A multi-party secure computing method and device for federated learning

Abstract

The embodiments of the present invention relate to the field of machine learning, and in particular to a multi-party secure computing method and device for federated learning, which is used to improve the efficiency and accuracy of model training on the basis of protecting the data security in the multi-party computing process. The embodiments of the present invention include: a data node divides all individual objects in the data node into multiple object sets according to a classification standard; for each object set, the data node determines the feature data of the object set according to the feature data of all individual objects in the object set; the data node sends the feature data of the object set to a model node, so that the model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes to obtain sample data of the object set, and trains the federated learning model according to the sample data of all object sets; the classification standards for dividing individual objects in multiple data nodes are the same.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a multi-party secure computing method and device for federated learning. Background Art

[0002] Federated learning is a machine learning framework that helps different organizations to jointly use and model data while meeting the requirements of user privacy protection, data security, and government regulations. Specifically, federated learning aims to solve the following problem: Under the premise that each enterprise's data does not leave the local area, a virtual shared model is established through parameter exchange and optimization under an encrypted mechanism. The performance of this shared model is similar to a model trained by aggregating data from all parties. This data joint modeling solution does not leak user privacy and complies with the principle of data security protection.

[0003] When currently federated learning is used to train models, it mainly adopts a data-end-model architecture. The model end is responsible for initializing the model parameters and then sending them to the data end. The data end updates the received model parameters locally and then returns them to the model end. The model end then aggregates the parameters, usually by averaging. The updated model is sent back to the data end, and the cycle repeats until convergence.

[0004] In the prior art, generally one model end corresponds to multiple data ends, each data end corresponds to a platform, and the platforms collect the personal data of their respective customers. However, due to the low user overlap between different platforms, there are problems such as poor data efficiency and low computing efficiency, and it is difficult to fully explore the value of multi-party data. Summary of the invention

[0005] The present application provides a multi-party secure computing method and device for federated learning, which are used to reduce the error caused by lossy compression of model parameters and improve the accuracy of the federated learning model.

[0006] An embodiment of the present invention provides a multi-party secure computing method for federated learning, including:

[0007] The data node divides all individual objects in the data node into multiple object sets according to the classification standard;

[0008] For each object set, the data node determines the feature data of the object set according to the feature data of all individual objects in the object set;

[0009] The data node sends the feature data of the object set to the model node, so that the model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes to obtain sample data of the object set, and trains the federated learning model according to the sample data of all object sets;

[0010] The classification standards for dividing individual objects in the multiple data nodes are the same.

[0011] Optionally, the feature data of the individual object includes continuous feature data, and the data node determines the feature data of the object set according to the feature data of all individual objects in the object set, including:

[0012] The data node calculates the mean of the continuous feature data of all individual objects as the first feature data of the object set.

[0013] Optionally, the feature data of the individual object includes discrete feature data, and the data node determines the feature data of the object set according to the feature data of all individual objects in the object set, including:

[0014] The data node converts the discrete feature data of the individual object into a feature vector by using one-hot encoding;

[0015] The feature data of the object set are determined by using the feature vectors of all individual objects in the object set.

[0016] Optionally, after the data node sends the feature data of the object set to the model node, the process further includes:

[0017] The data node receives the model calculation result sent by the model node, wherein the model calculation result includes a correspondence between an object set and a recommendation set; the correspondence between the object set and the recommendation set is determined by the model node using the federated learning model;

[0018] The data node determines a target object set corresponding to the target individual object according to the attribute data of the target individual object;

[0019] The data node determines the recommended content of the target individual object according to the recommendation set corresponding to the target object set.

[0020] The embodiment of the present invention also provides a multi-party secure computing method for federated learning, including:

[0021] The model node receives feature data of an object set sent by a plurality of data nodes, and the plurality of data nodes classify individual objects into different object sets according to the same classification standard;

[0022] The model node performs sample alignment on all feature data of the same object set to obtain sample data of the object set;

[0023] The model node trains the federated learning model based on sample data of all object sets.

[0024] The embodiment of the present invention further provides a multi-party secure computing device for federated learning, comprising:

[0025] A partitioning unit is used to partition all individual objects in a data node into multiple object sets according to a classification standard;

[0026] a calculation unit, configured to determine, for each object set, feature data of the object set according to feature data of all individual objects in the object set;

[0027] A transceiver unit, used to send feature data of an object set to a model node, so that the model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes to obtain sample data of the object set, and trains the federated learning model according to the sample data of all object sets;

[0028] Among them, the classification standard for dividing individual objects in multiple data nodes is the same.

