Data mining method, system, device and storage medium based on privacy computing

Abstract

This application relates to a data mining method, system, device, and computer-readable storage medium based on privacy computing. The method includes: a first node sends a first message to a second node, the first message carrying the user identifier of a target user; the first node aligns first sample data provided by the first node with second sample data provided by the second node based on the user identifier, obtaining first aligned data of the first node and second aligned data of the second node; the first node fuses the first aligned data and the second aligned data to obtain a virtual fused data set; the first node performs metadata parsing on the first aligned data to obtain first parsed data; and the first node receives the second parsed data sent by the second node. The first node configures operator parameters for the target user based on the target user's label data, the first parsed data, and the second parsed data, and constructs a data mining workflow based on the operator parameters. This application can improve the overall efficiency of privacy data mining.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to data mining methods, systems, devices, and storage media based on privacy computing. Background Art

[0002] Data mining technology is essentially the process of revealing implicit, previously unknown, and potentially valuable information from large amounts of data. However, this process is often difficult to complete through a one-time, fixed processing flow for all data mining and analysis. Due to the special nature of private data, mining private data is significantly different from mining ordinary data. This is manifested in that if private data is not processed and the same approach as for mining ordinary data is used, data mining for private data will be more time-consuming, more difficult, and more expensive. In related technologies, a data mining method based on a process orchestration canvas can be adopted. Specifically, a user (usually a data scientist) drags operators from an operator library into the canvas and connects the operators using lines representing data flow and task execution order to form a workflow. Thereafter, executing the workflow completes a data mining task.

[0003] As mentioned above, both general data mining and privacy-preserving data mining involve a repetitive, exploratory nature. For privacy-preserving data mining, the aforementioned technologies lack timely feedback on the results of each workflow task to guide the orchestration and configuration of the next workflow. Furthermore, insufficient understanding of computing power and data can lead to misoperations, requiring frequent workflow modifications, significantly reducing the efficiency of privacy-preserving data mining. Summary of the Invention

[0004] In order to solve or partially solve the problems existing in the related technologies, the present application provides a data mining method, system, device and computer-readable storage medium based on privacy computing to improve the overall efficiency of data mining based on privacy computing.

[0005] In a first aspect, the present application provides a data mining method based on privacy computing, which is applied to a secure multi-party privacy computing system. The secure multi-party privacy computing system includes multiple nodes participating in privacy computing, wherein the multiple nodes include a first node and a second node. The method includes:

[0006] The first node sends a first message to the second node, where the first message carries a user identifier of a target user and is used to request initiation of privacy computing for the target user;

[0007] The first node aligns first sample data provided by the first node and second sample data provided by the second node according to the user identifier to obtain first aligned data of the first node and second aligned data of the second node; wherein the second sample data is original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing;

[0008] The first node fuses the first aligned data and the second aligned data based on a privacy-preserving computing protocol to obtain a virtual fused data set;

[0009] The first node performs meta-information parsing on the first aligned data to obtain first parsed data; notifies the second node of an alignment result of the second aligned data of the second node, and receives second parsed data sent by the second node, wherein the second parsed data is obtained by the second node performing meta-information parsing on the second aligned data according to the alignment result;

[0010] The first node configures operator parameters for the target user based on the target user's label data, the first parsed data, and the second parsed data, and constructs a data mining workflow based on the operator parameters.

[0011] In one embodiment, the first parsed data and the second parsed data include distribution characteristics of target users, and the method further includes:

[0012] According to the distribution characteristics of the target user included in the first parsed data and the second parsed data, an outlier filtering logic and a configuration information value are configured for the target user to determine a binning result.

[0013] In one embodiment, the method further comprises:

[0014] Processing the first analytical data and the second analytical data to obtain global characteristic extreme values, global variance, and global mean;

[0015] The configuring, based on the distribution characteristics of the target user included in the first parsed data and the second parsed data, outlier filtering logic and configuration information values for the target user to determine a binning result includes:

[0016] Based on the global variance and the global mean, configuring outlier filtering logic for the target user;

[0017] Based on the global feature extreme value and the preset binning ratio, counting the first statistical data quantity and the second statistical data quantity;

[0018] determining a target quantile based on a global statistical data quantity, the first statistical data quantity, and the second statistical data quantity;

[0019] Based on the target quantile line, feature binning is performed on the feature data corresponding to the target user in the virtual fusion dataset to obtain a data binning result.

[0020] In one embodiment, configuring outlier filtering logic and configuration information values for the target user based on the distribution characteristics of the target user included in the first parsed data and the second parsed data to determine the binning result includes:

[0021] Sorting the feature values of the virtual fusion data set corresponding to the target user to obtain a sorting result;

[0022] Based on the preset binning strategy, the feature values in the sorting results are binned to obtain multiple feature bins;

[0023] Based on multiple feature bins, calculate the evidence weight value of each feature bin;

[0024] Calculating the information value of each feature of the target user based on the evidence weight value and the label data of the target user;

[0025] Based on the information value of the feature of each target user, user features corresponding to information values greater than or equal to a preset value threshold are screened.

[0026] In one embodiment, when configuring the parameters of the operator for the target user, field attributes of the field are configured for each operator, wherein the field attributes include one or a combination of user identification, feature, and label.

[0027] In one embodiment, when constructing a data mining workflow based on the parameters of the operator, data preprocessing is performed on the virtual fusion dataset, and labels and features are marked on the virtual fusion dataset.

[0028] In one embodiment, the method further comprises:

[0029] When configuring the field attributes of a field for each operator, the first node records the node from which the field comes;

[0030] If the source node is the first node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are stored on the first node; if the source node is the second node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are sent to the second node for storage;

[0031] The first node or the second node uses the encoder stored in itself to encode the information value and the evidence weight value of the field.

[0032] A second aspect of the present application provides a secure multi-party privacy-preserving computing system, comprising a plurality of nodes participating in privacy computing, wherein the plurality of nodes include a first node and a second node:

[0033] The first node is used to send a first message to the second node, where the first message carries the user identifier of the target user and is used to request to initiate privacy computing for the target user; align the first sample data provided by the first node and the second sample data provided by the second node according to the user identifier to obtain the first aligned data of the first node and the second aligned data of the second node; wherein the second sample data is the original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing; fuse the first aligned data and the second aligned data of the second node based on the privacy computing protocol to obtain a virtual fused data set; perform metadata parsing on the first aligned data to obtain first parsed data; notify the second node of the alignment result of the second aligned data of the second node, and receive the second parsed data sent by the second node, wherein the second parsed data is obtained by the second node performing metadata parsing on the second aligned data according to the alignment result; configure the parameters of the operator for the target user based on the label data of the target user, the first parsed data and the second parsed data of the second node, and construct a data mining workflow based on the parameters of the operator;

[0034] The second node is used to receive the first message, obtain the original data of the target user according to the user identifier, perform setting processing to obtain second sample data, and then provide it to the first node; perform metadata parsing on the second aligned data according to the alignment result to obtain second parsed data and send it to the first node.

[0035] In one embodiment, when configuring parameters of an operator for the target user, the first node configures field attributes of a field for each operator, wherein the field attributes include one or a combination of a user identifier, a feature, and a tag.

[0036] In one embodiment, when configuring the field attributes of a field for each operator, the first node records the node from which the field comes;

[0037] If the source node is the first node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are stored on the first node; if the source node is the second node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are sent to the second node for storage;

[0038] The first node or the second node uses the encoder stored in itself to encode the information value and the evidence weight value of the field.

[0039] A third aspect of the present application provides a computing device, including: a processor; and

[0040] The memory stores executable codes thereon, and when the executable codes are executed by the processor, the processor is caused to execute the method described above.

[0041] A fourth aspect of the present application provides a computer-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of a computing device, the processor is caused to execute the method described above.

