Anonymization processing method and system for federated learning

Abstract

This invention provides an anonymization method and system for federated learning, relating to the field of anonymization technology. The method includes: establishing communication channels between each participating party and the federated platform, configuring anonymization strategies, establishing a bidirectional compromise channel between the data end and the anonymization end (including multiple parallel channels mapped to multi-level anonymity limits), delegating to each participating party, performing adaptive anonymization training based on locally shared sensitive data, generating a local anonymization model, transmitting it back to the federated platform, performing a platform trust verification checkpoint, filtering the local trustworthy model, performing multi-layer aggregation processing to generate a global anonymization model, transmitting it back to each participating party and updating the local anonymization model, and performing data anonymization. This invention solves the technical problem of traditional methods that centrally store sensitive data on a central server, making sensitive information vulnerable to attacks and posing a privacy leakage risk in collaborative scenarios.

Description

Technical Field

[0001] This invention relates to the field of anonymization technology, and more specifically to anonymization methods and systems for federated learning. Background Technology

[0002] In many industries, organizations dealing with large amounts of sensitive data need to conduct data analysis and model training to drive innovation and provide better services. However, traditional centralized data analysis and model training methods pose privacy and security risks. On the one hand, in traditional methods, sensitive data is centrally stored on a central server, which may make sensitive information vulnerable to attack or misuse. In addition, data owners cannot control the risk of their data being misused. On the other hand, in collaborative scenarios, participating parties may have their own private knowledge and data. How to protect the privacy of all parties while sharing data is a significant challenge.

[0003] Therefore, a new approach is needed that allows participants to perform computations locally and share only the model update results, thereby protecting the privacy and security of the original data while allowing participants to retain control over their own data to ensure the protection of knowledge and privacy. Summary of the Invention

[0004] This application provides an anonymization processing method and system for federated learning, aiming to solve the technical problems of traditional methods that centrally store sensitive data on a central server, making sensitive information vulnerable to attacks, and the risk of privacy leakage when participants share data in collaborative scenarios.

[0005] In view of the above problems, this application provides an anonymization method and system for federated learning.

[0006] The first aspect disclosed in this application provides an anonymization method for federated learning, the method comprising: establishing communication channels between each participant and a federated platform, and configuring anonymization strategies; establishing a bidirectional compromise channel between the data end and the anonymization end in conjunction with the anonymization strategies, wherein the bidirectional compromise channel includes multiple parallel channels mapped to multi-level anonymity limits; delegating the bidirectional compromise channel to each participant in conjunction with the communication channels, performing adaptive anonymization training based on locally shared sensitive data, and generating a local anonymization model; transmitting the local anonymization model back to the federated platform in conjunction with the communication channels, performing a platform trust verification checkpoint, and filtering local trustworthy models; performing multi-layer aggregation processing based on the local trustworthy models to generate a global anonymization model, wherein each aggregation processing layer corresponds to a different source model category; transmitting the global anonymization model back to each participant in conjunction with the communication channels and updating the local anonymization model, and performing data anonymization processing.

[0007] Another aspect of this application discloses an anonymization processing system for federated learning. The system, used in the aforementioned method, includes: a communication channel establishment module for establishing communication channels between each participant and the federated platform, and configuring anonymization processing strategies; a compromise channel establishment module for establishing bidirectional compromise channels between the data end and the anonymization end in conjunction with the anonymization processing strategies, wherein the bidirectional compromise channels include multiple parallel channels mapped to multi-level defined anonymity limits; and a processing model training module for distributing the bidirectional compromise channels to each participant in conjunction with the communication channels, and performing local shared sensitivity... The system includes: an adaptive anonymization training module for sensing data to generate a local anonymization model; a trusted model filtering module, which, in conjunction with the communication channel, transmits the local anonymization model back to the federated middle platform, executes the middle platform's trusted verification checkpoint, and filters the local trusted model; a multi-layer aggregation processing module, which performs multi-layer aggregation processing based on the local trusted model to generate a global anonymization model, wherein each aggregation processing layer corresponds to a different source model category; and an anonymization processing module, which, in conjunction with the communication channel, transmits the global anonymization model back to each participant and updates the local anonymization model, performing data anonymization processing.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] By establishing a communication channel between participants and the federated platform and configuring anonymization strategies, the anonymization of sensitive data by participants is ensured, enhancing data confidentiality. A two-way compromise channel allows for bidirectional communication between the data source and the anonymization end, ensuring the effectiveness and accuracy of data sharing. Multiple parallel channels with multi-level defined anonymity limits contribute to improved data transmission efficiency and stability. Participants perform adaptive anonymization training locally based on shared sensitive data, generating local anonymization models. This reduces direct exposure to the original data and protects privacy while maintaining data accuracy. By transmitting the local anonymization models back to the federated platform and performing a trust verification checkpoint, trusted local models are selected, ensuring the quality and credibility of the models provided by participants. Multi-layer aggregation processing based on the trusted local models generates a global anonymized model. Each aggregation layer corresponds to a different source model category, facilitating the integration of knowledge and information from various participants and improving the performance and accuracy of the global model. In summary, this anonymization method for federated learning solves the privacy leakage and data sharing problems of traditional methods, protects the privacy of sensitive data, and promotes effective data analysis and sharing.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] Figure 1 This application provides a schematic flowchart of an anonymization process for federated learning in its embodiments.

