User group classification method and device, electronic equipment, storage medium and product

Abstract

The invention discloses a user group classification method and device, electronic equipment, a storage medium and a product, and relates to the field of artificial intelligence and the field of privacy computing. The method comprises the following steps: under the condition that authorization of a target user is obtained, obtaining feature data of the target user; feature encryption processing is carried out on the feature data of the target user to obtain an encryption result, and the feature encryption processing is used for encrypting the feature data in a data segmentation and confusion mode; and determining a target group type to which the target user belongs according to an encryption result. According to the method and the device, the technical problem of poor data security caused by information transmission among a plurality of data sources in the process of determining the user group type in the prior art is solved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and privacy computing, and specifically, to a method, device, electronic device, storage medium, and product for classifying user groups. Background Art

[0002] In customer classification management, companies need to identify user group types to optimize service strategies. Traditional methods rely on a single data source or limited data sharing, but there are significant data security vulnerabilities in the information transmission process involving multiple data sources.

[0003] Under the current technical system, the process of determining the user group type involves information transmission between multiple data sources. This cross-organizational data exchange information transmission is vulnerable to network attacks and does not comply with strict data protection regulations, leading to technical problems that increase the risk of customer privacy exposure.

[0004] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0005] The embodiments of the present application provide a method, device, electronic device, storage medium, and product for classifying user groups to at least solve the technical problem of poor data security caused by information transmission between multiple data sources in the process of determining the user group type in the prior art.

[0006] According to one aspect of an embodiment of the present application, a method for classifying user groups is provided, including: obtaining feature data of a target user with authorization from the target user; performing feature encryption processing on the feature data of the target user to obtain an encryption result, wherein the feature encryption processing is used to encrypt the feature data by data segmentation and obfuscation fusion; and determining the target group type to which the target user belongs based on the encryption result.

[0007] Optionally, determining the target group type to which the target user belongs based on the encryption result includes: inputting the encryption result into a target model, decrypting the encryption result through the target model, and determining the probability that the target user belongs to each of N group types based on the decryption result, wherein the target model is a model jointly trained by M data sources, and the transmitted data between the target model and each data source is in an encrypted state, and M and N are both integers greater than 1; based on the probability that the target user belongs to each of the N group types, the group type to which the target user has the highest probability of belonging is taken as the target group type.

[0008] Optionally, the target model training step includes the following steps:

[0009] Obtain user feature data provided by M data sources; divide the user feature data provided by each data source into X data slices, where X is an integer greater than 1; summarize the X data slices corresponding to each data source into a data set corresponding to the data source, thereby obtaining M data sets corresponding to the M data sources; perform obfuscation encryption on the data slices in each data set to obtain an obfuscated encryption result for each data set; perform data fusion on the obfuscated encryption results of the M data sets to obtain a data fusion result; and train a neural network deployed on a central server based on the data fusion result to obtain a target model, wherein the central server and the M data sources remain independent of each other.

[0010] Optionally, the neural network deployed on the central server is trained according to the data fusion result to obtain a target model, including: obtaining global clock information and metadata identifiers between M data sources; generating a key according to the global clock information and metadata identifiers; decrypting the data fusion result according to the key, and obfuscating the decrypted data; re-encrypting the obfuscated data to generate a common data set between the M data sources; and training the neural network deployed on the central server according to the common data set to obtain the target model.

[0011] Optionally, a neural network deployed on a central server is trained according to a public data set to obtain a target model, including: copying the neural network on the central server to the memory of each data source through a controller; for each data source, randomly selecting part of the data from the public data set, combined with the private data within each data source, as the training data of each data source; after locking the model training chip through the controller, transmitting the training data of each data source to the model training chip for training; obtaining the model training results of each data source from the model training chip through the controller, and performing an aggregation operation on the model training results of each data source; encrypting and storing the aggregated data obtained after the aggregation operation in the shared memory of M data sources; and training the target model based on the stored data in the shared memory.

[0012] Optionally, a target model is obtained by training based on the stored data in the shared memory, including: updating the neural network in the central server based on the stored data in the shared memory; encrypting the updated neural network in the central server according to the global clock information and metadata identifier, and distributing it to M data sources; training the received neural network through the M data sources until the neural network is in a convergence state, thereby obtaining the target model.

[0013] Optionally, in the process of training the neural network deployed on the central server according to the data fusion results to obtain the target model, a task operation is triggered at a fixed time, wherein the task operation is used to generate M key shards, and encrypt the current latest training result data based on the M key shards; the M key shards are distributed to M data sources, wherein different data sources store different key shards.