[0042] The technical solution provided by this application may have the following beneficial effects:

[0043] On the one hand, a first node aligns first sample data provided by the first node and second sample data provided by the second node based on a user identifier, thereby obtaining first aligned data of the first node and second aligned data of the second node; wherein the second sample data is the original data of the target user obtained by the second node based on the user identifier and provided to the first node after setting processing; the first node then fuses the first aligned data and the second aligned data based on a privacy computing protocol to obtain a virtual fused data set; the first node then performs metadata parsing on the first aligned data to obtain first parsed data; the alignment result of the second aligned data of the second node is notified to the second node, and the second parsed data sent by the second node is received, wherein the second parsed data is obtained by the second node performing metadata parsing on the second aligned data based on the alignment result. Through the above processing, on the one hand, the mining processing speed can be improved through distributed processing; on the other hand, because the first parsed data and the second parsed data have privacy characteristics and do not expose the original data, the virtual fused data set obtained by the first node has both privacy characteristics and can represent the characteristics of the target user. When subsequently configuring operator parameters for the target user, it is possible to understand the data characteristics analyzed by each participant (including the first node and the second node) of the privacy computing on the target user over a historical period. Therefore, by using the computing solution provided in this application, the number of debugging times for data analysis, i.e., joint data insight tasks, can be reduced while protecting the data privacy of all parties, and the data processing speed can be increased, thereby improving the overall debugging efficiency of data mining based on privacy computing.

[0044] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and other objects, features and advantages of the present application will become more apparent by describing in more detail exemplary embodiments of the present application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the present application.

[0046] Figure 1 Schematic diagram of the application environment of the data mining method based on privacy computing shown in an embodiment of the present application;

[0047] Figure 2 This is a flow chart of a data mining method based on privacy computing shown in an embodiment of the present application;

[0048] Figure 3 Schematic diagram of the application framework of the data mining method based on privacy computing shown in an embodiment of the present application;

[0049] Figure 4 This is a schematic diagram of data alignment provided in an embodiment of the present application;

[0050] Figure 5 is a schematic diagram of obtaining a virtual fusion data set in an embodiment of the present application;

[0051] Figure 6 This is a schematic diagram of a workflow for connecting operators to form an operator in an embodiment of the present application;

[0052] Figure 7 This is a schematic diagram of the structure of a secure multi-party privacy computing system shown in an embodiment of the present application;

[0053] Figure 8 is a schematic diagram of the structure of a computing device shown in an embodiment of the present application;

[0054] Figure 9 It is a structural diagram of the server shown in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The following describes embodiments of the present application in more detail with reference to the accompanying drawings. Although the accompanying drawings illustrate embodiments of the present application, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0056] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. As used in this application and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0057] It should be understood that although the terms "first", "second", "third", etc. may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0058] The embodiment of the present application provides a data mining method based on privacy computing, and the solution of the embodiment of the present application can be used for mining private data.

[0059] The mining of private data encompasses data mining and privacy computing. Data mining generally refers to the process of using algorithms to search for hidden information within large amounts of data. Data mining typically encompasses processes such as data cleaning, data transformation, and actual data mining. Data mining tasks include association analysis, cluster analysis, classification analysis, anomaly analysis, specific group analysis, and evolution analysis. Data mining techniques reveal hidden, previously unknown, and potentially valuable information from large amounts of data in databases. This process is often exploratory, meaning it is difficult to complete mining and analysis of all data using a one-time, fixed processing flow. Data mining techniques typically fall into two categories: data processing and data visualization. Each step in data processing relies on data visualization to display a comprehensive overview of the data before and after processing, such as metadata, distribution, and correlations. This helps guide data miners in further data processing or allows them to modify parameters in data processing steps to correct errors. With the gradual advancement of data mining technology, AI-related technologies are also being incorporated into the data mining process. Privacy computing, a combination of underlying technologies designed to ensure data is "available but invisible," encompasses federated learning, multi-party secure computation, and homomorphic encryption. Privacy computing addresses security concerns during data collaboration and breaks down data silos. It enables direct collaboration between multiple data sources while ensuring mutual data invisibility through cryptographic encryption, protecting data privacy and security. This creates a better collaborative environment for the development of data applications such as AI and data mining. Privacy computing has become a mainstream technology for protecting the privacy of personal, commercial, and public data. Based on underlying cryptographic protocols, privacy computing addresses the issue of data being "available but invisible," while data mining technology addresses the question of "how to use" data. Currently, mainstream privacy computing platforms on the market are all technology fusion products that use privacy computing as their foundation and data mining as upper-level tools. These products offer users essentially the same user experience as traditional data mining products (such as data middleware platforms, big data platforms, and data mining platforms that don't utilize privacy computing architectures), and the data mining process for users (typically data scientists) on these platforms is largely the same. However, due to security considerations, privacy-focused computing technology requires orders of magnitude more computation and communication than plaintext big data platforms. Consequently, privacy-focused computing platforms require tens to millions of times more time than plaintext operations. While mainstream privacy-focused computing platforms have made significant efforts to improve efficiency, achieving the same performance as plaintext computing remains elusive.On this basis, if users use the privacy computing platform in an exploratory trial-and-error manner, just like traditional data middle-end operation technology, and a single attempt takes hours or even days, then the difficulty and cost of data mining tasks under the privacy computing framework will increase significantly, which poses a great challenge to users.

[0060] Regarding data mining methods based on privacy computing, the related technology is a data mining method based on a process orchestration canvas. Specifically, the user (usually a data scientist) drags operators from the operator library into the canvas and connects the operators using lines representing the data flow and task execution order to form a workflow. After that, executing the workflow completes a data mining task. As mentioned above, both ordinary data mining and data mining based on privacy computing are exploratory and iterative. For the data mining based on privacy computing in the above-mentioned related technology, since users cannot obtain timely feedback on the results of each workflow task to guide the orchestration and configuration of the next workflow, and due to incorrect operations caused by insufficient understanding of computing power and insufficient understanding of data, users modify the workflow more frequently, which will seriously affect the efficiency of data mining based on privacy computing.

[0061] In response to the above problems, the embodiments of the present application provide a data mining method, system, device and storage medium based on privacy computing to improve the overall efficiency of data mining based on privacy computing.

[0062] In some implementations, the present invention can be applied to Figure 1 An application environment is shown. The application environment can be a secure multi-party privacy computing system, which includes multiple nodes participating in privacy computing, including a first node and a second node. The first node and the second node are the initiators (host) or participants (guest) of federated learning and multi-party secure computing. The nodes can be servers or terminals, etc. Each node, such as a server, can be equipped with a database, and the terminal can be a personal computer, etc.

[0063] When based on Figure 1When the application environment shown implements the above-mentioned method of the embodiment of the present application, the first node sends a first message to the second node, the first message carries the user identifier of the target user, and is used to request to initiate privacy computing for the target user; the first node aligns the first sample data provided by the first node and the second sample data provided by the second node according to the user identifier to obtain the first aligned data of the first node and the second aligned data of the second node; wherein the second sample data is the original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing; the first node fuses the first aligned data and the second aligned data based on the privacy computing protocol to obtain a virtual fused data set; the first node performs metadata parsing on the first aligned data to obtain first parsed data; notifies the second node of the alignment result of the second aligned data of the second node, and receives the second parsed data sent by the second node, wherein the second parsed data is obtained by the second node performing metadata parsing on the second aligned data according to the alignment result; the first node configures the parameters of the operator for the target user based on the label data of the target user, the first parsed data, and the second parsed data, and constructs a data mining workflow based on the parameters of the operator.