[0012] Figure 2 This application provides a schematic diagram of the anonymization processing system structure for federated learning in its embodiments;

[0013] Figure 3 An internal structural diagram of a computer device is provided for embodiments of this application.

[0014] Explanation of reference numerals in the attached diagram: Communication channel establishment module 10, compromise channel establishment module 20, processing model training module 30, trustworthy model screening module 40, multi-layer aggregation processing module 50, anonymization processing module 60. Detailed Implementation

[0015] This application provides an anonymization processing method for federated learning, which solves the technical problem that traditional methods centrally store sensitive data on a central server, making sensitive information vulnerable to attacks, and that participants face privacy leakage risks while sharing data in collaborative scenarios.

[0016] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0017] Example 1

[0018] like Figure 1 As shown in the embodiments of this application, an anonymization method for federated learning is provided, the method comprising:

[0019] Establish communication channels between each participating party and the federated platform, and configure anonymization processing strategies;

[0020] The participants refer to the data providers. The federated platform is the central platform responsible for coordinating and managing federated learning tasks, establishing a reliable and secure communication channel between the participants and the federated platform to facilitate the transmission of data and models.

[0021] Anonymization strategies are a series of measures taken to protect the privacy of sensitive data. These strategies include the selection of data anonymization technologies, the definition of anonymity levels, and the setting of anonymity restrictions. When configuring anonymization strategies, it is necessary to consider the data protection needs and data utilization requirements based on the actual situation. Applying the configured anonymization strategies to communication channels means that data will be affected by anonymization during transmission to protect data privacy and sensitivity.

[0022] In summary, establishing communication channels and configuring anonymization strategies are to ensure secure communication between participants and the federated platform, and to take anonymization measures during data transmission to protect data privacy.

[0023] In conjunction with the aforementioned anonymization processing strategy, a two-way compromise channel is established between the data end and the anonymization end, wherein the two-way compromise channel includes multiple parallel channels mapped to multi-level anonymity limits;

[0024] A reliable connection channel is established between the data end and the anonymization end through network protocols, security encryption, and authentication. Based on specific needs, multi-level anonymity limits are defined, i.e., multi-level anonymity restrictions. Anonymity restrictions can be at the data field level or at the data sample level, with different degrees of data processing applied according to different anonymity levels. Based on the requirements of multi-level anonymity limits, the data channel is divided into multiple parallel channels, each responsible for a different level of anonymization processing to achieve privacy protection during data transmission and processing. In each parallel channel, data is processed and transformed according to the set anonymity limits, using techniques such as k-anonymity, Laplace noise addition, and data perturbation to ensure appropriate anonymization during transmission. In the bidirectional channel, a compromise mechanism and negotiation method are established between the data end and the anonymization end, including conditions for data sharing, the selection of anonymization processing methods, and the stipulation of privacy protection measures, to ensure the fairness and security of data use.

[0025] By using the above methods, a two-way compromise channel is established between the data end and the anonymity end, and a parallel channel is implemented based on multi-level definition of anonymity limits. This can protect data privacy while meeting the needs of federated learning tasks.

[0026] Furthermore, the bidirectional compromise channel includes multiple parallel channels mapped to multi-level defined anonymity limits, including:

[0027] Reasonably balance the dimensions of data protection and data utilization, and determine multiple metrics for defining the degree of data anonymity;

[0028] Configure a multi-level defined anonymity limit based on the aforementioned multiple metrics;

[0029] Based on the aforementioned multi-level anonymity limits, the anonymization processing strategy is adaptively matched and adjusted to determine multiple pre-adjustment strategies.

[0030] By combining the aforementioned pre-adjustment strategies, a two-way compromise channel is established between the data end and the anonymity end.

[0031] When determining the degree of anonymity of data, consider the following metrics to achieve a reasonable balance between data protection and data utilization dimensions, that is, to find an appropriate balance point to ensure data privacy protection and effective data utilization, including information loss degree, identity risk, differential privacy protection degree, protection degree of sensitive attributes, and data correlation degree.

[0032] Specifically, information loss measures the extent to which sensitive information in the original data is deleted or distorted after anonymization. A higher information loss may reduce data usability but also improve data privacy protection. Identity risk assesses whether anonymized data still retains the identifiability of certain individuals. A higher identity risk means that anonymized data can still be associated with a specific individual, increasing the risk of privacy leakage. Data correlation assesses the degree of correlation between anonymized data. A higher data correlation may increase the risk of recovering individual information through data correlation attacks, while a lower correlation may limit the usability of the data.