[0014] Optionally, if an abnormality is detected in the training process of the target model at the target moment, the task operation triggered most recently at the target moment is determined as the target task operation; based on the M key shards generated by the target task operation, the training result data encrypted when executing the target task operation is decrypted, and the target model is retrained based on the decrypted training result data.

[0015] According to another aspect of the present application, a user group classification device is also provided, which includes: an acquisition unit for acquiring feature data of a target user with authorization from the target user; an encryption processing unit for performing feature encryption processing on the feature data of the target user to obtain an encryption result, wherein the feature encryption processing is used to encrypt the feature data by data segmentation and confusion fusion; and a determination unit for determining the target group type to which the target user belongs based on the encryption result.

[0016] According to another aspect of the present application, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium. When the computer program runs, the device where the computer-readable storage medium is located executes the above-mentioned method for classifying user groups.

[0017] According to another aspect of the present application, an electronic device is also provided, wherein the electronic device includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors execute the above-mentioned method for classifying user groups.

[0018] According to another aspect of the present application, a computer program product is further provided, wherein the computer program product includes a computer program or instructions, and the computer program or instructions implement the above-mentioned method for classifying user groups when executed by a processor.

[0019] In this application, the target user's feature data is first obtained with the target user's authorization. Feature encryption is then performed on the target user's feature data to obtain an encrypted result. Feature encryption is used to encrypt the feature data through data segmentation and obfuscation fusion. Finally, the target group type to which the target user belongs is determined based on the encryption result.

[0020] As can be seen from the above, according to the technical solution of this application, through feature encryption processing, data segmentation and obfuscation fusion technology, even if the data is intercepted during transmission, the specific user information cannot be deciphered, thereby greatly improving data security. Ultimately, this application can accurately determine the target group type to which the user belongs without sacrificing user privacy. This method not only improves the efficiency of data processing, but also maintains data security, providing strong support for enterprise customer management and risk control.

[0021] It can be seen that the technical solution of the present application effectively improves the security issues during multi-source data fusion by implementing encryption and privacy protection measures in various links of data collection, transmission, processing and storage, and solves the technical problem of poor data security caused by information transmission between multiple data sources in the process of determining user group types in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0023] Figure 1 is a flowchart of an optional method for classifying user groups according to an embodiment of the present application;

[0024] Figure 2 is a flow chart of training an optional target model according to an embodiment of the present application;

[0025] Figure 3 This is an optional model training flow chart based on a public training set according to an embodiment of the present application;

[0026] Figure 4 This is a schematic diagram of an optional device for classifying user groups according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] It should also be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) collected by this application are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data comply with the relevant laws, regulations and standards of the relevant regions, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse. For example, an interface is set up between this system and relevant users or institutions. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or institution through the interface, and obtain relevant information after receiving the consent information fed back by the aforementioned user or institution.

[0030] According to an embodiment of the present application, an embodiment of a method for classifying user groups is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0031] Optionally, according to an embodiment of the present application, a user group classification system (hereinafter referred to as the classification system) is provided as the execution subject of the user group classification method of the embodiment of the present application, wherein the system can be a software system or an embedded system combining software and hardware. Of course, the method execution subject in the embodiment of the present application can also be other forms of execution subjects, such as devices, equipment, etc. Those skilled in the art should know that this application does not specifically limit the specific form of expression of the method execution subject.

[0032] Figure 1 is a flow chart of an optional method for classifying user groups according to an embodiment of the present application, such as Figure 1 As shown, the method includes the following steps:

[0033] Step S101: Acquire characteristic data of the target user with authorization from the target user.

[0034] Optionally, feature data may include, but is not limited to, a user's spending habits, credit score, transaction history, and social network behavior. This data, taken together, can depict a user's behavioral patterns and identity attributes, which is crucial for determining the target user's user group. However, this data also contains a significant amount of personal privacy information and must be obtained through legitimate means and with the user's knowledge and consent.

[0035] User authorization can be obtained through a variety of means, such as electronic signatures, online consent forms, or by selecting a checkbox within an application. Importantly, the authorization process must be clear and concise, ensuring that the intended user fully understands how their data will be used and their rights in this process, such as the ability to review, modify, and withdraw authorization. Once authorization is obtained from the intended user, the business or organization may legally collect the required profile data in accordance with the pre-agreed terms and conditions. At this stage, data collection should be minimized to the extent possible, collecting only the information necessary to complete a specific function or service, and avoiding excessive collection that could lead to unnecessary privacy risks.

