parameter encoding, and the cross-mutation processing in this embodiment is preferably implemented based on genetic algorithms. Genetic Algorithm (Genetic Algorithm) is a method of solving optimization problems through search. It first randomly generates a certain amount of population, and then the algorithm includes: Selection, Crossover and Mutation. The selection is the process of defining individual fitness, evaluating the fitness of each individual and selecting the higher fitness as the next round of population members; crossover is the process of encoding the chromosomes of the members of the population, and crossing the chromosome codes of the two or two groups of members The process of mutation; mutation is the process of mutating the chromosome code after crossover with a certain probability. In this embodiment, the process of randomly generating multiple sets of first parameter codes is a process of randomly generating a certain amount of population in the genetic algorithm and encoding the chromosomes of the members of the population. The federated model is used to process the preset test data, and the accuracy of the generated processing results is used as the fitness in the genetic algorithm. The process of screening multiple sets of first parameter codes based on the accuracy is the selection process of the genetic algorithm. After that, through the crossover and mutation in the genetic algorithm, the selected multiple sets of second parameter codes are processed to obtain multiple sets of third parameter codes, and based on the multiple sets of third reference codes, the federated learning process is performed again on the target model. Until the federated model whose accuracy rate meets the preset conditions among multiple federated models converges, the optimal federated model parameters of the federated model are determined.

Step S40: If the federated models whose accuracy meets the preset conditions among the plurality of federated models converge, determine the first target parameter encoding group corresponding to the converged federated model, and combine the first target parameter encoding group corresponding to the first target parameter encoding group. A target parameter is determined as the federated model parameter.

In one embodiment, if it is determined that the federated model whose accuracy rate meets the preset condition among the multiple federated models has converged, it means that the federated model processes the preset test data with the parameters corresponding to the first parameter encoding, and the resulting test The result has a higher accuracy rate. In this way, the first parameter encoding group in the convergent federated model is obtained as the first target parameter encoding group, and the first target parameter corresponding to the first target parameter encoding is found according to the conversion relationship between the encoding and the parameter; or directly The first target parameter encoding is inversely encoded to generate a first target parameter corresponding to the first target parameter encoding. The first target parameter is determined as the federated model parameter of the federated model, and the federated model runs with the first target parameter, and the processing result obtained by processing the data has high accuracy.

The method for determining the parameters of the federated model of the present application firstly performs random coding combinations on the acquired multiple parameters of the target model to generate multiple sets of first parameter codes, and respectively execute the federated learning process on the target model based on the multiple sets of first parameter codes. Obtain multiple federated models, and determine the accuracy of multiple federated models; after that, if the federated model whose accuracy is determined to meet the preset conditions among multiple federated models does not converge, then based on the accuracy of multiple federated models, the accuracy of multiple federated models is determined. In the first parameter encoding, multiple sets of second parameter encodings are selected for cross-mutation processing to obtain multiple sets of third parameter encodings, and based on the multiple sets of third parameter encodings, continue the federated learning process for the target model; if multiple federated models are determined If the federated model whose medium accuracy meets the preset condition converges, the first target parameter encoding group corresponding to the converged federated model is determined, and the first target parameter corresponding to the first target parameter encoding group is determined as the federated model parameter. The accuracy of multiple federated models characterizes the accuracy of the data processing of the federated model with the corresponding parameters of each group of first parameter encoding; the higher the accuracy, the higher the degree of applicability of the parameters; thus, the accuracy of each parameter Combine the convergent federated model with the rate to determine the federated model parameters. In the process of determining the parameters of the federation model, multiple sets of second parameter codes are selected from multiple sets of first parameter codes for cross-mutation processing to obtain the third parameter code for continuing training, which avoids the difference between each batch The model parameters are trained using flexible training data, which leads to the problem of poor adaptability of the federated model, which improves the accuracy of the federated model's prediction for different scenarios.

In one embodiment, based on the first embodiment of the method for determining federal model parameters of the present application, a second embodiment of the method for determining federal model parameters of the present application is proposed.

