CN111368196A - Method, apparatus, device and readable storage medium for updating model parameters - Google Patents

Abstract

The invention discloses a method, device, device and readable storage medium for updating model parameters, and relates to the field of financial technology. The method includes the steps of: a second terminal receives encrypted first feature data sent by a first terminal, and obtains The second feature data and the label data corresponding to the second feature data, the first feature data is obtained by the first terminal and at least one terminal of the same type through horizontal federated learning; calculated according to the first feature data, the second feature data and the label data Obtain the residual, and send the residual to the first terminal for the first terminal to calculate the first gradient value according to the residual, and update the first model parameter corresponding to the first feature data according to the first gradient value; calculate according to the residual For the second gradient value, the second model parameter corresponding to the second feature data is updated according to the second gradient value. The invention realizes model training by integrating the data of two different terminals, so as to improve the performance of the model obtained by training.

Description

Method, apparatus, device and readable storage medium for updating model parameters

technical field

The present invention relates to the technical field of data processing in financial technology (Fintech), and in particular, to a method, apparatus, device and readable storage medium for updating model parameters.

Background technique

With the development of computer technology, more and more technologies are applied in the financial field. The traditional financial industry is gradually transforming into financial technology (Fintech), and data processing technology is no exception. However, due to the security and real-time requirements of the financial industry, It also puts forward higher requirements for data processing technology.

A typical application is a recommendation system, but it is often difficult to obtain label data directly. For example, in the recommendation scenario of e-commerce, users do not directly specify whether they like an item or not. We can only rely on the user's behavior, and Defined according to some artificial rules, for example, if a user buys an item, then we think that the user likes it (the label is 1); on the contrary, if the user only browses and does not buy, then we think that the user does not like it (the label is 0). However, this artificial rule definition is not real label data after all, and sometimes it may be far from the actual situation.

At present, the joint application of the mobile terminal and the enterprise terminal is a trend in the current development of AI. With the development of the mobile Internet, the mobile terminal has become the most important tool for people to obtain information. Therefore, the mobile terminal has very rich user behavior data, but as As described above, the authenticity of the data labels on the mobile phone is not high, so a lot of data preprocessing work needs to be done to use the user behavior on the mobile phone; on the other hand, on the enterprise side, the user information data is relatively small, but usually It is very sensitive and important private data, such as bank personal credit information, loans and other information, and these data are much more authentic than user behavior data on mobile phones. Therefore, if these important and real data can be used and combined with mobile phone data, the performance of the recommendation model will be greatly improved.

It can be seen that how to combine the data of two different terminals for model training to improve the performance of the trained model is an urgent problem to be solved.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a method, device, device and readable storage medium for updating model parameters, which aims to solve the problem of how to combine the data of two different terminals for model training in order to improve the performance of the model obtained by training. technical problem.

In order to achieve the above object, the present invention provides a method for updating model parameters, the method for updating model parameters includes the steps:

The second terminal receives the encrypted first characteristic data sent by the first terminal, and acquires the second characteristic data and the label data corresponding to the second characteristic data, wherein the first characteristic data is obtained by the first terminal obtained through horizontal federated learning with at least one terminal of the same type, the first terminal and the second terminal are terminals of different types;

The residual is calculated according to the first characteristic data, the second characteristic data and the label data, and the residual is sent to the first terminal for the first terminal to use the residual according to the residual. calculating a first gradient value, and updating the first model parameter corresponding to the first feature data according to the first gradient value;

A second gradient value is calculated according to the residual, and a second model parameter corresponding to the second feature data is updated according to the second gradient value.

Preferably, after the second terminal receives the encrypted first feature data sent by the first terminal and acquires the second feature data and the label data corresponding to the second feature data, the method further includes:

Obtain the encrypted first feature parameter corresponding to the first feature data, and obtain the second feature parameter corresponding to the second feature data, wherein the device corresponding to the first feature data and the second feature data same identification;

Calculate the loss value according to the first feature data, the second feature data, the label data, the first feature parameter and the second feature parameter;

sending the loss value to the third terminal, so that the third terminal can judge whether the first model parameter and the second model parameter meet the stop updating condition according to the loss value, and return prompt information;

If it is determined according to the prompt information that the first model parameter and the second model parameter do not meet the stop updating condition, return to executing the second terminal to receive the encrypted first feature data sent by the first terminal, and the steps of acquiring the second feature data and the label data.

Preferably, the step of sending the loss value to a third terminal, so that the third terminal can judge whether the first model parameter and the second model parameter satisfy the update stop condition according to the loss value includes: :

sending the loss value to the third terminal, so that after the third terminal receives the loss value, the loss history value of the last update of the first model parameter and the second model parameter is obtained, And when it is detected that the absolute value of the difference between the loss value and the historical loss value is smaller than a preset threshold, it is determined that the first model parameter and the second model parameter meet the stop updating condition, and when it is detected that the When the absolute value is greater than or equal to the preset threshold, it is determined that the first model parameter and the second model parameter do not satisfy the update stop condition.

Preferably, before the step of calculating the second gradient value according to the residual, the method further includes:

acquiring training data corresponding to the second characteristic data, and acquiring a third characteristic parameter corresponding to the second characteristic data;

The step of calculating the second gradient value according to the residual includes:

A second gradient value is calculated based on the training data, the third feature parameter and the residual.

Preferably, the step of updating the second model parameter corresponding to the second feature data according to the second gradient value includes:

obtaining an update coefficient corresponding to the second terminal;

The second model parameter corresponding to the second feature data is updated according to the second gradient value and the update coefficient.

Preferably, the step of calculating the residual according to the first feature data, the second feature data and the label data includes:

Calculate the difference between the second characteristic data and the second characteristic data label data, and encrypt the difference to obtain the encrypted difference;

Calculate the encrypted residual according to the encrypted difference and the encrypted first characteristic data;

The step of calculating the second gradient value according to the residual, and updating the second model parameter corresponding to the second feature data according to the second gradient value includes:

Calculate the encrypted second gradient value according to the encrypted residual, send the encrypted second gradient value to the third terminal for the third terminal to decrypt the encrypted second gradient value, and return the decrypted second gradient value. two gradient values;

Receive the decrypted second gradient value returned by the third terminal, and update the second model parameter corresponding to the second feature data according to the decrypted second gradient value.