[0033] Based on the importance of data protection and data utilization, appropriate weights are assigned to each metric, reflecting their relative importance in defining the level of anonymity. Multiple levels of anonymity restrictions are established based on these weights, each corresponding to a specific set of anonymization requirements and conditions. For example, the lowest level requires a low degree of information loss and identity risk, while higher levels can tolerate a higher degree of information loss but require a lower risk of identity loss.

[0034] Based on the anonymity restrictions at each level, multiple preset anonymization strategies are developed. Each strategy should meet specific anonymity requirements, and these strategies can employ different algorithms, techniques, and parameter settings to meet the needs of different anonymization levels. The preset anonymization strategies are matched with specific datasets and scenarios, and adjusted according to actual conditions. This involves adaptive selection and modification of strategies based on different balance standards between protection and utilization, primarily adjusting the processing level to determine multiple suitable strategies that meet the defined levels. Each strategy corresponds to a specific anonymization level. In this way, in practical applications, an appropriate strategy can be selected as needed, ensuring a balance between data privacy protection and effective utilization.

[0035] Two compromise directions are determined: one from the data source to the anonymity source, and the other from the anonymity source to the data source. For the first compromise direction from the data source to the anonymity source, a pre-tuning strategy is extracted and selected for training. One of the multiple pre-tuning strategies is chosen as the first pre-tuning strategy. Using data from the first compromise direction, the model is trained according to the first pre-tuning strategy, and the trend characteristics during the data transformation process based on the first compromise direction are recorded for use in subsequent steps. Similarly, for the second compromise direction from the anonymity source to the data source, another pre-tuning strategy is selected for training.

[0036] By integrating the first and second compromise branches in parallel, a bidirectional compromise channel is generated between the data end and the anonymous end, which enables bidirectional data transmission while meeting the requirements of security and privacy protection.

[0037] Furthermore, by combining the aforementioned multiple pre-tuning strategies, a two-way compromise channel is established between the data end and the anonymity end, including:

[0038] The two-way compromise channel includes a first compromise direction from the data end to the anonymous end and a second compromise direction from the anonymous end to the data end.

[0039] Based on the multiple pre-tuning strategies, the first pre-tuning strategy is extracted and the first bidirectional compromise channel is trained.

[0040] Based on the multiple pre-adjustment strategies, the Nth bidirectional compromise channel based on the Nth pre-adjustment strategy is completed;

[0041] The first bidirectional compromise channel is integrated in parallel up to the Nth bidirectional compromise channel to generate the bidirectional compromise channel.

[0042] Define the first compromise direction from the data end to the anonymous end, which represents the compromise measures and security guarantees required when transmitting data from the data end to the anonymous end; determine the second compromise direction from the anonymous end to the data end, which represents the compromise measures and security guarantees required when returning from the anonymous end to the data end.

[0043] A first pre-tuning strategy is selected from multiple pre-tuning strategies. This strategy is used for the first compromise direction. The model for the first compromise direction is trained using the selected first pre-tuning strategy to implement security protection when transmitting data from the data end to the anonymous end. When training the first bidirectional compromise channel, appropriate encryption algorithms, authentication mechanisms or other security measures are combined to ensure the privacy and integrity of the data.

[0044] Similarly, among multiple pre-tuning strategies, the Nth pre-tuning strategy is selected for the Nth compromise direction. The selected Nth pre-tuning strategy is used to train the model for the Nth compromise direction in order to implement security protection when transmitting data between the data end and the anonymization end.

[0045] The first bidirectional compromise channel is integrated in parallel with the second through N-1th bidirectional compromise channels. This involves integrating the functions, algorithms, and security measures of each channel into a unified system. During parallel integration, the mutual influence and collaboration between channels are considered to ensure conflict-free data transmission and processing between channels, while meeting the overall system's security and privacy protection requirements. Finally, the Nth bidirectional compromise channel is integrated in parallel with the previously integrated channels to generate a complete bidirectional compromise channel. This satisfies the need for secure data transmission between the data end and the anonymous end.

[0046] Furthermore, based on the multiple pre-tuning strategies, extracting a first pre-tuning strategy and training a first bidirectional compromise channel includes:

[0047] Extract the first pre-tuning strategy based on the multiple pre-tuning strategies, and train the first compromise branch based on the first compromise direction;

[0048] Record the transformation trend features based on the first compromise direction, and train the second compromise branch based on the second compromise direction;

[0049] Configure the encryption protocol and management permissions, and configure the second compromise branch;

[0050] A first bidirectional compromise channel is generated based on the first compromise branch and the second compromise branch.

[0051] Based on the multiple pre-tuning strategies, one is extracted as the first pre-tuning strategy. Using the extracted first pre-tuning strategy, the first compromise branch is trained. Specifically, data related to the first compromise direction is collected, and iterative training is performed on the training data using the extracted first pre-tuning strategy and the prepared data to optimize the weights and parameters of the first compromise branch to maximize its performance. During training, the trained first compromise branch is evaluated and validated to measure its performance and effectiveness. Based on the evaluation results, the first compromise branch is improved and adjusted to obtain the final first compromise branch.