[0036] It is important to reiterate that obtaining explicit authorization from the target user includes informing the target user of the purpose of data collection, how it will be used, and how their privacy will be protected. The target user is also clearly informed of the privacy policy, including policies for data collection, use, storage, and deletion, as well as how the target user can exercise their rights, such as accessing, correcting, or deleting their data. The target user has the right to choose whether to consent to the collection and use of their data. Furthermore, when collecting user data from the target user, the system will use encryption technology to ensure data security during transmission and storage. Data use will strictly comply with relevant laws and regulations and the scope of the target user's authorization. The target user's user data will be stored in a secure database and undergo regular security audits to ensure data security and the effectiveness of privacy protection measures. When the target user's user data is no longer needed, the system will securely delete the data according to the target user's request and legal requirements to ensure that the data cannot be recovered.

[0037] Step S102: Perform feature encryption processing on the feature data of the target user to obtain an encryption result.

[0038] In step S102, feature encryption processing is used to encrypt feature data by means of data segmentation and confusion fusion.

[0039] Optionally, data segmentation means breaking down the complete user feature data into multiple parts or fragments, each of which contains only a portion of the complete information. This segmentation method is extremely common in privacy computing, as it ensures that even if the data is intercepted during transmission or storage, the attacker cannot directly obtain the complete user information. Each data slice is independent, and the original information cannot be reconstructed without additional keys or context. Based on data segmentation, obfuscation fusion is a technical means to enhance data security. It mixes data fragments with other data (which can be noise data or slices from different data sources) so that the data fragments lose their clear directionality after fusion. Obfuscation fusion can be achieved through a variety of mathematical or statistical methods, such as homomorphic encryption, secret sharing, or differential privacy, to ensure that even if the data is fused, it is difficult for attackers to identify the original user features.

[0040] Optionally, after data segmentation and obfuscation fusion, the original user feature data is converted into an encrypted result. This encrypted result is a key indicator of data security, ensuring data confidentiality throughout the entire processing process. The encrypted result can be used for subsequent analysis and modeling without compromising data privacy, thereby achieving the goal of both protecting user privacy and realizing data value mining. Compared to traditional encryption technologies, feature encryption processing that combines data segmentation with obfuscation fusion can more effectively address privacy and security challenges in data sharing. It not only reduces the risk of data leakage but also establishes a collaborative rather than competitive relationship between data processors, promotes cross-domain data integration, and helps establish a more accurate and comprehensive user group classification system.

[0041] Step S103: determining the target group type to which the target user belongs according to the encryption result.

[0042] Alternatively, while the encryption results mask the specific details of the original data, carefully designed privacy-preserving computing techniques, such as multi-party secure computation, homomorphic encryption, and secure aggregation, still allow data analysis and modeling without revealing individual data. This means that, despite the obscured encryption results, they can still be used to infer and analyze the characteristic patterns and behavioral trends of user groups.

[0043] Optionally, group type refers to dividing users into different categories based on certain standards or characteristics, such as "first-category user group" represents high-value customers, and "second-category user group" represents potential high-risk customers. This application analyzes the encryption results and uses machine learning or statistical methods to identify patterns and correlations in the encrypted data, thereby determining whether the user has the characteristics of the first or second-category user group. For example, the encryption results may include some encrypted form of scores or indicators, which can be set through thresholds to distinguish whether the user should be included in a particular group.

[0044] In an optional embodiment, determining the target group type to which the target user belongs based on the encryption result includes: the classification system first inputs the encryption result into a target model, decrypts the encryption result using the target model, and determines the probability that the target user belongs to each of N group types based on the decrypted result, wherein the target model is a model jointly trained by M data sources, and the data transmitted between the target model and each data source is encrypted, and M and N are both integers greater than 1. Then, based on the probability that the target user belongs to each of the N group types, the classification system selects the group type to which the target user has the highest probability of belonging as the target group type.

[0045] Optionally, the classification system first receives previously encrypted user feature data, i.e., encryption results. These encryption results contain an encrypted version of the user data, ensuring data security during transmission and model processing.

[0046] Optionally, the target model is an advanced model trained with M different data sources. It utilizes a specialized decryption algorithm that allows it to process and analyze data directly in its encrypted state, without converting it to plaintext. This means that even while the model is running, the data remains securely encrypted, further enhancing privacy. The model analyzes the encrypted results and outputs a probability distribution for each of N predefined group types.

[0047] Optionally, the probability distribution output by the target model reflects the likelihood that the target user belongs to different group types. N group types can be predefined based on different business needs and scenarios. By analyzing these probabilities, the classification system can accurately estimate which group type the target user is most likely to belong to. Based on the probabilities output by the model, the classification system automatically selects the group type with the highest probability as the target group type to which the target user belongs. This automated and efficient process can quickly provide group classification for a large number of users while ensuring classification accuracy.