[0029] Optionally, the computing unit is specifically used to:

[0030] The mean of the continuous feature data of all individual objects is calculated as the first feature data of the object set.

[0031] Optionally, the computing unit is specifically used to:

[0032] Using one-hot encoding, the discrete feature data of the individual object is converted into a feature vector;

[0033] The feature data of the object set are determined by using the feature vectors of all individual objects in the object set.

[0034] Optionally, the transceiver unit is further used to receive a model calculation result sent by the model node, wherein the model calculation result includes a correspondence between an object set and a recommendation set; the correspondence between the object set and the recommendation set is determined by the model node using the federated learning model;

[0035] It also includes a recommendation unit, which is used to determine a target object set corresponding to the target individual object according to the attribute data of the target individual object; and determine the recommended content of the target individual object according to the recommendation set corresponding to the target object set.

[0036] An embodiment of the present invention further provides an electronic device, including:

[0037] at least one processor; and,

[0038] a memory communicatively connected to the at least one processor; wherein,

[0039] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0040] An embodiment of the present invention further provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the method as described above.

[0041] The federated learning system of the embodiment of the present invention includes a model node and multiple data nodes. Each data node collects feature data of individual objects, and sends it to the model node after processing. The model node stores a federated learning model, and performs model training based on the received feature data. Specifically, each data node divides all individual objects in the data node into multiple object sets according to classification criteria, wherein the classification criteria for dividing individual objects in multiple data nodes are the same. For each object set, the data node determines the feature data of the object set based on the feature data of all individual objects in the object set. Each data node sends the feature data of the object set to the model node. The model node can perform sample alignment on all feature data of the same object set based on the feature data of the object sets sent by multiple data nodes, obtain sample data of the object set, and train the federated learning model based on the sample data of all object sets. In the embodiment of the present application, individual objects are divided into multiple object sets, and the feature data of each object set is determined based on the feature data of all individual objects in the object set. The model node performs sample alignment on the feature data of the object sets of multiple data nodes to obtain the sample data of the object set, and then trains the federated learning model based on the sample data of the object set, thereby solving the data sparsity problem caused by the low user overlap between different data nodes and increasing the accuracy of model training. At the same time, since each data node calculates the feature data of the object set based on the feature data of the individual object before sending the feature data to the model node, in this way, during the process of the data node sending the feature data of the object set and the model node processing the feature data of the object set, it is impossible to deduce the feature data of each individual object based on the feature data of the object set, thereby better protecting the privacy and security of data in the multi-party computing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1 A schematic diagram of a possible system architecture provided by an embodiment of the present invention;

[0044] Figure 2 A flowchart of a multi-party secure computing method for federated learning provided by an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of sample alignment of feature data by a model node provided in a specific embodiment of the present invention;

[0046] Figure 4 A flowchart of a multi-party secure computing method for federated learning provided in Example 1 of the present invention;

[0047] Figure 5 A flowchart of a multi-party secure computing method for federated learning provided in Embodiment 2 of the present invention;

[0048] Figure 6 A schematic diagram of the structure of a multi-party secure computing device for federated learning provided by an embodiment of the present invention;

[0049] Figure 7 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] like Figure 1As shown, a system architecture applicable to an embodiment of the present invention includes a model node server 101 and multiple data node servers 102. The data node server 102 obtains feature data of individual objects and divides different individual objects into object sets according to classification criteria. The data node server 102 determines the feature data of the object set based on the feature data of all individual objects in the object set, which can usually be an average operation, etc. Each data node server 102 sends the feature data of the object set to the model node server 101, so that the model node server 101 can obtain sample data of the object set based on the feature data of the object sets of multiple data node servers 102, and train the federated learning model based on the sample data of all object sets.

[0052] The model node server 101 can be a network device such as a computer. The model node server 101 can be an independent device or a server cluster formed by multiple servers. Preferably, the model node server 101 can use cloud computing technology for information processing. The data node server 102 can be an independent device or a server cluster formed by multiple servers. Preferably, the model node server 101 can use cloud computing technology for information processing.

[0053] The data node server 102 can communicate with the model node server 101 through the INTERNET network, and can also communicate with the model node server 101 through mobile communication systems such as the Global System for Mobile Communications (GSM) and the Long Term Evolution (LTE) system.

