CN115409095A

CN115409095A - Multi-party Poisson regression privacy computation model training method and device and storage medium

Info

Publication number: CN115409095A
Application number: CN202210986434.8A
Authority: CN
Inventors: 赵珙炜; 薛瑞东; 田�健
Original assignee: Beijing Rongshulianzhi Technology Co ltd
Current assignee: Beijing Rongshulianzhi Technology Co ltd
Priority date: 2022-08-17
Filing date: 2022-08-17
Publication date: 2022-11-29
Anticipated expiration: 2042-08-17
Also published as: CN115409095B

Abstract

The embodiment of the invention provides a method, a device and a storage medium for training a multi-party Poisson regression privacy computation model, wherein the method comprises the following steps: step 121, calculating according to the held sample variables and the label side model training parameters to obtain a label side local model result; step 122, receiving local model results of other data source sides locally generated by the other data source sides according to the data source side model training parameters and local sample variables held by the other data source sides; step 123, obtaining a predicted value under the coordination of the data source side representatives; step 124, calculating the parameter gradient of the label side through residual errors under the cooperation of the data source side representation, and adjusting the training parameters of the label side model through the parameter gradient of the label side; 125. and calculating respective data source party parameter gradients through residual errors by matching with the data source parties, so as to realize model training iterative calculation. The count of events is modeled statistically without the use of a trusted third party.

Description

Multi-party Poisson regression privacy computation model training method and device and storage medium

Technical Field

The invention relates to the field of model training, in particular to a method and a device for training a multi-party Poisson regression privacy calculation model and a storage medium.

Background

The Privacy Computation is called ' Computation for protecting data Privacy ', or ' Privacy Preserving Computation ' (Privacy Preserving Computation) ', and refers to a technical system for realizing Computation and data value mining on the premise of ensuring data security and Privacy of each party when a plurality of parties perform joint Computation. The privacy calculation can complete data applications such as fusion calculation and combined modeling of data of all parties under the condition that plaintext data of all parties do not need to be exported, flow and share of data 'value' and 'knowledge' are achieved on the basis of meeting data privacy safety, and 'data is available and invisible' really is achieved. Currently, the most widely focused application field in the field of privacy computing is federal learning and a machine learning method based on MPC (Secure multi-party computing). In the process of implementing the invention, the applicant finds that at least the following problems exist in the prior art: when modeling the "count" of events from a statistical perspective, a trusted third party must be employed to ensure that the multi-party data is not revealed.

Disclosure of Invention

The embodiment of the invention provides a method, a device and a storage medium for training a multiparty Poisson regression privacy computation model, wherein a homomorphic encryption means is adopted to encrypt a local model result, a trusted third party is not needed, multiparty data transmission can be ensured not to be leaked, and privacy is protected; while the count of events can be modeled statistically.

In order to achieve the above object, in a first aspect, an embodiment of the present invention provides a longitudinal federal multi-party poisson regression privacy computation model training method, which is applied to a label party, and includes:

step 11, initiating a request for training a multi-party Poisson regression privacy calculation initial model to a data source side; the multi-party Poisson regression privacy calculation initial model comprises variables for measuring and counting;

step 12, carrying out model iterative training on the initial model of the multiparty Poisson regression privacy calculation until the convergence condition of the initial model of the multiparty Poisson regression privacy calculation is met, and obtaining a trained multiparty Poisson regression privacy calculation model; the multi-party Poisson regression privacy calculation model is used for predicting the counting of events in the same scene as the scene to which the training data belongs;

when the model iterative training is carried out on the initial model of the multi-party Poisson regression privacy calculation each time, the method specifically comprises the following steps:

step 121, calculating according to the held sample variable and the label side model training parameter to obtain a label side local model result; the local model result of the label side is encrypted by adopting a first public key, and the first public key belongs to a first key pair which is generated by a data source side representative and is used for homomorphic encryption; the label square model training parameters adopted for the first time are initial label square model training parameters;

step 122, receiving corresponding data source side local model results generated locally by other data source sides according to the data source side model training parameters and local sample variables held by the other data source sides, wherein the data source side local model results are encrypted by adopting a first public key; the method comprises the following steps that firstly, data source side model training parameters are used as initial parameters for training each data source side model; the data source side represents a data source side selected arbitrarily from the data source sides, and the other data source sides are data source sides except the data source side;

step 123, sending the sum of the local model results of the other data source sides and the local model result of the label side to a data source side representative, and obtaining a predicted value under the coordination of the data source side representative; the sum of the local model results of other data sources and the local model result of the label side is encrypted by the label side; the method for obtaining the predicted value under the cooperation of the data source side representation comprises the following steps: the data source side representative inputs the sum of the local model results of other data source sides and the local model result of the tag side and the local model result of the data source side representative into the multi-side Poisson regression privacy calculation initial model to obtain a predicted value, and the predicted value is sent to the tag side;

step 124, calculating partial derivatives of the overall linear result according to the predicted values, wherein the partial derivatives are equal to residual errors in Poisson's regression, calculating label-side parameter gradients through the residual errors under the coordination of data source-side representatives, and adjusting label-side model training parameters through the label-side parameter gradients to realize model training iterative calculation;

and step 125, calculating the data source side parameter gradient of each data source side by matching with each data source side through residual errors, wherein the data source side parameter gradient of each data source side is used for adjusting the data source side model training parameters of each data source side locally, and model training iterative calculation is realized.