Preferably, after the step of calculating the second gradient value according to the residual, and updating the second model parameter corresponding to the second feature data according to the second gradient value, the method further includes:

Using the updated second model parameters as the model parameters of the loan prediction model in the second terminal to obtain the loan prediction model;

After receiving the data to be predicted, the data to be predicted is input into the loan prediction model, the loan probability corresponding to the data to be predicted is obtained, and the information is pushed according to the loan probability.

In addition, in order to achieve the above object, the present invention also provides a device for updating model parameters, the device for updating model parameters includes:

The receiving module is configured to receive the encrypted first feature data sent by the first terminal, wherein the first feature data is obtained by the first terminal and at least one terminal of the same type through horizontal federated learning, and the first A terminal and the second terminal are different types of terminals;

an acquisition module for acquiring the second characteristic data and the label data corresponding to the second characteristic data;

a calculation module, configured to calculate and obtain a residual according to the first characteristic data, the second characteristic data and the label data;

a sending module, configured to send the residual to the first terminal, so that the first terminal can calculate a first gradient value according to the residual, and update the first feature according to the first gradient value the first model parameter corresponding to the data;

The calculation module is further configured to calculate a second gradient value according to the residual;

An update module, configured to update the second model parameter corresponding to the second feature data according to the second gradient value.

In addition, in order to achieve the above object, the present invention also provides a device for updating model parameters, the device for updating model parameters includes a memory, a processor, and model parameters stored on the memory and running on the processor The updating program of the model parameter, when the updating program of the model parameter is executed by the processor, implements the steps of the updating method for the model parameter corresponding to the federated learning server.

In addition, in order to achieve the above object, the present invention also provides a computer-readable storage medium, on which an update program of model parameters is stored, and the update program of the model parameters is implemented as described above when executed by a processor. The steps of the update method for the model parameters described above.

In the present invention, the second terminal performs joint training by combining the encrypted data of the first terminal to update the model parameters in the prediction model, so as to ensure the privacy of the data of the first terminal, that is, on the basis that the first terminal does not reveal its own original data On the other hand, the second terminal can integrate its own data and the encrypted data of the first terminal for joint training, update the model parameters, that is, update the corresponding prediction model, so as to realize the integration of the data of two different terminals for model training, so as to improve the The performance of the trained model.

Description of drawings

1 is a schematic flowchart of a first embodiment of a method for updating model parameters of the present invention;

2 is a schematic flowchart of the second embodiment of the method for updating model parameters of the present invention;

3 is a schematic flowchart of a model parameter update process in an embodiment of the present invention;

4 is a functional schematic block diagram of a preferred embodiment of the device for updating model parameters of the present invention;

FIG. 5 is a schematic structural diagram of a hardware operating environment involved in the solution of an embodiment of the present invention.

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

Detailed ways

It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The present invention provides a method for updating model parameters. Referring to FIG. 1 , FIG. 1 is a schematic flowchart of the first embodiment of the method for updating model parameters according to the present invention.

This embodiment of the present invention provides an embodiment of a method for updating model parameters. It should be noted that although a logical sequence is shown in the flowchart, in some cases, the sequence shown here may be performed in a different sequence. or the described steps.

The method for updating model parameters is applied to a server or a terminal, and the terminal may include mobile terminals such as mobile phones, tablet computers, notebook computers, handheld computers, and personal digital assistants (Personal Digital Assistant, PDA), as well as fixed terminals such as digital TVs and desktop computers. terminal. In each embodiment of the method for updating model parameters, for convenience of description, the execution body is omitted to describe each embodiment. Update methods for model parameters include:

Step S10, the second terminal receives the encrypted first characteristic data sent by the first terminal, and acquires the second characteristic data and the label data corresponding to the second characteristic data, wherein the first characteristic data is obtained by the The first terminal and at least one terminal of the same type are obtained through horizontal federated learning, and the first terminal and the second terminal are terminals of different types.

The second terminal receives the encrypted first feature data sent by the first terminal. Specifically, after the first terminal obtains the first characteristic data by encryption, the first terminal sends the encrypted first characteristic data to the second terminal. Or when the second terminal needs to update the second model parameter, the second terminal generates an acquisition request, and sends the acquisition request to the first terminal. After the first terminal receives the acquisition request, the first terminal performs horizontal federated learning with at least one terminal of the same type to obtain first feature data, and encrypts the first feature data to obtain encrypted first feature data. The first terminal sends the encrypted first characteristic parameter to the second terminal. In this embodiment, the first terminal and the second terminal are different types of terminals. For example, the first terminal is a mobile terminal, and the terminal that performs horizontal federated learning with the first terminal is also a mobile terminal. At this time, the mobile terminal has rich feature data. , but the label data is small, and the data authenticity is not high; the second terminal is the enterprise end, and the enterprise end has less data but high data authenticity. The first terminal may be one terminal or multiple terminals. Corresponding prediction models exist in both the first terminal and the second terminal, where the prediction model may be a linear regression model, a machine learning model, or a deep learning model, or the like. In this embodiment, for the convenience of distinction, the prediction model in the first terminal is recorded as the first prediction model, and the prediction model in the second terminal is recorded as the second prediction model.