[0054] Based on the above architecture, the embodiment of the present invention provides a multi-party secure computing method for federated learning, such as Figure 2 As shown, the multi-party secure computing method for federated learning provided by the embodiment of the present invention includes the following steps:

[0055] Step 201: A data node divides all individual objects in the data node into multiple object sets according to a classification standard.

[0056] Step 202: For each object set, the data node determines feature data of the object set according to feature data of all individual objects in the object set.

[0057] Step 203: The data node sends the feature data of the object set to the model node.

[0058] Step 204: The model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes to obtain sample data of the object set, and trains the federated learning model based on the sample data of all object sets.

[0059] Among them, the classification standard for dividing individual objects in multiple data nodes is the same.

[0060] The federated learning system of the embodiment of the present invention includes a model node and multiple data nodes. Each data node collects feature data of individual objects, and sends it to the model node after processing. The model node stores a federated learning model, and performs model training based on the received feature data. Specifically, each data node divides all individual objects in the data node into multiple object sets according to classification criteria, wherein the classification criteria for dividing individual objects in multiple data nodes are the same. For each object set, the data node determines the feature data of the object set based on the feature data of all individual objects in the object set. Each data node sends the feature data of the object set to the model node. The model node can perform sample alignment on all feature data of the same object set based on the feature data of the object sets sent by multiple data nodes, obtain sample data of the object set, and train the federated learning model based on the sample data of all object sets. In the embodiment of the present application, individual objects are divided into multiple object sets, and the feature data of each object set is determined based on the feature data of all individual objects in the object set. The model node performs sample alignment on the feature data of the object sets of multiple data nodes to obtain the sample data of the object set, and then trains the federated learning model based on the sample data of the object set, thereby solving the data sparsity problem caused by the low user overlap between different data nodes and increasing the accuracy of model training. At the same time, since each data node calculates the feature data of the object set based on the feature data of the individual object before sending the feature data to the model node, in this way, during the process of the data node sending the feature data of the object set and the model node processing the feature data of the object set, it is impossible to deduce the feature data of each individual object based on the feature data of the object set, thereby better protecting the privacy and security of data in the multi-party computing process.

[0061] In the embodiment of the present invention, for a data node, one data node corresponds to one merchant platform, and an individual object may be a user of the merchant platform, and the data node obtains characteristic data of the user of the corresponding merchant platform. It should be noted that the individual object in the embodiment of the present invention may be a user, or may be other objects, such as a terminal device, a software application, etc. The embodiment of the present invention only takes the user as an example for introduction, and no limitation is made here.

[0062] The data node obtains the user's portrait feature data in the corresponding merchant platform, such as age, gender, occupation, etc., as well as interactive behavior feature data, such as clicks, recommendations, ratings, etc. Different data nodes negotiate to determine the classification standards, and divide users into different user sets according to the portrait feature data according to the unified classification standards. In order to facilitate the distinction between different user sets, a set identifier, i.e., a set ID, can also be assigned to each user set. For example, the division standard of user set ID=1 is set to "age 20-30, female, financial industry, Xuhui District, Shanghai", then platform A will classify the users on the platform that meet the above division standards into the user set of platform A's ID=1, and platform B will classify the users on the platform that meet the above division standards into the user set of platform B's ID=1.

[0063] Through a unified classification standard, users of different data nodes can be divided into different user sets. For the same user set, different data nodes can collect the same interactive behavior characteristics or different interactive behavior characteristics. For example, platform A obtains the number of clicks and pushes of users, and platform B obtains the evaluation scores of users.

[0064] Since data is generally divided into continuous feature data and discrete feature data, the embodiments of the present invention adopt different processing methods for different types of data.

[0065] If the characteristic data of the individual object is continuous characteristic data, the data node determines the characteristic data of the object set according to the characteristic data of all individual objects in the object set, including:

[0066] The data node calculates the mean of the continuous feature data of all individual objects as the first feature data of the object set.

[0067] In the specific implementation process, for the single-value features of users, such as click-through rate, browsing time and other feature data, if the users are divided into the same user set, the mean processing is used to characterize the feature data of the same user set, that is:

[0068]

[0069] Among them, F group is the characteristic data of the user set, F 个体 It is the characteristic data of each user in the user set.

