CN110633805A

CN110633805A - Longitudinal federated learning system optimization method, device, equipment and readable storage medium

Info

Publication number: CN110633805A
Application number: CN201910918262.9A
Authority: CN
Inventors: 程勇; 刘洋; 陈天健
Original assignee: WeBank Co Ltd
Current assignee: WeBank Co Ltd
Priority date: 2019-09-26
Filing date: 2019-09-26
Publication date: 2019-12-31
Anticipated expiration: 2039-09-26
Also published as: CN110633805B; WO2021004551A1

Abstract

The invention discloses a method, a device, equipment and a readable storage medium for optimizing a longitudinal federated learning system, wherein the method comprises the following steps: the method comprises the steps that sample alignment is carried out on first equipment and second equipment to obtain first sample data of the first equipment, wherein the data characteristics of the first sample data are different from those of second sample data, and the second sample data are obtained by carrying out sample alignment on the second equipment and the first equipment; and performing cooperative training by adopting the first sample data and the second equipment to obtain an interpolation model, wherein the interpolation model is used for inputting data belonging to the data characteristics corresponding to the first equipment and outputting prediction data belonging to the data characteristics corresponding to the second equipment. When the longitudinal federal learning trained model is used by the participants of longitudinal federal learning, the model can be independently used without the cooperation of other participants, and the application range of the longitudinal federal learning is expanded.

Description

Longitudinal federated learning system optimization method, device, equipment and readable storage medium

Technical Field

The invention relates to the technical field of machine learning, in particular to a method, a device, equipment and a readable storage medium for optimizing a longitudinal federated learning system.

Background

With the development of artificial intelligence, people provide a concept of 'federal learning' for solving the problem of data islanding, so that both federal parties can train a model to obtain model parameters without providing own data, and the problem of data privacy disclosure can be avoided.

In the longitudinal federated learning, under the condition that the data features of the participants are overlapped less and the users are overlapped more, the part of the users and the data with the same users and different user data features of the participants are taken out to jointly train the machine learning model. For example, there are two participants a and B belonging to the same region, where participant a is a bank and participant B is an e-commerce platform. Participants a and B have more users in the same area, but a and B have different services and different recorded user data characteristics. In particular, the user data characteristics of the a and B records may be complementary. In such a scenario, vertical federated learning may be used to help a and B build a joint machine learning predictive model, helping a and B provide better service to their customers.

However, when using a model trained through longitudinal federal learning, participant a needs to collaborate with participant B to use the model for prediction. For example, participant a has data characteristics X3, X4 and X5 and participant B has data characteristics X1 and X2, and when participant a needs to predict a new client, participant a needs to communicate with participant B to query whether participant B also has the client because participant a does not have all of the data characteristics, i.e., no data for the client under data characteristics X1 and X2, and participant a cannot predict the client if participant B does not have the client's data. Even though participant B has the client's data, participant a and B need to cooperate to make predictions about the client.

Disclosure of Invention

The invention mainly aims to provide a method, a device, equipment and a readable storage medium for optimizing a longitudinal federal learning system, aiming at realizing that a participant in longitudinal federal learning can independently use a model without the cooperation of other participants when using the model trained through longitudinal federal learning.

In order to achieve the above object, the present invention provides a longitudinal federal learning system optimization method, which is applied to a first device, wherein the first device is in communication connection with a second device, and the longitudinal federal learning system optimization method includes the following steps:

performing sample alignment with the second device to obtain first sample data of the first device, wherein the first sample data has different data characteristics from second sample data obtained by performing sample alignment between the second device and the first device;

and obtaining an interpolation model by adopting the first sample data and the second equipment for cooperative training, wherein the interpolation model is used for inputting data belonging to the data characteristics corresponding to the first equipment and outputting data belonging to the data characteristics corresponding to the second equipment.

Optionally, the step of obtaining an interpolation model by using the first sample data and the second device for training in cooperation includes:

inputting the first sample data into a first partial model preset in the first equipment to obtain a first output;

sending the first output to the second equipment, so that the second equipment obtains a second output of a preset second partial model according to the first output, calculates a first loss function and first gradient information according to the second sample data and the second output, and updates parameters of the second partial model according to gradient information related to the second partial model in the first gradient information;

updating parameters of the first part of models according to gradient information related to the first part of models in the first gradient information received from the second equipment, and iteratively training until a preset stopping condition is met, and receiving the second part of models sent by the second equipment;

and combining the first partial model and the second partial model to obtain the interpolation model.

Optionally, after the step of combining the first partial model and the second partial model to obtain the interpolation model, the method further includes:

inputting local sample data belonging to the data characteristics corresponding to the first equipment into the interpolation model to obtain prediction sample data belonging to the data characteristics corresponding to the second equipment;

and locally training a preset machine learning model to be trained by adopting the local sample data and the prediction sample data to obtain a target machine learning model.

Optionally, a first Trusted Execution Environment (TEE) module is included in the first device, a second TEE module is included in the second device,

the sending the first output to the second device, so that the second device obtains a second output of a preset second partial model according to the first output, calculates a first loss function and first gradient information according to the second sample data and the second output, and updates parameters of the second partial model according to gradient information related to the second partial model in the first gradient information, where the parameters include:

encrypting the first output to obtain a first encrypted output;

sending the first encrypted output to the second device, so that the second device decrypts the first encrypted output in the second TEE module to obtain the first output, obtains a second output preset with a second partial model according to the first output, calculates a first loss function and first gradient information according to the second sample data and the second output, updates parameters of the second partial model according to gradient information related to the second partial model in the first gradient information, and encrypts gradient information related to the first partial model in the first gradient information to obtain encrypted gradient information;

the step of updating the parameters of the first partial model according to the gradient information related to the first partial model in the first gradient information received from the second device comprises:

and receiving the encrypted gradient information sent by the second device, decrypting the encrypted gradient information in the first TEE module to obtain gradient information related to the first part model in the first gradient information, and updating parameters of the first part model according to the gradient information related to the second part model.