[0064] The solutions of the embodiments of the present application can be implemented based on technologies such as artificial intelligence (AI), natural language processing (NLP), and machine learning (ML), and are specifically described through the following embodiments:

[0065] AI refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, artificial intelligence is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. Artificial intelligence is the study of the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0066] AI technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0067] NLP is a key area of research in computer science and artificial intelligence. It studies the theories and methods that enable effective communication between humans and computers using natural language. Natural language processing (NLP) integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language we use in everyday life—and is closely linked to linguistics. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robotic question answering, and knowledge graphs.

[0068] Among them, it should be noted that the server involved in the embodiment of the present application (such as the node in the above figure) can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal involved in the embodiment of the present application can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, personal digital assistant, etc., but is not limited to this. The nodes of the present application can be directly or indirectly connected by wired or wireless communication, and the embodiment of the present application is not limited here.

[0069] The technical solutions of the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0070] See also Figure 2 , is a flow chart of a data mining method based on privacy computing shown in an embodiment of the present application. The method is applied to a secure multi-party privacy computing system. The secure multi-party privacy computing system includes multiple nodes participating in privacy computing, and the multiple nodes include a first node and a second node. At the same time, you can refer to Figure 3 , is a schematic diagram of the application framework of the data mining method based on privacy computing shown in an embodiment of the present application.

[0071] like Figure 3 As shown, the technical solution of the embodiment of the present application can be applied to data mining scenarios, and can be based on Figure 3 The framework shown is implemented. The traditional data mining process in related technologies includes data alignment, data preprocessing, feature engineering, model training, model evaluation, and model launch. The embodiment of this application adds processing steps such as virtual fusion datasets for metadata analysis (i.e., privacy data insights), participant matching, ID (user identification) column labeling, label column labeling, and feature column labeling to the traditional data mining and privacy computing processes.

[0072] Among them, marking the feature column means marking the "feature column" from all columns, so that the downstream links always know that it is a feature column. The virtual fusion data set is subjected to meta-information analysis, including parsing the meta-information such as the data type, number of rows, distribution characteristics, mean, variance, quantile, missing rate, number of unique values of each field (i.e., each column) of the virtual fusion data set. In the embodiment of the present application, the field attributes of each field include features, labels, and ids. The characteristic of the id column is that each row value is different, which is often used for data alignment and should not enter the subsequent data preprocessing, feature engineering, model training and other links. Among them, the feature columns are usually divided into numerical and text types, and they need to go through different data conversion operators to be processed into usable features that can enter model training. Among them, the label column is usually labeled with integers such as 0 and 1 in the classification task. If it is a text type value label, it needs to be converted. Generally, the label column is a continuous value in regression tasks, while the label column is generally not required in clustering tasks. By marking the feature columns, label columns, and ID columns, the ID and label columns can be excluded as user-selectable options in components that require feature columns for calculation, thereby avoiding misoperation. Therefore, the solution of the embodiment of the present application, through the unified marking of feature columns, label columns, and ID columns throughout the entire process and the participant matching technology, reduces the possibility of users defining invalid data mining processes while protecting data privacy, thereby improving task debugging efficiency.

[0073] in, Figure 2 The exemplary method mainly includes steps S201 to S205, which are described as follows:

[0074] Step S201: The first node sends a first message to the second node, wherein the first message carries a user identifier of a target user and is used to request to initiate privacy computing for the target user.

[0075] It is assumed here that privacy computing includes two underlying technologies: federated learning and multi-party secure computing. The first node and the second node mentioned in the above embodiment are the initiators (host) or participants (guest) of federated learning and multi-party secure computing.

[0076] The initiator may refer to a party that initiates federated learning and provides a label field in the federated learning, for example, the first node.

[0077] A host may refer to a party that provides user data for joint modeling.

[0078] In the embodiment of the present application, the first node is used as the initiator and the second node is used as the participant. That is, the first node can be the initiator of federated learning and multi-party secure computing, and the second node can be the participant of federated learning and multi-party secure computing. The first message sent by the first node to the second node carries the ID (user identification) of the target user and is used to request the initiation of privacy computing for the target user. That is, at this time, the first node initiates privacy computing for the target user to the second node. The number of initiators can generally be one, and the number of participants can be one or more, which is not limited in the embodiment of the present application.

[0079] For example, if a bank needs to leverage data from insurance companies to assess the eligibility of its own clients, or target users, it can act as the initiator of federated modeling, with the insurance companies acting as additional participants. This additional participant, the insurance company, can be expanded to include multiple parties. This allows the bank to indirectly leverage data from a wider range of participants, providing more precise scores on client loan eligibility and further improving the accuracy of the model.

[0080] Step S202: The first node aligns the first sample data provided by the first node and the second sample data provided by the second node according to the user identifier to obtain the first aligned data of the first node and the second aligned data of the second node; wherein the second sample data is the original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing.

[0081] Privacy computing technology is a technology that allows joint computing without leaving the original data domain. Figure 3 In the embodiment of the present application, original data 1 and original data 2 can come from two parties, but the original data do not leave the domain of the two parties. In order to help with joint calculations, it is necessary to align the data on both sides, that is, data alignment can be performed according to id (user identification) (this can also be called vertical alignment). After data alignment, only a part of the original data on both sides is available, and the corresponding relationship with another data set is maintained. For example, the first node aligns the first sample data provided by the first node and the second sample data provided by the second node to obtain the first aligned data of the first node and the second aligned data of the second node; wherein the second sample data is the original data of the target user obtained by the second node according to the user identification and provided to the first node after setting processing.

[0082] Federated learning is a distributed machine learning technology or machine learning framework based on data privacy protection. Information related to federated learning can be transmitted in an encrypted manner, ensuring that no participant can infer the original data of the other party. Federated learning enables multi-party joint learning and training through the circulation and processing of intermediate encrypted data without releasing local original data from the warehouse. The encryption processing can be, for example, homomorphic encryption. Federated learning involves data alignment (also known as sample alignment). For example, if there are two participants whose respective user groups do not completely overlap, encryption-based user sample alignment technology can be used to confirm the common users of both parties (without exposing overlapping users) without disclosing their respective original data. Modeling can then be performed based on the characteristics of the joint common users.

[0083] Federated learning includes vertical federated learning (also known as sample-aligned federated learning) and horizontal federated learning (also known as feature-aligned federated learning).

[0084] The core of horizontal federated learning is the union of samples. It is applicable to scenarios where participants have the same industry status but reach different customers, that is, scenarios with a lot of feature overlap and little user overlap. For example, shopping malls in different regions have similar services and products (similar features) but different customers (different samples). The name of horizontal federated learning comes from the "horizontal division" of training data, that is, the (horizontal) division of the data matrix or table by row. Data in different rows have the same data features (samples), that is, the data features are aligned.

[0085] The core of vertical federated learning is the federation of features. Unlike horizontal federated learning, this approach is applicable to scenarios where participants have different industry status but the same clientele. For example, a shopping mall and a restaurant in the same region share local customers (identical samples), but offer different services and products (different features). Vertical federated learning gets its name from the "vertical partitioning" of training data, which involves dividing a data matrix or table into columns (vertical division). Data in different columns share the same sample labels (features), meaning the training samples are aligned.

[0086] Vertical federated learning involves splitting two datasets vertically (i.e., along the feature dimension) when there's significant user overlap but limited overlap in user characteristics. Training is then performed on the portion of data where users are common but their characteristics differ. For example, consider two different institutions: a bank in a certain location, and an e-commerce company in the same location. Their user bases likely encompass the majority of the local population, resulting in a significant overlap in user characteristics. However, since the bank records user spending behavior and credit ratings, while the e-commerce company maintains user browsing and purchase history, the overlap in user characteristics is minimal. Vertical federated learning aggregates these different features in an encrypted state to enhance model capabilities.

[0087] Because samples in vertical federated learning belong to different organizations, and the sample coverage of each organization varies, it is generally necessary to first align samples across domains to find the intersection. This can be done, for example, by using sample IDs, such as mobile phone numbers.

