Trusted data dynamic desensitization access method and system based on federal learning

Abstract

The invention relates to a credible data dynamic desensitization access method and system based on federated learning, and relates to the field of data processing, and the method comprises the steps: obtaining a data desensitization processing agent based on federated learning training; according to the number proportion of K types of desensitized data in the multiple sample desensitized data sets and the number proportion of desensitized data of P types of user permissions, a plurality of static contribution degrees of the multiple subjects are obtained through distribution; and obtaining a real-time trusted access data set and an access user set in the trusted data space, dynamically adjusting the plurality of static contribution degrees to obtain a plurality of dynamic contribution degrees, adjusting the user permissions of the plurality of subjects to obtain an access user permission set, and performing data desensitization processing and display when the user accesses the trusted data space. The technical problems that data desensitization processing is not accurate, and access control lacks flexibility and dynamics are solved, and fine management and security improvement of data desensitization access are achieved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method and system for dynamically desensitizing access to trusted data based on federated learning. Background Art

[0002] In today's digital age, data has become a vital asset for business and social development. However, with the massive growth and widespread use of data, data security and privacy protection issues are becoming increasingly prominent. In particular, ensuring data security and privacy while fully realizing its value in the process of data sharing and collaboration has become a key issue that needs to be addressed. Existing data desensitization methods often adopt a static approach, desensitizing data according to preset rules. While this method protects data privacy to a certain extent, it suffers from problems such as inaccurate desensitization and inability to adapt to dynamic data changes. Furthermore, existing data access control mechanisms are often based on fixed permission settings, lacking flexibility and dynamism, making it difficult to meet complex and changing data access needs. Summary of the Invention

[0003] The present invention aims to solve the technical problems of inaccurate data desensitization processing and lack of flexibility and dynamism in access control in the existing technology, and provides a trusted data dynamic desensitization access method and system based on federated learning to solve them.

[0004] The technical solution of the present invention to solve the above technical problems is as follows:

[0005] In a first aspect, the present invention provides a dynamic desensitizing access method for trusted data based on federated learning, the method comprising: obtaining a data desensitizing processing agent trained based on federated learning, wherein the data desensitizing processing agent is obtained by federated learning of multiple sample desensitized data sets of multiple subjects; allocating multiple static contribution degrees for each subject according to the proportion of the number of K-type desensitized data in each sample desensitized data set, and the proportion of the number of desensitized data with P-type user permissions, where K and P are positive integers; obtaining a real-time trusted access data set and an access user set in a trusted data space, dynamically adjusting each static contribution degree, and obtaining multiple dynamic contribution degrees; adjusting the user permissions of the multiple subjects according to the multiple dynamic contribution degrees, and obtaining an access user permission set, and when a user accesses the trusted data space, performing data desensitization processing and display according to the access user permissions by the data desensitizing processing agent.

[0006] In a second aspect, the present invention provides a dynamic desensitizing access system for trusted data based on federated learning, the system comprising: an access activation module for obtaining a data desensitizing processing agent trained based on federated learning, wherein the data desensitizing processing agent is obtained through federated learning of multiple sample desensitized data sets of multiple subjects; a weight allocation module for allocating multiple static contribution degrees of each subject according to the proportion of the number of K-type desensitized data in each sample desensitized data set and the proportion of the number of desensitized data with P-type user permissions, where K and P are positive integers; a dynamic adjustment module for obtaining a real-time trusted access data set and an access user set in a trusted data space, dynamically adjusting each static contribution degree to obtain multiple dynamic contribution degrees; a result display module for adjusting the user permissions of the multiple subjects according to the multiple dynamic contribution degrees to obtain an access user permission set, and when a user accesses the trusted data space, the data desensitizing processing and display are performed according to the access user permissions by the data desensitizing processing agent.

[0007] The beneficial effects of the present invention are: through federated learning to train the data desensitization processing intelligent agent, static contribution is allocated based on the proportion of the number of sample desensitized data sets, and the contribution is dynamically adjusted according to real-time access data, thereby adjusting user permissions, achieving refined management of data desensitization access and improved security, ensuring that users can only access desensitized data within their authority. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flowchart of the method for dynamic desensitized access to trusted data based on federated learning provided by the present invention.

[0009] Figure 2 This is a structural diagram of the trusted data dynamic desensitizing access system based on federated learning provided by the present invention.

[0010] Description of the reference numerals: access activation module 11 , weight allocation module 12 , dynamic adjustment module 13 , result display module 14 . DETAILED DESCRIPTION

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0012] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0013] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0014] Example 1:

[0015] like Figure 1 As shown, an embodiment of the present invention provides a method for dynamic desensitized access to trusted data based on federated learning, the method comprising:

[0016] S10: Obtain a data desensitization processing agent based on federated learning training, wherein the data desensitization processing agent is obtained by performing federated learning on multiple sample desensitization data sets of multiple subjects.