[0052] Based on the first compromise direction, key features of the data transformation process are recorded, including the methods, techniques, and algorithms used for data transformation, as well as the rules and constraints for achieving anonymization. The recorded transformation trend features are then used to train the second compromise branch. Specifically, based on the transformation trend features of the first compromise direction, personal information is anonymized, including using encryption algorithms and data desensitization techniques to ensure data anonymity. According to the requirements of the second compromise direction, the anonymized data is used to train the second compromise branch. During training, a correlation is established between the anonymized data and the original personal information. To protect data security, the execution permissions of the second branch are strictly configured, including access control, authentication, and permission management measures, to ensure that only authorized users can perform related operations.

[0053] Choose an appropriate encryption protocol to protect the confidentiality of data, based on specific needs and security requirements, such as AES (Advanced Encryption Standard) and RSA (Rivest-Shamir-Adleman). Select a suitable encryption algorithm and key length according to the application scenario and specific requirements.

[0054] Using the selected encryption protocol and algorithm, the data involved in the second compromise branch is encrypted, including steps such as data segmentation, key generation, and encryption. At the same time, in order to protect data security, the management permissions of the second compromise branch are configured, including authentication and access control, permission management, audit logs and monitoring, to ensure that only authorized users or users with application permissions can perform related operations.

[0055] A communication connection is established between the first and second compromise branches, thus creating a first bidirectional compromise channel. This allows for bidirectional data transmission and processing between the first and second compromise branches. Such a channel ensures data integrity during processing and meets security and privacy protection requirements under specific needs.

[0056] Combined with the aforementioned communication channel, the two-way compromise channel is distributed to each participating party to perform adaptive anonymization processing training based on locally shared sensitive data, thereby generating a local anonymization processing model;

[0057] Delegating the configuration and setup of the two-way compromise channel to each participating party means that each party is responsible for establishing a secure communication channel with the federated platform and configuring anonymization policies. Subject to privacy regulations and compliance requirements, each participating party can share its sensitive data locally, which can be raw data or data that has undergone a certain degree of preprocessing, based on task requirements.

[0058] Participants use locally shared sensitive data for adaptive anonymization training. This means that participants perform appropriate anonymization transformations and processing on sensitive data based on its characteristics and anonymization strategies, including adding noise, perturbation, and sampling. After the adaptive anonymization training phase is complete, each participant generates a local anonymization model, which is used for subsequent data sharing and federated learning tasks. This allows for the generation of local anonymization models while protecting data privacy, for use in subsequent federated learning tasks.

[0059] Combined with the communication channel, the local anonymization processing model is sent back to the federated middle platform to perform the middle platform's trusted verification checkpoint and filter the local trusted model.

[0060] Participants use a secure communication channel to transmit their locally anonymized models back to the federated platform, ensuring the security and integrity of the model transmission. A trust verification checkpoint is set up at the federated platform to verify the trustworthiness of the transmitted locally anonymized models. This checkpoint includes multiple layers of verification methods, such as model integrity checks, model parameter consistency verification, and model performance evaluation. Through the trust verification checkpoint, the federated platform selects verified and trusted locally anonymized models, which can then be used for subsequent federated learning tasks. This process ensures the security and quality of the models and provides reliable model resources for subsequent federated learning tasks.

[0061] Furthermore, the trusted verification checkpoint of the execution platform includes:

[0062] The federal platform is equipped with a trusted verification unit.

[0063] In conjunction with the aforementioned trusted verification unit, the fidelity of the local anonymization processing model is determined, and the trust coefficient is identified.

[0064] Identify the credibility coefficient, filter out the local anonymization processing models that do not meet the coefficient threshold, and generate an interception and repatriation instruction;

[0065] Based on the interception and repatriation instruction, the release and repatriation operations of the local anonymization processing model are executed.

[0066] The federated middle platform sets up a trusted verification unit, which is a component used to verify the trustworthiness of the model and to evaluate and verify the trustworthiness of the local anonymization processing model involved.

[0067] The Trusted Verification Unit uses a series of evaluation and verification methods to determine the fidelity of the local anonymization processing model. For example, it performs model architecture analysis to analyze and verify the model's structure, hierarchy, and parameters; it performs experimental result verification by comparing the model with known datasets or standard models to verify the model's performance and accuracy on the task; and it conducts security audits to assess the model's security, including assessments of protecting personal privacy and preventing malicious attacks on the model.

[0068] Based on the results of the fidelity assessment, a credibility coefficient is assigned to the local anonymization processing model. The credibility coefficient is an indicator of the model's credibility. For example, a continuous value between 0 and 1 is used to represent the model's credibility level, where 1 represents complete credibility and 0 represents complete untrustworthiness, for subsequent screening and processing operations.

[0069] Obtain the credibility coefficient of each local anonymization model. Based on the actual situation and specific needs, set a threshold to determine which credibility coefficients are considered to meet the requirements for a trustworthy model. Judge the credibility coefficients and filter out local anonymization models that do not meet the threshold requirements. These are the models that need to be intercepted and deported. For these models, generate corresponding interception and deportation instructions. These instructions may include notifying local participants to stop using the model, deleting the model, or conducting further evaluation.