[0070] If the feature data of the individual object is discrete feature data, the data node determines the feature data of the object set according to the feature data of all individual objects in the object set, including:

[0071] The data node converts the discrete feature data of the individual object into a feature vector by using one-hot encoding;

[0072] The feature data of the object set are determined by using the feature vectors of all individual objects in the object set.

[0073] In the specific implementation process, for the discrete features of users, such as which hospitals they have visited in the medical scenario, their individual features are first expressed using onehot vectors. For example, if a user has been to Huashan Hospital, Zhongshan Hospital, and Ruijin Hospital, their onehot vector expression is F A =(1 华山 , 0 中山 , 0 瑞金 ).

[0074] For a user set, it is the average representation of the onehot vector expression of the individual features corresponding to each user in the user set:

[0075]

[0076] In the embodiment of the present invention, each data node first generates a user list through the above operations. For example, for the user set with ID=1, platform A generates a user list:

[0077] A platform {user U A1 +Feature ListF A1 , user U A2 +Feature ListF A2 , user U A3 +Feature List F A3 …};

[0078] Platform B generates a user list:

[0079] Platform B {User U B1 +Feature ListF B1 , user U B2 +Feature List F B2 , user U B3 +Feature ListF B3 …}.

[0080] Then, by processing the platform feature values ​​and the feature data of each user in the user set, the average feature data of platform A and platform B corresponding to the user set with ID=1 is formed. A =(Feature List F A1 +Feature ListF A2 +Feature List F A3 +…+Feature List F AN ) / N

[0081] ...Formula 3

[0082] F B =(Feature List F B1 +Feature List F B2 +Feature List F B3 +…+Feature List F BM ) / M

[0083] ...Formula 4

[0084] Among them, F A is the characteristic data of the user set with ID=1 in platform A, N is the number of users in the user set with ID=1 in platform A; F B is the characteristic data of the user set with ID=1 in platform B, and M is the number of users in the user set with ID=1 in platform B.

[0085] Through the above operations, the user feature data in all user sets are subjected to the above operations, and the average feature matrix of each user set is calculated by different data nodes. For example, the feature matrix X is obtained for platform A and platform B respectively. A ={ID A , F A}, and the feature matrix X B ={ID B , F B}.

[0086] Further, in the embodiment of the present invention, after each data node sends the feature data to the model node, the model node receives the feature data of the object set sent by the multiple data nodes;

[0087] The model node performs sample alignment on all feature data of the same object set to obtain sample data of the object set;

[0088] The model node trains the federated learning model based on sample data of all object sets.

[0089] In the specific implementation process, the object collection is based on the same assigned ID, which can facilitate the model node to align samples of all received feature data, such as Figure 3 For example, for platform A (Yunsuanfu) and platform B (Meituan), the model node merges the feature data of the object set with the same ID as the feature data of the ID.

[0090] For example, after data processing, the data nodes of platform A obtain the click data feature matrix and marketing push strategy matrix. For each type of user set, the click data feature matrix of the merchant can be obtained. The feature matrix structure is shown in Table 1. The value of each row of the matrix is ​​the number of clicks, frequency and other related features of all fixed merchants on a certain day.

[0091] Table 1

[0092]

[0093]

[0094] The push strategy matrix is ​​shown in Table 2. This matrix represents the number of times each merchant is pushed to each type of user set, the average push frequency, etc. under different marketing plans. In joint modeling, this matrix will be used as feature input, and finally its push optimization strategy will be obtained in the reinforcement learning framework.

[0095] Table 2

[0096]

[0097] After data processing, the data nodes of platform B obtain the review data feature matrix, as shown in Table 3. This matrix represents the average review scores of various groups of people on each merchant on platform B. This score can quantify the user's preference for the merchant. In joint modeling, this score and click behavior are used as user feedback and as reward values ​​in reinforcement learning training.

[0098] Table 3

[0099]

[0100] The federated learning model in the model node can be various models, such as linear models such as support vector machine (SVM), logistic regression (LR), etc., and can also support deep neural network (DNN) models, etc., without any restriction here.

[0101] The sample data of the object set after sample alignment is used as the training sample of the federated learning model to establish the model framework of reinforcement learning.

[0102] Based on the average click volume matrix component U of platform A and the average rating matrix component P of platform B, the reward value function R=U·P of reinforcement learning is set. This reward indicator comprehensively considers the user's click-to-visit tendency preference and the merchant's evaluation index. The reward value is used as the learning goal of reinforcement learning to select the appropriate marketing strategy.