In a second aspect, an embodiment of the present invention provides a method for training a multi-party poisson regression privacy computation model, which is applied to a data source side, and includes:

step 21, receiving a request for training an initial model of multi-party Poisson regression privacy calculation initiated by a tag party; the multi-party Poisson regression privacy calculation initial model comprises variables for measuring and counting;

step 22, carrying out model iterative training on the initial model of the multiparty Poisson regression privacy calculation until the convergence condition of the initial model of the multiparty Poisson regression privacy calculation is met, and obtaining a trained multiparty Poisson regression privacy calculation model; the multi-party Poisson regression privacy calculation model is used for predicting the counting of events in the same scene as the scene to which the training data belongs;

step 221, other data source sides locally generate corresponding local model results of the data source sides according to the model training parameters of the data source sides and local sample variables held by the other data source sides, and send the local model results to the label side; the data source side local model result is encrypted by adopting a first public key; the method comprises the following steps that firstly, data source side model training parameters are used as initial parameters for training each data source side model; the data source side represents a data source side selected arbitrarily from the data source sides, and the other data source sides are data source sides except the data source side;

step 222, the data source side represents the sum of the local model result of the other data source side and the local model result of the label side sent by the receiving label side;

step 223, the data source side represents and locally generates a corresponding local model result of the data source side according to the data source side representative model training parameters and local sample variables held by the data source side representative model training parameters and the local sample variables; the data source side representative inputs the sum of the local model results of other data source sides and the local model result of the label side and the local model result of the data source side representative into the multi-party Poisson regression privacy calculation initial model to obtain a predicted value; sending the predicted value to a label party;

step 224, after the label side calculates the residual according to the predicted value, matching the label side to calculate the residual to obtain a label side parameter gradient, wherein the label side parameter gradient is used for adjusting the label side model training parameters to realize the model multi-side poisson regression privacy calculation initial model training iterative calculation;

step 225, the data source side represents the residual error sent by the receiving label side, and the data source side represents the data source side parameter gradient calculated by the residual error; the residual error is sent to other data source parties, the other data source parties calculate respective data source party parameter gradients through the residual error, each data source party adjusts data source party model training parameters locally through the respective data source party parameter gradients, and multi-party Poisson regression privacy calculation initial model training iterative calculation is achieved, wherein each data source party comprises: each data source side comprises a data source side representative and other data source sides.

In a third aspect, an embodiment of the present invention further provides a method for training a multi-party poisson regression privacy computation model, including:

step 31, a label direction data source side initiates a request for training a multi-party Poisson regression privacy calculation initial model; the multi-party Poisson regression privacy calculation initial model comprises a variable for measuring and counting;

step 32, carrying out model iterative training on the initial model of the multiparty Poisson regression privacy calculation until the convergence condition of the initial model of the multiparty Poisson regression privacy calculation is met, and obtaining a trained multiparty Poisson regression privacy calculation model; the multi-party Poisson regression privacy calculation model is used for predicting the counting of events in the same scene as the scene to which the training data belongs;

the specific steps of carrying out model iterative training on the initial model of the multi-party Poisson regression privacy computation each time comprise:

step 321, the label side calculates according to the held sample variables and the label side model training parameters to obtain a label side local model result; the label side local model result is encrypted by adopting a first public key to obtain an encrypted label side local model result, and the first public key belongs to a first key pair which is generated by a data source side representative and is used for homomorphic encryption; the label square model training parameters adopted for the first time are initial label square model training parameters;

322, other data source parties locally generate corresponding data source party local model results according to the respective data source party model training parameters and local sample variables held by the other data source parties, and send the data source party local model results to the label party; the data source side local model result is encrypted by adopting a first public key; the method comprises the following steps that firstly, data source side model training parameters are used as initial parameters for training each data source side model; the data source party represents any selected data source party, and the rest data source parties are called other data source parties;

step 323, the label party sends the sum of the label party local model result and the other data source party local model results to the data source party representative, the data source party representative inputs the sum of the label party local model result and the other data source party local model results and the label party local model result and the local model result of the data source party into the multi-party poisson regression privacy calculation initial model to obtain a predicted value, and the predicted value is sent to the label party;

324, calculating partial derivatives of the overall linear result by the label side according to the predicted value, wherein the partial derivatives are equal to residual errors in the Poisson regression, calculating label side parameter gradients through the residual errors under the cooperation of the data source side representatives, adjusting label side model training parameters through the label side parameter gradients, and realizing the training iterative calculation of the multi-side Poisson regression privacy calculation initial model;

step 325, the label party cooperates with each data source party to calculate the self data source party parameter gradient of each data source party through the residual error, each data source party adjusts the corresponding data source party model training parameter locally through the self data source party parameter gradient, and the multi-party poisson regression privacy calculation initial model training iterative calculation is achieved, wherein each data source party comprises a data source party representative and other data source parties.

In a fourth aspect, an embodiment of the present invention provides a multi-party poisson regression privacy computation model training apparatus, which is applied to a labeler, and includes:

a processor; and the number of the first and second groups,

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the method of the first aspect.

In a fifth aspect, an embodiment of the present invention provides a multi-party poisson regression privacy computation model training apparatus, which is applied to a data source side, and includes:

a processor; and (c) a second step of,

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the method of the second aspect.

In a sixth aspect, an embodiment of the present invention provides a computer-readable storage medium applied to a label side, where the computer-readable storage medium stores one or more programs that, when executed by an electronic device including a plurality of application programs, cause the electronic device to perform the method of the first aspect.

In a seventh aspect, an embodiment of the present invention provides a computer-readable storage medium applied to a data source side, where the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by an electronic device including a plurality of application programs, the electronic device is caused to execute the method according to the second aspect.

The technical scheme has the following beneficial effects: the local model result is encrypted by adopting a homomorphic encryption means, so that the multi-party data transmission is ensured not to be leaked under the condition of not adopting a trusted third party, and the privacy is protected; the potential safety hazard caused by the scurrying of internal personnel can be effectively avoided, the compliance of data application is ensured, and the use cost of the product is greatly reduced from the aspects of node deployment and performance requirements. The counting of events is modeled statistically.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the prior art descriptions will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.