[0088] The embodiment of the present application takes the data alignment of vertical federated learning as an example but is not limited thereto. In the process of data alignment in the embodiment of the present application, the first node may determine the intersection of the user identifiers of the first node and the second node based on the user identifier of the target user.

[0089] The intersection of user IDs refers to the set containing the same user ID. As described above, each participant in privacy computing can be a node in federated learning, and each participant in privacy computing has different sample data.

[0090] For example, small and micro Internet financial companies have sample lending history data and loan performance data, mobile operators have all users' call records, text message records and other data, and banks have users' deposit, loan and financial management data.

[0091] For example, the participants in privacy computing include bank A, merchant B and online payment platform C. The sample data owned by bank A includes the asset status of the users corresponding to bank A, the sample data owned by merchant B includes the product purchase data of the users corresponding to merchant B, and the sample data owned by online payment platform C includes the transaction records of the users of online payment platform C, etc.

[0092] The purpose of obtaining this intersection is to provide an accurate data foundation for subsequent modeling. However, due to the data privacy requirements of each participant, the first node aligns the sample data provided by each participant in the privacy-focused computation, assuming that the participants interacting with each other based on their secure privacy-focused computing protocol are unable to perceive the shared sample data. This intersection is achieved by aligning the sample data provided by each participant. The sample data provided by each participant can be raw data that has undergone specific processing, such as filtering some data and applying homomorphic encryption, before being provided to the first node.

[0093] Figure 4 This is a schematic diagram of aligning the data of samples from all parties involved in privacy computing, as provided in an embodiment of the present application. Figure 4 Taking Bank A, Merchant B, and Online Payment Platform C as an example, when Bank A, Merchant B, and Online Payment Platform C jointly conduct privacy computing, they need to filter out the common user group of Bank A, Merchant B, and Online Payment Platform C. Only the sample data corresponding to this common user group among the three participants is meaningful for model training. Here, it is assumed that before privacy computing participants 1 and 2 conduct vertical federated learning model training, they need to align their sample data and filter out the intersection of their sample data with the same user ID. In other words, they need to find the intersection of the sample data with the same user ID in sets D1 and D2. For example, the feature information X1 and X2 of the same bank customer with the same user ID, held by participants 1 and 2 respectively, need to be aligned. That is, they need to be combined during model training to form a training sample (X1, X2); combining the feature information of different bank customers is meaningless and cannot construct a training sample.

[0094] As can be seen from the above description, sample data alignment first requires mutual discovery between at least two parties (for example, Participant 1 and Participant 2). This can be achieved through a registration server located on a central server. In other words, Participant 1 first needs to obtain the IP address and port of the other party to be aligned before establishing a network connection. Any participant wishing to participate in federated training must first register with the central server, specifying information such as the type of dataset provided, local IP address, and listening port. The registration server maintains a globally unified metadata hash table for each participant. This table contains metadata related to each participant, such as the local IP address and current listening port, as well as information such as whether the participant is currently ready for federated modeling. Each party wishing to participate in private computing must first register its information with the registration server. Only in this way can the private computing participants be aware of each other. Otherwise, each private computing participant is isolated, and other participants cannot align their sample data with it. Figure 4A schematic diagram is given for determining the intersection of the user identifiers of the first node and the second node and the intersection of the sample data based on the user identifier of the target user. Participant 1 has sample data X3, X4 and X5 with user identifiers u1, u2, u4, u6 and u7, and participant 2 has sample data X1, X2 and Y with user identifiers u1, u2, u3, u5 and u7. Then, the intersection of the user identifiers of participant 1 and participant 2 is {u1, u2, u7}. When the first node aligns the sample data of participant 1 and participant 2, the intersection of the sample data with user identifiers u1, u2 and u7 is {X1, X2, X3, X4, X5, Y}.

[0095] It should be noted that the above-mentioned data alignment is based on an example in which each participant sends their own sample data to the first node, and the first node performs sample alignment to obtain a sample alignment result, but it is not limited to this. It can also be that each participant performs sample alignment in pairs, and then summarizes the sample alignment results to the first node. The first node integrates the results of each sample alignment to obtain an overall sample alignment result.

[0096] Step S203: The first node fuses the first aligned data and the second aligned data based on the privacy computing protocol to obtain a virtual fused data set.

[0097] See also Figure 5 , is a data flow diagram for obtaining a virtual fused data set. After the first node performs data alignment, the first aligned data of the first node after data alignment and the second aligned data of the second node after data alignment are fused based on the privacy computing protocol to obtain a virtual fused data set. At this time, for the initiator of the calculation, such as the first node, what is seen is some basic information of the data set after the "fusion" of the two original data after alignment, but in fact this data set is scattered on both sides, so it is called a "virtual fused data set". In other words, after data alignment, the example solution of this application can obtain a "virtual fused data set".

[0098] The first node can encrypt and aggregate the first aligned data and the second aligned data according to a preset encryption algorithm to obtain a virtual fused dataset. Subsequent steps can parse the metadata of the virtual fused dataset, such as the type, number of rows, number of columns, mean, variance, extreme values, quantiles, missingness rate, and number of unique values of the data in the virtual fused dataset, to obtain parsed data.

[0099] It can be seen that in the embodiment of the present application, the first node first determines the intersection of the user identifiers of the first node and the second node based on the user identifier of the target user, and then encrypts and aggregates the first aligned data of the first node and the second aligned data of the second node corresponding to the intersection according to the preset encryption algorithm to obtain a virtual fusion data set. This can reduce the amount of data mining processing, only analyze and process the required data, and improve the execution success rate of the data mining process.

[0100] Step S204: the first node performs metadata parsing on the first aligned data to obtain first parsed data; notifies the second node of the alignment result of the second aligned data of the second node, and receives the second parsed data sent by the second node, wherein the second parsed data is obtained by the second node performing metadata parsing on the second aligned data according to the alignment result.

[0101] After data alignment, the first aligned data of the first node in the virtual fused dataset contains metadata such as the original data type, number of rows, number of columns, mean, variance, extreme values, quantiles, missing rate, and number of unique values. For example, the original data could be a user's bank loan history. Assuming this loan history includes 24 months of repayment records, with each monthly repayment record represented as a separate row, the total number of rows is 24. Each column in the repayment record corresponds to a different item field, including 10 columns, including fields such as the loan date, repayment date, each repayment amount, and whether the payment is overdue. By performing operations based on the repayment record data according to specified indicators, the average monthly repayment value and average annual repayment value for the loan history data can be obtained.

[0102] In some embodiments, after the first node determines the virtual fusion data set, it can parse the type, number of rows, number of columns, mean, variance, characteristic extreme value, quantile, missing rate, and number of unique values of the first aligned data of the first node in the virtual fusion data set in its own domain to obtain first parsed data. It should be noted that since the first node parses the metadata of the first aligned data in its own domain, it complies with the principle that the data provided by each participant in privacy computing does not go beyond their respective domains. Therefore, even if the first parsed data obtained by parsing the first aligned data is provided to a third party (such as a second node as a participant), the third party can only obtain the characteristic extreme values of the original data, but cannot obtain the characteristics of the original data, so the original data itself will not be exposed to the third party.

[0103] For example, assuming that the first aligned data of the first node in the virtual fusion dataset is a user's loan history data at a bank, after parsing the metadata, the parsed data obtained includes the loan history data in 24 rows and 10 columns, the average monthly repayment value, the average annual repayment value, etc.