[0103] The model node uses the data received from each data node to build and train the federated learning model. For example, the merchant review data of platform A and the user click data of platform B are integrated to generate the reward value index of reinforcement learning. The marketing strategy matrix under different samples provided by platform A is used as the action feature of learning to build a model and learn the optimal marketing strategy, that is, the optimal push weight matrix for merchants in each group of people.

[0104] In the model node, the data of each party is distributed through the third-party coordination of federated learning to strengthen the learning model and distribute encrypted calculations to ensure that the data is not leaked. Based on the third-party coordination calculation gradient of the basic federated learning framework machine learning, considering that the reward value function is a function value that combines the clicks on platform A and the comments on platform B, one party can reverse the other party's data based on this value and its own value. This parameter must be distributed to each data source as a gradient, and cannot be directly output to both parties.

[0105] The specific implementation process is as follows:

[0106] Platform A and Platform B extract their respective features according to user types. Platform A generates a feature matrix of user review data for all merchants under different marketing strategies for each user type as the sample input for reward value (Reward) calculation component of reinforcement learning; Platform B generates a feature matrix of merchant push notifications for each user type under different marketing strategies as the action input (Action) of reinforcement learning, and uses the click behavior characteristics of each user type on each merchant as the sample input for reward value (Reward) calculation component of reinforcement learning.

[0107] The model node establishes a federated learning model, aligns the IDs of each user set, and distributes the federated reinforcement learning model through a third-party coordinator. Data sources from all parties upload data to the model according to the reinforcement learning framework to calculate distributed gradients, including feature self-learning gradients and gradient components of reward values ​​on all parties:

[0108]

[0109]

[0110] in,

[0111] The distributed distribution and calculation of the model are achieved through component scheduling coordination, and finally the optimal merchant recommendation matrix for each user class is obtained, that is, the merchant set corresponding to each user set is used as its recall selection pool.

[0112] Furthermore, after the data node sends the feature data of the object set to the model node, the method further includes:

[0113] The data node receives the model calculation result sent by the model node, wherein the model calculation result includes a correspondence between an object set and a recommendation set; the correspondence between the object set and the recommendation set is determined by the model node using the federated learning model;

[0114] The data node determines a target object set corresponding to the target individual object according to the attribute data of the target individual object;

[0115] The data node determines the recommended content of the target individual object according to the recommendation set corresponding to the target object set.

[0116] In the specific implementation process, the above steps can be used to obtain a personalized recommended merchant pool for users in each user set. For each user, combined with the device information and real-time behavior characteristics of the mobile terminal, the recommendation is further refined to the user granularity to make personalized recommendations for each user.

[0117] For example, based on user transaction behavior, platform A creates user portraits for users and merchant portraits for merchants, and groups users according to user classification standard fields. Based on the user's historical transaction data and the historical behavior data of platform A, a recall strategy is formulated according to scenario requirements, and the user's recall merchant list under each recall strategy is calculated. The marketing strategy obtained through federated reinforcement learning is used as a review preference recall merchant. Since it does not consider the user's real-time behavior, the merchants recalled through the reinforced federated model are scalable and applicable to different scenarios. Due to the characteristics of the recommendation system, the purpose of most recommendation scenarios is to enhance the interaction between users and the platform. Users can achieve multiple rounds of interaction with the recommendation system by turning pages. In this process, the recommendation system needs to be able to perceive the user's real-time behavior and understand the user's intention. Each page turn is a reinforcement learning feedback.

[0118] In order to more clearly understand the present invention, the above process is described in detail with specific embodiments below. The process of specific embodiment 1 is as follows: Figure 4 As shown, including:

[0119] After the user opens the app, the recommendation system completes the following operations:

[0120] The real-time user portrait, real-time merchant portrait, real-time user characteristics (exposure, click, favorite and other behaviors between the last refresh and this refresh, the business district, etc.), and user group tags are transmitted back. The recommendation system recalls the preferred merchants, sorts the recalled merchants according to the real-time characteristics, adjusts the results according to the user's location, card holding status and other conditions, and presents the recommendation results to the user.

[0121] The real-time click rate sent back by the user is used as the reward value to calculate the model recommendation effect, update and strengthen the federated learning model, adjust the user group push strategy in real time, and update the user group label according to the real-time user portrait.

