WO2020029585A1

WO2020029585A1 - Neural network federation modeling method and device employing transfer learning, and storage medium

Info

Publication number: WO2020029585A1
Application number: PCT/CN2019/078522
Authority: WO
Inventors: 刘洋; 杨强; 成柯葳; 范涛; 陈天健
Original assignee: 深圳前海微众银行股份有限公司
Priority date: 2018-08-10
Filing date: 2019-03-18
Publication date: 2020-02-13
Also published as: CN109165725A; CN109165725B

Abstract

A neural network federation modeling method and device employing transfer learning, and a storage medium. The method comprises: a first terminal inputting a feature vector of first sample data to a first neural network so as to acquire a first neural network vector, determining a first gradient value and a first loss value according to the first neural network vector, and encrypting the first gradient value and the first loss value (S101); combining the encrypted first gradient value and the encrypted first loss value with a received encrypted second gradient value and encrypted second loss value sent by a second terminal so as to acquire an encrypted third loss value and encrypted third gradient value (S102); sending the encrypted third loss value and the encrypted third gradient value to a third terminal, and determining, according to the third loss value and a historical loss value decrypted and returned by the third terminal, whether a model to be trained converges (S103); and if the model converges, using a model parameter at the time of convergence to establish the model (S104).

Description

Neural network federation modeling method, equipment and storage medium based on transfer learning Ranch

Technical field

The present invention relates to the technical field of machine learning, and in particular, to a neural network federation modeling method, device, and storage medium based on transfer learning.

Background technique

With the rapid development of machine learning, machine learning can be applied to various fields, such as data mining, computer vision, natural language processing, biometric recognition, medical diagnosis, detection of credit card fraud, securities market analysis, and DNA sequence sequencing. Machine learning includes a learning part and an execution part. The learning part uses sample data to modify the system's knowledge base to improve the efficiency of the system's execution part to complete the task. The execution part completes the task according to the knowledge base, and at the same time feeds the obtained information to the learning part.

At present, because the sample data of all parties are closely related, if machine learning uses only one party's sample data, the learned model is inaccurate. To solve the above problems, by combining the sample data of all parties, applying logistic regression or decision trees, etc. Single layer simple model for machine learning. However, due to the need to combine the sample data of the parties, there is a case where the sample data of one party is known by the other party. In addition, the current joint learning is mainly based on the common sample data, and the common sample data of the parties is limited, making the parties Unique sample data cannot be effectively used.

Therefore, how to improve the privacy and utilization of the sample data of all parties is an urgent problem.

Summary of the invention

The main purpose of the present invention is to provide a neural network federation modeling method, equipment and storage medium based on transfer learning, which aims to improve the privacy and utilization of sample data of all parties.

To achieve the above object, the present invention provides a neural network federation modeling method based on transfer learning. The neural network federation modeling method based on transfer learning includes the following steps:

The first terminal inputs a feature vector of the first sample data to a first neural network to obtain a first neural network vector, and determines a first gradient value and a first loss value according to the first neural network vector, and applies the The first gradient value and the first loss value are encrypted;

Combining the encrypted first gradient value and the first loss value with the received encrypted second gradient value and the second loss value sent by the second terminal to obtain the encrypted third loss value and the third gradient value, The second terminal inputs the second sample data to the second neural network to obtain a second neural network vector, determines a second gradient value and a second loss value according to the first neural network vector, and applies the After the second gradient value and the second loss value are encrypted, and transmitted to the first terminal, the feature dimensions of the first neural network vector and the second neural network vector are the same;

Sending the encrypted third loss value and the third gradient value to a third terminal, and determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption;

If the model to be trained converges, the model parameters to be converged are used to establish the model to be trained.

Further, when the third terminal receives the encrypted third loss value sent by the first terminal, it obtains the encrypted historical loss value sent by the first terminal last time, and according to the pre-stored private key pair The encrypted third loss value, the historical loss value, and the third gradient value are decrypted, and the decrypted third loss value, the historical loss value, and the third gradient value are returned to the first terminal.

Further, the step of determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption includes:

Receiving a third loss value and a historical loss value returned by the third terminal decryption;

Calculating a difference between the third loss value and the historical loss value returned by the decryption, and determining whether the difference is less than or equal to a preset threshold;

If the difference is less than or equal to a preset threshold, it is determined that the model to be trained converges, otherwise it is determined that the model to be trained does not converge.