[0048] Furthermore, because the target model is trained jointly by M data sources, it can integrate data insights from multiple sources to build a more comprehensive and accurate user group classification model. This multi-source data integration not only enhances the model's generalization capabilities but also prevents data silos and promotes cross-organizational data collaboration and knowledge sharing, all while maintaining encryption and privacy protection.

[0049] It's important to note that during model training and prediction, all data transmission is encrypted, ensuring data security during transmission. This includes updating model parameters, sharing intermediate results, and delivering final results. Encrypted transmission is a key component of privacy protection, ensuring that data cannot be eavesdropped or cracked by third parties even as it flows across the network.

[0050] In summary, the classification system demonstrates advanced and privacy-friendly user group identification technology by receiving encrypted results, decrypting and analyzing them using a multi-party trained model, determining the probability distribution of group types, and ultimately automatically selecting the group type with the highest probability as the user classification. This series of steps not only demonstrates the complexity and efficiency of the technology, but also highlights the importance of protecting user privacy in data-driven decision-making.

[0051] In an optional embodiment, Figure 2 This is a training flow chart of an optional target model according to an embodiment of the present application, such as Figure 2 As shown, the training steps of the target model include the following steps:

[0052] Step S201: Obtain user feature data provided by M data sources.

[0053] Step S202: Divide the user feature data provided by each data source into X data slices, where X is an integer greater than 1.

[0054] Step S203 : summarize the X data slices corresponding to each data source into a data set corresponding to the data source, and obtain M data sets corresponding to the M data sources.

[0055] Step S204: Obfuscate and encrypt the data slices in each data set to obtain an obfuscation encryption result for each data set.

[0056] Step S205 , performing data fusion on the obfuscated encryption results of the M data sets to obtain a data fusion result.

[0057] Step S206 , training the neural network deployed on the central server according to the data fusion result to obtain a target model, wherein the central server and the M data sources remain independent of each other.

[0058] Optionally, the classification system can first obtain user feature data from M different data sources (such as banks, telecom operators, e-commerce platforms, etc.). These data sources each contain some key information about the user, but do not share the original data with each other. To enhance data protection, the user feature data provided by each data source is divided into X data slices (X>1). This process ensures that a single data slice cannot reconstruct a complete user profile. Even if a data slice is intercepted during transmission or processing, it cannot directly identify the user information.

[0059] Optionally, X data slices from each data source are combined into one dataset. This allows the classification system to generate M datasets corresponding to M data sources, each containing all slices of user feature data provided by that data source. The classification system then performs obfuscation encryption on the data slices in each dataset. This is a technique for enhancing data security by adding random noise or other obfuscation mechanisms to make the encrypted data more difficult to reverse engineer. The obfuscated data slices form a new obfuscated encryption result that preserves the statistical characteristics of the data but hides specific user information. Finally, the classification system can fuse the obfuscated encryption results from the M data sources. In the context of federated learning, this is typically achieved through a distributed training algorithm, in which only encrypted gradients or model parameters, rather than raw data, are exchanged between the central server and the data sources. The result of data fusion is a summarized, encrypted state that can be used for model training without exposing user privacy.

[0060] Optionally, based on the data fusion results, the neural network model on the central server begins training. Due to the independence of the model and the data source, model training can be conducted securely without leaking sensitive data from the data source. After multiple rounds of training and optimization, the resulting target model is able to predict the user's group type based on their characteristic data without requiring access to the original data source. The greatest advantage of this process is that it allows M data sources to participate in model training without directly sharing or exposing user data, thereby fully leveraging the unique data value of each data source and increasing the diversity and accuracy of the model. At the same time, privacy protection measures such as data segmentation, obfuscation encryption, and data fusion ensure the security and privacy of user data.

[0061] It should be noted that model training data is also data authorized for use by users. Moreover, during the use of model training data, sensitive data needs to be desensitized and subject to audit by third-party auditing agencies at any time to ensure the compliance of model training data.

[0062] In an optional embodiment, a neural network deployed on a central server is trained based on the data fusion results to obtain a target model. This includes: first, a classification system obtains global clock information and metadata identifiers between M data sources, and then generates a key based on the global clock information and metadata identifiers. Subsequently, the classification system can decrypt the data fusion results using the key and obfuscate the decrypted data. The classification system can also re-encrypt the obfuscated data to generate a common dataset between the M data sources. Finally, the classification system trains the neural network deployed on the central server based on the common dataset to obtain the target model.