[0122] Through the above process, personalized recommendations can be made for users, which can not only recall users' review preferences for merchants by combining data from multiple parties while protecting user privacy, but also ensure the diversity of recommendations, so that the recommendation system can better understand user intentions and improve user experience.

[0123] The scenario of the second specific embodiment is the development of a medical dispatch system based on federated learning based on user group alignment. The purpose is to reduce the user's waiting time while satisfying the user's satisfaction with diagnosis and treatment, and recommend the most suitable hospital and department for the user to register, so as to maximize the user's overall satisfaction. Cooperating with various hospitals to conduct multi-federated learning can protect the user's privacy and reasonably allocate social medical resources. The process is as follows Figure 5 As shown, including:

[0124] The users on the payment (UnionPay) side and the hospital side are divided into multiple user groups based on the basic attributes (age, gender, region, purchasing power, etc.) and treatment attributes (treatment items, treatment time, treatment location, treatment results, etc.) negotiated by both parties. A group ID is assigned to each user group, and features are extracted from each user group to form a feature data matrix. Among them, the data on the hospital side comes from the user's registration and treatment data, and the data on the payment side comes from the payment information of the registration and treatment. The maximum value of the numerical feature is used as the user group feature value, and the categorical feature is expressed using a onehot vector.

[0125] In some categories, there may be fewer users, which means there is a risk of leaking user privacy. Therefore, before the data fusion model is built, the data of both parties still need to be encrypted and aligned.

[0126] After the data from all parties are integrated, the federated dual-model collaborative training is performed:

[0127] According to the user's medical record characteristics and hospital attribute characteristics, a federated regression model is trained to obtain the user's satisfaction with the treatment in the current hospital and clinic;

[0128] Based on the user's basic attributes and historical reservation status, the federated classification model is trained to determine whether the user's reservation status at different levels is acceptable.

[0129] The multi-label, dual-model collaborative training is performed by integrating diagnosis and treatment satisfaction and appointment acceptability. The loss function is to maximize the user's overall satisfaction:

[0130] R (用户满意度) =f (诊疗满意度，预约可接受)

[0131] The saved model is used to process real-time user requests. Basic attribute characteristics and diagnosis and treatment characteristics are calculated through real-time requests to obtain the user's overall satisfaction prediction for different hospitals and different clinics.

[0132] When users view various hospitals and clinics through the UnionPay medical channel, the basic attribute information of the user is first obtained, and the user fills in the real-time condition description and medical treatment items requested. The user characteristics are calculated and the overall satisfaction of the user with each hospital and clinic is calculated through the model. The top N clinics are taken as the candidate diagnosis and treatment list, and then they are screened according to distance and other information to recommend the most suitable hospitals and clinics for users.

[0133] The embodiment of the present invention also provides a multi-party secure computing device for federated learning, such as Figure 6 As shown, including:

[0134] A division unit 601 is used to divide all individual objects in the data node into multiple object sets according to a classification standard;

[0135] A calculation unit 602 is used to determine, for each object set, feature data of the object set according to feature data of all individual objects in the object set;

[0136] The transceiver unit 603 is used to send the feature data of the object set to the model node, so that the model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes to obtain sample data of the object set, and trains the federated learning model according to the sample data of all object sets;

[0137] Among them, the classification standard for dividing individual objects in multiple data nodes is the same.

[0138] Optionally, the computing unit 602 is specifically configured to:

[0139] The mean of the continuous feature data of all individual objects is calculated as the first feature data of the object set.

[0140] Optionally, the computing unit 602 is specifically configured to:

[0141] Using one-hot encoding, the discrete feature data of the individual object is converted into a feature vector;

[0142] The feature data of the object set are determined by using the feature vectors of all individual objects in the object set.

[0143] Optionally, the transceiver unit 603 is further used to receive a model calculation result sent by the model node, where the model calculation result includes a correspondence between an object set and a recommendation set; the correspondence between the object set and the recommendation set is determined by the model node using the federated learning model;

[0144] The system further comprises a recommendation unit 604, which is used to determine a target object set corresponding to the target individual object according to the attribute data of the target individual object; and determine recommended content for the target individual object according to a recommendation set corresponding to the target object set.