FIG. 1 is a flow chart of a multi-party Poisson regression privacy computation model training method (applied to a laber) according to an embodiment of the present invention;

FIG. 2 is a flowchart of a multi-party Poisson regression privacy computation model training method (applied to a data source side) according to an embodiment of the present invention;

FIG. 3 is a flowchart of a multi-party Poisson regression privacy computation model training method (applied to a tag party and a data source party) according to an embodiment of the present invention;

FIG. 4 is a training flow diagram of a multi-party longitudinal federated Poisson regression model according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

As shown in fig. 1, in combination with the embodiment of the present invention, a multi-party poisson regression privacy computation model training method is provided, which is applied to a label party, and includes:

step 11, initiating a request for training a multi-party Poisson regression privacy calculation initial model to a data source side; the multi-party Poisson regression privacy calculation initial model comprises a variable for measuring and counting;

step 121, calculating according to the held sample variables and the label side model training parameters to obtain a label side local model result; the label side local model result is encrypted by adopting a first public key, and the first public key belongs to a first key pair which is generated by a data source side representative and is used for homomorphic encryption; the label square model training parameters adopted for the first time are initial label square model training parameters;

step 123, sending the sum of the local model results of the other data source sides and the local model result of the label side to a data source side representative, and obtaining a predicted value under the coordination of the data source side representative; the sum of the local model results of other data sources and the local model result of the label side is encrypted by the label side; the method for obtaining the predicted value under the coordination of the data source side representatives comprises the following steps: the data source side representative inputs the sum of the local model results of other data source sides and the local model result of the tag side and the local model result of the data source side representative into the multi-party Poisson regression privacy calculation initial model to obtain a predicted value, and the predicted value is sent to the tag side;

and step 125, calculating the data source side parameter gradient of each data source side by matching with each data source side through residual errors, wherein the data source side parameter gradient of each data source side is used for adjusting the model training parameters of the data source side locally by each data source side, so as to realize model training iterative calculation.

Preferably, the method further comprises the following steps:

step 126, before a request of training the multi-side poisson regression privacy calculation initial model is sent to a data source, generating a second key pair for homomorphic encryption, wherein the second key pair comprises a second public key and a second private key; receiving a first public key sent by a data source representative;

step 123, specifically including:

taking the sum of the local model results of other data source sides and the local model result of the label side as a first result, and encrypting the first result by adopting a first public key;

generating a first random number as a first salt, adding the first salt into a first result, and sending the first result of adding the first salt to a data source side representative;

receiving a model result which is sent by a data source party and encrypted by a first public key, and calculating the model result encrypted by the first public key to obtain a predicted value; the model result which is sent by the data source side and encrypted by the first public key is obtained by the following method: after receiving the first result of adding the first salt sent by the label party, the data source party representative decrypts the first result of adding the first salt by using a first private key to obtain a decrypted first result of adding the first salt, and adds the local model result of the data source party representative into the decrypted first result of adding the first salt to obtain a second result of adding the first salt; and the data source party representative inputs the second result added with the first salt into the multi-party Poisson regression privacy calculation initial model to perform power operation to obtain a multi-party Poisson regression privacy calculation model result, encrypts the multi-party Poisson regression privacy calculation model result by adopting a first public key to obtain a model result encrypted by adopting the first public key, and sends the model result encrypted by adopting the first public key to the tag party.

Preferably, step 124 specifically includes:

generating a second random number, taking the second random number as a second salt, and adding the second salt into the label side parameter gradient to form a label side parameter gradient added with the second salt;

sending the parameter gradient of the label side added with the second salt to a data source side representative, and receiving the parameter gradient of the label side added with the second salt after decryption by adopting a first private key returned by the data source side representative; the data source side represents the returned label side parameter gradient added with the second salt after being decrypted by the first private key, and the label side parameter gradient is obtained by the following method: after receiving the parameter gradient of the label side added with the second salt, the data source side decrypts the parameter gradient of the label side added with the second salt by using the first private key to obtain the decrypted parameter gradient of the label side added with the second salt, and returns the decrypted parameter gradient of the label side added with the second salt to the label side;

removing the second salt in the decrypted label side parameter gradient added with the second salt to obtain a plaintext of the label side parameter gradient;

and adjusting the training parameters of the label side model according to the plain text of the gradient of the label side parameters.

Preferably, step 125 specifically includes:

generating a third random number, taking the third random number as third salt, and encrypting the third random number by adopting a second public key to obtain third encrypted salt;

adding the third salt into the residual error, and sending the residual error added with the third salt and the third encrypted salt to a data source side for representation;

receiving the parameter gradients of the data sources, wherein the parameter gradients of the data sources are added with the corresponding salts; wherein, the parameter gradient of each data source is obtained by the following method:

after the residual error added with the third salt and the third encrypted salt are sent to the data source side representative, the data source side representative decrypts the residual error added with the third salt by adopting a first private key to obtain the residual error containing the third encrypted salt;

the data source side represents that the third encrypted salt in the residual error containing the third encrypted salt is removed to obtain the residual error encrypted by the second public key, and the residual error encrypted by the second public key is sent to other data source sides;

generating a random number locally for each data source party to serve as respective salt, calculating the parameter gradient of the data source party according to the residual error, adding the salt of each data source party into the respective parameter gradient to obtain the parameter gradient of the salt added by each data source party, and sending the parameter gradient of the salt added to the label party;

and decrypting the salted parameter gradients of the data source sides by adopting a second private key respectively to obtain the decrypted salted parameter gradients of the data source sides, and sending the decrypted salted parameter gradients to the data source side representatives.

As shown in fig. 2, in combination with the embodiment of the present invention, there is provided a multi-party poisson regression privacy computation model training method, applied to a data source side, including:

step 221, other data source parties locally generate corresponding data source party local model results according to the data source party model training parameters and local sample variables held by the other data source parties, and send the data source party local model results to the label party; the data source side local model result is encrypted by adopting a first public key; the method comprises the following steps that firstly, data source side model training parameters are used as initial parameters for training each data source side model; the data source party represents a data source party selected at will from the data source parties, and the other data source parties are data source parties except the data source party;

step 222, the data source party represents the sum of the local model result of the other data source party and the local model result of the label party sent by the receiving label party;

step 223, the data source side representative locally generates a corresponding local model result of the data source side representative according to the data source side representative model training parameters and local sample variables held by the data source side representative model training parameters and the local sample variables; the data source side representative inputs the sum of the local model results of other data source sides and the local model result of the label side and the local model result of the data source side representative into the multi-party Poisson regression privacy calculation initial model to obtain a predicted value; sending the predicted value to a label party;

step 225, the data source side represents the residual sent by the receiving label side, and the data source side represents the data source side parameter gradient calculated by the residual; the residual error is sent to other data source parties, the other data source parties calculate respective data source party parameter gradients through the residual error, and each data source party locally adjusts data source party model training parameters through the respective data source party parameter gradients to achieve multi-party Poisson regression privacy calculation initial model training iterative calculation, wherein each data source party comprises: each data source side comprises a data source side representative and other data source sides.