Optionally, after the step of sending the first output to the second device, the method further includes:

receiving the second output and the first loss function sent by the second device;

inputting the first sample data and the second output into a preset machine learning model to be trained to obtain predicted label data;

calculating a second loss function and second gradient information of the machine learning model to be trained according to the predicted label data and pre-stored local actual label data;

the step of receiving the second partial model sent by the second device when the iterative training is detected to meet the preset stop condition comprises:

and updating parameters of the machine learning model to be trained according to the second gradient information, performing iterative training to minimize a fusion loss function until a target machine learning model is obtained when a preset stopping condition is met, and receiving the second part of model sent by the second equipment, wherein the first equipment fuses the first loss function and the second loss function to obtain the fusion loss function.

Optionally, the first device includes a TEE module, and the step of obtaining an interpolation model by using the first sample data and the second device for training in cooperation includes:

receiving second encryption sample data sent by the second device, wherein the second device encrypts the second sample data to obtain the second encryption sample data;

and decrypting the second encrypted sample data in the TEE module to obtain second sample data, and training the interpolation model to be trained according to the first sample data and the second sample data to obtain the interpolation model.

Optionally, the target machine learning model is configured to predict a purchase intention of a user, and after the step of performing local training on a preset machine learning model to be trained by using the local sample data and the predicted sample data to obtain the target machine learning model, the method further includes:

inputting first data of a target user into the interpolation model to obtain second data, wherein the data characteristics of the first data comprise user identity characteristics, and the data characteristics of the second data comprise user purchase characteristics;

and inputting the first data and the second data into the target machine learning model to obtain the purchase intention of the target user.

In order to achieve the above object, the present invention further provides a longitudinal federal learning system optimization device, where the longitudinal federal learning system optimization device is disposed on a first device, and the first device is in communication connection with a second device, and the longitudinal federal learning system optimization device includes:

an alignment module, configured to perform sample alignment with the second device to obtain first sample data of the first device, where data characteristics of the first sample data are different from those of second sample data, and the second sample data is obtained by performing sample alignment between the second device and the first device;

and the training module is used for obtaining an interpolation model by adopting the first sample data and the second equipment for cooperative training, wherein the interpolation model is used for inputting data belonging to the data characteristics corresponding to the first equipment and outputting prediction data belonging to the data characteristics corresponding to the second equipment.

In order to achieve the above object, the present invention further provides a longitudinal federal learning system optimization device, including: a memory, a processor, and a longitudinal federated learning system optimization program stored on the memory and executable on the processor, the longitudinal federated learning system optimization program when executed by the processor implementing the steps of the longitudinal federated learning system optimization method as described above.

In addition, to achieve the above object, the present invention further provides a computer readable storage medium, on which a longitudinal federal learning system optimization program is stored, and the longitudinal federal learning system optimization program, when executed by a processor, implements the steps of the longitudinal federal learning system optimization method as described above.

In the invention, the first device performs sample alignment with the second device, and performs collaborative training by using the aligned sample data to obtain an interpolation model capable of complementing the missing data characteristics of the first device relative to the second device, so that when the first device uses the machine learning model obtained by training the data characteristics of the first device and the second device, even if the first device does not have the data of the second device, the first device can independently predict the data belonging to the corresponding data characteristics of the second device through the interpolation model, thereby completing the prediction by using the machine learning model through the complemented data, expanding the application range of longitudinal federal learning, and avoiding that the first device cannot use the machine learning model obtained by longitudinal federal learning under the scene that the second device cannot provide data for the first device.

Drawings

FIG. 1 is a schematic diagram of a hardware operating environment according to an embodiment of the present invention;

FIG. 2 is a schematic flow chart of a first embodiment of a longitudinal federated learning system optimization method of the present invention;

FIG. 3 is a schematic diagram of sample data according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a cutting method of a mold according to an embodiment of the present invention;

FIG. 5 is a diagram illustrating a first device alone training a machine learning model according to an embodiment of the present invention;

fig. 6 is a schematic flowchart of a process of training an interpolation model by cooperation of a first device and a second device and training a machine learning model by the first device alone according to an embodiment of the present invention;

fig. 7 is a schematic diagram illustrating a first device and a second device cooperatively training an interpolation model in a respective TEE module environment according to an embodiment of the present invention;

fig. 8 is a schematic diagram illustrating a first device and a second device cooperatively training an interpolation model and a machine learning model according to an embodiment of the present invention;

FIG. 9 is a block diagram of a preferred embodiment of the longitudinal federated learning system optimization apparatus of the present invention.

The implementation, functional features and advantages of the objects of the present invention will be further explained with reference to the accompanying drawings.

Detailed Description

It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

As shown in fig. 1, fig. 1 is a schematic device structure diagram of a hardware operating environment according to an embodiment of the present invention.

It should be noted that, in the embodiment of the present invention, the longitudinal federal learning system optimization device may be a smart phone, a personal computer, a server, and other devices, which are not specifically limited herein.

As shown in fig. 1, the longitudinal federal learning system optimization device may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, a communication bus 1002. Wherein a communication bus 1002 is used to enable connective communication between these components. The user interface 1003 may include a Display screen (Display), an input unit such as a Keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface, a wireless interface. The network interface 1004 may optionally include a standard wired interface, a wireless interface (e.g., WI-FI interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory (e.g., a magnetic disk memory). The memory 1005 may alternatively be a storage device separate from the processor 1001.

Those skilled in the art will appreciate that the configuration of the apparatus shown in FIG. 1 does not constitute a limitation on the longitudinal Federal learning System optimization apparatus, and may include more or fewer components than shown, or some components in combination, or a different arrangement of components.

As shown in fig. 1, the memory 1005, which is a type of computer storage medium, may include an operating system, a network communication module, a user interface module, and a longitudinal federal learning system optimization program therein. A TEE (Trusted execution environment) module is also included. The operating system is a program for managing and controlling hardware and software resources of the equipment and supports the running of a federally learned private data processing program and other software or programs. The TEE is a secure area within the host processor, running in a separate environment and running in parallel with the operating system, which ensures that the confidentiality and integrity of the code and data loaded in the TEE are protected. Trusted applications running in the TEE can access all functions of the device main processor and memory, while hardware isolation protects these components from user-installed applications running in the main operating system. In this embodiment, the TEE module may be implemented in various ways, such as Software guard extensions (SGX) based on Intel, Secure Encrypted Virtualization (SEV) of AMD, Trust Zone of ARM, or santtum of MIT. Authentication and authorization of the TEE module may be accomplished through a third party secure server. For example, when a TEE is an SGX using Intel, the TEE may be authenticated by the security server of Intel, i.e., the TEE is secured.