[0104] The first node may notify the second node of the alignment result of the second node's second aligned data and receive second parsed data from the second node. The second parsed data is obtained by the second node performing metadata parsing on the second aligned data based on the alignment result. Based on the notification from the first node, the second node determines the second aligned data of the second node after data alignment. The metadata of the second aligned data of the second node may include the type of the original data, number of rows, number of columns, mean, variance, eigenvalues, quantiles, missing rate, and number of unique values. For example, the original data may be a user's purchase data on a shopping platform. Assume that the purchase data includes purchase records for 100 items, with each purchase record represented as a separate row, resulting in 100 rows. Each column in the purchase record corresponds to a different item field, including six columns, including product name, unit price, quantity, and purchase date. By performing operations based on the product purchase data according to predefined indicators, average daily purchase costs, average monthly purchase costs, and other data can be obtained.

[0105] In some embodiments, the second node parses the type, number of rows, number of columns, mean, variance, characteristic extreme values, quantiles, missing rate, and number of unique values of the second aligned data of the second node in its own domain to obtain second parsed data. It should be noted that since the second node parses the metadata of the second aligned data in its own domain, it complies with the principle that the data provided by each participant in privacy computing does not exceed their respective domains. Therefore, even if the second aligned data is provided to a third party (such as the first node as the initiator), the third party can only obtain the characteristic extreme values of the original data, but cannot obtain the characteristics of the original data, so the original data itself is not exposed to the third party.

[0106] For example, assuming that the second aligned data of the second node in the virtual fusion dataset is the product purchase data of a user on a shopping platform, after parsing the metadata therein, the parsed data obtained includes product purchase data in 100 rows and 6 columns, average daily purchase cost, average monthly purchase cost, etc.

[0107] The metadata parsing tasks of the first node and the second node are calculated separately in each domain involved in the data, and then aggregated based on the privacy computing protocol to obtain an aggregated result. That is, the first node receives the second parsed data sent by each second node and aggregates it with the first parsed data to obtain an aggregated result. These metadata of the virtual fusion dataset do not expose the original data themselves, but they have important guiding significance for the subsequent selection and arrangement of data mining operators. Among them, after the data in the virtual fusion dataset is subjected to metadata parsing, it can also include parsed data after the first parsed data and the second parsed data are aggregated, including global results obtained by calculating the metadata, such as the global mean, global variance, and global characteristic extreme values.

[0108] In this embodiment of the present application, a first node can obtain parsed data from multiple nodes in a secure multi-party privacy-preserving computing system. The parsed data for each node is obtained by parsing the original data of each node. The first node can be the initiator of the privacy-preserving computing and thus act as a service server.

[0109] Step S205: The first node configures operator parameters for the target user based on the target user's label data, the first parsed data, and the second parsed data, and constructs a data mining workflow based on the operator parameters.

[0110] When configuring operator parameters for target users, field attributes of fields may be configured for each operator, where the field attributes include one or a combination of id (user ID), distribution characteristics, and labels.

[0111] The characteristic of the id column is that the values in each row are different. It is often used for data alignment and should not be used in subsequent data preprocessing, feature engineering, model training, and other links. Feature columns are usually divided into numerical and text types, and each needs to go through different data conversion operators to be processed into usable features that can be used for model training. Generally, the label column in regression tasks is a continuous value, while the label column is generally not required in clustering tasks. By configuring the field attributes of the field for each operator and marking the feature column, label column, and id column, in the component that needs to select the feature column for calculation, the possibility of the id column and label column as user optional options can be excluded, thereby avoiding misoperation.

[0112] Among them, see Figure 6 The workflow formation process can include: connecting each operator with lines representing the data flow and task execution order to form an operator workflow, that is, a data mining workflow. At this point, executing the workflow can complete a data mining task.

[0113] For example, the feature engineering in the data mining workflow is actually a series of many feature engineering operators. The selection of operators and the order of execution are determined by the user. Generally, the effect of feature engineering can be adjusted through debugging. In the data mining workflow of the embodiment of the present application, after data alignment, in the data preprocessing and feature engineering links, when it is necessary to specify the columns required for calculation, the user can be guided again to configure the field attributes. After the process is arranged and connected, the field attributes specified by the preceding component will be globally recorded, refreshed, and inherited to the subsequent process. Therefore, in the subsequent components or modules that need to select feature columns for calculation, the possibility of id columns and label columns as user optional options can be excluded, thereby avoiding misoperation. Moreover, the global inheritance of field attributes can also ensure the consistency of the goals of preceding data preprocessing and subsequent data mining.

[0114] When constructing a data mining workflow based on operator parameters, data preprocessing and labeling of the virtual fused dataset with labels and features can be performed on the virtual fused dataset. For example, the preprocessing can include one-hot encoding, numerical normalization, missing value filling, mapping encoding, hash encoding, scaling, or discretization of the data in the virtual fused dataset. The present application can separately label the user identifier, label, and feature fields corresponding to the target user in the virtual fused dataset.

[0115] In some embodiments of the present application, considering that in the related art, due to the lack of labeling of the fields of the virtual fusion dataset, some fields are input into the task links that should not be input for privacy computing. For example, for the task link of feature engineering of privacy computing, id (user identification) is an unnecessary field, and in the task link of clustering, label is an unnecessary field, and so on. Once the fields that are not required for these task links are selected and input by the user, they will trigger meaningless but time-consuming and computationally expensive invalid tasks, thereby reducing debugging efficiency. Therefore, after obtaining the virtual fusion dataset, in order to avoid triggering meaningless but time-consuming and computationally expensive invalid tasks, the virtual fusion dataset can also be preprocessed and the user identification, label and feature fields corresponding to the target user can be marked respectively.

[0116] In the embodiments of the present application, by separately marking the field attributes of fields in the virtual fusion dataset, such as user ID, label, and feature fields, it is possible to prevent unnecessary fields from being selected by the user and entered into the task. For example, after marking the user ID field, it is only used for sample data alignment and is not entered into privacy computing tasks such as data preprocessing, feature engineering, and model training. For another example, after marking both the user ID and label fields, these two fields can be avoided from being entered into the feature engineering task, and only the feature field can be entered into the feature engineering task.

[0117] This application can also record the source of the privacy computing participants of the user identification, label and feature fields, that is, record the node from which they come, such as from the first node or the second node.

[0118] For example, when the first node configures the field attributes of each operator, it can record the node from which the field comes; if the node from which the field comes is the first node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are stored in the first node; if the node from which the field comes is the second node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are sent to the second node for storage; the first node or the second node uses the encoder stored in itself to encode the information value and evidence weight value of the field.

[0119] The technical solution of the embodiment of the present application can be applied to the application scenario of joint financial anti-fraud and avoiding multiple loans. For example, each party of the data source can provide a part of the features of the desensitized user respectively, and the calculation initiator provides the fraud label of the desensitized user. The data of each party is aligned through the unique identification ID of the desensitized user. After obtaining the virtual fusion data set, the data mining workflow is jointly debugged, and finally a joint anti-fraud model can be trained. For example, the XGBoost algorithm can be used to construct the anti-fraud model. The technical solution of the embodiment of the present application can assist in configuring the outlier filtering logic and the IV value calculation bin number by providing the distribution characteristics of the target user in the virtual fusion data set in the process of constructing the data mining task of joint anti-fraud; and by specifying the global ID (user unique identification ID), features (such as SDK (Software Development Kit, software development kit) information, consumption information, reception information, etc.) and labels (whether there is fraud) in advance to avoid configuration errors of operator parameters, thereby reducing the task error rate and improving the debugging efficiency of the anti-fraud task.

[0120] It should be noted that in the data mining workflow formed by the embodiment of the present application, that is, the process orchestration task, some by-products of data processing and visualization, that is, intermediate processing results, such as IV (Information Value), WOE (Weight Of Evidence, Weight Of Evidence) results, etc., will be saved. These intermediate processing results will be used as input again in the subsequent data processing link. For example, a by-product can be generated in a certain step of feature engineering, and will be used as input in other steps of feature engineering. The participant matching method of the embodiment of the present application can ensure that the IV value and WOE value of each field are used to encode the features from this party in the subsequent processing flow. In the background, the encoders corresponding to these intermediate processing results are also stored in the participant to which the field belongs, and are executed at the participant, thereby avoiding invalid data processing and feature engineering processing, while also ensuring the distributed processing of the data set, avoiding additional communication and data synchronization, and ensuring data security.