Preferably, the method further comprises:

step 226, before receiving a request of training the multi-party poisson regression privacy computation initial model initiated by the labeler, generating a first key pair for homomorphic encryption by the data source representative, wherein the first key pair comprises a first public key and a first private key, and sending the first public key to the labeler and other data source representatives;

step 225, specifically comprising:

the data source side represents the residual error added with the third salt and the third encrypted salt sent by the receiving label side; the residual error added with the third salt and the third encrypted salt sent by the label side are obtained by the following method: the tag party generates a third random number, the third random number is used as a third salt, and a second public key is adopted to encrypt the third random number to obtain a third encrypted salt; adding third salt into the residual error by the label party to obtain the residual error added with the third salt;

the data source side adopts a first private key to decrypt the residual error added with the third salt to obtain the residual error containing the third encrypted salt;

the data source party removes the third encrypted salt in the residual error containing the third encrypted salt to obtain the residual error encrypted by the second public key, and sends the residual error encrypted by the second public key to other data source parties;

for each data source side, generating a random number locally and using the random number as respective salt, calculating the parameter gradient of the data source side according to the residual error, adding the respective salt into the parameter gradient of the data source side to obtain the parameter gradient of the data source side with salt, and sending the parameter gradient of the data source side with salt to the tag side;

the data source party represents a data source party parameter gradient which is sent by the receiving label party and encrypted by the first public key; the data source party parameter gradient which is sent by the label party and encrypted by the first public key is obtained by the following method: for the salted data source party parameter gradient of each data source party, the tag party decrypts each salted data source party parameter gradient by using a second private key, encrypts the salted data source party parameter gradient by using a first public key to obtain the salted data source party parameter gradient encrypted by using the first public key, and sends the salted data source party parameter gradient encrypted by using the first public key to the data source party representative;

for the data source party parameter gradient of each data source party encrypted by the first public key, the data source party uses the first private key to decrypt each data source party parameter gradient to obtain respective salted data source party parameter gradient, and the data source party parameter gradient is sent to the corresponding data source party;

each data source party locally removes salt in the self-salted data source party parameter gradient to obtain the self-salted data source party parameter gradient; and adjusting the model training parameters of the self through the gradient of the parameters of the data source of the self.

Preferably, in step 222, the data source side represents the sum of the local model result of the other data source side and the local model result of the tag side sent by the receiving tag side, and specifically includes:

the data source side represents a first result which is sent by the receiving label side and added with the first salt; wherein, the first result of adding the first salt sent by the label side is obtained by the following method: the label side takes the sum of the label side local model result and other data source side local model results as a first result, and encrypts the first result by adopting a first public key to obtain a first result encrypted by adopting the first public key; generating a first random number as a first salt, adding the first salt into a first result encrypted by a first public key to obtain a first result added with the first salt encrypted by a first private key, and sending the first result to a data source side representative;

step 223 specifically includes:

the data source side represents decrypts the first result added with the first salt by adopting a first private key to obtain a decrypted first result added with the first salt, and adds the local model result represented by the data source side into the decrypted first result added with the first salt to obtain a second result added with the first salt;

the data source party representative inputs a second result added with the first salt into the multi-party Poisson regression privacy calculation initial model for power operation to obtain a multi-party Poisson regression privacy calculation model result, encrypts the multi-party Poisson regression privacy calculation model result by adopting a first public key and sends the multi-party Poisson regression privacy calculation model result to the tag party; and the result of the multi-party Poisson regression privacy calculation model is used for the tag party to calculate the result to obtain a predicted value.

Preferably, step 224 specifically includes:

and after the data source party represents and receives the parameter gradient of the label party added with the second salt, which is sent by the label party, the parameter gradient of the label party added with the second salt is decrypted by adopting the first private key, and the parameter gradient of the label party of the second salt decrypted by adopting the first private key is returned to the label party.

As shown in fig. 3, in combination with the embodiment of the present invention, there is further provided a method for training a multi-party poisson regression privacy computation model, including:

step 31, a label direction data source side initiates a request for training a multi-side Poisson regression privacy calculation initial model; the multi-party Poisson regression privacy calculation initial model comprises variables for measuring and counting;

322, other data source parties locally generate corresponding data source party local model results according to the respective data source party model training parameters and local sample variables held by the other data source parties, and send the data source party local model results to the label party; the data source side local model result is encrypted by adopting a first public key; the method comprises the following steps that firstly, data source side model training parameters are used as initial parameters for training each data source side model; the data source party is arbitrarily selected from the data source parties, and the rest data source parties are called other data source parties;

step 323, the tag party sends the sum of the tag party local model result and the other data source party local model results to the data source party representative, the data source party representative inputs the sum of the tag party local model result and the other data source party local model results and the tag party local model result and the data source party local model result into the multi-party poisson regression privacy computation initial model to obtain a predicted value, and the predicted value is sent to the tag party;

Preferably, the method further comprises:

step 326, before the data source party of the labeler initiates a request for training the multi-party poisson regression privacy computation initial model, the data source party generates a first key pair for homomorphic encryption, including a first public key and a first private key, and sends the first public key to the labeler and other data source parties; and the tag side generates a second key pair for homomorphic encryption, the second key pair comprising: a second public key and a second private key;

step 323, specifically including:

the label side takes the sum of the local model results of other data source sides and the local model result of the label side as a first result, and encrypts the first result by adopting a first public key to obtain a first result encrypted by adopting the first public key;

generating a first random number as a first salt, adding the first salt into a first result encrypted by a first public key to obtain the first result encrypted by a first private key and added with the first salt, and sending the first result to a data source side for representation;

the data source side represents decrypts the first result added with the first salt by adopting a first private key to obtain a decrypted first result added with the first salt, and adds the local model result represented by the data source side to obtain a second result added with the first salt;

the data source party representative inputs the second result added with the first salt into a multi-party Poisson regression privacy calculation initial model to perform power operation to obtain a multi-party Poisson regression privacy calculation model result, encrypts the model result by adopting a first public key to obtain the multi-party Poisson regression privacy calculation model result encrypted by adopting the first public key, and sends the model result encrypted by adopting the first public key to the tag party;

and the label party calculates the model result encrypted by the first public key to obtain a predicted value.