第一训练数据记为

第一特征数据记为

则第一终端计算第一特征数据的公式可表示为：

第一终端在计算得到一特征数据后，第一终端采用预先设置好的加密算法加密第一特征数据，得到加密后的第一特征数据发送给第二终端。在本发明实施例中，用[[a]]表示加密后的a，如加密后的第一特征数据表示为

在本实施例中，不限制第一终端中的加密算法，如第一终端可采用DES(Data Encryption Standard，数据加密标准)、IDEA(International Data EncryptionAlgorithm国际数据加密算法)和AES(Advanced Encryption Standard，高级加密标准)等加密算法对第一特征数据进行加密。可以理解的是，第一终端将第一特征数据加密后发送给第二终端，保证了第一终端中数据的私密性，即保证了第一终端中数据的安全性。Specifically, the first terminal calculates and obtains the first feature data according to the first model parameters of the first prediction model and the first training data. If the first model parameter is denoted as

The first training data is recorded as

The first characteristic data is recorded as

Then the formula for calculating the first characteristic data by the first terminal can be expressed as:

After the first terminal calculates and obtains a characteristic data, the first terminal encrypts the first characteristic data by using a preset encryption algorithm, and obtains the encrypted first characteristic data and sends it to the second terminal. In the embodiment of the present invention, the encrypted a is represented by [[a]], for example, the encrypted first feature data is represented as

In this embodiment, the encryption algorithm in the first terminal is not limited. For example, the first terminal may use DES (Data Encryption Standard, data encryption standard), IDEA (International Data Encryption Algorithm) and AES (Advanced An encryption algorithm such as Advanced Encryption Standard) encrypts the first characteristic data. It can be understood that the first terminal encrypts the first characteristic data and sends it to the second terminal, which ensures the privacy of the data in the first terminal, that is, the security of the data in the first terminal.

第二训练数据记为

第二特征数据记为

则第二终端计算第二特征数据的公式可表示为：

需要说明的是，本实施例是为了便于描述，采用两个模型参数来举例。在具体应用过程中，第一预测模型和第二预测模型中模型参数的个数可根据具体需要而设置。The second terminal acquires the second feature data and the label data corresponding to the second feature data. It should be noted that the process for the second terminal to obtain the second feature data is similar to the process for the first terminal to obtain a feature data, that is, the second terminal obtains the second model parameters, and obtains the second training data, and then calculates the second parameter and The second feature data is obtained by multiplying the second training data. If the second model parameter is denoted as

The second training data is denoted as

The second characteristic data is recorded as

Then the formula for calculating the second characteristic data by the second terminal can be expressed as:

It should be noted that, in this embodiment, for the convenience of description, two model parameters are used as an example. In a specific application process, the number of model parameters in the first prediction model and the second prediction model can be set according to specific needs.

The label data corresponds to the second training data, and each second training data has corresponding label data. In this embodiment, only the second training data in the second terminal has label data, and the first training data in the first terminal has no corresponding label data. Specifically, if the second training data indicates that the user likes item A, if the user indicates that he likes item A, the label data is "1"; if the user indicates that he does not like item A, the label data is "0". In this embodiment, the representation form of the tag data is not limited.

Step S20: Calculate and obtain a residual error according to the first feature data, the second feature data and the label data, and send the residual error to the first terminal for the first terminal to use according to the selected data. A first gradient value is calculated from the residual, and a first model parameter corresponding to the first feature data is updated according to the first gradient value.

After the second terminal obtains the encrypted first feature data, the second feature data and the label data, the second terminal calculates and obtains the residual according to the first feature data, the second feature data and the label data, and sends the residual to the first a terminal. It should be noted that the residual sent by the second terminal to the first terminal is the encrypted residual. After the first terminal receives the encrypted residual, the first terminal calculates and obtains the first gradient value according to the encrypted residual, the first training data and the first model parameter.

则计算加密后的第一梯度值的公式可表示为：Specifically, the first terminal calculates the first product between the encrypted residual and the first training data, and calculates the second product between the first model parameter and the gradient coefficient, and encrypts the second product to obtain the encrypted After the second product, the encrypted first gradient value is obtained by calculating the first product and the encrypted second product. If the encrypted residual is denoted as [[d _i ]], the gradient coefficient is denoted as λ, and the size of the gradient coefficient can be set according to specific needs. This embodiment does not impose specific restrictions on the size of the gradient coefficient. A gradient value is denoted as

Then the formula for calculating the encrypted first gradient value can be expressed as:

则第一终端计算更新后的第一模型参数的公式可表示为：After the first terminal calculates and obtains the encrypted first gradient value, the first terminal sends the encrypted first gradient value to the third terminal. It should be noted that the third terminal is a terminal trusted by the first terminal and the second terminal, the third terminal knows the encryption algorithm of the encrypted data of the first terminal and the second terminal, and the third terminal can decrypt the data sent by the first terminal and the second terminal. , encrypted data. When the third terminal receives the encrypted first gradient value sent by the first terminal, the third terminal decrypts the encrypted first gradient value, obtains the decrypted first gradient value, and deciphers the decrypted first gradient value. The value is sent to the first terminal. After the first terminal receives the decrypted first gradient value, the first terminal obtains the first update coefficient, calculates the product of the first update coefficient and the decrypted first gradient value, and then subtracts the first model parameter The product of the first update coefficient and the decrypted first gradient value obtains the updated first model parameter. In this embodiment, the size of the first update coefficient is not limited, and the user can set the size of the first update coefficient according to specific needs. If the first update coefficient is denoted as η ₁ , the updated first model parameter is denoted as

Then the formula for calculating the updated first model parameters by the first terminal can be expressed as:

It can be seen from the above formula that in this embodiment, the gradient descent algorithm is used to update the model parameters.

Step S30: Calculate a second gradient value according to the residual, and update a second model parameter corresponding to the second feature data according to the second gradient value.

After the second terminal calculates and obtains the residual, the second terminal calculates the second gradient value according to the residual error, and updates the second model parameter corresponding to the second feature data according to the second gradient value, that is, updates the second parameter of the second prediction model. model parameters to obtain the updated second model parameters. When the updated second model parameters are obtained, the second terminal inputs the updated second model parameters into the second prediction model to obtain a second prediction model that can be used for data prediction.