In the device shown in fig. 1, the user interface 1003 is mainly used for data communication with a client; the network interface 1004 is mainly used for establishing communication connection with other terminal devices participating in federal learning, such as a second device which is a participant in longitudinal federal learning; and the processor 1001 may be configured to invoke a longitudinal federated learning system optimization program stored in the memory 1005 and perform the following operations:

Further, the step of obtaining an interpolation model by using the first sample data and the second device for training in cooperation includes:

Further, after the step of combining the first partial model and the second partial model to obtain the interpolation model, the processor 1001 may be further configured to call a federal learned privacy data processing program stored in the memory 1005, and perform the following steps:

Further, a first TEE module is included in the first device, a second TEE module is included in the second device,

encrypting the first output to obtain a first encrypted output;

and receiving the encrypted gradient information sent by the second device, decrypting the encrypted gradient information in the first TEE module to obtain gradient information related to the first part model in the first gradient information, and updating parameters of the first part model according to the gradient information related to the first part model.

Further, after the step of sending the first output to the second device, the processor 1001 may be further configured to call a federally learned privacy data handler stored in the memory 1005, and perform the following steps:

Further, the first device includes a TEE module, and the step of obtaining an interpolation model by using the first sample data and the second device for training in cooperation includes:

Further, the target machine learning model is configured to predict a purchase intention of a user, and after the step of performing local training on a preset machine learning model to be trained by using the local sample data and the prediction sample data to obtain the target machine learning model, the processor 1001 may be further configured to call a federal learning privacy data processing program stored in the memory 1005, and execute the following steps:

Based on the structure, various embodiments of the longitudinal federal learning system optimization method are provided.

While a logical order is shown in the flow chart, in some cases, the steps shown or described may be performed in an order different than presented herein. The first device and the second device related in the embodiment of the present invention may be participating devices participating in federal learning of longitudinal federal learning, and the participating devices may be devices such as a smart phone, a personal computer, and a server.

Referring to fig. 2, fig. 2 is a schematic flow chart of a first embodiment of the longitudinal federated learning system optimization method of the present invention. In this embodiment, the longitudinal federated learning system optimization method includes:

step S10, performing sample alignment with the second device to obtain first sample data of the first device, where the first sample data and second sample data have different data characteristics, and the second sample data is obtained by performing sample alignment between the second device and the first device;

it should be noted that, in the following description, the first device and the second device distinguish two participating devices, and a model required for training and using the first device is taken as an example. The model required by the second device can be trained and used only by exchanging the roles of the first device and the second device, and the method is applicable to a scene with a plurality of participating devices and coordination devices and participating devices according to similar principles.

The first device establishes a communication connection with the second device in advance. The local data of the first device and the second device have an overlapping part in user dimension, and have different parts (possibly completely different) in data characteristics, the first device and the second device perform sample alignment by adopting respective local data to determine common users and different data characteristics of the two devices, the first device takes the data of the common users in the local data as first sample data, and the second device takes the data of the common users in the local data, which is different from the data characteristics of the first device, as second sample data, that is, the finally determined first sample data and the second sample data have the same user and different data characteristics. The first device and the second device may perform sample alignment by using an existing sample alignment technique, and if the first device and the second device are not trusted, an encrypted sample alignment technique may be used, which is not described in detail herein. For example, fig. 3 is a schematic diagram of sample data in a first device and a second device, where the local data of the first device includes 3 users { U1, U2, U3}, the data characteristics include { X1, X2, X3}, the local data of the second device includes 3 users { U1, U2, U4}, and the data characteristics include { X4, X5 }. After sample alignment, the first sample data determined by the first device is data of users U1 and U2 under data characteristics X1, X2 and X3, and the second sample data determined by the second device is data of users U1 and U2 under data characteristics X4 and X5.

Step S20, obtaining an interpolation model by using the first sample data and the second device for collaborative training, where the interpolation model is used to input data belonging to the data feature corresponding to the first device and output data belonging to the data feature corresponding to the second device.

The first device adopts the first sample data to cooperatively train with the second device to obtain an interpolation (interpolation) model. The interpolation model is used for complementing data of the missing data characteristics of the first device relative to the second device. In the above specific example, the interpolation model is used to complement the data of the first sample data of the first device under the data characteristics X4 and X5 of the users U1 and U2. Specifically, the interpolation model may be configured to input data belonging to a data feature corresponding to the first device and output data belonging to a data feature corresponding to the second device, that is, data of the user in the first sample data is input into the interpolation model, and data close to the user in the second sample data is output, for example, data of the user U1 under data features X1, X2, and X3 is input, and the interpolation model outputs data close to data of the user U1 under data features X4 and X5 in the second sample data, so as to complement data of the missing user U1 under data features X4 and X5 for the first device. The interpolation model can adopt a model structure such as a variational automatic encoder (variational automatic encoder), a generation countermeasure Network (general adaptive Network), a Pixel-RNN, a Pixel-CNN or a Restricted Boltzmann Machine (Restricted Boltzmann Machine), and the like, and the first device and the second device can adopt a Split Learning (Split Learning) mode to cooperatively train the interpolation model.