Further, the encrypted first gradient value and the first loss value are combined with the received encrypted second gradient value and the second loss value sent by the second terminal to obtain an encrypted third loss value and After the step of the third gradient value, the method further includes:

The second terminal combines the encrypted second gradient value with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sends the encrypted fourth gradient value. The third terminal;

After the step of determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption, the method further includes:

If the model to be trained does not converge, a gradient update instruction is sent to the third terminal, and the third terminal decrypts the encrypted third gradient value and the fourth gradient value according to the gradient update instruction, and decrypts the decrypted first A three gradient value is returned to the first terminal, and a decrypted fourth gradient value is returned to the second terminal;

The first terminal updates the local gradient of the first neural network according to the third gradient value returned by the third terminal decryption, and after the update is completed, returns to the execution step: the first terminal updates the characteristics of the first sample data The vector is input to the first neural network to obtain a first neural network vector, and a first gradient value and a first loss value are determined according to the first neural network vector, and the first gradient value and the first loss value are determined. encryption;

The second terminal updates the local gradient of the second neural network according to the fourth gradient value decrypted by the third terminal, and after the update is completed, returns to the execution step: the second terminal encrypts the second gradient Value, combined with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sending the encrypted fourth gradient value to the third terminal.

Further, the third terminal generates a set of public key and private key, and transmits the public key to the first terminal and the second terminal, and the first terminal and the first terminal The two terminals respectively store the public key in respective preset storage areas.

Further, the third terminal generates a set of public key and private key at a preset interval, and transmits the generated public key to the first terminal and the second terminal, and the first terminal and the The second terminal updates the public key stored in the respective preset storage area according to the received public key, respectively.

Further, the step of encrypting the first gradient value and the first loss value includes:

The first terminal obtains a public key from a preset storage area, and performs homomorphic encryption on the first gradient value and the first loss value according to the public key.

Further, the neural network federation modeling method based on transfer learning further includes:

When the configuration instruction of the initial weight is detected, counting the number of synapses in the first neural network, and calling a preset random number generator to generate a set of random numbers corresponding to the number of synapses;

Configure the initial weights of each synapse in the first neural network according to the generated set of random numbers.

In addition, in order to achieve the above object, the present invention also provides a neural network federation modeling device based on transfer learning. The neural network federation modeling device based on transfer learning includes: a memory, a processor, and a memory stored on the memory and A neural network federation modeling program based on transfer learning that can be run on the processor, and the neural network federation modeling program based on transfer learning implements the neural network based on transfer learning as described above when executed by the processor Steps in federated modeling method.

The present invention also provides a storage medium storing a neural network federation modeling program based on transfer learning, and implementing the neural network federation modeling program based on transfer learning to implement The steps of a neural network federation modeling method for transfer learning.

The present invention provides a neural network federation modeling method, device, and storage medium based on transfer learning. The present invention inputs feature vectors of sample data of two parties into two neural networks, and two parties correspondingly obtain two neural networks with the same feature dimension. Vector, and get the respective gradient and loss values according to the neural network vectors with the same feature dimensions, and one of them encrypts the gradient and loss values, and then combines the encrypted gradient and loss values sent by the other party To obtain the encrypted total loss value and total gradient value, and transmit the encrypted total loss value to a third party, and finally determine whether the model to be trained converges based on the decrypted total loss value and historical loss value returned by the third party. When the training model converges, the model parameters to be trained are used to establish the model to be trained. Because the data to be transmitted by both parties is encrypted, and joint training can be performed in an encrypted form, the privacy of the sample data of all parties is effectively improved. , Multi-layer neural network of joint parties for machine learning can effectively use the sample of all parties Data, improve the utilization of the parties to the sample data.

BRIEF DESCRIPTION OF THE DRAWINGS

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

2 is a schematic flowchart of a first embodiment of a neural network federation modeling method based on transfer learning according to the present invention;

3 is a schematic flowchart of a second embodiment of a neural network federation modeling method based on transfer learning according to the present invention.

The realization of the purpose, functional characteristics and advantages of the present invention will be further explained with reference to the embodiments and the drawings.

detailed description

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

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

The neural network federation modeling device based on the transfer learning in the embodiment of the present invention may be a PC, or a mobile terminal device with a display function such as a smart phone, a tablet computer, a portable computer, or the like.