[0121] In the embodiments of the present application, the first and second parsed data in the virtual fused dataset may include the target user's distribution characteristics. In other words, the distribution characteristics in the fused dataset are part of the fused dataset's metadata. Furthermore, based on the target user's distribution characteristics contained in the first and second parsed data in the virtual fused dataset, outlier filtering logic and configuration information values can be configured for the target user to determine the binning results.

[0122] The process of configuring outlier filtering logic and configuring information values for the target user to determine the binning results may be implemented through steps S2061 to S2064, but is not limited thereto, as described below:

[0123] Step S2061: Based on the global variance and the global mean, configure outlier filtering logic for the target user.

[0124] In the above embodiment, global characteristic extreme values, global variance, and global mean can be obtained by summarizing and processing the first parsed data and the second parsed data. The calculation method for obtaining the global characteristic extreme values, global variance, and global mean can be implemented using the algorithm of the relevant technology, and this application is not limited thereto. This application can configure outlier filtering logic for the target user based on the global variance and the global mean. For example, it can filter out data exceeding 3sigma or 2sigma in the target user's original data based on the global variance and the global mean, and so on.

[0125] Step S2062: Based on the global feature extreme value and the preset binning ratio, count the first statistical data quantity and the second statistical data quantity.

[0126] In an embodiment of the present application, the global feature extreme values include the global feature maximum value and the global feature minimum value. Based on these global feature maximum values and global feature minimum values and a preset binning ratio, a target feature value can be calculated, and then the number of data with feature values less than the target feature value in the feature data of the original data is counted to obtain a first statistical data quantity, and the number of data with feature values greater than the target feature value in the feature data of the original data is counted to obtain a second statistical data quantity. The target feature value can be set to the product of the preset binning ratio and the sum of the global feature extreme values, and the sum of the global feature extreme values is the sum of the global feature maximum value and the global feature minimum value.

[0127] Step S2063: Determine a target quantile based on the global statistical data quantity, the first statistical data quantity, and the second statistical data quantity.

[0128] The target quantile refers to the characteristic value of the boundary between data bins. Specifically, step S2063 can be implemented by calculating the sum of the first statistical data quantities to obtain the total first statistical data, and calculating the sum of the second statistical data quantities to obtain the total second statistical data; calculating a first global data ratio based on the total first statistical data and the global statistical data quantity, and calculating a second global data ratio based on the total second statistical data and the global statistical data quantity; determining the quantile distribution position based on the first global data ratio, the second global data ratio, and the preset bin ratio; and determining the target quantile based on the quantile distribution position.

[0129] Step S2064: Based on the target quantile line, feature binning is performed on the feature data corresponding to the target user in the virtual fusion dataset to obtain a data binning result.

[0130] Specifically, the implementation of step S2064 can be: based on the target quantile line, each data characteristic value in the characteristic data of the original data is divided into a preset number of data bins to obtain each bin. For example, if the number of target quantile lines is 1, each data with a data characteristic value less than the target quantile line is taken as the first data bin, and each data with a data characteristic value not less than the target quantile line is taken as the second data bin.

[0131] The process of configuring outlier filtering logic and configuring information values for the target user to determine the binning results can also be implemented through steps S'2061 to S'2065, as described below:

[0132] Step S'2061: sorting the feature values of the virtual fusion data set corresponding to the target user to obtain a sorting result.

[0133] The target user has a characteristic value. After the characteristic value of the target user is determined, for example, the virtual fusion data set corresponding to the characteristic value of the target user can be sorted in ascending order to obtain a sorting result.

[0134] Step S'2062: Based on a preset binning strategy, the feature values in the sorting result are binned to obtain a plurality of feature bins.

[0135] The preset binning strategy here can be equal frequency binning or equal distance binning.

[0136] Step S'2063: Based on the multiple feature bins, calculate the evidence weight value of each feature bin.

[0137] The so-called Weight of Evidence (WOE) is the logarithm of the ratio of positive and negative values for a certain value of a character variable or a segment of a continuous variable. It represents a form of encoding for character variables. For example, in a risk control model, the ratio of risky customers to the ratio of normal customers is compared to measure the difference between the two. The logarithm is then taken. The greater the difference, the clearer the risky customers are distinguished.

[0138] Step S'2064: Calculate the information value of each feature of the target user based on the evidence weight value and the target user's label data.

[0139] The Weight of Evidence (WOE) is the logarithm of the ratio of positive and negative values for a given value of a character variable or a segment of a continuous variable. The Information Value (IV) can be used to measure the predictive power of a feature for a label. IV and WOE are typically used together.

[0140] Step S'2065: Based on the information value of the characteristics of each target user, user characteristics corresponding to information values greater than or equal to a preset value threshold are screened.

[0141] After filtering the information value of the features of each target user, feature selection can be performed based on the IV value (based on the information value, filtering features corresponding to information values greater than or equal to the preset value threshold to obtain a filtered feature set), and unimportant features (i.e., features corresponding to information values less than the preset value threshold) are eliminated. For example, the preset value threshold can be set to 0.02, so IV < 0.02 indicates that the corresponding feature is unimportant.

[0142] It can be seen that the embodiment of the present application filters the user features corresponding to the information value greater than or equal to the preset value threshold based on the information value of the characteristics of each target user, so that only important user features are processed. This can also reduce the amount of data mining processing, and only analyze and process the required data, thereby improving the accuracy of the data mining process.

[0143] In this embodiment of the present application, by recording the privacy computing participant source of the user identifier, label, and feature field, that is, recording which node it comes from, the first node can provide the intermediate processing results such as IV and WOE of the virtual fusion dataset to the corresponding privacy computing participant, so that the privacy computing participant can apply the intermediate processing results to the encoding process of its own feature field. Because the encoders corresponding to the intermediate processing results such as IV and WOE are also stored in the privacy computing participant to which the field belongs and are executed at the privacy computing participant, invalid data processing and feature engineering tasks can be avoided, while also ensuring distributed processing of the dataset, as well as additional communication and data synchronization, thereby ensuring data security.

[0144] On the other hand, since the privacy computing participant sources of the specified fields are recorded after the virtual fusion dataset is obtained, in one embodiment of the present application, matching the intermediate processing results of each field of the virtual fusion dataset to the feature fields provided by the privacy computing participant corresponding to the intermediate processing results for encoding can be: according to the recorded privacy computing participant sources of the specified fields, only the feature fields provided by the privacy computing participant corresponding to the feature fields whose information values and / or evidence weights exceed a preset threshold are encoded. Since the privacy computing participant sources of the fields are recorded, the information values and / or evidence weights can be used only for encoding at the corresponding privacy computing participant, thereby avoiding invalid data processing and feature engineering tasks.

[0145] From the above-mentioned example of the data mining method based on privacy computing, it can be seen that the first node aligns the first sample data provided by the first node and the second sample data provided by the second node according to the user identifier to obtain the first aligned data of the first node and the second aligned data of the second node; wherein the second sample data is the original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing; then the first node fuses the first aligned data and the second aligned data based on the privacy computing protocol to obtain a virtual fused data set; then the first node performs metadata parsing on the first aligned data to obtain first parsed data; the alignment result of the second aligned data of the second node is notified to the second node, and the second parsed data sent by the second node is received, wherein the second parsed data is obtained by the second node performing metadata parsing on the second aligned data according to the alignment result. Through the above-mentioned processing, on the one hand, the mining processing speed can be improved through distributed processing; on the other hand, since the first parsed data and the second parsed data have privacy characteristics and do not expose the original data, the virtual fused data set obtained by the first node has both privacy characteristics and can represent the characteristics of the target user. When the parameters of the operator are subsequently configured for the target user, the data characteristics of the target user analyzed by each participant in the privacy computing (including the first node and the second node) over a historical period can be understood. Therefore, by using the computing solution provided in this application, the number of debugging times for data analysis, i.e., joint data insight tasks, can be reduced while protecting the data privacy of all parties, and the data processing speed can be increased, thereby improving the overall debugging efficiency of data mining based on privacy computing.