Preferably, in step 324, the labeler calculates a labeler parameter gradient through the residual error under the cooperation of the data source representation, and adjusts a labeler model training parameter through the labeler parameter gradient, which specifically includes:

generating a second random number by the label side, taking the second random number as a second salt, and adding the second salt into the label side parameter gradient to form a label side parameter gradient added with the second salt;

the tagger sends the tagger parameter gradient added with the second salt to the data source representative, the data source representative decrypts the tagger parameter gradient added with the second salt by adopting a first private key to obtain the decrypted tagger parameter gradient added with the second salt, and the decrypted tagger parameter gradient added with the second salt is returned to the tagger;

removing the second salt added into the parameter gradient of the second salt by the label side to obtain a plaintext of the parameter gradient of the label side;

and the label side adjusts the label side model training parameters according to the plain text of the label side parameter gradient.

Preferably, step 325 specifically includes:

the tag party generates a third random number, the third random number is used as third salt, and a second public key is adopted to encrypt the third random number to obtain third encrypted salt;

the label party adds the third salt into the residual error, and sends the residual error added with the third salt and the third encrypted salt to the data source party for representation;

for each data source side, generating a corresponding random number locally and using the random number as respective salt, calculating the parameter gradient of the data source side according to the residual error to obtain the parameter gradient of the salt added to each data source side, and sending the parameter gradient of the salt added to each data source side to the tag side;

the tag party decrypts the salted data source party parameter gradients of the data source parties by adopting a second private key to obtain the decrypted salted data source party parameter gradients of the data source parties, encrypts the decrypted salted data source party parameter gradients of the data source parties by adopting a first public key to obtain respective data source party parameter gradients of the data source parties, and sends the salted data source party parameter gradients encrypted by adopting the first public key to the data source party representative;

for the salted data source party parameter gradient encrypted by the first public key of each data source party, the data source party performs decryption by adopting a first private key to obtain respective salted data source party parameter gradients, and sends the data source party parameter gradients to the corresponding data source parties;

each data source party locally removes salt in the data source party parameter gradient added with salt to obtain the data source party parameter gradient of the data source party; and adjusting the model training parameters of the self through the gradient of the parameters of the self data source side.

An embodiment of the present invention also provides a computer-readable storage medium applied to a label side, where the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by an electronic device including a plurality of application programs, the electronic device is caused to execute the method shown in fig. 1.

Step 13, step 14, step 15 embodiments of the present invention further provide a multi-party poisson regression privacy computation model training apparatus, which is applied to a label party, and includes:

a processor; and the number of the first and second groups,

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the method shown in figure 1.

An embodiment of the present invention also provides a computer-readable storage medium applied to a data source side, where the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by an electronic device including a plurality of application programs, the electronic device executes the method illustrated in fig. 2.

The embodiment of the invention also provides a device for training the multi-party Poisson regression privacy computation model, which is applied to a data source side and comprises the following steps:

a processor; and (c) a second step of,

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the method shown in figure 2.

The above technical solutions of the embodiments of the present invention are described in detail below with reference to specific application examples, and reference may be made to the foregoing related descriptions for technical details that are not described in the implementation process.

The invention discloses a method for training a longitudinal federated multi-party Poisson regression privacy calculation model without a third party, belongs to the field of privacy calculation, and is used for performing longitudinal federated Poisson regression modeling on a plurality of participants through federated learning and homomorphic encryption in privacy calculation. Regarding the multiple parties, one party is a tag party, and the other parties are data source parties.

The method is used for modeling the counting type Y label safely by all parties in the federal modeling under the condition of no trusted third party. Such as: taking the estimation of the number of vehicle insurance claims as an example, the federal modeling label party is an insurance company, and the number of subordinate persons such as gender, age and the like and historical claims recorded when a client applies for insurance are held; the data source side 1 is a vehicle data provider and can provide vehicle data such as vehicle model, seat number, purchase price and the like; the data source side 2 is an in-vehicle sensor data provider and can provide vehicle driving habit data and the like. Through a federal modeling mode, an insurance company can perform poisson regression modeling by using own Y labels and variables and various variables of a vehicle data provider and a vehicle-mounted sensor data provider under the condition that a plurality of data are not exported, so that the number of times of claims in a future period of time of each sample can be estimated, and pricing of vehicle insurance rates of each customer can be assisted. The method can meet the requirements of high data security of the industries such as finance, government affairs and the like, can provide definite cost estimation, and simultaneously has basically no loss in all aspects of model results compared with plaintext calculation.

The technical scheme adopted by the invention is that a Poisson regression model is as follows:

Y＝exp(Z)；

Z＝Xw。

z = Xw + ln (exposure) if there is an exposure variable for measuring the unit time, area or volume of "count" Y. Wherein exp is an exponential function with a natural constant e as the base; ln, i.e. the natural logarithm lna = log a, the base e logarithm is usually used for ln, and e is also an overriding number.

The multi-party poisson regression can be written as follows:

if there is an exposure variable, then

Subscript G denotes the label side (Guest), subscript H _k Indicating the kth data source side (Host). X _G Variables (characteristics) representing label side, X _H Variables (characteristics) representing the data source side; z is a linear or branched member _G Intermediate results, Z, representing the linear part corresponding to the label side variable _H Intermediate result, exp (Z), representing the linear part corresponding to the data source side variable _G ) And representing the local model result corresponding to the data source side variable.

As shown in fig. 4, the poisson regression federal modeling procedure without a third party is as follows:

and step 1, generating and sending a key.

Step 1.1, the label side generates a homomorphic encryption key pair (pk) _G ，sk _G )，pk _G Is the second public key, sk _G Is a second private key;

step 1.2, the data source side randomly selects one as a representative, and the representative is recorded as H ₁ Generating a homomorphic cryptographic key pair (pk) _H ，sk _H )，pk _H Is the first public key, sk _H As a first private key, a first public key pk _H Sending to label side and other data source side H ₂ ，...，H _n 。

The homomorphic encryption algorithm used herein is fully homomorphic encryption or finite series fully homomorphic encryption, such as CKKS. In the following steps

Both represent homomorphic encrypted x.