As shown in FIG. 1, the neural network federation modeling device based on transfer learning may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen, an input unit such as a keyboard, and the optional user interface 1003 may further include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory. memory), such as disk storage. The memory 1005 may optionally be a storage device independent of the foregoing processor 1001.

Those skilled in the art can understand that the structure of the neural network federation modeling device based on transfer learning shown in FIG. 1 does not constitute a limitation on the neural network federation modeling device based on transfer learning, and may include more or more than illustrated Fewer components, or some components combined, or different component arrangements.

As shown in FIG. 1, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a neural network federation modeling program based on transfer learning.

In the neural network federation modeling device based on transfer learning shown in FIG. 1, the network interface 1004 is mainly used to connect to the background server and perform data communication with the background server; the user interface 1003 is mainly used to connect to the client (user), and The client performs data communication; and the processor 1001 can be used to call a neural network federation modeling program based on transfer learning stored in the memory 1005, and execute the following steps:

Further, the processor 1001 may be used to call a neural network federation modeling program based on transfer learning stored in the memory 1005, and further perform the following steps:

If the difference is less than or equal to a preset threshold, it is determined that the model to be trained is in a convergence state, otherwise it is determined that the model to be trained is not in a convergence state.

The specific embodiments of the neural network federation modeling device based on transfer learning of the present invention are basically the same as the specific embodiments of the neural network federation modeling method based on transfer learning described below, and will not be repeated here.

Referring to FIG. 2, FIG. 2 is a schematic flowchart of a first embodiment of a neural network federation modeling method based on transfer learning according to the present invention.

In step S101, a first terminal inputs a feature vector of first sample data to a first neural network, obtains a first neural network vector, and determines a first gradient value and a first loss value according to the first neural network vector. A gradient value and a first loss value are encrypted;

In this embodiment, the present invention can combine multi-party sample data to train the model to be trained. The following uses the joint two-party sample data as an example to explain, where one sample data is the first sample data and is stored in the first terminal. And the first neural network is deployed at the first terminal, and the other party ’s sample data is the second sample data, which is stored in the second terminal, and the second neural network is deployed at the second terminal, and the first terminal is connected to the second terminal, In addition, in order to ensure the privacy of the sample data of both parties, a third terminal is introduced. The third terminal stores a set of public and private keys required for encryption, and the first terminal is connected to the third terminal. When transmitting data, the second terminal is connected to the third terminal and can transmit data. The labeling of the sample data of both parties includes labeling the first sample data without labeling the second sample data, not labeling the first sample data, and labeling the second sample data, the first sample data, and the second sample data are labeled or Neither the first sample data nor the second sample data are labeled. It should be noted that, in this embodiment, the labeling of the sample data of both parties is not specifically limited. It should be noted that the network parameters of the first neural network and the second neural network can be set by those skilled in the art based on the actual situation and the actual situation, which is not specifically limited in this embodiment. The network parameters include, but are not limited to, the number of network nodes in each layer, the number of hidden layers, the initial weight of each synapse, the learning rate, dynamic parameters, allowable errors, the number of iterations, and the activation function.

In this embodiment, in the process of training the model to be trained, the first terminal inputs the first sample data to the first neural network, and when the last layer of the first neural network is reached, the characteristics of the first sample data are obtained. Expression, that is, the first neural network vector, and the first gradient value and the first loss value are determined according to the first neural network vector, that is, the first gradient value is the gradient function of the model to be trained. For the gradient of the first common feature vector, the first A loss value is a loss of the loss function of the model to be trained for the first common feature vector, and the first gradient value and the first loss value are encrypted.

Specifically, the third terminal generates a set of public key and private key, and transmits the public key to the first terminal and the second terminal, and the first terminal and the second terminal store the public key respectively in respective presets. In the storage area, after obtaining the first gradient value and the first loss value, the first terminal obtains a public key from a preset storage area, and homomorphizes the first gradient value and the first loss value according to the public key. Encrypt, and send the encrypted first gradient value and first loss value to the second terminal. Among them, the encryption method is homomorphic encryption (processing the homomorphically encrypted data to obtain an output, and decrypting this output, the result is the same as the output obtained by processing the unencrypted original data in the same method) , Can be calculated in the form of cipher text, does not affect the results obtained by the calculation.