Further, the step of calculating the residual according to the first feature data, the second feature data and the label data includes:

Step a: Calculate the difference between the second characteristic data and the label data corresponding to the second characteristic data, and encrypt the difference to obtain an encrypted difference.

In step b, the encrypted residual is calculated according to the encrypted difference and the encrypted first characteristic data.

Specifically, the second terminal calculates the difference between the second characteristic data and the label data corresponding to the second characteristic data, and encrypts the calculated difference to obtain the encrypted difference. It can be understood that the label data corresponding to the second feature data is the label data corresponding to the second training data. In this embodiment, the encryption algorithm for encrypting the difference by the second terminal is not limited. In this embodiment, in order to improve the update efficiency of the model parameters, the encryption algorithm used by the second terminal and the first terminal to encrypt data is the same. In other embodiments, the encryption algorithms used by the first terminal and the second terminal to encrypt data may also be different.

After the second terminal obtains the encrypted difference, the second terminal calculates and obtains the encrypted residual according to the encrypted difference and the encrypted first characteristic data. If the label data is denoted as y _i , the calculation formula of the encrypted residual by the second terminal can be expressed as:

The step S30 includes:

Step c, calculating the encrypted second gradient value according to the encrypted residual, sending the encrypted second gradient value to the third terminal for the third terminal to decrypt the encrypted second gradient value, and returning the decryption after the second gradient value.

Step d, receiving the decrypted second gradient value returned by the third terminal, and updating the second model parameter corresponding to the second feature data according to the decrypted second gradient value.

After the second terminal calculates and obtains the encrypted residual, the second terminal calculates and obtains the encrypted second gradient value according to the encrypted residual, and sends the encrypted second gradient value to the third terminal. After the third terminal receives the encrypted second gradient value, the third terminal decrypts the encrypted second gradient value, and sends the decrypted second gradient value to the second terminal. After the second terminal receives the decrypted second gradient value sent by the third terminal, the second terminal updates the second model parameter corresponding to the second feature data according to the decrypted second gradient value.

Further, the method for updating the model parameters also includes:

Step e, acquiring training data corresponding to the second characteristic data, and acquiring third characteristic parameters corresponding to the second characteristic data.

The step of calculating the second gradient value according to the residual includes:

Step f: Calculate a second gradient value according to the training data, the third feature parameter and the residual.

梯度系数表示为λ，第三特征参数表示为

则第二终端计算得到加密后的第二梯度值的计算公式可表示为：After the second terminal obtains the second feature data, the second terminal obtains the training data corresponding to the second feature data, that is, obtains the second training data corresponding to the second feature data, and obtains the third feature parameter corresponding to the second feature data. Specifically, the second terminal obtains the gradient coefficient, calculates the product between the gradient coefficient and the second model parameter, records the product between the gradient coefficient and the second model parameter as the third product, and encrypts the third product, Obtain the encrypted third product, and calculate the product between the second training data and the encrypted residual, and record the product between the second training data and the encrypted residual as the fourth product, and then according to the first The encrypted second gradient value is obtained by calculating the fourth product and the encrypted third product. Among them, if the encrypted second gradient value is expressed as

The gradient coefficient is denoted as λ, and the third characteristic parameter is denoted as

Then the calculation formula of the encrypted second gradient value obtained by the second terminal can be expressed as:

Further, the step of updating the second model parameter corresponding to the second characteristic data according to the second gradient value includes:

Step g, acquiring the update coefficient corresponding to the second terminal.

Step h, updating the second model parameter corresponding to the second feature data according to the second gradient value and the update coefficient.

则第二终端计算更新后的第二模型参数的计算公式可为：Specifically, it should be noted that in the process of updating the second model parameter by the second terminal, the second gradient value used by the second terminal is the decrypted second gradient value. The second terminal obtains and updates the update coefficient corresponding to the second model parameter, that is, obtains the second update coefficient, and updates the second model parameter corresponding to the second feature data according to the decrypted second gradient value and the second update coefficient. Specifically, the second terminal calculates the product between the second update coefficient and the decrypted second gradient value, and then subtracts the product between the second update coefficient and the decrypted second gradient value from the second model parameter to obtain The updated second model parameters to update the second model parameters. The second update coefficient may be equal to the first update coefficient, or may not be equal to the first update coefficient. If the second update coefficient is denoted as η ₂ , the updated second model parameter is denoted as

Then the calculation formula for calculating the updated second model parameter by the second terminal may be:

In this embodiment, the second terminal performs joint training by combining the encrypted data of the first terminal to update the model parameters in the prediction model, so as to ensure the privacy of the data of the first terminal, that is, on the basis that the first terminal does not disclose its own original data , so that the second terminal can integrate its own data and the encrypted data of the first terminal for joint training, update the model parameters, that is, update the corresponding prediction model, so as to realize the integration of the data of two different terminals for model training to improve training. The performance of the resulting model.

Further, a second embodiment of the method for updating model parameters of the present invention is proposed. The difference between the second embodiment of the method for updating the model parameters and the first embodiment of the method for updating the model parameters is that, referring to FIG. 2 , the method for updating the model parameters further includes:

Step S40, obtaining the encrypted first feature parameter corresponding to the first feature data, and obtaining the second feature parameter corresponding to the second feature data, wherein the first feature data and the second feature data The corresponding device IDs are the same.

The second terminal obtains the encrypted first feature parameter corresponding to the first feature data, and obtains the second feature parameter corresponding to the second feature data, wherein the encrypted first feature parameter can be sent by the first terminal to the second feature data to When the second terminal is used, it is sent to the second terminal together. The manner in which the first terminal sends the encrypted first characteristic parameter to the second terminal may be the same as or different from that of sending the encrypted first characteristic data. It should be noted that the device identifiers corresponding to the first feature data and the second feature data are the same, that is, the device identifiers of the first training data and the second training data in this embodiment are the same. The device identifier can uniquely represent a certain terminal, and different terminals can be identified through the device identifier. In this embodiment, the device identification can be represented by the package name or other information of the terminal, or by the ID card number or phone number of the user corresponding to the terminal.