The difference between the second embodiment of the federated model parameter determination method and the first embodiment of the federated model parameter determination method is that the accuracy of the federated model is based on multiple sets of the first parameter encoding. The steps of selecting multiple sets of second parameter codes for cross-mutation processing, and obtaining multiple sets of third parameter codes include:

Step S31, comparing the accuracy rates of the multiple federated models with a preset threshold, and determine a target accuracy rate of the accuracy rates of the multiple federated models that is greater than the preset threshold;

Step S32, screening multiple sets of the first parameter codes according to each target accuracy rate to obtain multiple sets of second parameter codes;

Step S33: Perform cross-mutation processing on multiple sets of the second parameter codes to obtain multiple sets of third parameter codes.

In this embodiment, in order to screen multiple sets of first parameter codes according to the accuracy rate of each federation model, preset thresholds representing the accuracy rate are preset, and each accuracy rate is compared with the preset thresholds to find Among them, the target accuracy rate is greater than the preset threshold. Each accuracy rate is generated by the federated model according to each first parameter code. After the accuracy rate of each target is determined, the multiple sets of first parameter codes are screened and searched, and the first parameter code that generates each target accuracy rate is found. Encode as the second parameter. After that, cross-mutation processing is performed on multiple sets of second parameter codes to obtain multiple sets of third parameter codes for re-executing the federated learning process on the target model. Specifically, performing cross mutation processing on multiple sets of second parameter codes to obtain multiple sets of third parameter codes includes:

Step S311: randomly divide the multiple groups of the second parameter codes into multiple data group categories according to the preset number of cross-terms;

Step S312, according to the preset number of crossover bits, cross each of the second parameter encoding groups in each of the data group classes to generate multiple sets of to-be-mutated second parameter codes;

Step S313: Perform mutation processing on multiple sets of the second parameter codes to be mutated to generate multiple sets of third parameter codes.

In one embodiment, a preset number of cross items is set in advance according to requirements to represent the number of groups for performing crossover on the second parameter encoding, for example, two sets of second parameter codes are crossed, or three sets of third parameter codes are crossed. Perform crossover and so on. In this embodiment, preferably, two sets of second parameter codes are used for crossover, that is, two by two crossovers. During the crossover process, multiple sets of second parameter codes are randomly divided according to the preset number of crossover groups, and each two groups of second parameter codes are divided into a data group type, and the second parameter codes in each data group type are individually coded Make a cross.

In one embodiment, a preset number of crossover bits is set in advance according to requirements to characterize the number of data bits used for crossover in each set of second parameter codes. Each group of the second parameter codes of each group extracts one-half of the codes for crossover to form a new parameter code; or sets the number of crossover bits to four or six crosses, that is, from a group of second parameter codes in the data group class Two-fifths of the code is extracted from the second parameter code, and three-fifths of the code is extracted from another group of second parameter codes, and the two are crossed to form a new parameter code. In this embodiment, it is preferable to perform crossover with a preset crossover number of five to five crosses. According to the preset crossover number, a reading operation is performed on each set of second parameter codes in each data group; The cross-digit number embodies part of the parameter codes with the same number of codes, and the position of the read part of the parameter codes in the second parameter codes is not limited, that is, the second parameter codes are read randomly. After that, the read parameter codes are combined, and the position of the combination is not limited; that is, the read parameter codes are randomly combined to generate the second parameter code to be mutated in the data group type. The second parameter codes in each data group category are all crossed, and after the multiple sets of to-be-mutated second parameter codes corresponding to each data group category are obtained, the multiple sets of to-be-mutated second parameter codes are subjected to mutation processing, and through the mutation The processing generates multiple sets of third parameter codes. Specifically, performing mutation processing on multiple sets of second parameter codes to be mutated, and the step of generating multiple sets of third parameter codes includes:

Step S3131, randomly selecting the parameter codes to be mutated in each of the second parameter coding groups to be mutated according to the preset mutation ratio;

Step S3132: mutate each parameter code to be mutated in the multiple sets of second parameter codes to be mutated to generate multiple sets of third parameter codes.

In one embodiment, in order to make the sample data used for federated model training more abundant and accurate, this embodiment performs mutation processing on the second parameter code to be mutated based on the mutation process in the genetic algorithm, so as to generate other non-regularities through mutation. Featured parameter coding for federated training. Specifically, preset mutation ratios, such as 20%, 30%, etc., are set according to requirements in advance to represent the demand for encoding the number of mutations in the second parameter to be mutated. According to the preset mutation ratio, each group of to-be-encoded second parameter codes is randomly screened to obtain the number of to-be-mutated parameter codes in each of the to-be-mutated second parameter encoding groups that is consistent with the number represented by the mutation ratio.