[0070] Based on the interception and repatriation instructions, the local anonymization processing model is executed for release and repatriation operations. Specifically, the local participants receive the interception and repatriation instructions generated by the federation platform. For local anonymization processing models that are recognized as meeting the threshold requirements, the local participants execute the release operation and continue to use the model for subsequent tasks. For local anonymization processing models that are determined not to meet the threshold requirements, the local participants execute the repatriation operation, including stopping the use of the model, deleting the model file, or sending relevant information to the federation platform.

[0071] Perform multi-layer aggregation processing based on the local trusted model to generate a global anonymized model, wherein each aggregation processing layer corresponds to a different source model category;

[0072] Models submitted by various participants are collected from a pool of locally trusted models. These models come from different source model categories, representing the contributions of different participants. A multi-layered aggregation process is performed, systematically aggregating the models from each participant layer by layer. This effectively aggregates related information, improving aggregation efficiency and effectiveness. For example, a model's functional mechanism, while existing in current applications, can be adapted to a novel level, thus enhancing its capabilities. Through multi-layered aggregation, a globally anonymized model is generated, integrating the contributions of different participants. This systematic, layer-by-layer aggregation effectively aggregates related information, improving both aggregation efficiency and effectiveness.

[0073] Furthermore, performing multi-layer aggregation processing based on the local trust model, the method includes:

[0074] The local trusted models are classified into similar categories, and horizontal aggregation based on X types of model groups is performed to determine X first-level models.

[0075] The X first-level models are collaboratively associated and classified, and vertical aggregation based on the Y-type model group is performed to determine Y second-level models, where Y is a positive integer greater than or equal to 1 and Y < X;

[0076] If Y is greater than 1, the Y two-layer models are aggregated and expanded to generate the global anonymized model.

[0077] Based on the model type, the local trusted models are classified into categories. From each of the classified model categories, X model groups are selected as the basis for horizontal aggregation. X can be a pre-set fixed value or can be flexibly adjusted according to specific needs.

[0078] Within each selected model group, the models are horizontally aggregated, such as through model fusion or ensemble learning, to integrate models of the same category to obtain a more powerful single-layer model. After horizontal aggregation, the best single-layer model is selected from each model group, for a total of X models as the final result. The selection criteria can be based on performance metrics, validation results, or other evaluation criteria.

[0079] Based on the correlation, complementarity, or other correlation indicators among the models, X first-level models are classified into synergistic categories to group models with synergistic effects into the same category for subsequent vertical aggregation. From each model category after synergistic classification, Y model groups are selected as the basis for vertical aggregation. Y should be a positive integer greater than or equal to 1 and less than X to ensure the effectiveness and diversity of vertical aggregation.

[0080] Within each selected model group, the models are vertically aggregated, such as through model stacking or ensemble learning, to integrate the models within the same group to obtain a more powerful two-layer model. After vertical aggregation, the optimal two-layer model is selected from each model group, for a total of Y models as the final result. The selection criteria can be based on performance metrics, validation results, or other evaluation criteria.

[0081] For Y two-layer models, different aggregation methods, such as model fusion and ensemble learning, are employed to integrate them. Aggregation helps improve model performance, reduce bias, and increase model robustness. After aggregation, novel functional mechanisms are introduced to extend the model. This means that the model is endowed with more complete functions in applications where it is already adapted, while also adapting to another novel level. For example, a function can be migrated from the original application domain to other domains, enabling the model to perform well in multiple tasks or applications.

[0082] Through aggregation and migration expansion, a globally anonymized model is obtained. This model has more powerful functions and performance, can adapt to diverse tasks and scenarios, and, due to the use of anonymization technology, protects data privacy, making the model more secure during use.

[0083] Using the communication channel, the global anonymization model is sent back to each participant and the local anonymization processing model is updated to perform data anonymization processing.

[0084] Using a secure communication channel, the global anonymization model is transmitted back to each participant. Upon receiving the global anonymization model, each participant applies it to update its local anonymization model. This means that participants use the global model to improve and update their local models to meet better anonymization requirements. Using the updated local anonymization model, each participant performs data anonymization processing, including appropriate transformations, noise addition, perturbations, or other anonymization methods on sensitive data to protect data privacy. This process ensures model synchronization and consistency while protecting the privacy and security of sensitive data.

[0085] Furthermore, performing anonymization of the data includes:

[0086] The source data is analyzed to define the anonymity limit, the target two-way compromise channel of the global anonymization model is activated, the source data is anonymized, and the anonymization result is generated.

[0087] Perform identity and permission authentication on the user, and determine data restoration permission;

[0088] An auxiliary encryption protocol is used to activate the target bidirectional compromise channel of the global anonymization model, perform the restoration process of the anonymized data, and generate the restoration result.