Specifically, when the first terminal needs to send the encrypted first characteristic parameter to the second terminal, the first terminal obtains the first characteristic data, then calculates the square of the first characteristic data, and obtains the first model parameter and gradient coefficient , the characteristic parameter to be calculated is obtained by calculating according to the first model parameter and the gradient coefficient, and the first characteristic parameter is obtained according to the square of the characteristic parameter to be calculated and the first characteristic data. If the intersection of the device identification in the first training data and the device identification in the second training data is denoted as D, the first characteristic parameter encrypted by the first terminal can be expressed as:

则第一终端可通过加密后的第一特征数据和加密后的第一特征数梯度系数据的平方计算得到加密后的第一特征参数的过程可表示为：Further, the first terminal may obtain the encrypted first characteristic parameter by calculating the square of the encrypted first characteristic data and the encrypted first characteristic data. If the square of the encrypted first feature data is recorded as

Then, the process that the first terminal can obtain the encrypted first characteristic parameter through the square calculation of the encrypted first characteristic data and the encrypted first characteristic number gradient coefficient data can be expressed as:

需要说明的是，本实施例是通过第二特征参数计算损失值，而损失值是要发送给第三终端的，因此为了保护第二终端数据的安全性，需要对计算损失值的数据进行加密。在本发明实施例中，为了便于计算，第一终端加密数据的加密算法和第二终端加密数据的加密算法相同，在其它实施例中，第一终端加密数据的加密算法和第二终端加密数据的加密算法也可以不相同。When the second terminal acquires the second characteristic parameter, the second terminal acquires the gradient coefficient and the second model parameter, and obtains the second characteristic parameter according to the calculation with the second model parameter. Specifically, the encrypted second characteristic parameter can be expressed as

It should be noted that, in this embodiment, the loss value is calculated by the second characteristic parameter, and the loss value is to be sent to the third terminal. Therefore, in order to protect the security of the data of the second terminal, it is necessary to encrypt the data for calculating the loss value. . In this embodiment of the present invention, in order to facilitate calculation, the encryption algorithm of the encrypted data of the first terminal is the same as the encryption algorithm of the encrypted data of the second terminal. In other embodiments, the encryption algorithm of the encrypted data of the first terminal and the encrypted data of the second terminal are the same. The encryption algorithm can also be different.

Step S50, calculating a loss value according to the first feature data, the second feature data, the label data, the first feature parameter and the second feature parameter.

After the second terminal obtains the first characteristic parameter and the second characteristic parameter, the second terminal calculates the loss value according to the first characteristic data, the second characteristic data, the label data, the first characteristic parameter and the second characteristic parameter. It should be noted that, since each data for calculating the loss value is encrypted, the calculated loss value is also encrypted.

Specifically, if the encrypted loss value is denoted as [[L]], the loss function for calculating the loss value by the second terminal can be expressed as:

Specifically, the derivation process of the loss function is as follows:

因此：because

therefore:

Step S60, sending the loss value to the third terminal, so that the third terminal can judge whether the first model parameter and the second model parameter meet the stop updating condition according to the loss value, and return a prompt message .

After the second terminal calculates and obtains the encrypted loss value, the second terminal sends the encrypted loss value to the third terminal. It should be noted that the second terminal may send the loss value to the third terminal while sending the residual to the first terminal, or may not send the loss value to the third terminal after sending the residual to the first terminal , or send the loss value to the third terminal before sending the residual to the first terminal. After the third terminal receives the encrypted loss value, the third terminal decrypts the encrypted loss value, obtains the decrypted loss value, and judges whether the first model parameter and the second model parameter satisfy the threshold according to the decrypted loss value. Update conditions. Wherein, the stop update condition can be set according to specific needs. For example, it can be set to determine that the first model parameter and the second model parameter meet the stop update condition when the decrypted loss value is less than the preset loss value; when the decrypted loss value When it is greater than or equal to the preset loss value, it is determined that the first model parameter and the second model parameter do not satisfy the stop updating bar, wherein the size of the preset loss value can be set according to specific needs.

After the third terminal determines whether the first model parameter and the second model parameter satisfy the condition for stopping updating, it generates prompt information, and sends the prompt information to the second terminal, and may also send the prompt information to the first terminal at the same time. Wherein, the prompt information carries a keyword, through which it can be determined whether the first model parameter and the second model parameter satisfy the stop updating condition. For example, when the keyword in the prompt information is "true", it indicates that the first model parameter and the second model parameter meet the conditions for stopping updating; when the keyword in the prompt information is "false", it indicates that the first model parameter and the second model parameter Model parameters do not meet the stop update condition. It should be noted that the examples of keywords in this embodiment are only for the convenience of understanding, and do not constitute a limitation on keywords. For example, keywords can be set to "0" and "1".

Further, step S60 includes:

Step g, sending the loss value to the third terminal, so that the third terminal can obtain the loss when the first model parameter and the second model parameter were updated last time after receiving the loss value history value, and when it is detected that the absolute value of the difference between the loss value and the loss history value is less than a preset threshold, it is determined that the first model parameter and the second model parameter meet the stop updating condition, and When it is detected that the absolute value is greater than or equal to the preset threshold, it is determined that the first model parameter and the second model parameter do not satisfy the update stop condition.

Further, when the second terminal sends the encrypted loss value to the third terminal, the third terminal receives the encrypted loss value and decrypts to obtain the decrypted loss value, the third terminal obtains the last updated first model. The loss history value when the parameter and the second model parameter. It should be noted that each time the model parameters are updated, a loss value will be generated, that is, each time the model parameters are updated, the third terminal will receive the loss value sent by the second terminal, and store the loss value. After the third terminal obtains the historical loss value, the third terminal calculates the difference between the currently received loss value and the historical loss value, calculates the absolute value of the difference, and detects whether the absolute value of the difference is smaller than the preset value threshold. If the third terminal detects that the absolute value of the difference is less than the preset threshold, the third terminal determines that the first model parameter and the second model parameter meet the conditions for stopping the update; if the third terminal detects that the absolute value of the difference is greater than or equal to the preset threshold, the third terminal determines that the first model parameter and the second model parameter do not meet the condition for stopping updating. The size of the preset threshold can be set according to specific needs, for example, it can be set to 0.00001.