In another embodiment, the code of the parameter to be mutated in each group of second parameter codes to be mutated is mutated, and the respective unmutated codes in each group of the second parameter to be mutated are generated together as the third parameter code. For the target model to perform a federated learning process based on multiple sets of third parameter codes to determine the federated model parameters. It should be noted that, in this embodiment, mutation may also be performed by pre-setting mutation positions, and the mutation position represents the position where the mutation code is located, so as to realize the mutation of the code at the fixed position. It can be multiple positions or single-out positions, such as the 3rd and 5th positions of the code, or the 2nd and 3rd positions. According to the preset mutation positions, the codes of the parameters to be mutated in the coding groups of the second parameter to be mutated are screened out for mutation, and multiple sets of third parameter samples are obtained.

In this embodiment, after multiple sets of second parameter codes are selected from multiple sets of first parameter codes, the multiple sets of second parameter codes are subjected to cross-mutation processing to generate multiple sets of third parameter codes used in the federated learning process to avoid The problem of poor adaptability caused by federated learning training based on the training data collected by each sensor ensures the accuracy of the federated model's prediction for different scenarios.

The application also provides a device for determining the parameters of the federation model.

Referring to Fig. 3, Fig. 3 is a schematic diagram of the functional modules of the first embodiment of the device for determining the parameters of the federation model of this application. The device for determining the parameters of the federation model includes:

The generating module 10 is used to obtain a module, which is used to obtain multiple parameters of the target model, and perform random coding and combination processing on the multiple parameters to obtain multiple sets of first parameter codes;

The execution module 20 is configured to execute a federated learning process on the target model based on multiple sets of the first parameter codes to obtain multiple federated models, and determine the accuracy of the multiple federated models;

The selecting module 30 is configured to select multiple sets of the first parameter codes based on the accuracy rates of the multiple federation models if the federation models whose accuracy rates meet the preset conditions among the multiple federation models do not converge. Perform cross-mutation processing on the set of second parameter codes to obtain multiple sets of third parameter codes, and then perform a federated learning process on the target model based on the multiple sets of third parameter codes;

The determining module 40 is configured to determine the first target parameter encoding group corresponding to the convergent federated model if the federal model whose accuracy rate meets the preset condition among the plurality of federated models converges, and combine the first target parameter encoding group The corresponding first target parameter is determined as the federated model parameter.

In an embodiment, the selection module 30 further includes:

A screening unit, configured to screen multiple sets of the first parameter codes according to each target accuracy rate to obtain multiple sets of second parameter codes;

In an embodiment, the processing unit is further configured to:

The comparison module is used to compare the accuracy rates of multiple federated models and determine the target accuracy rate with the largest value among the accuracy rates of the multiple federated models;

In an embodiment, the execution module 20 further includes:

An obtaining unit, configured to obtain test results generated by processing preset test data on a plurality of federated models in a federated learning process;

The specific implementation of the federal model parameter determination device of the present application is basically the same as each embodiment of the aforementioned federal model parameter determination method, and will not be repeated here.

In addition, the embodiment of the present application also proposes a computer storage medium.

The computer storage medium stores a federation model parameter determination program, and the federation model parameter determination program is executed by the processor to realize the steps of the federation model parameter determination method as described above.

The computer storage medium of the present application may be a computer-readable storage computer storage medium, and its specific implementation is basically the same as each embodiment of the above-mentioned federated model parameter determination method, and will not be repeated here.

The embodiments of the application are described above with reference to the accompanying drawings, but the application is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are only illustrative and not restrictive. Those of ordinary skill in the art are Under the enlightenment of this application, without departing from the purpose of this application and the scope of protection of the claims, many forms can be made, any equivalent structure or equivalent process transformation made by using the content of the description and drawings of this application, or It is directly or indirectly used in other related technical fields, and these all fall within the protection of this application.