Step 2, removing H ₁ The external data source side (Host) synchronizes the linear results of the own-party variables.

Step 2.1, data Source side H ₁ Calculating local model result of own-square linearity ₁ (ii) a While other data sources H ₂ ，...，H _n Line for calculating own squareLocal model results of sex Z _k With the first public key pk, respectively _H Encrypting to obtain ciphertext of respective linear results

And each will encrypt the text

And sending the data to the label side.

And 3, calculating the gradient and updating the parameters by the label party (Guest).

Step 3.1, label side calculation

Step 3.1.1 tag side calculates own Z _G And then calculates the sum

Step 3.1.2, the label side generates a random number R _G As the first salt, to be added

Sent to the data Source Party H ₁ ；

Step 3.1.3, data Source side represents H ₁ Using the first private key sk _H Decryption

To obtain Z _t +R _G And then with its own model results

Adding to obtain Z + R _G ；

Step 3.1.4, data Source side represents H ₁ To Z + R _G Performing exponentiation operation, and encrypting with the first public key

Sending to a label party;

step (ii) of3.1.5, calculating the predicted value by the label side

Step 3.2, calculating the ciphertext of partial derivative dZ of the overall linear result by the label side

Y is a real label value of a label side, and the partial derivative dZ of the integral linear result in Poisson regression is equal to a residual error;

and, the tag side generates a random number R _G Will become a second random number R _G As a second salt, compute the ciphertext of the entire model gradient

Adding a second salt to obtain a salted capron parameter gradient

To the data source representative.

The data source side representative uses the first private key sk _H Decryption

To give dW' _G To tag side, dW' _G Is the label square gradient after salting.

Step 3.3, calculating gradient dW of own variable by the label side _G ＝dW′ _G -R _G ；

Step 3.4, updating parameters on the label side, e.g. by gradient descent W _G ←W _G -learning_rate·dW _G 。

And 4, calculating the gradient and updating the parameters by the data source side (Host).

Step 4.1, the label side generates a third random number R _G (as a third salt) with a second public key pk _G For the third random number R _G Encrypting to obtain third encrypted salt

Step 4.2, add third salt to the residual (i.e. the result of 3.2) by the label side:

will be provided with

And a third encrypted salt

Sent to the data Source Party H ₁ ；

Step 4.3, data Source side represents H ₁ With the first private key sk _H Decryption

Obtaining dZ ' of the plaintext containing the ' third encryption salt ';

step 4.4, data Source side represents H ₁ Removal of "first encryption salt":

i.e. encrypted with the key of the label party (second public key)

Sent to other data source H ₂ ，...，H _n (ii) a Step 4.5, each data source side H _k Respectively generating random numbers

Random number is added

As respective salts, ciphertext of respective gradients are calculated

Gradient of each added salt

Sending to a label party;

step 4.6, the second private key sk is used by the label party _G Decrypting each data source side separately

To obtain

Sent to each data source side H ₁ ，H ₂ ，...，H _n ；

Step 4.7, data sources H _k Calculating the gradient of the own-square parameter by removing the salt

Step 4.8, each data source side updates the parameters, for example, by adopting a gradient descent method

And 5, repeating the steps 2 to 4 until convergence or the configured Poisson iteration number is reached.

The embodiment of the invention has the following beneficial effects:

according to the invention, a Poisson regression iteration of an algorithm flow with symmetric CKKS homomorphic encryption means, label sides and data source sides is adopted; meanwhile, by adopting an exchangeable homomorphic encryption means and a salt adding mode, the data transmission of each party is ensured not to be leaked, and the privacy is protected, so that a 'trusted third party' is successfully removed, the potential safety hazard caused by the fleeing of internal personnel is effectively avoided, the compliance of data application is ensured, and the use cost of the product is greatly reduced in terms of node deployment and performance requirements. Meanwhile, the problems that in the prior art, when a secret shared MPC protocol is adopted, the consumption of used resources is high, the communication complexity is high, and the cost for landing implementation is correspondingly high are solved, so that the problem that the MPC protocol must be relied on in other longitudinal federated Poisson regression methods without a trusted third party is solved.

In the multi-party Poisson technology, all data are encrypted homomorphically when being exported, part of steps are accompanied by a salt adding method, and no clear text local model result is transmitted. The nature of salt addition (random mask addition) is also an encryption mode, the random number of salt addition can be discarded after only one time, the other party cannot guess the random number, the plaintext cannot be calculated, and the safety of the plaintext can be ensured. The key steps were analyzed as follows: for a data source side, the Z value calculated by the method is not exposed to a label side, and the label side can update the gradient under the condition of protecting data by means of homomorphic encryption; for the tag side, calculating the ciphertext

The method contains important information of the Y label, protects the label information through homomorphic encryption and salting, and helps a data source side to update the gradient.

The deployment of a 'trusted third party' coordination node is reduced in the project deployment, and the deployment cost is reduced; because all parties participating in the calculation are parties who actually own the data, the cost required by calculation implementation is more conveniently estimated, the two parties can set a relatively reasonable local calculation power and network bandwidth according to the scale, the variable number and the like of local data of the two parties, the situation of power waste is avoided, high delay is avoided, and the feasibility of business cooperation is greatly improved.

In addition, no matter the homomorphic encryption or the salt adding mode, the safety can be expanded, and a user can configure the length of a homomorphic encryption key and the size range of a random number according to the application scene of private calculation. For scenes with higher safety requirements, the length of a larger homomorphic encryption key and the size range of a random number can be selected to meet the requirement of data compliance; for the scenes with lower safety requirements, the length of a smaller homomorphic encryption key can be selected or only a simple salt adding mode is used, the transmission quantity is reduced while the operation speed is improved, the balance of safety and physical resources such as calculation speed, time, bandwidth and the like is realized, and the requirements of different scenes are better met. With respect to accuracy, the same parameters are used on the same data set, and the calculation result of the technology is basically lossless compared with the plaintext.