Further, the step S20 includes:

step S201, inputting the first sample data into a first part of models preset in the first equipment to obtain a first output;

the method comprises the steps of presetting the structure of an interpolation model to be trained, setting the input characteristic of the interpolation model to be trained as the data characteristic corresponding to first equipment, setting the output characteristic as the data characteristic corresponding to second equipment, and cutting the interpolation model to be trained into two parts, namely a first part model and a second part model. The first partial model is pre-positioned in a first device and the second partial model is pre-positioned in a second device. Specifically, as shown in fig. 4, when the interpolation model to be trained is a neural network model including a plurality of layers, there are two ways to cut the interpolation model to be trained; the first method is to select two layers in the neural network model, cut the two layers, take the front part of the neural network model as the first part model and the back part as the second part model, and the two layers are called as cut layers; the second is to select two layers in the neural network model, add one layer between the two layers, the input of the layer is the output of the previous layer in the two layers, the output is the input of the next layer in the two layers, that is, the previous layer and the added layer are directly connected, cut the neural network model from the position between the previous layer and the added layer, the front part is used as the first part model, the back part is used as the second part model, and the position between the previous layer and the added layer is called as the cut layer; compared with the first cutting mode, the second cutting mode has less connection between the first part model and the second part model, so that the data volume needing to be transmitted between the first equipment and the second equipment is reduced, and which cutting mode is selected according to the specific situations of the dimension of the data characteristics of the first equipment and the second equipment, the structural complexity of the interpolation model to be trained and the like.

The first device inputs the first sample data into a first partial model of the first device, the first partial model outputting a first output, i.e. an output of a last layer of the first partial model.

Step S202, sending the first output to the second device, so that the second device can obtain a second output of a preset second partial model according to the first output, calculate a first loss function and first gradient information according to the second sample data and the second output, and update parameters of the second partial model according to gradient information related to the second partial model in the first gradient information;

the first device sends the first output to the second device. It should be noted that, if the interpolation model to be trained is cut according to the first cutting mode, the first device multiplies the first output by the corresponding connection weight according to the connection relationship between the cut layers to convert the first output into the input of the first layer of the second partial model, and sends the converted result to the second device, so that the second device inputs the converted result into the first layer of the second partial model, and the second partial model outputs the second output; and if the interpolation model to be trained is cut according to the second cutting mode, the first equipment directly sends the first output to the second equipment, the second equipment directly inputs the first output into the first layer of the second part model after receiving the first output, and the second part model outputs the second output.

And after obtaining second output of the second partial model according to the first output, the second equipment calculates a first loss function and first gradient information according to second sample data and the second output. The first loss function is a loss function of the interpolation model to be trained, and the first gradient information may include gradient information related to the second partial model, such as gradient information of the first loss function on each parameter of the second partial model, and may also include gradient information related to the first partial model, such as gradient information of the first loss function on an input variable of the second partial model.

The second device may detect whether a preset stop condition is met, where the preset stop condition may be a preset stop condition, such as stopping when the first loss function convergence is detected, or stopping when the number of iterative training times reaches a maximum number, or stopping when the time of iterative training reaches a maximum training time. Specifically, the second device may determine that the first loss function converges when detecting that the loss value of the first loss function is smaller than a preset value, and determine that the first loss function does not converge if the loss value of the first loss function is not smaller than the preset value. And when the preset stopping condition is not met, the second equipment updates the parameters in the second part of the model according to the gradient information related to the second part of the model in the first gradient information. The second device may send the gradient information related to the first partial model in the first gradient information to the first device, or may send both the first gradient information to the first device.

Step S203, updating parameters of the first part model according to gradient information related to the first part model in the first gradient information received from the second device, and receiving the second part model sent by the second device when iterative training is detected to meet a preset stop condition;

the first device receives the first gradient information from the second device, or receives the gradient information related to the first partial model in the first gradient information. And the first equipment updates each parameter of the first part model according to the gradient information related to the first part model, for example, according to the gradient information of the first loss function on the input variable of the second part model, the gradient information of the first loss function on the output variable of the first part model is deduced, the gradient information of the first loss function on each parameter of the first part model is deduced, and each parameter is updated according to the gradient information corresponding to the parameter. After the parameters of the first part of models are updated, the first device continues to input the first part of models by adopting first sample data to obtain first output, the first output is sent to the second device, the second device calculates a first loss function and gradient information, the process is circulated until the preset stopping condition is met, the training of the interpolation model to be trained is stopped, and the current parameters are used as the final parameters of the first part of models and the second part of models. The second device sends the second partial model, which determines the final parameters, to the first device. It should be noted that the second device may detect whether the preset stop condition is met, or the first device detects whether the preset stop condition is met, and if the preset stop condition is met, the second device may send the calculation result of the first loss function to the first device, and the first device detects whether the first loss function converges, or the second device sends the result of detecting whether the first loss function converges to the first device.

Step S204, combining the first partial model and the second partial model to obtain the interpolation model.

And the first equipment receives the second partial model which is sent by the second equipment and determines the final parameters, and combines the second partial model with the first partial model determining the final parameters to obtain the trained interpolation model. Specifically, the combination mode may correspond to a cutting mode, and when a first cutting mode is adopted, the trained first partial model and the trained second partial model are directly spliced, and when a second cutting mode is adopted, the first layer of the second partial model is removed and then spliced with the first partial model.

In this embodiment, the first device performs sample alignment with the second device, performs collaborative training by using aligned sample data, and obtains an interpolation model capable of complementing missing data features of the first device relative to the second device, so that when the first device uses a machine learning model obtained by training data features of the first device and the second device, even if there is no data of the second device, the first device can independently predict data belonging to corresponding data features of the second device through the interpolation model locally, and thus the complemented data is used to complete prediction by using the machine learning model, thereby expanding the application range of longitudinal federal learning, and avoiding that the first device cannot use the machine learning model obtained by longitudinal federal learning in a scene where the second device cannot provide data for the first device.