[0089] The source data is analyzed to determine the specific requirements and limitations of anonymization, including defining anonymization metrics, privacy protection needs, and data availability. By analyzing the characteristics of the source data and relevant laws and policies, the boundaries and constraints of anonymization are established. A two-way compromise mechanism ensures that anonymization processing meets the model's needs while maintaining data privacy and data availability. Based on the defined anonymization limits and the two-way compromise mechanism, anonymization of the source data begins to ensure the protection of personal identity and sensitive information. After anonymization, the anonymized results are obtained. These results no longer contain personally identifiable information, and the anonymized data still meets the model's analytical needs.

[0090] User authentication is required to ensure their authenticity and permissions, using methods such as username / password verification, biometric identification, and two-factor authentication. In addition to identity authentication, permission authentication is also necessary to determine if the user has permission to restore data. This includes defining and managing permission levels for different roles or organizations, and verifying the permissions of the user's role or organization during the authentication process.

[0091] By employing identity and authorization authentication, the system determines which users have permission to restore data and allows them to proceed with the restoration process. Before restoring anonymous data, auxiliary encryption protocols, such as homomorphic encryption and secure multi-party computation, are used to ensure data security, protecting sensitive data during the restoration process.

[0092] Activating the target two-way compromise channel of the global anonymization model ensures a balance between the needs of the anonymization model and the protection of data privacy during the restoration process. With the auxiliary encryption protocol and the target two-way compromise channel in place, the anonymized data restoration process is performed. This involves reversing the anonymized data according to a specific algorithm to recover part or all of the original data. The result of the restoration process is a data form similar to or close to the data before anonymization. This allows for the restoration and use of anonymized data while protecting data privacy.

[0093] In summary, the anonymization method and system for federated learning provided in this application have the following technical effects:

[0094] 1. By establishing a communication channel between the participants and the federated platform and configuring anonymization processing strategies, the anonymization of the participants' sensitive data is ensured, thereby enhancing data confidentiality;

[0095] 2. By establishing a two-way compromise channel, bidirectional communication between the data end and the anonymous end is allowed to ensure the effectiveness and accuracy of data sharing, and multiple parallel channels with multi-level defined anonymity limits help improve the efficiency and stability of data transmission;

[0096] 3. Participants perform adaptive anonymization training on shared sensitive data locally to generate a local anonymization model. This reduces direct exposure to the original data and protects privacy while maintaining data accuracy.

[0097] 4. By sending the local anonymization processing model back to the federated platform and performing a trust verification checkpoint, the local trustworthy models are selected, ensuring the quality and trustworthiness of the models provided by the participants.

[0098] 5. Multi-layer aggregation processing based on the local trusted model generates a global anonymized model. Each aggregation processing layer corresponds to a different source model category, which helps to integrate the knowledge and information of all participants and improve the performance and accuracy of the global model.

[0099] In summary, this anonymization method for federated learning addresses the privacy and data sharing issues inherent in traditional methods, protects the privacy of sensitive data, and promotes effective data analysis and sharing.

[0100] Example 2

[0101] Based on the same inventive concept as the anonymization method used for federated learning in the foregoing embodiments, such as Figure 2 As shown, this application provides an anonymization processing system for federated learning, the system comprising:

[0102] The communication channel establishment module 10 is used to establish communication channels between each participant and the federation platform, and to configure anonymization processing strategies.

[0103] The compromise channel establishment module 20 is used to establish a two-way compromise channel between the data end and the anonymity end in combination with the anonymization processing strategy. The two-way compromise channel includes multiple parallel channels mapped to multi-level anonymity limits.

[0104] The processing model training module 30 is used to combine the communication channel, decentralize the two-way compromise channel to each participating party, perform adaptive anonymization processing training based on locally shared sensitive data, and generate a local anonymization processing model.

[0105] Trusted model filtering module 40 is used to combine the communication channel to send the local anonymization processing model back to the federated middle platform, execute the middle platform's trust verification checkpoint, and filter the local trusted model.

[0106] A multi-layer aggregation processing module 50 is used to perform multi-layer aggregation processing based on the local trusted model to generate a global anonymized model, wherein each aggregation processing layer corresponds to a different source model category;

[0107] An anonymization processing module 60 is used to combine the communication channel to send the global anonymization model back to each participant and update the local anonymization processing model, and perform data anonymization processing.

[0108] Furthermore, the system also includes a two-way compromise channel establishment module to perform the following operational steps:

[0109] Reasonably balance the dimensions of data protection and data utilization, and determine multiple metrics for defining the degree of data anonymity;

[0110] Configure a multi-level defined anonymity limit based on the aforementioned multiple metrics;

[0111] Based on the aforementioned multi-level anonymity limits, the anonymization processing strategy is adaptively matched and adjusted to determine multiple pre-adjustment strategies.

[0112] By combining the aforementioned pre-adjustment strategies, a two-way compromise channel is established between the data end and the anonymity end.

[0113] Furthermore, the system also includes a bidirectional compromise channel generation module to perform the following operational steps:

[0114] The two-way compromise channel includes a first compromise direction from the data end to the anonymous end and a second compromise direction from the anonymous end to the data end.