[0063] Optionally, global clock information is crucial for synchronizing the operations of various data sources, ensuring time consistency for data processing and model training. Metadata identifiers are used to distinguish and identify data slices provided by each data source, which is crucial for tracking data flows and correctly merging data. These two pieces of information are used to generate a secure key for subsequent data decryption, obfuscation, and re-encryption processes. The classification system can decrypt the data fusion results based on the key generated by the global clock information and metadata identifiers. The decryption process allows the classification system to access the obfuscated and encrypted data, but this access is carried out under strict control to ensure data privacy.

[0064] Optionally, the decrypted data needs to be further obfuscated to increase its anonymity and protect user privacy. The obfuscated data is then encrypted again to generate a public dataset. This public dataset includes contributions from all data sources but does not expose sensitive information about individual users in any form, ensuring data security during training. The public dataset is then used to train a neural network model deployed on a central server. The training process can be performed in an encrypted state. Leveraging the distributed architecture of federated learning, the central server can coordinate the computing power of various data sources to jointly optimize model parameters without directly accessing the raw data from any source. This allows efficient data analysis and pattern recognition through data fusion and model training, even if the central server and data sources remain independent.

[0065] After multiple rounds of iterative training, the resulting target model can accurately predict the user's group type based on their characteristic data. This model not only leverages the combined power of M data sources but also ensures the privacy of the training data through full encryption and obfuscation techniques, making it a powerful and secure classification tool.

[0066] In an optional embodiment, Figure 3 This is an optional model training flow chart based on a public training set according to an embodiment of the present application, such as Figure 3 As shown, the following steps are included:

[0067] Step S301: The neural network on the central server is copied to the memory of each data source through the controller.

[0068] In step S302 , for each data source, a portion of data is randomly selected from the public data set and combined with the private data within each data source to serve as training data for each data source.

[0069] Step S303: After locking the model training chip through the controller, the training data of each data source is transmitted to the model training chip for training.

[0070] In step S304 , the controller obtains the model training result of each data source from the model training chip, and performs an aggregation operation on the model training result of each data source.

[0071] Step S305: encrypt and store the aggregated data obtained after the aggregation operation in the shared memory of the M data sources.

[0072] Step S306: training the target model based on the data stored in the shared memory.

[0073] Optionally, first, the controller in the classification system copies the neural network model initialized on the central server to the memory of each data source. This step ensures that all data sources have the same model starting point, which facilitates distributed training. Secondly, for each data source, the classification system randomly extracts a portion of data from the public data set. This portion of data is obfuscated and encrypted to ensure that data sharing between data sources does not leak user privacy. In addition, each data source will combine its own private data (that is, data that is unique to the source and not shared) to form its own training data set. Such a data preparation strategy maximizes the diversity and richness of model training while protecting sensitive information.

[0074] The controller then locks specific hardware resources, such as a GPU or dedicated AI accelerator, for model training. The training data from each data source is then transferred to the model training chip, initiating the local model training process. Data transmission is also encrypted to prevent unauthorized third-party access during transmission. The model training chip processes the data in an encrypted environment, trains the local model, and obtains model training results. These results typically include updated model parameters or gradient information. The controller then retrieves these model training results, i.e., the local model update information for each data source, from the model training chip.

[0075] Optionally, the controller aggregates the acquired model training results, a process that aims to combine the contributions of different data sources to form a global model update. Aggregation typically uses a weighted average or a consensus algorithm to ensure the validity and accuracy of the model update. The aggregated data, known as the global model update, is encrypted and stored in shared memory between the M data sources. Shared memory is a mechanism that allows multiple data sources to exchange information in a secure and trusted environment. Encrypted storage ensures data security even if shared memory is accessed without authorization. Based on the encrypted aggregated data stored in shared memory, the neural network model on the central server is updated and optimized. This process is repeated in multiple rounds until the model converges and reaches the desired performance level.

[0076] It should be noted that distributed model training coordinated by the controller not only improves the accuracy and generalization ability of the model, but also greatly enhances the privacy protection of user data through data encryption, obfuscation and the use of shared memory.

[0077] In an optional embodiment, training a target model based on data stored in a shared memory includes: the classification system first updates a neural network in a central server based on the data stored in the shared memory; then, based on global clock information and metadata identifiers, encrypts the updated neural network in the central server and distributes it to M data sources; and finally, the M data sources train the received neural networks until the neural networks converge, thereby obtaining the target model.

[0078] Optionally, the classification system extracts encrypted aggregated data from the shared memory between the M data sources. This data contains model gradients or parameter updates refined after multiple rounds of distributed computation. Based on this stored data, the classification system updates the neural network on the central server. This operation essentially optimizes the model parameters to improve the model's predictive accuracy. A global clock synchronizes the training process across all participants, ensuring that each model update and distribution occurs within a unified timeframe, preventing performance degradation caused by desynchronization. Metadata identification facilitates the correct identification and tracking of different model versions and their sources, which is crucial for maintaining model consistency and the continuity of the training process. The updated neural network on the central server is re-encrypted based on the global clock information and metadata identification. This step ensures that even as the model is distributed to the M data sources, the model parameters remain completely confidential and cannot be cracked or tampered with during transmission. The encrypted model is then distributed to each data source for the next round of local model training.