Step S102: Combine the encrypted first gradient value and the first loss value with the received encrypted second gradient value and the second loss value sent by the second terminal to obtain the encrypted third loss value and the third gradient value. ;

In this embodiment, the second terminal inputs the second sample data to the second neural network for iteration, and when it reaches the last layer of the second neural network, it obtains the characteristic expression of the second sample data, that is, the second neural network vector, and Determine the second gradient value and the second loss value according to the first neural network vector, that is, the second gradient value is the gradient function of the model to be trained for the gradient of the second common feature vector, and the second loss value is the loss function of the model to be trained for The second common feature vector is lost, and the second gradient value and the second loss value are encrypted and sent to the first terminal, that is, the public key in the pre-stored storage area is obtained, and the second gradient value and the second loss value are the same. State encryption, and send the encrypted second gradient value and the second loss value to the first terminal. Wherein, the feature dimensions of the first neural network vector and the second neural network vector are the same.

The first terminal combines the encrypted first gradient value and the first loss value with the encrypted second gradient value and the second loss value sent by the second terminal to obtain the encrypted third loss value and the third gradient value. That is, the first terminal receives the encrypted second gradient value and the second loss value sent by the second terminal, and combines the encrypted first gradient value and the second gradient value to obtain an encrypted third gradient value, and combines the encrypted first A loss value and a second loss value to obtain an encrypted third loss value.

Further, in order to further improve the security of the data of both parties, during the training of the model, the first terminal and the second terminal obtain a public key from the third terminal at a preset interval to update the local storage in the pre- The public key in the storage area is set. Specifically, a timer is set in the third terminal. When the model is trained, the timer starts timing. When the timer reaches a preset time, the third terminal generates a group of public keys. Key and private key, and sends the public key to the first terminal and the second terminal, and the timer restarts, and the first terminal and the second terminal update the public key stored in the preset storage area . It should be noted that the preset time can be set by a person skilled in the art based on the actual situation, which is not specifically limited in this embodiment.

Step S103: Send the encrypted third loss value and the third gradient value to the third terminal, and determine whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption;

In this embodiment, the first terminal sends the encrypted third loss value and the third gradient value to the third terminal, and the third terminal receives the encrypted third loss value and the third gradient value sent by the first terminal, and obtains The first encrypted historical loss value sent by the first terminal, and the encrypted third loss value, historical loss value, and third gradient value are decrypted according to the pre-stored private key, and the decrypted third loss value and history are decrypted. The loss value and the third gradient value are returned to the first terminal, and the first terminal determines whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption;

Specifically, the first terminal receives the third loss value and the historical loss value returned by the third terminal, and then calculates a difference between the third loss value and the historical loss value returned by the decryption, and determines whether the difference is less than or equal to A preset threshold. If the difference is less than or equal to the preset threshold, it is determined that the model to be trained converges, otherwise it is determined that the model to be trained does not converge. It should be noted that the preset threshold may be set by a person skilled in the art based on actual conditions, and this embodiment does not specifically limit this.

Step S104: If the model to be trained converges, establish the model to be trained based on the model parameters during convergence.

In this embodiment, if the model to be trained converges, the model parameter to be trained is used to establish the model to be trained. In specific implementation, the operation of determining whether the model to be converged can also be performed by the third terminal. Specifically, the third terminal receives the encrypted third loss value sent by the first terminal, and obtains the encryption history sent by the first terminal last time. Loss value, and then decrypt the encrypted third loss value and historical loss value according to the pre-stored private key, and determine whether the model to be trained converges based on the decrypted third loss value and historical loss value, and determine the model convergence Deploying on the third terminal can reduce the resource occupation of the second terminal or the third terminal, and improve the resource utilization of the third terminal.

Further, after step S102, the method further includes:

In step a, the second terminal combines the encrypted second gradient value with the encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sends the encrypted fourth gradient value to the first Three terminals

In this embodiment, when the first terminal performs the determination of the gradient value and the loss value, the second terminal combines the encrypted second gradient value with the encrypted first gradient value sent by the first terminal to obtain encryption. The third terminal that sends the encrypted fourth gradient value, that is, the second terminal receives the encrypted first gradient value sent by the first terminal, and combines the encrypted second gradient value to obtain the encrypted first gradient value. Four gradient values.

After step S103, the method further includes:

Step b, if the model to be trained does not converge, send a gradient update instruction to the third terminal, and the third terminal decrypts the encrypted third gradient value and the fourth gradient value according to the gradient update instruction, and decrypts the decrypted third The gradient value is returned to the first terminal, and the decrypted fourth gradient value is returned to the second terminal;

If the model to be trained is not in convergence, the local gradients of the first neural network and the second neural network need to be updated, that is, the first terminal sends a gradient update instruction to the third terminal, and the third terminal sends the encrypted first The three gradient values and the fourth gradient value are decrypted, and the decrypted third gradient value is returned to the first terminal, and the decrypted fourth gradient value is returned to the second terminal.