It should be understood that the specific order or hierarchy of steps in the processes disclosed is an example of exemplary approaches. Based upon design preferences, it is understood that the specific order or hierarchy of steps in the processes may be rearranged without departing from the scope of the present disclosure. The accompanying method claims present elements of the various steps in a sample order, and are not intended to be limited to the specific order or hierarchy presented.

In the foregoing detailed description, various features are grouped together in a single embodiment for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the claimed embodiments of the subject matter require more features than are expressly recited in each claim. Rather, as the following claims reflect, invention lies in less than all features of a single disclosed embodiment. Thus, the following claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. To those skilled in the art; various modifications to these embodiments will be readily apparent, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the disclosure. Thus, the present disclosure is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

What has been described above includes examples of one or more embodiments. It is, of course, not possible to describe every conceivable combination of components or methodologies for purposes of describing the aforementioned embodiments, but one of ordinary skill in the art may recognize that many further combinations and permutations of various embodiments are possible. Accordingly, the embodiments described herein are intended to embrace all such alterations, modifications and variations that fall within the scope of the appended claims. Furthermore, to the extent that the term "includes" is used in either the detailed description or the claims, such term is intended to be inclusive in a manner similar to the term "comprising" as "comprising" is interpreted when employed as a transitional word in a claim. Furthermore, any use of the term "or" in the specification of the claims is intended to mean a "non-exclusive or".

Those of skill in the art will also appreciate that the various illustrative logical blocks, elements, and steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate the interchangeability of hardware and software, various illustrative components, elements, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design requirements of the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present embodiments.

The various illustrative logical blocks or elements described in connection with the embodiments disclosed herein may be implemented or performed with a general purpose processor, a digital signal processor, an Application Specific Integrated Circuit (ASIC), a field programmable gate array or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a digital signal processor core, or any other similar configuration.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may be stored in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. For example, a storage medium may be coupled to the processor such the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC, which may be disposed in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.

In one or more exemplary designs, the functions described in the embodiments of the present invention may be implemented in hardware, software, firmware, or any combination of the three. If implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media that facilitate transfer of a computer program from one place to another. Storage media may be any available media that can be accessed by a general purpose or special purpose computer. For example, such computer-readable media can comprise, but is not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store program code in the form of instructions or data structures and that can be read by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. Additionally, any connection is properly termed a computer-readable medium, and, thus, is included if the software is transmitted from a website, server, or other remote source via a coaxial cable, fiber optic cable, twisted pair, digital Subscriber Line (DSL), or wirelessly, e.g., infrared, radio, and microwave. Such discs (disk) and disks (disc) include compact disks, laser disks, optical disks, DVDs, floppy disks and blu-ray disks where disks usually reproduce data magnetically, while disks usually reproduce data optically with lasers. Combinations of the above may also be included in the computer-readable medium.

The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are merely exemplary embodiments of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalent substitutions, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims

1. A multi-party Poisson regression privacy computation model training method is applied to a label party and is characterized by comprising the following steps:

when model iterative training is carried out on the initial model of the multi-side Poisson regression privacy calculation each time, the method specifically comprises the following steps:

step 121, calculating according to the held sample variables and the label side model training parameters to obtain a label side local model result; the local model result of the label side is encrypted by adopting a first public key, and the first public key belongs to a first key pair which is generated by a data source side representative and is used for homomorphic encryption; the label square model training parameters adopted for the first time are initial label square model training parameters;

step 124, calculating partial derivatives of the overall linear result according to the predicted value, wherein the partial derivatives are equal to a residual error in Poisson's regression, calculating a label side parameter gradient through the residual error under the coordination of the data source side representation, and adjusting label side model training parameters through the label side parameter gradient to realize model training iterative calculation;

2. The multi-party poisson regression privacy computation model training method of claim 1, further comprising:

step 126, before a request of training the multiparty Poisson regression privacy computation initial model is initiated to the data source, generating a second key pair for homomorphic encryption, wherein the second key pair comprises a second public key and a second private key; receiving a first public key sent by a data source side representative;

step 123, specifically including:

receiving a model result which is sent by a data source party and encrypted by a first public key, and calculating the model result encrypted by the first public key to obtain a predicted value; the model result which is sent by the data source party and encrypted by the first public key is obtained by the following method: after receiving the first result of adding the first salt sent by the label party, the data source party representative decrypts the first result of adding the first salt by using a first private key to obtain a decrypted first result of adding the first salt, and adds the local model result of the data source party representative into the decrypted first result of adding the first salt to obtain a second result of adding the first salt; and the data source party representative inputs the second result added with the first salt into the multi-party Poisson regression privacy calculation initial model for power operation to obtain a multi-party Poisson regression privacy calculation model result, encrypts the multi-party Poisson regression privacy calculation model result by adopting the first public key to obtain a model result encrypted by adopting the first public key, and sends the model result encrypted by adopting the first public key to the tag party.

3. The multi-party poisson regression privacy computation model training method according to claim 2, wherein the step 124 specifically comprises:

sending the parameter gradient of the tag party added with the second salt to a data source party representative, and receiving the parameter gradient of the tag party added with the second salt after decryption by adopting a first private key returned by the data source party representative; the data source side represents the returned label side parameter gradient which is decrypted by the first private key and added with the second salt, and the label side parameter gradient is obtained by the following method: after receiving the parameter gradient of the label side added with the second salt, the data source side decrypts the parameter gradient of the label side added with the second salt by using the first private key to obtain the decrypted parameter gradient of the label side added with the second salt, and returns the decrypted parameter gradient of the label side added with the second salt to the label side;

4. The multi-party poisson regression privacy computation model training method according to claim 2, wherein the step 125 specifically comprises:

generating a third random number, taking the third random number as a third salt, and encrypting the third random number by adopting a second public key to obtain a third encrypted salt;

receiving the self parameter gradient sent by each data source party, wherein the self parameter gradient of each data source party is added with the corresponding salt; wherein, the parameter gradient sent by each data source is obtained by the following method:

and decrypting the salted parameter gradients of the data source sides by adopting a second private key respectively to obtain the decrypted salted parameter gradients of the data source sides, and sending the decrypted salted parameter gradients to the data source side representative.