[0079] Optionally, after receiving the encrypted neural network, each of the M data sources performs model training using its own local private data. After each round of training, the updated model parameters or gradient information are encrypted and uploaded to shared memory, awaiting the next round of centralized update and distribution. Local model training iterations continue until the neural network model converges, meaning that the model performance no longer changes significantly after multiple rounds of training, meeting a predetermined performance standard or stopping condition. Model convergence is typically determined based on global clock information to ensure consistent training progress across all data sources.

[0080] When the neural network model converges, the central server decrypts the final neural network version (if necessary) and integrates it into a complete target model. This target model is trained using the updated information contributed by M data sources, reflecting the collective knowledge of the data while strictly protecting the privacy and security of each data source.

[0081] In an optional embodiment, while training a neural network deployed on a central server based on the data fusion results to obtain a target model, the classification system can periodically trigger a task operation. The task operation is used to generate M key shards and encrypt the latest training result data based on the M key shards. The classification system then distributes the M key shards to M data sources, where different data sources store different key shards.

[0082] Optionally, the classification system automatically initiates a task at predetermined intervals or training iterations to encrypt the latest training result data and securely distribute it to M data sources. This timing mechanism ensures timely data processing and consistent training. Furthermore, during each encryption process, the classification system generates M key shards, where M equals the number of data sources participating in federated learning. Each key shard is a portion of the complete key; a single shard is insufficient to decrypt the data.

[0083] Optionally, the classification system uses a combination of these M key shards to encrypt the latest training result data, namely updated model parameters or gradient information. This encryption process ensures that even when the data is transmitted over the network, it cannot be eavesdropped or interpreted by unauthorized third parties. The encrypted training result data is securely stored and managed.

[0084] Subsequently, the M key shards are distributed to M data sources, each of which holds only its corresponding key shard. This approach prevents any single data source from decrypting the data on its own. Only when the key shards from all data sources are aggregated under specific conditions (such as abnormal recovery) can the original key be recovered and the data decrypted. This distribution of key shards establishes a mutually dependent trust relationship between different data sources. Each data source participates in the data encryption and decryption process, but no single source has sole access to or control over all data. This promotes data value sharing while maintaining data privacy and security.

[0085] In an optional embodiment, if an anomaly is detected during the training of the target model at the target time, the task operation most recently triggered before the target time is determined as the target task operation. Then, based on the M key shards generated by the target task operation, the training result data encrypted during the execution of the target task operation is decrypted, and the target model is retrained based on the decrypted training result data.

[0086] Optionally, at the target moment, which is a specific training iteration or time point, if the classification system detects that the training of the target model has an anomaly (such as sudden performance degradation, data loss, or communication failure, etc.), the classification system will take immediate action to prevent the anomaly from affecting the quality of the model and the stability of the training process. Specifically, the system will look for the task operation record that was successfully triggered most recently since the target moment and define it as the target task operation. The purpose of this step is to find the time point when the last training result data was correctly encrypted and stored, so as to start recovery and retraining from there. The classification system then notifies the M data sources and asks them to provide the key shards that they each saved when executing the target task operation. Once all M key shards are collected, the classification system can aggregate these shards and recover the complete key. This complete key is then used to decrypt the training result data encrypted in the target task operation to obtain the original form of the data.

[0087] Additionally, once the training result data is successfully decrypted, the classification system uses this data to retrain the target model on the central server. This step rolls back the training process to the state before the anomaly occurred and then continues until the model reaches the predetermined convergence standard or target performance.

[0088] This application uses key sharding and data encryption to not only ensure data security and privacy during training, but also establishes an effective exception recovery mechanism. Even if an unexpected situation occurs during training, the system can roll back to the most recent successful training state, recover the data, and restart training, thus avoiding the long-term adverse effects of training interruptions on model performance.

[0089] According to another aspect of the present application, a device for classifying user groups is provided, wherein: Figure 4 is a schematic diagram of an optional user group classification device according to an embodiment of the present application, such as Figure 4 As shown, the user group classification device includes: an acquisition unit 401 , an encryption processing unit 402 , and a determination unit 403 .