The first terminal updates the local gradient of the first neural network according to the third gradient value decrypted by the third terminal, and after the update is completed, returns to step S101, that is, the first terminal inputs the feature vector of the first sample data to the first A neural network obtains a first neural network vector, determines a first gradient value and a first loss value according to the first neural network vector, and encrypts the first gradient value and the first loss value.

The second terminal updates the local gradient of the second neural network according to the fourth gradient value decrypted by the third terminal, and after the update is completed, returns to execute step a, that is, the second terminal encrypts the second gradient value and the received second gradient value. The encrypted first gradient value sent by the first terminal is combined to obtain an encrypted fourth gradient value, and the encrypted fourth gradient value is sent to a third terminal.

In specific implementation, if the structure of the neural network layer after the labeled neural network layer in the first neural network and the second neural network is completely the same, the first terminal transmits the weight parameter value WA of the first neural network to the second after being encrypted. Terminal, and the second terminal transmits the weight parameter value WB of the second neural network to the first terminal, and the first terminal trains the first neural network according to the encrypted weight parameter values WA and WB, until convergence, and the second terminal The second neural network is trained according to the encrypted weight parameter values WA and WB until convergence, and when both the first neural network and the second neural network converge, a model to be trained is established according to the weight parameter values WA and WB in the convergence state.

In this embodiment, the present invention inputs the feature vectors of the sample data of the two parties into two neural networks respectively, and the two parties correspondingly obtain two neural network vectors with the same feature dimension, and obtain the respective neural network vectors according to the respective neural network vectors with the same feature dimension. Gradient value and loss value, and one of them encrypts the gradient value and loss value, and then combines the encrypted gradient value and loss value sent by the other party to obtain the encrypted total loss value and total gradient value, and encrypts the encrypted value. The total loss value is transmitted to a third party. Finally, based on the decrypted total loss value and the historical loss value returned by the third party, it is determined whether the model to be trained converges. If the model to be trained converges, the model parameters at the time of convergence are used to establish the model to be trained. Because the data that the two parties need to transmit is encrypted, and joint training can be performed in an encrypted form, the privacy of the sample data of all parties is effectively improved, and at the same time, the multi-layer neural network of the parties is used for machine learning, which can effectively Use the sample data of all parties to improve the utilization of the sample data of all parties.

Further, referring to FIG. 3, based on the first embodiment described above, a second embodiment of the neural network federation modeling method based on transfer learning of the present invention is proposed. The difference from the foregoing embodiment is that the neural network federation based on transfer learning The modeling method also includes:

Step 105: when the configuration instruction of the initial weight is detected, count the number of synapses in the first neural network, and call a preset random number generator to generate a set of random numbers corresponding to the number of synapses;

In this embodiment, before training the model to be trained, the initial weights of each synapse in the model to be trained need to be configured. When a configuration instruction of the initial weight is detected, the first terminal counts the synapses in the first neural network. Number of contacts, and call a preset random number generator to generate a set of random numbers corresponding to the number of synapses, while the second terminal counts the number of synapses in the second neural network, and calls the preset random number generator To generate another set of random numbers corresponding to the number of synapses. It should be noted that the value range of the random number can be set by a person skilled in the art based on the actual situation, which is not specifically limited in this embodiment. Preferably, the value range of the random number is -0.5 to +0.5.

Step 106: Configure an initial weight of each synapse in the first neural network according to the generated set of random numbers.

In this embodiment, the first terminal configures the initial weight of each synapse in the first neural network according to the generated set of random numbers, that is, from the generated set of random numbers, according to the generated sequence of random numbers A random number is sequentially selected as the initial weight and assigned to a synapse in the first neural network; the second terminal configures the initial weight of each synapse in the second neural network according to another set of random numbers generated, that is, based on The order of the size of the generated another set of random numbers. From the generated another set of random numbers, a random number is sequentially selected as the initial weight and assigned to a synapse in the second neural network. Each synapse is configured once. Initial weight.

In this embodiment, the present invention uses a random number generator to assign random initial weights to each synapse of the first neural network and the second neural network in the model to be trained to prevent the initial weights of the synapses from being the same, resulting in training During the process, the weight of each synapse is always kept equal, which effectively improves the accuracy of the model obtained by training.