Step S70, if it is determined according to the prompt information that the first model parameter and the second model parameter do not meet the stop updating condition, then return to executing the encrypted first data sent by the second terminal and sent by the first terminal. feature data, and the steps of acquiring the second feature data and the label data.

When the second terminal receives the prompt information and determines according to the prompt information that the first model parameter and the second model parameter do not meet the conditions for stopping updating, the second terminal receives the encrypted first feature data sent by the first terminal again, and obtains the The second feature data and label data, that is, continue to update the model parameters. Further, when the second terminal receives the prompt information, and determines according to the prompt information that the first model parameter and the second model parameter meet the conditions for stopping the update, the update operation of the model parameters is no longer performed, that is, the iteration is no longer performed, and the second terminal The updated second model parameters are input into the second prediction model to obtain a trained second prediction model. After the first terminal determines according to the received prompt information that the first model parameter and the second model parameter satisfy the condition for stopping the update, the first terminal inputs the updated first model parameter into the first prediction model, and obtains the trained first model parameter. prediction model. It can be understood that, when the first model parameter and the second model parameter satisfy the update stop condition, it indicates that the update of the model parameter satisfies the convergence condition.

In this embodiment, the loss value is used to determine whether to end the update operation of the first model parameter and the second model parameter, and when it is determined not to end the update operation of the first model parameter and the second model parameter, that is, the first model parameter and the second model parameter are determined. When the model parameters do not meet the conditions for stopping the update, continue to perform the update operation of the first model parameters and the second model parameters, thereby improving the model performance of the prediction model obtained according to the updated model parameters, and improving the accuracy of the prediction model for data prediction. .

Specifically, referring to FIG. 3 , the entire process of the method for updating model parameters is as follows: Step 1: Initialize the weight parameters first, that is, initialize the first model parameter and the second model parameter. In this embodiment, the user can choose how to initialize as needed. For the first model parameters and the second model parameters, for example, all the first model parameters can be initialized to the same value, or each first model parameter can be initialized to different values; step 2, in the first model parameter and the second model parameter. After the model parameters are initialized, the device ID is checked for duplicates, that is, the first training data and the second training data with the same device ID are selected; in step 3, the training data of some mobile devices are selected for training, and the set of mobile devices is the first training data. A terminal (PartA), A ₁ , A _i , A _k , etc. represent different mobile terminal devices; Steps 4 and 5 are the calculation process performed on the first terminal (Part B); Step 6 is the calculation and encryption of the second terminal (Part B) The process of loss function and encrypted residual, that is, the process of calculating loss value and residual; Steps 7 and 8 are the operation process of the first terminal after receiving the encrypted residual. In step 8, the first terminal and The second terminal will calculate the encrypted gradient, that is, calculate the first gradient value and the second gradient value; in step 9, the third terminal (third party C) will obtain the decrypted gradient value according to the gradient values sent by the first terminal and the second terminal. gradient value, and return it accordingly; in step ten, the first terminal and the second terminal will update their respective model parameters according to the gradient value returned by the third terminal; in step eleven, the third terminal will decrypt the loss function to Judging whether the iteration is over, that is, according to the loss value to determine whether the iteration is over, where the end of the iteration means that the first model parameter and the second model parameter meet the stop update condition, and the iteration is not over, it means that the first model parameter and the second model parameter are not If the update stop condition is satisfied, if the iteration ends, the update operation of the model parameters is completed; if the iteration is not ended, return to step 3 to continue the update operation of the model parameters.

It should be noted that this embodiment is an extension of the existing federated learning scheme. Federated learning is a technology for collaborative modeling under the premise of protecting the privacy of user data. Still able to model the data effectively. According to the division of data and features, it can generally be divided into horizontal federation and vertical federation. Horizontal federation requires that each client has the same characteristics and has label data at the same time. The problem to be solved by vertical federation is that only one side has label data, and the rest have label data. The parties have only characteristic situations.

Further, a third embodiment of the method for updating model parameters of the present invention is proposed.

The difference between the third embodiment of the method for updating the model parameters and the first and/or second embodiments for the method for updating the model parameters is that the method for updating the model parameters further includes:

Step h, using the updated second model parameters as the model parameters of the loan prediction model in the second terminal to obtain the loan prediction model.

Step i, after receiving the data to be predicted, input the data to be predicted into the loan prediction model, obtain the loan probability corresponding to the data to be predicted, and push the information according to the loan probability.

After the second terminal obtains the updated second model parameters, the second terminal uses the updated second model parameters as the model parameters of the loan prediction model. It can be understood that, at this time, the second prediction model in the second terminal is a loan prediction model, and the user can also set the second prediction model as a text recognition model or a picture recognition model as required. It should be noted that the training data used above is different depending on the type of model. For example, the training data corresponding to the loan prediction model is the data generated when the user takes a loan, and the training data corresponding to the image recognition model is the picture-related data.

After the second terminal obtains the loan prediction model, the second terminal inputs the data to be predicted into the loan prediction model, and obtains the loan probability corresponding to the data to be predicted. It can be understood that the loan probability is the output result of the loan prediction model. The second terminal may push loan-related information according to the loan probability.

In this embodiment, a loan prediction model is obtained by using the updated second model parameters as the model parameters of the loan prediction model in the second terminal, and then the loan probability information is pushed according to the result obtained by the loan prediction model, so as to improve the pushing of loan-related information 's accuracy.