[0090] Optionally, the acquisition unit 401 is used to obtain the characteristic data of the target user with the authorization of the target user; the encryption processing unit 402 is used to perform characteristic encryption processing on the characteristic data of the target user to obtain an encryption result, wherein the characteristic encryption processing is used to encrypt the characteristic data through data segmentation and confusion fusion; the determination unit 403 is used to determine the target group type to which the target user belongs based on the encryption result.

[0091] Optionally, the determination unit 403 includes: a first processing subunit, used to input the encryption result into the target model, decrypt the encryption result through the target model, and determine the probability that the target user belongs to each group type of N group types based on the decryption result, wherein the target model is a model jointly trained by M data sources, and the transmission data between the target model and each data source is in an encrypted state, and M and N are both integers greater than 1; a second processing subunit, used to take the group type to which the target user has the highest probability of belonging as the target group type based on the probability that the target user belongs to each group type of the N group types.

[0092] Optionally, the user group classification device also includes: a first acquisition unit, used to acquire user feature data provided by M data sources; a first processing unit, used to divide the user feature data provided by each data source into X data slices, where X is an integer greater than 1; a second processing unit, used to summarize the X data slices corresponding to each data source into a data set corresponding to the data source, and obtain M data sets corresponding to the M data sources; an obfuscation encryption unit, used to obfuscate and encrypt the data slices in each data set to obtain an obfuscated encryption result for each data set; a data fusion unit, used to fuse the obfuscated encryption results of the M data sets to obtain a data fusion result; and a training unit, used to train a neural network deployed on a central server according to the data fusion result to obtain a target model, wherein the central server and the M data sources remain independent of each other.

[0093] Optionally, the training unit includes: a first acquisition subunit, used to obtain global clock information and metadata identifiers between M data sources; a first processing subunit, used to generate a key based on the global clock information and metadata identifier; a decryption subunit, used to decrypt the data fusion result based on the key and obfuscate the decrypted data; an encryption subunit, used to re-encrypt the obfuscated data to generate a common data set between the M data sources; and a training subunit, used to train the neural network deployed on the central server based on the common data set to obtain a target model.

[0094] Optionally, the training subunit includes: a first processing module, which is used to copy the neural network on the central server to the memory of each data source through the controller; a second processing module, which is used to randomly select part of the data from the public data set for each data source, and combine it with the private data within each data source as the training data of each data source; a third processing module, which is used to transfer the training data of each data source to the model training chip for training after locking the model training chip through the controller; a fourth processing module, which is used to obtain the model training results of each data source from the model training chip through the controller, and perform aggregation operations on the model training results of each data source; a fifth processing module, which is used to encrypt and store the aggregated data obtained after the aggregation operation is completed in the shared memory of M data sources; and a sixth processing module, which is used to train the target model based on the stored data in the shared memory.

[0095] Optionally, the sixth processing module includes: an updating submodule, used to update the neural network in the central server based on the stored data in the shared memory; an encryption submodule, used to encrypt the updated neural network in the central server according to the global clock information and metadata identifier, and distribute it to M data sources; a training submodule, used to train the received neural network through M data sources until the neural network is in a convergence state and a target model is obtained.

[0096] Optionally, the user group classification device also includes: a task triggering unit, used to trigger task operations at a fixed time, wherein the task operation is used to generate M key shards and encrypt the current latest training result data based on the M key shards; a key distribution unit, used to distribute the M key shards to M data sources, wherein different data sources store different key shards.

[0097] Optionally, the user group classification device also includes: a first determination unit, which is used to determine the task operation triggered most recently from the target moment as the target task operation if an abnormality is detected in the training process of the target model at the target moment; a retraining unit, which is used to decrypt the training result data encrypted when executing the target task operation based on the M key shards generated by the target task operation, and retrain the target model based on the decrypted training result data.

[0098] According to another aspect of the present application, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed, the device where the computer-readable storage medium is located executes the above-mentioned method for classifying user groups.

[0099] According to another aspect of the present application, an electronic device is also provided, wherein the electronic device includes one or more processors and a memory, and the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the above-mentioned user group classification method.

[0100] According to another aspect of the present application, a computer program product is further provided, wherein the computer program product includes a computer program or instructions, and the computer program or instructions implement the above-mentioned method for classifying user groups when executed by a processor.

[0101] The above-mentioned embodiments or examples disclosed in this application are not exhaustive, but are only illustrations of some embodiments or examples, and are not intended to be specific limitations on the scope of protection disclosed in this application. In the absence of contradiction, each step in a certain embodiment or example in this application can be implemented as an independent example, and the steps can be arbitrarily combined. For example, the solution after removing some steps in a certain embodiment or example can also be implemented as an independent example, and the order of the steps in a certain embodiment or example can be arbitrarily exchanged. In addition, the optional methods or optional examples in a certain embodiment or example can be arbitrarily combined; in addition, the various embodiments or examples can be arbitrarily combined. For example, some or all of the steps in different embodiments or examples can be arbitrarily combined, and a certain embodiment or example can be arbitrarily combined with the optional methods or optional examples of other embodiments or examples.