In addition, an embodiment of the present invention further provides a storage medium that stores a neural network federation modeling program based on transfer learning. When the neural network federation modeling program based on transfer learning is executed by a processor, the storage medium is executed. The following steps:

Further, when the neural network federation modeling program based on transfer learning is executed by a processor, the following steps are also performed:

The specific embodiments of the storage medium of the present invention are basically the same as the above embodiments of the neural network federation modeling method based on transfer learning, and will not be repeated here.

It should be noted that, in this article, the terms "including", "including" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, It also includes other elements not explicitly listed, or elements inherent to such a process, method, article, or system. Without more restrictions, an element limited by the sentence "including a ..." does not exclude the existence of other identical elements in the process, method, article, or system that includes the element.

The sequence numbers of the foregoing embodiments of the present invention are only for description, and do not represent the superiority or inferiority of the embodiments.

Through the description of the above embodiments, those skilled in the art can clearly understand that the methods in the above embodiments can be implemented by means of software plus a necessary universal hardware platform, and of course, also by hardware, but in many cases the former is better. Implementation. Based on such an understanding, the technical solution of the present invention in essence or a part that contributes to the existing technology can be embodied in the form of a software product, which is stored in a storage medium such as ROM / RAM as described above , Magnetic disk, optical disc), including a number of instructions to enable a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in the embodiments of the present invention.

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

Claims

A neural network federation modeling method based on transfer learning, characterized in that the neural network federation modeling method based on transfer learning includes the following steps:

The first terminal inputs a feature vector of the first sample data to a first neural network to obtain a first neural network vector, and determines a first gradient value and a first loss value according to the first neural network vector, and applies the The first gradient value and the first loss value are encrypted;

Combining the encrypted first gradient value and the first loss value with the received encrypted second gradient value and the second loss value sent by the second terminal to obtain the encrypted third loss value and the third gradient value, The second terminal inputs the second sample data to the second neural network to obtain a second neural network vector, determines a second gradient value and a second loss value according to the first neural network vector, and applies the After the second gradient value and the second loss value are encrypted, and transmitted to the first terminal, the feature dimensions of the first neural network vector and the second neural network vector are the same;

Sending the encrypted third loss value and the third gradient value to a third terminal, and determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption;

If the model to be trained converges, the model parameters to be converged are used to establish the model to be trained.
The neural network federation modeling method based on transfer learning according to claim 1, wherein, when the third terminal receives the encrypted third loss value sent by the first terminal, the first terminal acquires the first The encrypted historical loss value sent by the terminal last time, and the encrypted third loss value, historical loss value, and third gradient value are decrypted according to the pre-stored private key, and the decrypted third loss value and historical loss value are decrypted. And a third gradient value is returned to the first terminal.
The neural network federation modeling method based on transfer learning according to claim 2, wherein the step of determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption comprises:

Receiving a third loss value and a historical loss value returned by the third terminal decryption;

Calculating a difference between the third loss value and the historical loss value returned by the decryption, and determining whether the difference is less than or equal to a preset threshold;

If the difference is less than or equal to a preset threshold, it is determined that the model to be trained converges, otherwise it is determined that the model to be trained does not converge.
The neural network federation modeling method based on transfer learning according to claim 1, wherein the encrypted first gradient value and the first loss value are compared with the encrypted encrypted data sent by the second terminal. After the step of combining the second gradient value and the second loss value to obtain the encrypted third loss value and the third gradient value, the method further includes:

The second terminal combines the encrypted second gradient value with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sends the encrypted fourth gradient value. The third terminal;

After the step of determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption, the method further includes:

If the model to be trained does not converge, a gradient update instruction is sent to the third terminal, and the third terminal decrypts the encrypted third gradient value and the fourth gradient value according to the gradient update instruction, and decrypts the decrypted first A three gradient value is returned to the first terminal, and a decrypted fourth gradient value is returned to the second terminal;

The first terminal updates the local gradient of the first neural network according to the third gradient value returned by the third terminal decryption, and after the update is completed, returns to the execution step: the first terminal updates the characteristics of the first sample data The vector is input to the first neural network to obtain a first neural network vector, and a first gradient value and a first loss value are determined according to the first neural network vector, and the first gradient value and the first loss value are determined. encryption;