In addition, the present invention also provides a device for updating model parameters. Referring to FIG. 4 , the device for updating model parameters includes:

The receiving module 10 is configured to receive the encrypted first characteristic data sent by the first terminal, wherein the first characteristic data is obtained by the first terminal and at least one terminal of the same type through horizontal federated learning, the The first terminal and the second terminal are different types of terminals;

an acquisition module 20, configured to acquire the second characteristic data and the label data corresponding to the second characteristic data;

a calculation module 30, configured to calculate and obtain a residual according to the first characteristic data, the second characteristic data and the label data;

A sending module 40, configured to send the residual to the first terminal, so that the first terminal can calculate a first gradient value according to the residual, and update the first gradient according to the first gradient value the first model parameter corresponding to the feature data;

The calculation module 30 is further configured to calculate a second gradient value according to the residual;

The updating module 50 is configured to update the second model parameter corresponding to the second characteristic data according to the second gradient value.

Further, the obtaining module 20 is further configured to obtain the encrypted first feature parameter corresponding to the first feature data, and obtain the second feature parameter corresponding to the second feature data, wherein the first feature The data is the same as the device identification corresponding to the second feature data;

The calculation module 30 is further configured to calculate a loss value according to the first characteristic data, the second characteristic data, the label data, the first characteristic parameter and the second characteristic parameter;

The sending module 40 is further configured to send the loss value to a third terminal, so that the third terminal can judge whether the first model parameter and the second model parameter meet the stop updating condition according to the loss value. , and return the prompt information;

The device for updating the model parameters also includes:

The execution module is configured to return to execute the encrypted data sent by the second terminal to receive the encrypted data sent by the first terminal if it is determined according to the prompt information that the first model parameter and the second model parameter do not meet the stop updating condition. first feature data, and the step of acquiring the second feature data and the label data.

Further, the sending module 40 is further configured to send the loss value to a third terminal, so that after the third terminal receives the loss value, obtain the last update of the first model parameter and all parameters. The loss history value when the second model parameter is described, and when it is detected that the absolute value of the difference between the loss value and the loss history value is smaller than a preset threshold, determine the first model parameter and the second model parameter. The model parameter satisfies the update stop condition, and when it is detected that the absolute value is greater than or equal to the preset threshold, it is determined that the first model parameter and the second model parameter do not meet the stop update condition.

Further, the obtaining module 20 is further configured to obtain the training data corresponding to the second feature data, and obtain the third feature parameter corresponding to the second feature data;

The calculation module 30 is further configured to calculate a second gradient value according to the training data, the third characteristic parameter and the residual.

Further, the update module 50 includes:

an obtaining unit, configured to obtain an update coefficient corresponding to the second terminal;

An update unit, configured to update the second model parameter corresponding to the second feature data according to the second gradient value and the update coefficient.

Further, the computing module 30 includes:

a calculation unit for calculating the difference between the second characteristic data and the label data of the second characteristic data;

an encryption unit for encrypting the difference to obtain an encrypted difference;

The computing unit is also used to calculate and obtain the encrypted residual according to the encrypted difference and the encrypted first characteristic data; and calculate and obtain the encrypted second gradient value according to the encrypted residual;

a sending unit, configured to send the encrypted second gradient value to the third terminal, so that the third terminal can decrypt the encrypted second gradient value, and return the decrypted second gradient value;

The updating module 50 is further configured to receive the decrypted second gradient value returned by the third terminal, and update the second model parameter corresponding to the second feature data according to the decrypted second gradient value.

Further, the device for updating the model parameters also includes:

a processing module, configured to use the updated second model parameter as the model parameter of the loan prediction model in the second terminal to obtain the loan prediction model;

The input module is configured to input the data to be predicted into the loan prediction model after receiving the data to be predicted, obtain the loan probability corresponding to the data to be predicted, and push the information according to the loan probability.

The specific implementation manner of the apparatus for updating model parameters of the present invention is basically the same as that of the above-mentioned methods for updating model parameters, and details are not described herein again.

In addition, the present invention also provides a device for updating model parameters. As shown in FIG. 5 , FIG. 5 is a schematic structural diagram of a hardware operating environment involved in an embodiment of the present invention.

It should be noted that FIG. 5 can be a schematic structural diagram of the hardware operating environment of the device for updating model parameters. The device for updating the model parameters in the embodiment of the present invention may be a terminal device such as a PC and a portable computer.

As shown in FIG. 5 , the device for updating model parameters may include: a processor 1001 , such as a CPU, a memory 1005 , a user interface 1003 , a network interface 1004 , and a communication bus 1002 . Among them, the communication bus 1002 is used to realize the connection 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 and a wireless interface. Optionally, the network interface 1004 may include a standard wired interface and a wireless interface (eg, a WI-FI interface). The memory 1005 may be high-speed RAM memory, or may be non-volatile memory, such as 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 structure of the updating device for model parameters shown in FIG. 5 does not constitute a limitation on the updating device for model parameters, and may include more or less components than those shown in the figure, or combine some components, Or a different component arrangement.

As shown in FIG. 5 , the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an update program for model parameters. The operating system is a program that manages and controls the hardware and software resources of the updating device for model parameters, and supports the updating program of model parameters and the running of other software or programs.

In the model parameter updating device shown in FIG. 5 , the user interface 1003 is mainly used to connect the first terminal and the third terminal, and perform data communication with the first terminal and the third terminal respectively; the network interface 1004 is mainly used for the background server, Data communication is performed with the background server; the processor 1001 can be used to invoke the model parameter update program stored in the memory 1005, and execute the steps of the model parameter update method described above.

The specific implementation manner of the device for updating model parameters of the present invention is basically the same as that of the above-mentioned methods for updating model parameters, and details are not described herein again.

In addition, an embodiment of the present invention also provides a computer-readable storage medium, where an update program for model parameters is stored on the computer-readable storage medium, and the model parameter update program is executed by a processor to implement the above-mentioned model The steps of the parameter's update method.

The specific implementation manner of the computer-readable storage medium of the present invention is basically the same as the above-mentioned embodiments of the method for updating model parameters, and details are not described herein again.