Further, based on the first embodiment, a second embodiment of the optimization method for a longitudinal federal learning system of the present invention is provided, and after the step S204, the second embodiment of the optimization method for a longitudinal federal learning system of the present invention further includes:

step S205, inputting local sample data belonging to the corresponding data characteristics of the first device into the interpolation model to obtain prediction sample data belonging to the corresponding data characteristics of the second device;

after the interpolation model is obtained by the first device and the second device in cooperation training, the machine learning model with the input characteristics including the data characteristics of the first device and the second device can be independently trained. Specifically, the first device locally stores local sample data (which may include data of multiple users) belonging to corresponding data features of the first device. It should be noted that the local sample data may include user data that is not used in the training interpolation model, that is, the second device may not have data of the user. And the first equipment inputs the local sample data into the interpolation model to obtain the prediction sample data belonging to the corresponding data characteristics of the second equipment. Or the first device inputs the first sample data into the interpolation model to obtain the prediction sample data which is corresponding to the second sample data and belongs to the data characteristics corresponding to the second device. For example, the first device inputs the data of the user U1 under the data features X1, X2 and X3 into the interpolation model, and obtains the data under the data features X4 and X5 (called prediction sample data), because the data are the interpolation model trained in advance, the interpolation model outputs the obtained data under the data features X4 and X5, and the data are close to the real data of the user U1 under the data features X4 and X5.

And S206, locally training a preset machine learning model to be trained by adopting the local sample data and the prediction sample data to obtain a target machine learning model.

The first equipment carries out local training on a preset machine learning model to be trained by adopting local sample data and prediction sample data to obtain a target machine learning model. Or the first device performs local training on a preset machine learning model to be trained by adopting the first sample data and the prediction sample data to obtain a target machine learning model. The machine learning model to be trained may be a machine learning model with specific prediction or classification targets, such as a neural network model. Specifically, the first device combines the data features of the user data in the local sample data and the data features of the user data in the prediction sample data to serve as input features of the machine learning model to be trained, iterative training is performed on the machine learning model to be trained until the machine learning model to be trained is detected to be converged, and the target machine learning model is obtained after training is completed. The first device may use the trained target machine learning model to perform a prediction or classification task.

It should be noted that, if the machine learning model to be trained is a supervised machine learning model and only the second device has a data tag, the second device may share the data corresponding to the data tag with the first device, and the first device performs supervised training on the machine learning model to be trained by using the data corresponding to the data tag. If the first device and the second device are in an untrusted scene, the second device may share the data tag with the first device in an encrypted manner.

Fig. 5 is a schematic diagram of the first device training a machine learning model by itself. GM in the figure_AIs an interpolation model obtained by the cooperative training of a first device and a second device, wherein the first device uses the first sample data X_AInput GM_AAnd M_ALocally learning the model M to the machine by means of an interpolation model_ATraining is carried out, and the first sample data X is_AInput GM_AAnd M_A。

In this embodiment, the first device completes missing data features of the first device by using an interpolation model obtained through collaborative training with the second device, so that the first device can independently train locally to obtain a target machine learning model even without the help of the second device, and thus, the application scenario of longitudinal federal learning is expanded.

Further, the target machine learning model is used for predicting the purchase intention of the user, and after the step S206, the method further includes:

step S207, inputting first data of a target user into the interpolation model to obtain second data, wherein the data characteristics of the first data comprise user identity characteristics, and the data characteristics of the second data comprise user purchase characteristics;

the target machine learning model may be a machine learning model for predicting the purchasing intention of the user, that is, the output label of the target machine learning model may be the purchasing intention, such as an output result of 0 or 1, where 1 indicates that the user will purchase and 0 indicates that the user will not purchase. The first device may be a device deployed with a banking institution, the second device may be a device deployed with an e-commerce institution, and the data characteristics of the user data in the first device may be different from those in the second device due to different businesses, and the data characteristics of the user data in the first device may include user identification characteristics such as age, deposit, monthly salary and the like, and the data characteristics of the user data in the second device may include user purchase characteristics such as purchase times, purchase preferences and the like. The first device and the second device cooperate in advance to train to obtain an interpolation model for predicting user purchase characteristic data according to the user identity characteristic data.

After the target machine learning model is obtained through training, the first device can independently adopt the interpolation model and the target machine learning model to predict the purchase intention of the target user under the condition that the second device does not provide data on the aspect of the user purchase characteristics. Specifically, the first device may input first data of the target user into the interpolation model, and the interpolation model outputs second data. The first data, namely the data of the target user recorded by the first device, includes the user identity characteristic, and the second data, namely the predicted purchase characteristic data of the target user.

Step S208, inputting the first data and the second data into the target machine learning model to obtain the purchase intention of the target user.

And after the first equipment obtains the second data, inputting the first data and the second data into the target machine learning model to obtain the purchase intention of the target user. The second data is predicted by a pre-trained interpolation model and is close to the real data of the user, so that the predicted purchasing intention error of the target user is small by inputting the first data and the second data into the target machine learning model.

It should be noted that the target machine learning model may also be used in other application scenarios besides the purchasing intention prediction, such as performance level prediction, paper value evaluation, and the like, and the embodiments of the present invention are not limited herein.

Fig. 6 is a schematic flow chart of a process of training an interpolation model by cooperation of a first device and a second device, and training a machine learning model by the first device alone.

In this embodiment, an interpolation model is obtained through cooperative training of the first device and the second device, and when a target machine learning model is used, the interpolation model is used to complement the missing data characteristics of the first device, so that the first device can use the target machine learning model to complete a prediction function even in a scene without cooperation of the second device, and thus the application range of longitudinal federal learning is expanded.

Further, based on the first and second embodiments, a third embodiment of the optimization method for a longitudinal federated learning system according to the present invention is provided, where in the third embodiment of the optimization method for a longitudinal federated learning system according to the present invention, the first device includes a first TEE module, the second device includes a second TEE module, and the step S202 includes:

step S2021, encrypting the first output to obtain a first encrypted output;

in this embodiment, to adapt to a scenario in which the first device and the second device are not trusted with each other, the first device and the second device may be respectively provided with a TEE module, and the private data is processed in the TEE module, and the private data outside the TEE module is encrypted, so that the private data of the other party cannot be acquired. Specifically, after obtaining the first output, the first device may encrypt the first output to obtain a first encrypted output, and the encryption manner is not limited in this embodiment. It should be noted that, because the first output is obtained according to the local data of the first device and is visible in the first device, the first output may be encrypted in the first TEE module or encrypted in the second TEE module, and the encryption is to ensure that the second device cannot obtain the original first output after sending the first output to the second device, so that it is ensured that the data information of the first device is not leaked to the second device in a scenario where the first device and the second device are not trusted.