[0146] Corresponding to the aforementioned application function implementation method embodiment, the present application also provides a secure multi-party privacy computing system, computing device, server and corresponding embodiments.

[0147] See also Figure 7 , is a schematic diagram of the structure of a secure multi-party privacy computing system shown in an embodiment of the present application. For ease of explanation, only the parts related to the embodiment of the present application are shown. Figure 7 The example secure multi-party privacy computing system 70 includes multiple nodes participating in privacy computing, including a first node 71 and a second node 72:

[0148] The first node 71 is used to send a first message to the second node 72, where the first message carries the user identifier of the target user and is used to request to initiate privacy computing for the target user; the first sample data provided by the first node 71 and the second sample data provided by the second node 72 are aligned according to the user identifier to obtain first aligned data of the first node 71 and second aligned data of the second node 72; the second sample data is the original data of the target user obtained by the second node 72 according to the user identifier and provided to the first node 71 after setting processing; the first aligned data and the second aligned data of the second node 72 are fused based on the privacy computing protocol to obtain a virtual fused data set; the first aligned data are meta-information parsed to obtain first parsed data; the alignment result of the second aligned data of the second node 72 is notified to the second node 72, and the second parsed data sent by the second node 72 is received, wherein the second parsed data is the meta-information parsed by the second node 72 on the second aligned data according to the alignment result; the parameters of the operator are configured for the target user based on the label data of the target user, the first parsed data and the second parsed data of the second node 72, and a data mining workflow is constructed based on the parameters of the operator.

[0149] For example, after the first node 71 determines the virtual fusion data set, it can parse the type, number of rows, number of columns, mean, variance, characteristic extreme values, quantiles, missing rate, and number of unique values of the first aligned data of the first node 71 in the virtual fusion data set in its own domain to obtain first parsed data. It should be noted that since the first node 71 parses the metadata of the first aligned data in its own domain, it complies with the principle that the data provided by each participant in privacy computing does not go beyond their respective domains. Therefore, even if the first parsed data obtained by parsing the first aligned data is provided to a third party (such as the second node as a participant), the third party can only obtain the characteristic extreme values of the original data, but cannot obtain the characteristics of the original data, so the original data itself will not be exposed to the third party.

[0150] The second node 72 is used to receive the first message, obtain the original data of the target user according to the user identifier, perform setting processing to obtain second sample data, and then provide it to the first node 71, perform metadata parsing on the second aligned data according to the alignment result to obtain second parsed data and send it to the first node 71.

[0151] When configuring the parameters of the operator for the target user, the first node 71 configures the field attributes of the field for each operator, wherein the field attributes include one or a combination of user identification, features and labels. In an embodiment of the present application, by marking the field attributes such as user identification, labels and features respectively, it is possible to avoid fields that are not needed in the task link being selected by the user and input into the task link. For example, after the user identification field is marked, it is only used for sample data alignment, and it will not be input into the task links of privacy calculation such as data preprocessing, feature engineering and model training; for another example, after marking the user identification and label fields, it is possible to avoid inputting these two fields into the task link of feature engineering, and only input the feature field into the task link of feature engineering. When configuring the field attributes of the field for each operator, the first node 71 records the node from which the field comes;

[0152] If the node from which the data is received is the first node 71, after determining the information value and the weight of evidence of the field, the information value and the weight of evidence of the field are stored in the first node; if the node from which the data is received is the second node, after determining the information value and the weight of evidence of the field, the information value and the weight of evidence of the field are sent to the second node for storage;

[0153] The first node 71 or the second node 72 uses the encoder stored in itself to encode the information value and the evidence weight value of the field.

[0154] In this embodiment of the present application, by recording the privacy computing participant source of the user identifier, label, and feature field, that is, recording which node it comes from, the first node can provide the intermediate processing results such as IV and WOE of the virtual fusion dataset to the corresponding privacy computing participant, so that the privacy computing participant can apply the intermediate processing results to the encoding process of its own feature field. Because the encoders corresponding to the intermediate processing results such as IV and WOE are also stored in the privacy computing participant to which the field belongs and are executed at the privacy computing participant, invalid data processing and feature engineering tasks can be avoided, while also ensuring distributed processing of the dataset, as well as additional communication and data synchronization, thereby ensuring data security.

[0155] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated again here.

[0156] Figure 8 It is a structural diagram of a computing device shown in an embodiment of the present application.

[0157] See also Figure 8 , the computing device 400 includes a memory 410 and a processor 420 .

[0158] The processor 420 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0159] Memory 410 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage. ROM may store static data or instructions required by processor 420 or other modules of the computer. Permanent storage may be a readable and writable storage device. A permanent storage device may be a non-volatile storage device that retains stored instructions and data even when the computer is powered off. In some embodiments, the permanent storage device uses a large-capacity storage device (e.g., a magnetic or optical disk, flash memory) as the permanent storage device. In other embodiments, the permanent storage device may be a removable storage device (e.g., a floppy disk, optical drive). System memory may be a readable and writable storage device or a volatile readable and writable storage device, such as dynamic random access memory. System memory may store some or all instructions and data required by the processor during operation. In addition, memory 410 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and magnetic disks and / or optical disks may also be used. In some embodiments, the memory 410 may include a readable and / or writable removable storage device, such as a compact disc (CD), a read-only digital versatile disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not include carrier waves and transient electronic signals transmitted wirelessly or wired.

[0160] The memory 410 stores executable codes. When the executable codes are processed by the processor 420 , the processor 420 may execute part or all of the above-mentioned methods.

[0161] Figure 9 This is a schematic diagram of a server structure provided in an embodiment of the present application.

[0162] The server 100 may vary greatly due to different configurations or performance, and may include one or more central processing units (CPUs) 1022 (e.g., one or more processors) and memory 1032, and one or more storage media 1030 (e.g., one or more mass storage devices) storing application programs 1042 or data 1044. The memory 1032 and storage medium 1030 may be either short-term storage or persistent storage. The program stored in the storage medium 1030 may include one or more modules (not shown), each module of which may include a series of instruction operations on the server. Furthermore, the central processing unit 1022 may be configured to communicate with the storage medium 1030 to execute a series of instruction operations in the storage medium 1030 on the server 100.

[0163] The server 100 may also include one or more power supplies 1026, one or more wired or wireless network interfaces 1050, one or more input and output interfaces 1058, and / or one or more operating systems 1041, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc.

[0164] In the above embodiment, the steps performed by the node, such as the first node, can be based on the Figure 9 The structure of the server 100 is shown. For example, the processor 1022 performs the following operations by calling the instructions in the memory 1032:

[0165] Sending a first message to the second node through the input / output interface 1058 , where the first message carries the user identifier of the target user and is used to request initiation of privacy computing for the target user;

[0166] The processor 1022 performs data alignment on the first sample data provided by the first node and the second sample data provided by the second node according to the user identifier, thereby obtaining first aligned data of the first node and second aligned data of the second node; wherein the second sample data is the original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing;

[0167] The processor 1022 fuses the first aligned data and the second aligned data based on the privacy computing protocol to obtain a virtual fused data set;

[0168] The processor 1022 performs metadata parsing on the first aligned data to obtain first parsed data; notifies the second node of an alignment result of the second aligned data of the second node, and receives the second parsed data sent by the second node, wherein the second parsed data is obtained by the second node performing metadata parsing on the second aligned data according to the alignment result;

[0169] The processor 1022 configures operator parameters for the target user based on the target user's tag data, the first parsed data, and the second parsed data, and constructs a data mining workflow based on the operator parameters.