[0115] Based on the multiple pre-tuning strategies, the first pre-tuning strategy is extracted and the first bidirectional compromise channel is trained.

[0116] Based on the multiple pre-adjustment strategies, the Nth bidirectional compromise channel based on the Nth pre-adjustment strategy is completed;

[0117] The first bidirectional compromise channel is integrated in parallel up to the Nth bidirectional compromise channel to generate the bidirectional compromise channel.

[0118] Furthermore, the system also includes a first bidirectional compromise channel generation module to perform the following operational steps:

[0119] Extract the first pre-tuning strategy based on the multiple pre-tuning strategies, and train the first compromise branch based on the first compromise direction;

[0120] Record the transformation trend features based on the first compromise direction, and train the second compromise branch based on the second compromise direction;

[0121] Configure the encryption protocol and management permissions, and configure the second compromise branch;

[0122] A first bidirectional compromise channel is generated based on the first compromise branch and the second compromise branch.

[0123] Furthermore, the system also includes a release and repatriation module to perform the following steps:

[0124] The federal platform is equipped with a trusted verification unit.

[0125] In conjunction with the aforementioned trusted verification unit, the fidelity of the local anonymization processing model is determined, and the trust coefficient is identified.

[0126] Identify the credibility coefficient, filter out the local anonymization processing models that do not meet the coefficient threshold, and generate an interception and repatriation instruction;

[0127] Based on the interception and repatriation instruction, the release and repatriation operations of the local anonymization processing model are executed.

[0128] Furthermore, the system also includes a global anonymization model generation module to perform the following steps:

[0129] The local trusted models are classified into similar categories, and horizontal aggregation based on X types of model groups is performed to determine X first-level models.

[0130] The X first-level models are collaboratively associated and classified, and vertical aggregation based on the Y-type model group is performed to determine Y second-level models, where Y is a positive integer greater than or equal to 1 and Y < X;

[0131] If Y is greater than 1, the Y two-layer models are aggregated and expanded to generate the global anonymized model.

[0132] Furthermore, the system also includes a restoration processing result generation module to perform the following operation steps:

[0133] The source data is analyzed to define the anonymity limit, the target two-way compromise channel of the global anonymization model is activated, the source data is anonymized, and the anonymization result is generated.

[0134] Perform identity and permission authentication on the user, and determine data restoration permission;

[0135] An auxiliary encryption protocol is used to activate the target bidirectional compromise channel of the global anonymization model, perform the restoration process of the anonymized data, and generate the restoration result.

[0136] Through the foregoing detailed description of the anonymization method for federated learning, those skilled in the art can clearly understand the anonymization system for federated learning in this embodiment. As the apparatus disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section description.

[0137] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores news data and data such as time decay factors. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements an anonymization processing method for federated learning.

[0138] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0139] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0140] Establish communication channels between each participating party and the federated platform, and configure anonymization processing strategies;

[0141] In conjunction with the aforementioned anonymization processing strategy, a two-way compromise channel is established between the data end and the anonymization end, wherein the two-way compromise channel includes multiple parallel channels mapped to multi-level anonymity limits;

[0142] Combined with the aforementioned communication channel, the two-way compromise channel is distributed to each participating party to perform adaptive anonymization processing training based on locally shared sensitive data, thereby generating a local anonymization processing model;

[0143] Combined with the communication channel, the local anonymization processing model is sent back to the federated middle platform to perform the middle platform's trusted verification checkpoint and filter the local trusted model.

[0144] Perform multi-layer aggregation processing based on the local trusted model to generate a global anonymized model, wherein each aggregation processing layer corresponds to a different source model category;

[0145] Using the communication channel, the global anonymization model is sent back to each participant and the local anonymization processing model is updated to perform data anonymization processing.

[0146] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0147] Establish communication channels between each participating party and the federated platform, and configure anonymization processing strategies;

[0148] In conjunction with the aforementioned anonymization processing strategy, a two-way compromise channel is established between the data end and the anonymization end, wherein the two-way compromise channel includes multiple parallel channels mapped to multi-level anonymity limits;

[0149] Combined with the aforementioned communication channel, the two-way compromise channel is distributed to each participating party to perform adaptive anonymization processing training based on locally shared sensitive data, thereby generating a local anonymization processing model;

[0150] Combined with the communication channel, the local anonymization processing model is sent back to the federated middle platform to perform the middle platform's trusted verification checkpoint and filter the local trusted model.

[0151] Perform multi-layer aggregation processing based on the local trusted model to generate a global anonymized model, wherein each aggregation processing layer corresponds to a different source model category;

[0152] Using the communication channel, the global anonymization model is sent back to each participant and the local anonymization processing model is updated to perform data anonymization processing.