It should be noted that, herein, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements, It also includes other elements not expressly listed or inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

From the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation. Based on this understanding, the technical solutions of the present invention can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products are stored in a storage medium (such as ROM/RAM, magnetic disk, CD-ROM), including several instructions to make a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of the present invention.

The above are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields , are similarly included in the scope of patent protection of the present invention.

Claims (10)

1. A method for updating model parameters is characterized by comprising the following steps:

the method comprises the steps that a second terminal receives encrypted first feature data sent by a first terminal, and obtains second feature data and label data corresponding to the second feature data, wherein the first feature data are obtained by the first terminal and at least one terminal of the same type through horizontal federal learning, and the first terminal and the second terminal are different types of terminals;

calculating to obtain a residual according to the first characteristic data, the second characteristic data and the label data, and sending the residual to the first terminal, so that the first terminal can calculate a first gradient value according to the residual, and update a first model parameter corresponding to the first characteristic data according to the first gradient value;

and calculating a second gradient value according to the residual error, and updating a second model parameter corresponding to the second characteristic data according to the second gradient value.

2. The method for updating model parameters according to claim 1, wherein after the steps of receiving the encrypted first feature data sent by the first terminal and acquiring the second feature data and the tag data corresponding to the second feature data, the second terminal further comprises:

acquiring an encrypted first characteristic parameter corresponding to the first characteristic data and acquiring a second characteristic parameter corresponding to the second characteristic data, wherein the device identifications corresponding to the first characteristic data and the second characteristic data are the same;

calculating a loss value according to the first characteristic data, the second characteristic data, the label data, the first characteristic parameter and the second characteristic parameter;

sending the loss value to a third terminal, so that the third terminal can judge whether the first model parameter and the second model parameter meet the update stop condition according to the loss value, and returning a prompt message;

and if the first model parameter and the second model parameter are determined not to meet the updating stop condition according to the prompt message, returning to the step of receiving the encrypted first characteristic data sent by the first terminal and acquiring the second characteristic data and the label data by the second terminal.

3. The method for updating model parameters according to claim 2, wherein the step of sending the loss value to a third terminal so that the third terminal can determine whether the first model parameter and the second model parameter satisfy the update stop condition according to the loss value comprises:

and sending the loss value to a third terminal, so that the third terminal obtains a loss history value when the first model parameter and the second model parameter are updated last time after receiving the loss value, determines that the first model parameter and the second model parameter meet an update stop condition when detecting that an absolute value of a difference value between the loss value and the loss history value is smaller than a preset threshold value, and determines that the first model parameter and the second model parameter do not meet the update stop condition when detecting that the absolute value is larger than or equal to the preset threshold value.

4. The method for updating model parameters of claim 1, wherein said step of calculating a second gradient value from said residual is preceded by the steps of:

acquiring training data corresponding to the second feature data, and acquiring a third feature parameter corresponding to the second feature data;

the step of calculating a second gradient value from the residual error comprises:

and calculating a second gradient value according to the training data, the third characteristic parameter and the residual error.

5. The method for updating model parameters according to claim 1, wherein the step of updating the second model parameters corresponding to the second feature data according to the second gradient values comprises:

acquiring an updating coefficient corresponding to the second terminal;

and updating a second model parameter corresponding to the second characteristic data according to the second gradient value and the updating coefficient.

6. The method for updating model parameters according to claim 1, wherein said step of calculating a residual from said first feature data, said second feature data and said tag data comprises:

calculating a difference value between the second characteristic data and the second characteristic data label data, and encrypting the difference value to obtain an encrypted difference value;

calculating to obtain an encrypted residual error according to the encrypted difference value and the encrypted first characteristic data;

the step of calculating a second gradient value according to the residual error, and updating a second model parameter corresponding to the second feature data according to the second gradient value includes:

calculating according to the encrypted residual error to obtain an encrypted second gradient value, sending the encrypted second gradient value to the third terminal so that the third terminal can decrypt the encrypted second gradient value, and returning the decrypted second gradient value;

and receiving the decrypted second gradient value returned by the third terminal, and updating the second model parameter corresponding to the second characteristic data according to the decrypted second gradient value.

7. The method for updating model parameters according to any one of claims 4 to 6, wherein said step of calculating a second gradient value according to said residual error and updating a second model parameter corresponding to said second feature data according to said second gradient value further comprises:

taking the updated second model parameter as the model parameter of the loan prediction model in the second terminal to obtain the loan prediction model;

after receiving data to be predicted, inputting the data to be predicted into the loan prediction model to obtain loan probability corresponding to the data to be predicted, and pushing information according to the loan probability.

8. An apparatus for updating model parameters, comprising:

the terminal comprises a receiving module, a sending module and a processing module, wherein the receiving module is used for receiving encrypted first characteristic data sent by a first terminal, the first characteristic data is obtained by the first terminal and at least one terminal of the same type through horizontal federal learning, and the first terminal and the second terminal are different types of terminals;

the acquisition module is used for acquiring second characteristic data and label data corresponding to the second characteristic data;

the calculation module is used for calculating to obtain a residual error according to the first characteristic data, the second characteristic data and the label data;

a sending module, configured to send the residual to the first terminal, so that the first terminal calculates a first gradient value according to the residual, and updates a first model parameter corresponding to the first feature data according to the first gradient value;

the calculation module is further used for calculating a second gradient value according to the residual error;

and the updating module is used for updating the second model parameter corresponding to the second characteristic data according to the second gradient value.

9. An updating apparatus of model parameters, characterized in that the updating apparatus of model parameters comprises a memory, a processor and an updating program of model parameters stored on the memory and executable on the processor, the updating program of model parameters implementing the steps of the updating method of model parameters as claimed in any one of claims 1 to 7 when executed by the processor.

10. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored thereon an update program of model parameters, which when executed by a processor implements the steps of the update method of model parameters according to any one of claims 1 to 7.

Images