[0170] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0171] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0172] In the several embodiments provided in the embodiments of the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0173] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of the present embodiment according to actual needs.

[0174] In addition, the functional modules in each embodiment of the present application can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into a module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0175] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0176] The computer program product includes one or more computer instructions. When the computer program is loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a storage medium, or transmitted from one storage medium to another storage medium. For example, the computer instructions can be transmitted from a website, a computer, a server, or a data center by wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode to another website, a computer, a server, or a data center. The storage medium can be any available medium that a computer can store or a data storage device such as a server or a data center that includes one or more available media integrations. The available medium can be a magnetic medium, (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid-state drive Solid State Disk (SSD)), etc.

[0177] In addition, the method according to the present application may also be implemented as a computer program or a computer program product, which includes computer program code instructions for executing some or all of the steps in the above method of the present application.

[0178] Alternatively, the present application can also be implemented as a computer-readable storage medium (or non-transitory machine-readable storage medium or machine-readable storage medium) on which executable code (or computer program or computer instruction code) is stored. When the executable code (or computer program or computer instruction code) is executed by a processor of a computing device (or server, etc.), the processor executes part or all of the steps of the above-mentioned method according to the present application.

[0179] The embodiments of the present application have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to the technology in the market, or to enable other persons skilled in the art to understand the embodiments disclosed herein.

Claims

1. A data mining method based on privacy computing, wherein the method is applied to a secure multi-party privacy computing system, wherein the secure multi-party privacy computing system includes multiple nodes participating in privacy computing, wherein the multiple nodes include a first node and a second node; characterized in that: The method comprises: The first node sends a first message to the second node, where the first message carries a user identifier of a target user and is used to request initiation of privacy computing for the target user; The first node aligns first sample data provided by the first node and second sample data provided by the second node according to the user identifier to obtain first aligned data of the first node and second aligned data of the second node; wherein the second sample data is original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing; The first node fuses the first aligned data and the second aligned data based on a privacy-preserving computing protocol to obtain a virtual fused data set; The first node performs metadata parsing on the first aligned data in the virtual fused data set to obtain first parsed data; notifies the second node of an alignment result of the second aligned data obtained by the second node, and receives second parsed data sent by the second node, wherein the second parsed data is obtained by the second node performing metadata parsing on the second aligned data in the virtual fused data set according to the alignment result; The first node configures operator parameters for the target user based on the target user's label data, the first parsed data, and the second parsed data, and constructs a data mining workflow based on the operator parameters.

2. The method according to claim 1, characterized in that The first parsed data and the second parsed data contain distribution characteristics of target users, and the method further includes: According to the distribution characteristics of the target user included in the first parsed data and the second parsed data, an outlier filtering logic and a configuration information value are configured for the target user to determine a binning result.

3. The method according to claim 2, characterized in that The method further comprises: Processing the first analytical data and the second analytical data to obtain global characteristic extreme values, global variance, and global mean; The configuring, based on the distribution characteristics of the target user included in the first parsed data and the second parsed data, outlier filtering logic and configuration information values for the target user to determine a binning result includes: Based on the global variance and the global mean, configuring outlier filtering logic for the target user; Based on the global feature extreme value and the preset binning ratio, counting the first statistical data quantity and the second statistical data quantity; determining a target quantile based on a global statistical data quantity, the first statistical data quantity, and the second statistical data quantity; Based on the target quantile line, feature binning is performed on the feature data corresponding to the target user in the virtual fusion dataset to obtain a data binning result.

4. The method according to claim 2, characterized in that The configuring, based on the distribution characteristics of the target user included in the first parsed data and the second parsed data, outlier filtering logic and configuration information values for the target user to determine a binning result includes: Sorting the feature values of the virtual fusion data set corresponding to the target user to obtain a sorting result; Based on the preset binning strategy, the feature values in the sorting results are binned to obtain multiple feature bins; Based on multiple feature bins, calculate the evidence weight value of each feature bin; Calculating the information value of each feature of the target user based on the evidence weight value and the label data of the target user; Based on the information value of the feature of each target user, user features corresponding to information values greater than or equal to a preset value threshold are screened.

5. The method according to claim 1, wherein: When configuring the parameters of the operator for the target user, field attributes of the field are configured for each operator, wherein the field attributes include one or a combination of user identification, feature, and label.

6. The method according to claim 5, characterized in that: When constructing a data mining workflow based on the parameters of the operator, data preprocessing is performed on the virtual fusion dataset, and labels and features are marked on the virtual fusion dataset.

7. The method according to claim 5, characterized in that The method further comprises: When configuring the field attributes of a field for each operator, the first node records the node from which the field comes; If the source node is the first node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are stored on the first node; if the source node is the second node, after determining the information value and evidence weight value of the field, the information value and evidence weight value of the field are sent to the second node for storage; The first node or the second node uses the encoder stored in itself to encode the information value and the evidence weight value of the field.

8. A secure multi-party privacy-preserving computing system, comprising a plurality of nodes participating in privacy computing, the plurality of nodes including a first node and a second node, characterized in that: The first node is configured to send a first message to the second node, the first message carrying a user identifier of a target user and requesting to initiate privacy computing for the target user; aligning first sample data provided by the first node and second sample data provided by the second node according to the user identifier to obtain first aligned data of the first node and second aligned data of the second node; The second sample data is the original data of the target user obtained by the second node according to the user identifier and provided to the first node after setting processing; fusing the first aligned data and the second aligned data of the second node based on a privacy-preserving computing protocol to obtain a virtual fused data set; performing metadata parsing on the first aligned data in the virtual fusion data set to obtain first parsed data; notifying the second node of an alignment result of the second aligned data of the second node, and receiving second parsed data sent by the second node, wherein the second parsed data is obtained by the second node performing meta-information parsing on the second aligned data in the virtual fused dataset according to the alignment result; Based on the target user's tag data, the first parsed data, and the second parsed data of the second node, configuring operator parameters for the target user, and building a data mining workflow based on the operator parameters; The second node is used to receive the first message, obtain the original data of the target user according to the user identifier, perform setting processing to obtain second sample data, and then provide it to the first node; perform metadata parsing on the second aligned data according to the alignment result to obtain second parsed data and send it to the first node.

9. The system according to claim 8, characterized in that: When configuring the parameters of the operator for the target user, the first node configures field attributes of each operator, where the field attributes include one or a combination of user identification, features, and tags.

10. The system according to claim 9, characterized in that: When configuring the field attributes of a field for each operator, the first node records the node from which the field comes; If the node from which the source is obtained is the first node, after determining the information value and the weight of evidence value of the field, the information value and the weight of evidence value of the field are stored in the first node; If the node from which the message is received is a second node, after determining the information value and the weight of evidence value of the field, the information value and the weight of evidence value of the field are sent to the second node for storage; The first node or the second node uses the encoder stored in itself to encode the information value and the evidence weight value of the field.

11. The system according to claim 9, wherein: When constructing a data mining workflow based on the parameters of the operator, the first node performs data preprocessing on the virtual fusion dataset and marks labels and features on the virtual fusion dataset.

12. A computing device, characterized in that include: processor; as well as A memory having executable codes stored thereon, which, when executed by the processor, causes the processor to perform the method according to any one of claims 1 to 7.

13. A computer-readable storage medium having executable codes stored thereon, which, when executed by a processor of a computing device, causes the processor to perform the method according to any one of claims 1 to 7.

Images