[0145] Based on the same principle, the present invention also provides an electronic device, such as Figure 7 As shown, including:

[0146] It includes a processor 701, a memory 702, a transceiver 703, and a bus interface 704, wherein the processor 701, the memory 702 and the transceiver 703 are connected via the bus interface 704;

[0147] The processor 701 is configured to read the program in the memory 702 and execute the following method:

[0148] Dividing all individual objects in the data node into multiple object sets according to classification criteria;

[0149] For each object set, determining feature data of the object set according to feature data of all individual objects in the object set;

[0150] Sending feature data of an object set to a model node, so that the model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes, obtains sample data of the object set, and trains a federated learning model based on the sample data of all object sets;

[0151] The classification standards for dividing individual objects in the multiple data nodes are the same.

[0152] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0153] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0154] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0155] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0156] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A multi-party secure computing method for federated learning, characterized in that: include: The data node divides all individual objects in the data node into multiple object sets according to the classification standard; wherein the individual objects are users or other objects, and the other objects include terminal devices and software applications; For each object set, the data node determines the feature data of the object set according to the feature data of all individual objects in the object set; The data node sends the feature data of the object set to the model node, so that the model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes, obtains the sample data of the object set, and trains the federated learning model according to the sample data of all object sets; The classification standards for dividing individual objects in the multiple data nodes are the same; The characteristic data of the individual objects include continuous characteristic data, and the data node determines the characteristic data of the object set according to the characteristic data of all individual objects in the object set, including: The data node calculates the mean of the continuous feature data of all individual objects as the first feature data of the object set; The characteristic data of the individual objects include discrete characteristic data, and the data node determines the characteristic data of the object set according to the characteristic data of all individual objects in the object set, including: The data node converts the discrete feature data of the individual object into a feature vector by using one-hot encoding; The feature data of the object set are determined by using the feature vectors of all individual objects in the object set.

2. The method according to claim 1, characterized in that After the data node sends the feature data of the object set to the model node, the method further includes: The data node receives the model calculation result sent by the model node, wherein the model calculation result includes a correspondence between an object set and a recommendation set; the correspondence between the object set and the recommendation set is determined by the model node using the federated learning model; The data node determines a target object set corresponding to the target individual object according to the attribute data of the target individual object; The data node determines the recommended content of the target individual object according to the recommendation set corresponding to the target object set.

3. A multi-party secure computing method for federated learning, characterized in that: include: The model node receives feature data of an object set sent by a plurality of data nodes, and the plurality of data nodes classify individual objects into different object sets according to the same classification standard; wherein the objects are users or other objects, and the other objects include terminal devices and software applications; The model node performs sample alignment on all feature data of the same object set to obtain sample data of the object set; The model node trains the federated learning model based on the sample data of all object sets; The characteristic data of the individual object includes continuous characteristic data, and the process of determining the characteristic data of the object set includes: The data node calculates the mean of the continuous feature data of all individual objects as the first feature data of the object set; The characteristic data of the individual object includes discrete characteristic data, and the process of determining the characteristic data of the object set includes: The data node converts the discrete feature data of the individual object into a feature vector by using one-hot encoding; The feature data of the object set are determined by using the feature vectors of all individual objects in the object set.

4. A multi-party secure computing device for federated learning, characterized in that: include: A partitioning unit, used to partition all individual objects in the data node into a plurality of object sets according to a classification standard; wherein the objects are users or other objects, and the other objects include terminal devices and software applications; a calculation unit, configured to determine, for each object set, feature data of the object set according to feature data of all individual objects in the object set; A transceiver unit, used to send feature data of an object set to a model node, so that the model node performs sample alignment on all feature data of the same object set according to the feature data of the object set sent by multiple data nodes to obtain sample data of the object set, and trains the federated learning model according to the sample data of all object sets; Among them, the classification standard for dividing individual objects in multiple data nodes is the same; Wherein, the computing unit is specifically used for: Calculating the mean of the continuous feature data of all individual objects as the first feature data of the object set; Wherein, the computing unit is specifically used for: Using one-hot encoding, the discrete feature data of the individual object is converted into a feature vector; The feature data of the object set are determined by using the feature vectors of all individual objects in the object set.

5. The device according to claim 4, characterized in that The transceiver unit is further used to receive a model calculation result sent by the model node, wherein the model calculation result includes a correspondence between an object set and a recommendation set; the correspondence between the object set and the recommendation set is determined by the model node using the federated learning model; It also includes a recommendation unit, which is used to determine a target object set corresponding to the target individual object according to the attribute data of the target individual object; and determine the recommended content of the target individual object according to the recommendation set corresponding to the target object set.

6. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1-2.

7. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute any one of the methods of claims 1 to 2.

Images