The second terminal updates the local gradient of the second neural network according to the fourth gradient value decrypted by the third terminal, and after the update is completed, returns to the execution step: the second terminal encrypts the second gradient Value, combined with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sending the encrypted fourth gradient value to the third terminal.
The neural network federation modeling method based on transfer learning according to claim 1, wherein the third terminal generates a set of public key and private key, and transmits the public key to the first A terminal and the second terminal, and the first terminal and the second terminal respectively store the public key in respective preset storage areas.
The neural network federation modeling method based on transfer learning according to claim 5, wherein the third terminal generates a set of public key and private key at a preset interval, and generates the generated public key And transmitting to the first terminal and the second terminal, and the first terminal and the second terminal update the public key stored in the respective preset storage areas according to the received public key, respectively.
The neural network federation modeling method based on transfer learning according to claim 5, wherein the step of encrypting the first gradient value and the first loss value comprises:

The first terminal obtains a public key from a preset storage area, and performs homomorphic encryption on the first gradient value and the first loss value according to the public key.
The neural network federation modeling method based on transfer learning according to claim 1, wherein the neural network federation modeling method based on transfer learning further comprises:

When the configuration instruction of the initial weight is detected, counting the number of synapses in the first neural network, and calling a preset random number generator to generate a set of random numbers corresponding to the number of synapses;

Configure the initial weights of each synapse in the first neural network according to the generated set of random numbers.
A neural network federation modeling device based on transfer learning, characterized in that the neural network federation modeling device based on transfer learning includes: a memory, a processor, and can be stored on the memory and can be on the processor A running neural network federation modeling program based on transfer learning, which implements the following steps when executed by the processor:

The first terminal inputs a feature vector of the first sample data to a first neural network to obtain a first neural network vector, and determines a first gradient value and a first loss value according to the first neural network vector, and applies the The first gradient value and the first loss value are encrypted;

Combining the encrypted first gradient value and the first loss value with the received encrypted second gradient value and the second loss value sent by the second terminal to obtain the encrypted third loss value and the third gradient value, The second terminal inputs the second sample data to the second neural network to obtain a second neural network vector, determines a second gradient value and a second loss value according to the first neural network vector, and applies the After the second gradient value and the second loss value are encrypted, and transmitted to the first terminal, the feature dimensions of the first neural network vector and the second neural network vector are the same;

Sending the encrypted third loss value and the third gradient value to a third terminal, and determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption;

If the model to be trained converges, the model parameters to be converged are used to establish the model to be trained.
The neural network federation modeling device based on transfer learning according to claim 9, characterized in that when the third terminal receives the encrypted third loss value sent by the first terminal, the first terminal acquires the first The encrypted historical loss value sent by the terminal last time, and the encrypted third loss value, historical loss value, and third gradient value are decrypted according to the pre-stored private key, and the decrypted third loss value and historical loss value are decrypted. And a third gradient value is returned to the first terminal.
The neural network federation modeling device based on transfer learning according to claim 10, wherein when the neural network federation modeling program based on transfer learning is executed by the processor, the following steps are further implemented:

Receiving a third loss value and a historical loss value returned by the third terminal decryption;

Calculating a difference between the third loss value and the historical loss value returned by the decryption, and determining whether the difference is less than or equal to a preset threshold;

If the difference is less than or equal to a preset threshold, it is determined that the model to be trained converges, otherwise it is determined that the model to be trained does not converge.
The neural network federation modeling device based on transfer learning according to claim 9, wherein when the neural network federation modeling program based on transfer learning is executed by the processor, the following steps are further implemented:

The second terminal combines the encrypted second gradient value with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sends the encrypted fourth gradient value. The third terminal;

After the step of determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption, the method further includes:

If the model to be trained does not converge, a gradient update instruction is sent to the third terminal, and the third terminal decrypts the encrypted third gradient value and the fourth gradient value according to the gradient update instruction, and decrypts the decrypted first A three gradient value is returned to the first terminal, and a decrypted fourth gradient value is returned to the second terminal;

The first terminal updates the local gradient of the first neural network according to the third gradient value returned by the third terminal decryption, and after the update is completed, returns to the execution step: the first terminal updates the characteristics of the first sample data The vector is input to the first neural network to obtain a first neural network vector, and a first gradient value and a first loss value are determined according to the first neural network vector, and the first gradient value and the first loss value are determined. encryption;