Step S2022, sending the first encrypted output to the second device, so that the second device decrypts the first encrypted output in the second TEE module to obtain the first output, obtains a second output preset with a second partial model according to the first output, calculates a first loss function and first gradient information according to the second sample data and the second output, updates parameters of the second partial model according to gradient information related to the second partial model in the first gradient information, and encrypts gradient information related to the first partial model in the first gradient information to obtain encrypted gradient information;

the first device outputs the first encryption to the second device. And after receiving the first encrypted output, the second equipment decrypts the first encrypted output in the second TEE module to obtain a first output. It should be noted that the decryption mode in the second TEE module of the second device corresponds to the encryption mode in the first TEE module of the first device, and if the first output is encrypted by using the public key in the first TEE module, the second TEE module decrypts by using the private key corresponding to the public key. After the second device decrypts the first output in the second TEE module, because the first output cannot be exposed to the part outside the second TEE module, the second device continues to obtain the first loss function and the first gradient information in the TEE module in a manner similar to that in step S202, and updates the parameters of the second partial model, and encrypts the gradient information related to the first partial model in the first gradient information to obtain encrypted gradient information. And after obtaining the encryption gradient information, the second equipment sends the encryption gradient information to the first equipment. Because the first output, the first loss function and the first gradient information are obtained on the basis of data in the first device, the original first output, the first loss function and the first gradient information are processed in the TEE module in the second device, and the data are encrypted outside the TEE module of the second device, so that the second device cannot obtain the private data of the first device.

The step of updating the parameters of the first partial model according to the gradient information related to the first partial model in the first gradient information received from the second device in the step SS203 includes:

step S2031, receiving the encrypted gradient information sent by the second device, decrypting the encrypted gradient information in the first TEE module to obtain gradient information related to the first partial model in the first gradient information, and updating parameters of the first partial model according to the gradient information related to the first partial model.

After the first device receives the encrypted gradient information sent by the second device, because the encrypted gradient information is obtained on the basis of the data of the second device, in order to ensure that the data of the second device is not leaked to the first device, the first device decrypts the encrypted gradient information in the first TEE module, and obtains the gradient information related to the first part model in the original first gradient information. The way of decrypting the data in the first TEE module corresponds to the way of encrypting the data in the second TEE module. After obtaining the gradient information related to the first partial model in the first gradient information, the first device updates the parameters of the first partial model according to the gradient information related to the first partial model.

After updating the parameters of the first part of models, the first equipment continues to input the first part of models by adopting first sample data to obtain first output, the first output is encrypted and sent to the second equipment, the second equipment calculates a first loss function and gradient information in the TEE module, the process is circulated until the preset stopping condition is met, the training of the interpolation model to be trained is stopped, and the current parameters are used as the final parameters of the first part of models and the second part of models. The second device sends a second partial model (encryptable) determining the final parameters to the first device. And the first equipment obtains an interpolation model according to the combination of the first partial model and the second partial model. Specifically, similar to the process of step S204, detailed description is omitted here.

FIG. 6 is a schematic diagram of a first device and a second device cooperatively training an interpolation model in a respective TEE module environment, where the first device uses a first sample data X in a first TEE module_AInputting interpolation model GM_AThe first partial model GM of_A-in Part1, the first output is encrypted and then transmitted to the second device, and the second device decrypts the first output in the second TEE module and obtains the second Part model GM according to the decryption result_ASecond output X of Part2_B' and calculating gradient information, encrypting the gradient information, then transmitting the gradient information to the first equipment, updating parameters by the first equipment according to the gradient information, and carrying out iterative training until meeting a preset stop condition, wherein the GM is processed by the second equipment_A-Part2 to the first device, the first device sending a GM_A-Part1 and GM_A-Part2 combination to obtain a trained interpolation model GM_A。

In this embodiment, the TEE modules are respectively arranged in the first device and the second device, the first device and the second device process original data in the TEE modules, and only encrypted data can be obtained outside the TEE modules, so that private data cannot be leaked to the other side in a mutually untrusted scene of the first device and the second device, and the safety of the data is guaranteed.

Further, based on the first, second, and third embodiments, a fourth embodiment of the optimization method for a longitudinal federated learning system according to the present invention is provided, where in the fourth embodiment of the optimization method for a longitudinal federated learning system according to the present invention, after the step S202, the method further includes:

a step a10 of receiving the second output and the first loss function sent by the second device;

in this embodiment, the first device and the second device may cooperatively train the machine learning model while cooperatively training the interpolation model. Specifically, on the basis of the first embodiment, after the first device sends the first output to the second device, the second device obtains the second output, and calculates the first loss function and the first gradient information, the second device sends the second output, the first loss function, and the second gradient information to the first device.

Step A20, inputting the first sample data and the second output into a preset machine learning model to be trained to obtain predicted label data;

the first equipment is provided with a machine learning model to be trained in advance. And the first equipment inputs the first sample data and the second output into the machine learning model to be trained, and the machine learning model to be trained outputs the prediction label data.

Step A30, calculating a second loss function and second gradient information of the machine learning model to be trained according to the predicted label data and pre-stored local actual label data;

and the first equipment calculates a second loss function and second gradient information of the machine learning model to be trained according to the predicted label data and the pre-stored local actual label data. It should be noted that the local actual tag data pre-stored in the first device may be tag data locally recorded by the first device, or the second device may share the locally recorded tag data to the first device, where the first device is stored locally.

The step of receiving the second partial model sent by the second device until the iterative training in step S203 detects that a preset stop condition is met includes:

step S2032, updating parameters of the machine learning model to be trained according to the second gradient information, performing iterative training to minimize a fusion loss function until a target machine learning model is obtained when a preset stop condition is detected to be satisfied, and receiving the second partial model sent by the second device, wherein the first device fuses the first loss function and the second loss function to obtain the fusion loss function.