Claims

A method for determining the parameters of a federation model, wherein the determination of the parameters of the federation model includes the following steps:

Acquire multiple parameters of the target model, and perform random coding and combination processing on the multiple parameters to obtain multiple sets of first parameter codes;

Performing a federated learning process on the target model based on multiple sets of the first parameter codes to obtain multiple federated models, and determine the accuracy of the multiple federated models;

If the federated model whose accuracy meets the preset condition among the multiple federated models does not converge, then based on the accuracy of the multiple federated models, select multiple sets of second parameter encoding from multiple sets of the first parameter encoding. Cross mutation processing to obtain multiple sets of third parameter codes, and based on multiple sets of third parameter codes, perform a federated learning process on the target model;

If the federated model whose accuracy meets the preset condition among the plurality of federated models converges, the first target parameter encoding group corresponding to the convergent federated model is determined, and the first target parameter corresponding to the first target parameter encoding group is determined , Determined as the federated model parameter.
The method for determining the parameters of a federation model according to claim 1, wherein said selecting a plurality of sets of second parameter codes from a plurality of sets of said first parameter codes for cross mutation processing based on the accuracy rates of a plurality of said federation models, The steps of obtaining multiple sets of third parameter codes include:

Comparing the accuracy rates of the multiple federated models with a preset threshold to determine a target accuracy rate of the accuracy rates of the multiple federated models that is greater than the preset threshold;

Screening multiple sets of the first parameter codes according to each target accuracy rate to obtain multiple sets of second parameter codes;

Cross-mutation processing is performed on multiple sets of the second parameter codes to obtain multiple sets of third parameter codes.
The method for determining the parameters of a federation model according to claim 2, wherein the step of performing cross-mutation processing on multiple sets of the second parameter codes to obtain multiple sets of third parameter codes comprises:

Randomly divide the multiple groups of the second parameter codes into multiple data group categories according to the preset number of cross items;

According to the preset number of crossover bits, cross each of the second parameter encoding groups in each of the data group classes to generate multiple sets of to-be-mutated second parameter codes; and

Perform mutation processing on multiple sets of the second parameter codes to be mutated to generate multiple sets of third parameter codes.
The method for determining the parameters of a federation model according to claim 3, wherein the step of performing mutation processing on multiple sets of the second parameter codes to be mutated, and generating multiple sets of third parameter codes comprises:

According to the preset mutation ratio, randomly select the parameter codes to be mutated in each of the second parameter codes to be mutated; and

Each of the multiple sets of the second parameter codes to be mutated is respectively mutated to generate multiple sets of third parameter codes.
The method for determining the parameters of a federation model according to claim 1, wherein, before the step of if the federation model whose accuracy rate meets the preset condition among the plurality of federation models does not converge, the method further comprises:

Comparing the accuracy rates of the multiple federated models to determine the target accuracy rate with the largest value among the accuracy rates of the multiple federated models;

The federated model with the target accuracy rate is determined as the federated model whose accuracy rate meets the preset condition among the plurality of federated models.
The method for determining the parameters of a federation model according to claim 5, wherein after the step of determining the federation model with the target accuracy rate as the federation model with the accuracy rate meeting a preset condition among the plurality of federation models, The method also includes:

It is determined whether the federated model whose accuracy rate meets the preset condition among the plurality of federated models has converged.
The method for determining the parameters of a federation model according to any one of claims 1 to 6, wherein the step of determining the accuracy of a plurality of federation models comprises:

Acquiring test results generated by processing preset test data for multiple federated models in the federated learning process;

The reference results corresponding to the preset test data are respectively compared with each of the test results to generate accuracy rates of multiple federated models.
The method for determining the parameters of a federation model according to claim 6, wherein after the step of judging whether the federated model whose accuracy rate meets the preset condition among the plurality of federated models has converged, the method further comprises

If it is determined that the federated model whose accuracy rate meets the preset condition among multiple federated models does not converge, re-execute federated learning on the target model to generate a new multiple federated model whose accuracy rate meets the preset condition to determine whether convergence.
The method for determining the parameters of a federation model according to claim 1, wherein:

The crossover mutation processing is implemented by genetic algorithm, where the genetic algorithm includes: selection, crossover and mutation.
The method for determining the parameters of a federation model according to claim 2, wherein:

The preset number of cross items includes: cross two sets of second parameter codes and cross three sets of third parameter codes;

The preset number of crossovers includes: setting the number of crossovers to five or five crosses, or four to six crosses.
The method for determining the parameters of a federation model according to claim 4, wherein the step of mutating each of the multiple sets of the second parameter codes to be mutated separately, and generating multiple sets of third parameter codes comprises:

The mutation position is preset for mutation, wherein the mutation position represents the position where the mutation code is located, and the position includes multiple positions and single-out positions.
A federation model parameter determination device, wherein the federation model parameter determination device includes:

The acquiring module is used to acquire multiple parameters of the target model, and perform random coding and combination processing on the multiple parameters to obtain multiple sets of first parameter codes;

An execution module, configured to execute a federated learning process on the target model based on multiple sets of the first parameter codes to obtain multiple federated models, and determine the accuracy of the multiple federated models;

The selection module is configured to select multiple groups from multiple sets of the first parameter codes based on the accuracy rates of the multiple federation models if the federation models whose accuracy rates meet the preset conditions among the multiple federation models do not converge Perform cross-mutation processing on the second parameter encoding to obtain multiple sets of third parameter encodings, and then perform a federated learning process on the target model based on the multiple sets of third parameter encodings; and

The determining module is used for determining the first target parameter encoding group corresponding to the convergent federated model if the federal model whose accuracy rate meets the preset condition among the plurality of federal models converges, and corresponding the first target parameter encoding group The first target parameter of is determined as the federated model parameter.
The device for determining federal model parameters according to claim 12, wherein the selection module further comprises:

A comparison unit, configured to compare the accuracy rates of the multiple federation models with a preset threshold, and determine a target accuracy rate of the accuracy rates of the multiple federation models that is greater than the preset threshold;

A screening unit, configured to screen multiple sets of the first parameter codes according to the target accuracy rates to obtain multiple sets of second parameter codes; and

The processing unit is configured to perform cross-mutation processing on multiple sets of the second parameter codes to obtain multiple sets of third parameter codes.
The device for determining federal model parameters according to claim 13, wherein the processing unit is further configured to:

Randomly divide the multiple groups of the second parameter codes into multiple data group categories according to the preset number of cross items;

According to the preset number of crossover bits, cross each of the second parameter encoding groups in each of the data group classes to generate multiple sets of to-be-mutated second parameter codes; and

Perform mutation processing on multiple sets of the second parameter codes to be mutated to generate multiple sets of third parameter codes.
The device for determining federal model parameters according to claim 13, wherein the processing unit is further configured to:

According to the preset mutation ratio, randomly select the parameter codes to be mutated in each of the second parameter codes to be mutated; and

Each of the multiple sets of the second parameter codes to be mutated is respectively mutated to generate multiple sets of third parameter codes.
The federation model parameter determination device according to claim 13, wherein the federation model parameter determination device further comprises:

The comparison module is used to compare the accuracy rates of multiple federated models and determine the target accuracy rate with the largest value among the accuracy rates of the multiple federated models; and

The determining module is further configured to determine the federation model with the target accuracy rate as a federation model whose accuracy rate meets a preset condition among the plurality of federation models.
The federated model parameter determining device according to claim 12, wherein the federated model parameter determining device further comprises:

The judging module is used for judging whether the federated model whose accuracy rate meets the preset condition among the multiple federated models has converged.
The device for determining federal model parameters according to claim 12, wherein the execution module further comprises:

An obtaining unit, configured to obtain test results generated by processing preset test data on a plurality of federated models in the federated learning process; and

The generating unit is configured to compare the reference results corresponding to the preset test data with each of the test results to generate the accuracy rates of multiple federated models.
A federated model parameter determining device, wherein the federated model parameter determining device includes a memory, a processor, and a federated model parameter determining program stored on the memory and running on the processor, and the federated model parameter When the determining program is executed by the processor, the steps of the method for determining the parameters of the federation model according to any one of claims 1-11 are implemented.
A computer storage medium, wherein a federated model parameter determination program is stored on the computer storage medium, and the federated model parameter determination program is executed by a processor to realize the federated model parameter according to any one of claims 1-11 Determine the method steps.