The second terminal updates the local gradient of the second neural network according to the fourth gradient value decrypted by the third terminal, and after the update is completed, returns to the execution step: the second terminal encrypts the second gradient Value, combined with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sending the encrypted fourth gradient value to the third terminal.
The neural network federation modeling device based on transfer learning according to claim 9, wherein the third terminal generates a set of public key and private key, and transmits the public key to the first A terminal and the second terminal, and the first terminal and the second terminal respectively store the public key in respective preset storage areas.
The neural network federation modeling device based on transfer learning according to claim 13, wherein the third terminal generates a set of public keys and private keys at preset intervals, and generates the generated public keys And transmitting to the first terminal and the second terminal, and the first terminal and the second terminal update the public key stored in the respective preset storage areas according to the received public key, respectively.
A storage medium is characterized in that the storage medium stores a neural network federation modeling program based on transfer learning, and the neural network federation modeling program based on transfer learning implements the following steps when executed by a processor:

The first terminal inputs a feature vector of the first sample data to a first neural network to obtain a first neural network vector, and determines a first gradient value and a first loss value according to the first neural network vector, and applies the The first gradient value and the first loss value are encrypted;

Combining the encrypted first gradient value and the first loss value with the received encrypted second gradient value and the second loss value sent by the second terminal to obtain the encrypted third loss value and the third gradient value, The second terminal inputs the second sample data to the second neural network to obtain a second neural network vector, determines a second gradient value and a second loss value according to the first neural network vector, and applies the After the second gradient value and the second loss value are encrypted, and transmitted to the first terminal, the feature dimensions of the first neural network vector and the second neural network vector are the same;

Sending the encrypted third loss value and the third gradient value to a third terminal, and determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption;

If the model to be trained converges, the model parameters to be converged are used to establish the model to be trained.
The storage medium according to claim 15, wherein when the third terminal receives the encrypted third loss value sent by the first terminal, the third terminal obtains an encryption history sent by the first terminal last time. Loss value, and decrypt the encrypted third loss value, historical loss value, and third gradient value according to the pre-stored private key, and return the decrypted third loss value, historical loss value, and third gradient value to the Mentioned first terminal.
The storage medium according to claim 16, wherein when the neural network federation modeling program based on transfer learning is executed by a processor, the following steps are further implemented:

Receiving a third loss value and a historical loss value returned by the third terminal decryption;

Calculating a difference between the third loss value and the historical loss value returned by the decryption, and determining whether the difference is less than or equal to a preset threshold;

If the difference is less than or equal to a preset threshold, it is determined that the model to be trained converges, otherwise it is determined that the model to be trained does not converge.
The storage medium according to claim 15, wherein when the neural network federation modeling program based on transfer learning is executed by a processor, the following steps are further implemented:

The second terminal combines the encrypted second gradient value with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sends the encrypted fourth gradient value. The third terminal;

After the step of determining whether the model to be trained converges according to the third loss value and the historical loss value returned by the third terminal decryption, the method further includes:

If the model to be trained does not converge, a gradient update instruction is sent to the third terminal, and the third terminal decrypts the encrypted third gradient value and the fourth gradient value according to the gradient update instruction, and decrypts the decrypted first A three gradient value is returned to the first terminal, and a decrypted fourth gradient value is returned to the second terminal;

The first terminal updates the local gradient of the first neural network according to the third gradient value returned by the third terminal decryption, and after the update is completed, returns to the execution step: the first terminal updates the characteristics of the first sample data The vector is input to the first neural network to obtain a first neural network vector, and a first gradient value and a first loss value are determined according to the first neural network vector, and the first gradient value and the first loss value are determined. encryption;

The second terminal updates the local gradient of the second neural network according to the fourth gradient value decrypted by the third terminal, and after the update is completed, returns to the execution step: the second terminal encrypts the second gradient Value, combined with the received encrypted first gradient value sent by the first terminal to obtain an encrypted fourth gradient value, and sending the encrypted fourth gradient value to the third terminal.
The storage medium according to claim 15, wherein the third terminal generates a set of public key and private key, and transmits the public key to the first terminal and the second terminal , The first terminal and the second terminal respectively store the public key in respective preset storage areas.
The storage medium according to claim 19, wherein the third terminal generates a set of a public key and a private key at a preset interval, and transmits the generated public key to the first terminal and The second terminal updates the public key stored in the respective preset storage area by the first terminal and the second terminal according to the received public key, respectively. Ranch