The first device fuses the first loss function and the second loss function to obtain a fusion loss function, and the fusion mode may be to calculate a weighted sum of the first loss function and the second loss function as the fusion loss function. The first device determines whether a preset stop condition is met, where the preset stop condition may be a preset stop condition, and for example, the first device stops when convergence of the fusion loss function is detected, or stops when the number of iterative training times reaches a maximum number, or stops when the time of iterative training reaches a maximum training time. Specifically, whether the fusion loss function converges may be determined according to a calculated loss value of the fusion loss function, and if the loss value of the fusion loss function is detected to be smaller than a preset value, it is determined that the fusion loss function converges, and if the loss value of the fusion loss function is not smaller than the preset value, it is determined that the fusion loss function does not converge. If the first device detects that the preset stop condition is not met, the parameters of the machine learning model to be trained can be updated according to the second gradient information, and the parameters of the first part of models can be updated according to the gradient information related to the first part of models in the first gradient information. After parameters of the first part model and the machine learning model to be trained are updated, the first equipment continues to input the first part model by adopting first sample data to obtain first output, the first output is sent to the second equipment, the second equipment calculates a first loss function and gradient information, iterative training is carried out to achieve the purpose of minimizing the fusion loss function, the training of the interpolation model to be trained and the machine learning model to be trained is stopped when the detection meets the preset stop condition, the current parameters are used as final parameters of the first part model, the second part model and the machine learning model to be trained, and the machine learning model to be trained with the finally determined parameters is used as a target machine learning model after the training is finished. And the second equipment sends the second part of model with the determined final parameters to the first equipment, and the first equipment combines the first part of model and the second part of model to obtain the trained interpolation model.

It should be noted that, on the basis of the above-mentioned collaborative training, if the first device and the second device are not trusted with each other, a TEE module may be respectively set in the first device and the second device, and the original private data is processed in the TEE module, and the private data is encrypted outside the TEE module, so as to ensure that the private data of the first device and the second device is not leaked to the other party.

As shown in fig. 7, a schematic diagram of a first device training an interpolation model and a machine learning model in cooperation with a second device is shown, and the first device uses first sample data X_AInputting interpolation model GM_AThe first partial model GM of_A-Part1, forwarding the first output to the second device, which derives a second partial model GM from the first output_ASecond output X of Part2_B', and calculates the gradient information and the loss function, forwards the second output to the first device, and backwards propagates the gradient information and the loss function to the first device. The first device updates the GM according to the gradient information_A-parameters of Part1, and taking the second output as the machine learning model M to be trained_AUntil the fusion loss function is converged, the trained machine learning model M is obtained_A. The second device sends GM_A-Part2 to the first device, the first device sending a GM_A-Part1 and GM_A-Part2 combination to obtain a trained interpolation model GM_A。

In this embodiment, the machine learning model is cooperatively trained while the interpolation model is cooperatively trained by the first device and the second device, so that the training efficiency of the machine learning model is improved, and after the interpolation model is trained by the first device, missing data features can be complemented by the interpolation model, so that the first device can use the trained machine learning model to complete a prediction task under the condition that the second device does not cooperate, and the application range of longitudinal federal learning is expanded.

Further, based on the first, second, third, and fourth embodiments, a fifth embodiment of the optimization method for a longitudinal federated learning system according to the present invention is provided, where in the fifth embodiment of the optimization method for a longitudinal federated learning system according to the present invention, the first device includes a TEE module, and the step S20 includes:

step B10, receiving second encryption sample data sent by the second device, where the second device encrypts the second sample data to obtain the second encryption sample data;

in this embodiment, a different method of cooperatively training an interpolation model is proposed from that in the first embodiment. Specifically, in order to adapt to a scenario in which the first device and the second device are not trusted with each other, a TEE module may be disposed in the first device, the first device performs processing on private data in the TEE module, and the private data is encrypted outside the TEE module, so that the private data of the second device cannot be acquired. And after the second equipment performs sample alignment with the first equipment to obtain second sample data, encrypting the second sample data to obtain second encrypted data, and sending the second encrypted data to the first equipment.

And step B20, decrypting the second encrypted sample data in the TEE module to obtain second sample data, and training the interpolation model to be trained according to the first sample data and the second sample data to obtain the interpolation model.

And the first equipment decrypts the second encrypted sample data in the TEE module to obtain second sample data. It should be noted that the way of decrypting the data in the TEE module corresponds to the way of encrypting the data by the second device. The method includes that a first device trains an interpolation model to be trained in a TEE module according to first sample data and second sample data to obtain the interpolation model, and the mode of training the interpolation model by the first device alone is the same as that of training the interpolation model by the traditional device alone, and is not limited specifically here. Further, the second device may train the machine learning model locally alone in the TEE module through the trained interpolation model; the first device may also train the machine learning model while training the interpolation model in the TEE module.

In this embodiment, the TEE module is arranged in the first device, the second device encrypts second sample data and sends the second sample data to the first device, and the first device decrypts the second sample data in the TEE module to enable the second device to train the interpolation model independently under the condition that private data are not leaked to the first device, so that the situation that the interpolation model cannot be trained and used by the first device under the condition that the second device cannot cooperate with the first device to train the interpolation model is avoided, and the application range of longitudinal federal learning is expanded.

In addition, an embodiment of the present invention further provides a longitudinal federal learning system optimization device, where the longitudinal federal learning system optimization device is disposed in a first device, and the first device is in communication with a second device, and referring to fig. 9, the longitudinal federal learning system optimization device includes:

an alignment module 10, configured to perform sample alignment with the second device to obtain first sample data of the first device, where data characteristics of the first sample data are different from those of second sample data, and the second sample data is obtained by performing sample alignment between the second device and the first device;

a training module 20, configured to perform collaborative training on the first sample data and the second device to obtain an interpolation model, where the interpolation model is used to input data belonging to a data feature corresponding to the first device and output prediction data belonging to a data feature corresponding to the second device.