[0153] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An anonymization method for federated learning, characterized in that, The method includes: Establish communication channels between each participating party and the federated platform, and configure anonymization processing strategies; In conjunction with the aforementioned anonymization processing strategy, a two-way compromise channel is established between the data end and the anonymization end, wherein the two-way compromise channel includes multiple parallel channels mapped to multi-level anonymity limits; Combined with the aforementioned communication channel, the two-way compromise channel is distributed to each participating party to perform adaptive anonymization processing training based on locally shared sensitive data, thereby generating a local anonymization processing model; Combined with the communication channel, the local anonymization processing model is sent back to the federated middle platform to perform the middle platform's trusted verification checkpoint and filter the local trusted model. Perform multi-layer aggregation processing based on the local trusted model to generate a global anonymized model, wherein each aggregation processing layer corresponds to a different source model category; Using the communication channel, the global anonymization model is sent back to each participant and the local anonymization processing model is updated to perform data anonymization processing.

2. The method as described in claim 1, characterized in that, The bidirectional compromise channel includes multiple parallel channels mapped to multi-level defined anonymity limits, and the method includes: Reasonably balance the dimensions of data protection and data utilization, and determine multiple metrics for defining the degree of data anonymity; Configure a multi-level defined anonymity limit based on the aforementioned multiple metrics; Based on the aforementioned multi-level anonymity limits, the anonymization processing strategy is adaptively matched and adjusted to determine multiple pre-adjustment strategies. By combining the aforementioned pre-adjustment strategies, a two-way compromise channel is established between the data end and the anonymity end.

3. The method as described in claim 2, characterized in that, By combining the aforementioned pre-tuning strategies, a two-way compromise channel is established between the data end and the anonymity end. include: The two-way compromise channel includes a first compromise direction from the data end to the anonymous end and a second compromise direction from the anonymous end to the data end. Based on the multiple pre-tuning strategies, the first pre-tuning strategy is extracted and the first bidirectional compromise channel is trained. Based on the multiple pre-adjustment strategies, the Nth bidirectional compromise channel based on the Nth pre-adjustment strategy is completed; The first bidirectional compromise channel is integrated in parallel up to the Nth bidirectional compromise channel to generate the bidirectional compromise channel.

4. The method as described in claim 3, characterized in that, Based on the multiple pre-tuning strategies, a first pre-tuning strategy is extracted and a first bidirectional compromise channel is trained. The method includes: Extract the first pre-tuning strategy based on the multiple pre-tuning strategies, and train the first compromise branch based on the first compromise direction; Record the transformation trend features based on the first compromise direction, and train the second compromise branch based on the second compromise direction; Configure the encryption protocol and management permissions, and configure the second compromise branch; A first bidirectional compromise channel is generated based on the first compromise branch and the second compromise branch.

5. The method as described in claim 1, characterized in that, The method of the trusted verification checkpoint in the execution platform include: The federal platform is equipped with a trusted verification unit. In conjunction with the aforementioned trusted verification unit, the fidelity of the local anonymization processing model is determined, and the trust coefficient is identified. Identify the credibility coefficient, filter out the local anonymization processing models that do not meet the coefficient threshold, and generate an interception and repatriation instruction; Based on the interception and repatriation instruction, the release and repatriation operations of the local anonymization processing model are executed.

6. The method as described in claim 1, characterized in that, Performing multi-layer aggregation processing based on the local trust model, the method includes: The local trusted models are classified into similar categories, and horizontal aggregation based on X types of model groups is performed to determine X first-level models. The X first-level models are collaboratively associated and classified, and vertical aggregation based on the Y-type model group is performed to determine Y second-level models, where Y is a positive integer greater than or equal to 1 and Y < X; If Y is greater than 1, the Y two-layer models are aggregated and expanded to generate the global anonymized model.

7. The method as described in claim 1, characterized in that, Perform anonymization processing on the data, the method including: The source data is analyzed to define the anonymity limit, the target two-way compromise channel of the global anonymization model is activated, the source data is anonymized, and the anonymization result is generated. Perform identity and permission authentication on the user, and determine data restoration permission; An auxiliary encryption protocol is used to activate the target bidirectional compromise channel of the global anonymization model, perform the restoration process of the anonymized data, and generate the restoration result.

8. An anonymization system for federated learning, characterized in that, Anonymization processing method for federated learning as described in any one of claims 1-7, comprising: A communication channel establishment module is used to establish communication channels between each participant and the federated platform, and to configure anonymization processing strategies. The compromise channel establishment module is used to establish a two-way compromise channel between the data end and the anonymization end in combination with the anonymization processing strategy. The two-way compromise channel includes multiple parallel channels mapped to multi-level anonymity limits. The processing model training module is used to combine the communication channel, decentralize the two-way compromise channel to each participating party, perform adaptive anonymization processing training based on locally shared sensitive data, and generate a local anonymization processing model. A trusted model filtering module is used to combine the communication channel to send the local anonymization processing model back to the federated middle platform, execute the middle platform's trusted verification checkpoint, and filter the local trusted model. A multi-layer aggregation processing module is used to perform multi-layer aggregation processing based on the local trusted model to generate a global anonymized model, wherein each aggregation processing layer corresponds to a different source model category; An anonymization processing module is used to combine the communication channel to send the global anonymization model back to each participant and update the local anonymization processing model, and perform data anonymization processing.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images