[0102] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0103] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0104] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of the units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units 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 through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0105] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0106] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0107] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0108] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for classifying user groups, characterized in that: include: Obtaining characteristic data of the target user with authorization from the target user; Performing feature encryption processing on the feature data of the target user to obtain an encryption result, wherein the feature encryption processing is used to encrypt the feature data by data segmentation and obfuscation fusion; The target group type to which the target user belongs is determined according to the encryption result.

2. The method according to claim 1, characterized in that Determining the target group type to which the target user belongs according to the encryption result includes: Inputting the encrypted result into a target model, decrypting the encrypted result using the target model, and determining the probability that the target user belongs to each of N group types based on the decrypted result, wherein the target model is a model jointly trained by M data sources, and the data transmitted between the target model and each data source is in an encrypted state, and M and N are both integers greater than 1; According to the probability that the target user belongs to each group type among the N group types, the group type to which the target user has the highest probability of belonging is taken as the target group type.

3. The method according to claim 2, characterized in that The target model training step includes the following steps: Obtaining user feature data provided by the M data sources; Divide the user feature data provided by each data source into X data slices, where X is an integer greater than 1; Summarize the X data slices corresponding to each data source into a data set corresponding to the data source, to obtain M data sets corresponding to the M data sources; Performing obfuscation encryption on the data slices in each data set to obtain an obfuscation encryption result of each data set; Perform data fusion on the obfuscated encryption results of the M data sets to obtain a data fusion result; The neural network deployed on the central server is trained according to the data fusion result to obtain the target model, wherein the central server and the M data sources remain independent of each other.

4. The method according to claim 3, characterized in that The neural network deployed on the central server is trained according to the data fusion result to obtain the target model, including: Obtaining global clock information and metadata identifiers between the M data sources; generating a key according to the global clock information and the metadata identifier; Decrypting the data fusion result according to the key, and obfuscating the decrypted data; Re-encrypting the obfuscated data to generate a common data set among the M data sources; The neural network deployed on the central server is trained according to the public data set to obtain the target model.

5. The method according to claim 4, characterized in that Training the neural network deployed on the central server according to the public data set to obtain the target model includes: Copying the neural network on the central server to the memory of each data source through the controller; For each data source, randomly select a portion of data from the public dataset and combine it with the private data within each data source to serve as training data for each data source; After locking the model training chip through the controller, transmitting the training data of each data source to the model training chip for training; Obtaining the model training result of each data source from the model training chip through the controller, and performing an aggregation operation on the model training result of each data source; Encrypting and storing the aggregated data obtained after the aggregation operation in the shared memory of the M data sources; The target model is obtained by training according to the stored data in the shared memory.

6. The method according to claim 5, characterized in that The target model is obtained by training according to the stored data in the shared memory, comprising: updating the neural network in the central server according to the stored data in the shared memory; Encrypting the updated neural network in the central server according to the global clock information and metadata identifier, and distributing it to the M data sources; The received neural network is trained through the M data sources until the neural network is in a convergence state, thereby obtaining the target model.

7. The method according to claim 3, characterized in that In the process of training the neural network deployed on the central server according to the data fusion result to obtain the target model, the method further includes: A timed triggering task operation is used to generate M key shards and encrypt the latest training result data based on the M key shards; The M key shards are distributed to the M data sources, wherein different data sources store different key shards.

8. The method according to claim 7, characterized in that The method further comprises: If an abnormality is detected in the training process of the target model at the target time, the task operation triggered most recently from the target time is determined as the target task operation; According to the M key shards generated by the target task operation, the training result data encrypted when the target task operation is executed is decrypted, and the target model is retrained based on the decrypted training result data.

9. A device for classifying user groups, characterized in that: include: an acquiring unit, configured to acquire characteristic data of a target user upon authorization by the target user; an encryption processing unit, configured to perform feature encryption processing on the feature data of the target user to obtain an encryption result, wherein the feature encryption processing is configured to encrypt the feature data by data segmentation and obfuscation fusion; A determination unit is configured to determine the target group type to which the target user belongs according to the encryption result.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes the user group classification method according to any one of claims 1 to 8.

11. An electronic device, characterized in that: It includes one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the user group classification method described in any one of claims 1 to 8.

12. A computer program product, characterized in that The method comprises a computer program or instructions, which implements the method for classifying user groups according to any one of claims 1 to 8 when executed by a processor.

Images