Further, the training module 20 includes:

the first input unit is used for inputting the first sample data into a first part of models preset in the first equipment to obtain first output;

a sending unit, configured to send the first output to the second device, so that the second device obtains a second output of a preset second partial model according to the first output, calculates a first loss function and first gradient information according to the second sample data and the second output, and updates a parameter of the second partial model according to gradient information related to the second partial model in the first gradient information;

a first receiving unit, configured to update parameters of the first partial model according to gradient information related to the first partial model in the first gradient information received from the second device, and perform iterative training until a preset stop condition is detected to be met, and receive the second partial model sent by the second device;

and the combination unit is used for combining the first partial model and the second partial model to obtain the interpolation model.

Further, the training module 20 further includes:

the second input unit is used for inputting local sample data belonging to the data characteristics corresponding to the first equipment into the interpolation model to obtain prediction sample data belonging to the data characteristics corresponding to the second equipment;

and the first training unit is used for carrying out local training on a preset machine learning model to be trained by adopting the local sample data and the prediction sample data to obtain a target machine learning model.

Further, the first device includes a first trusted execution environment TEE module therein, the second device includes a second TEE module therein, and the sending unit includes:

the encryption subunit is used for encrypting the first output to obtain a first encrypted output;

a sending subunit, configured to send the first encrypted output to the second device, so that the second device decrypts the first encrypted output in the second TEE module to obtain the first output, obtains a second output preset in a second partial model according to the first output, calculates a first loss function and first gradient information according to the second sample data and the second output, updates parameters of the second partial model according to gradient information related to the second partial model in the first gradient information, and encrypts gradient information related to the first partial model in the first gradient information to obtain encrypted gradient information;

the first receiving unit is further configured to receive the encrypted gradient information sent by the second device, decrypt the encrypted gradient information in the first TEE module to obtain gradient information in the first gradient information, where the gradient information is related to the first partial model, and update parameters of the first partial model according to the gradient information related to the first partial model.

Further, the training module 20 further includes:

a second receiving unit, configured to receive the second output and the first loss function sent by the second device;

the third input unit is used for inputting the first sample data and the second output data into a preset machine learning model to be trained to obtain predicted label data;

the calculation unit is used for calculating a second loss function and second gradient information of the machine learning model to be trained according to the predicted label data and pre-stored local actual label data;

the first receiving unit is further configured to update parameters of the machine learning model to be trained according to the second gradient information, perform iterative training to minimize a fusion loss function until a target machine learning model is obtained when a preset stop condition is detected to be met, and receive the second partial model sent by the second device, where the first device fuses the first loss function and the second loss function to obtain the fusion loss function.

Further, a TEE module is included in the first device, and the training module 20 includes:

a third receiving unit, configured to receive second encryption sample data sent by the second device, where the second device encrypts the second sample data to obtain the second encryption sample data;

and the second training unit is used for decrypting the second encrypted sample data in the TEE module to obtain second sample data, and training the interpolation model to be trained according to the first sample data and the second sample data to obtain the interpolation model.

Further, the target machine learning model is used for predicting the purchase intention of the user, and the longitudinal federal learning system optimization device further comprises:

the input module is used for inputting first data of a target user into the interpolation model to obtain second data, wherein the data characteristics of the first data comprise user identity characteristics, and the data characteristics of the second data comprise user purchase characteristics;

and the prediction module is used for inputting the first data and the second data into the target machine learning model to obtain the purchase intention of the target user.

The development content of the specific implementation mode of the longitudinal federal learning system optimization device is basically the same as that of each embodiment of the longitudinal federal learning system optimization method, and is not described herein again.

In addition, an embodiment of the present invention further provides a computer-readable storage medium, where a longitudinal federated learning system optimization program is stored on the storage medium, and when being executed by a processor, the longitudinal federated learning system optimization program implements the steps of the longitudinal federated learning system optimization method described below.

For the embodiments of the longitudinal federated learning system optimization apparatus and the computer-readable storage medium of the present invention, reference may be made to the embodiments of the longitudinal federated learning system optimization method of the present invention, which are not described herein again.

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments.

Through the above description of the embodiments, those skilled in the art will clearly understand that the method of the above embodiments can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware, but in many cases, the former is a better implementation manner. Based on such understanding, the technical solutions of the present invention may be embodied in the form of a software product, which is stored in a storage medium (such as ROM/RAM, magnetic disk, optical disk) and includes instructions for enabling a terminal device (such as a mobile phone, a computer, a server, an air conditioner, or a network device) to execute the method according to the embodiments of the present invention.

The above description is only a preferred embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes, which are made by using the contents of the present specification and the accompanying drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims

1. A longitudinal federated learning system optimization method is applied to a first device, the first device is in communication connection with a second device, and the longitudinal federated learning system optimization method comprises the following steps:

2. The method of claim 1, wherein the step of training an interpolation model using the first sample data in cooperation with the second device comprises:

3. The method for optimizing a longitudinal federal learning system as claimed in claim 2, wherein said step of combining said first model portion and said second model portion to obtain said interpolation model further comprises:

4. The longitudinal federated learning system optimization method of claim 2, wherein a first Trusted Execution Environment (TEE) module is included in the first device, a second TEE module is included in the second device,

encrypting the first output to obtain a first encrypted output;

5. The longitudinal federal learning system optimization method of claim 2, wherein said step of sending said first output to said second device is followed by further comprising:

6. The method of claim 1, wherein the first device includes a TEE module, and the step of training the interpolation model using the first sample data in cooperation with the second device includes:

7. The method for optimizing a longitudinal federated learning system of claim 3, wherein the target machine learning model is used for predicting purchasing intentions of a user, and after the step of locally training a preset machine learning model to be trained using the local sample data and the prediction sample data to obtain the target machine learning model, the method further comprises:

8. A longitudinal federated learning system optimization apparatus deployed on a first device in communication with a second device, the longitudinal federated learning system optimization apparatus comprising:

9. A longitudinal federated learning system optimization apparatus, comprising: a memory, a processor, and a longitudinal federated learning system optimization program stored on the memory and executable on the processor, the longitudinal federated learning system optimization program when executed by the processor implementing the steps of the longitudinal federated learning system optimization method of any one of claims 1 to 7.

10. A computer readable storage medium having stored thereon a longitudinal federated learning system optimization program that, when executed by a processor, performs the steps of the longitudinal federated learning system optimization method of any of claims 1-7.