5. A multi-party Poisson regression privacy computation model training method is applied to a data source side and is characterized by comprising the following steps:

step 225, the data source side represents the residual error sent by the receiving label side, and the data source side represents the data source side parameter gradient calculated by the residual error; the residual error is sent to other data source parties, the other data source parties calculate respective data source party parameter gradients through the residual error, and each data source party locally adjusts data source party model training parameters through the respective data source party parameter gradients to achieve multi-party Poisson regression privacy calculation initial model training iterative calculation, wherein each data source party comprises: each data source side comprises a data source side representative and other data source sides.

6. The multi-party Poisson regression privacy computation model training method of claim 5, further comprising:

step 225, specifically comprising:

for each data source side, generating a random number locally and using the random number as respective salt, calculating the parameter gradient of the data source side according to the residual error, adding the respective salt into the parameter gradient of the data source side to obtain the parameter gradient of the data source side with salt, and sending the parameter gradient of the data source side with salt to the label side;

the data source party represents a data source party parameter gradient which is sent by the receiving label party and encrypted by the first public key; the data source party parameter gradient encrypted by the first public key and sent by the label party is obtained by the following method: for the parameter gradient of the salted data source party of each data source party, the tag party decrypts the parameter gradient of each salted data source party by using a second private key, encrypts the parameter gradient of the salted data source party by using a first public key to obtain the parameter gradient of the salted data source party encrypted by using the first public key, and sends the parameter gradient of the salted data source party encrypted by using the first public key to the data source party representative;

for the data source party parameter gradient of each data source party encrypted by the first public key, the data source party decrypts the data source party parameter gradient by the first private key to obtain respective salted data source party parameter gradient, and sends the data source party parameter gradient to the corresponding data source party;

each data source party locally removes salt in the data source party parameter gradient added with the salt to obtain the data source party parameter gradient of the data source party; and adjusting the model training parameters of the self through the gradient of the parameters of the data source of the self.

7. The multi-party Poisson regression privacy computation model training method of claim 6, wherein,

step 222, the data source party represents the sum of the local model result of the other data source party and the local model result of the label party sent by the receiving label party, and specifically includes:

the data source side represents a first result which is sent by the receiving label side and added with the first salt; wherein, the first result of adding the first salt sent by the label side is obtained by the following method: the method comprises the following steps that a tag party takes the sum of a tag party local model result and other data source party local model results as a first result, and the first result is encrypted by a first public key to obtain a first result encrypted by the first public key; generating a first random number as a first salt, adding the first salt into a first result encrypted by a first public key to obtain a first result added with the first salt encrypted by a first private key, and sending the first result to a data source side representative;

step 223 specifically includes:

8. The method for training the multi-party poisson regression privacy computation model of claim 6, wherein step 224 specifically comprises:

9. A multi-party Poisson regression privacy computation model training method is characterized by comprising the following steps:

step 31, a label direction data source side initiates a request for training a multi-party Poisson regression privacy calculation initial model; the multi-party Poisson regression privacy calculation initial model comprises variables for measuring and counting;

step 322, other data source parties locally generate corresponding local model results of the data source parties according to the respective model training parameters of the data source parties and local sample variables held by the other data source parties, and send the local model results to the label party; the data source side local model result is encrypted by adopting a first public key; the method comprises the following steps that firstly, data source side model training parameters are used as initial parameters for training each data source side model; the data source party is arbitrarily selected from the data source parties, and the rest data source parties are called other data source parties;

324, calculating partial derivatives of the overall linear result by the label side according to the predicted value, wherein the partial derivatives are equal to a residual error in Poisson's regression, calculating a label side parameter gradient through the residual error under the coordination of the data source side, adjusting a label side model training parameter through the label side parameter gradient, and realizing the training iterative calculation of the multi-side Poisson's regression privacy calculation initial model;

10. The multi-party poisson regression privacy computation model training method of claim 9, further comprising:

step 323, specifically including:

the data source side representative adopts a first private key to decrypt the first result added with the first salt to obtain a decrypted first result added with the first salt, and adds the local model result represented by the data source side to obtain a second result added with the first salt;

the data source party representative inputs a second result added with the first salt into the multi-party Poisson regression privacy calculation initial model to perform power operation to obtain a multi-party Poisson regression privacy calculation model result, encrypts the model result by adopting a first public key to obtain the multi-party Poisson regression privacy calculation model result encrypted by adopting the first public key, and sends the model result encrypted by adopting the first public key to the tag party;

11. The method for training the multi-party poisson regression privacy computation model according to claim 10, wherein in step 324, the labeler computes the labeler parameter gradient through the residual error under the cooperation of the data source representative, and adjusts the labeler model training parameter through the labeler parameter gradient, specifically comprising:

12. The method for training the multi-party poisson regression privacy computation model of claim 10, wherein step 325 specifically comprises:

the tag party generates a third random number, the third random number is used as a third salt, and a second public key is adopted to encrypt the third random number to obtain a third encrypted salt;

generating corresponding random numbers locally for each data source party, taking the random numbers as respective salts, calculating the parameter gradient of the data source party according to the residual error to obtain the parameter gradient of the salting of each data source party, and sending the parameter gradient of the salting of each data source party to the tag party;

for the salted data source party parameter gradient encrypted by the first public key of each data source party, the data source party representation adopts the first private key to decrypt to obtain the respective salted data source party parameter gradient, and the data source party parameter gradients are sent to the corresponding data source parties;

13. A multi-party Poisson regression privacy computation model training device is applied to a label party and is characterized by comprising the following components:

a processor; and the number of the first and second groups,

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the method of any one of claims 1-4.

14. A multi-party Poisson regression privacy computation model training device is applied to a data source side and is characterized by comprising the following components:

a processor; and the number of the first and second groups,

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the method of any one of claims 5-8.

15. A computer readable storage medium applied to a label side, characterized in that the computer readable storage medium stores one or more programs that, when executed by an electronic device comprising a plurality of application programs, cause the electronic device to perform the method of any one of claims 1-4.

16. A computer-readable storage medium applied to a data source side, characterized in that the computer-readable storage medium stores one or more programs that, when executed by an electronic device including a plurality of application programs, cause the electronic device to perform the method of any one of claims 5-8.