[0017] For example, a data desensitization agent trained through federated learning is first obtained. This agent is essentially a data processing model, whose construction relies on multiple sample desensitized datasets provided by multiple entities. These sample desensitized datasets contain various types of desensitized data. They are converted from original datasets based on certain desensitization rules, aiming to protect data privacy while preserving data value. Federated learning refers to a distributed machine learning method that enables collaborative training and optimization of a global model through local computation and model updates by multiple participating entities (or clients) without directly transmitting the original data. For example, suppose multiple medical institutions wish to jointly train a disease prediction model, but due to data privacy and regulatory restrictions, they cannot directly share patient data. In this case, they can use federated learning to each locally train the model on the desensitized patient dataset and regularly transmit the model updates to a central server for aggregation, thereby obtaining a globally optimal disease prediction model, namely the data desensitization agent. In this process, the amount and quality of data contributed by each entity will affect the training effect of the global model. Therefore, when building the data desensitization agent, it is necessary to fully consider the contributions of each entity.

[0018] S20: Based on the proportion of K-type desensitized data in each sample desensitized data set and the proportion of P-type desensitized data with user permissions, multiple static contributions are allocated to each subject, where K and P are positive integers.

[0019] Furthermore, static contributions are assigned based on the proportion of K-category masked data in multiple sample masked datasets provided by multiple subjects, as well as the proportion of masked data categorized by P-category user permissions. K-category masked data refers to a collection of masked data categorized by certain standards or characteristics, while P-category user permissions represent different levels of user access rights. Together, they form the basis for evaluating subject contributions. The specific process is as follows: First, the number of each type of masked data in each sample masked dataset and the number of masked data categorized by user permissions are calculated. Next, for each subject, the ratio of the number of K-category masked data provided by the subject to the total number of corresponding categories, as well as the ratio of the number of masked data categorized by user permissions to the total number of corresponding permission categories, is calculated. These ratios reflect the subject's contribution to data diversity and user permission coverage. Finally, by averaging these ratios and comprehensively considering them, each subject is assigned a static contribution, which is a quantitative indicator of the value and importance of the data provided by the subject during the training of the data masking agent. For example, if a medical institution provides a desensitized patient dataset that includes both rich disease type data (i.e., diverse K-type desensitized data) and a wide range of user permission levels (i.e., comprehensive P-type user permissions), the institution may receive a higher score in the static contribution distribution.

[0020] S30: Acquire a real-time trusted access data set and access user set in the trusted data space, and dynamically adjust each static contribution to obtain multiple dynamic contributions.

[0021] Specifically, to adapt to real-time data changes within the trusted data space and the dynamic adjustments to the access user set, previously assigned static contribution degrees are further dynamically adjusted. This process first involves obtaining a real-time trusted access dataset within the trusted data space. This dataset contains the masked data currently being accessed or requiring access within the space. This data is also categorized into K categories to reflect the diversity and real-time nature of the data. Furthermore, a real-time access user set is obtained, which contains information about the users currently requesting access to the data. This information includes their permission levels, i.e., category P user permissions. After obtaining this real-time information, the proportion of K-category masked data in the real-time access dataset and the proportion of category P user permissions in the access user set are recalculated. Changes in these proportions reflect real-time shifts in data access patterns and user needs. Furthermore, based on these real-time calculated ratios, the original static contribution degrees are adjusted to reflect the actual contributions of each entity in the current data access environment, thereby generating dynamic contribution degrees. The calculation of dynamic contribution degrees not only considers the real-time nature of the data but also the dynamic changes in user needs, making contribution assessment more accurate and responsive to actual conditions. For example, if the K-category desensitized data provided by a subject accounts for a high proportion in the real-time access data set, and this data is frequently accessed by users with high authority levels, the subject may receive a higher score in the dynamic contribution adjustment, reflecting its important contribution in the current data access environment.

[0022] S40: According to the multiple dynamic contribution degrees, the user rights of the multiple subjects are adjusted to obtain an access user permission set. When the user accesses the trusted data space, the data desensitization processing agent performs data desensitization processing and display according to the access user permissions.

[0023] Specifically, multiple dynamic contribution scores are used as a basis for adjusting the user permissions of multiple entities, thereby generating a more reasonable and secure set of user permissions. This process first evaluates and adjusts each entity's permission level based on the previously calculated dynamic contribution score. Dynamic contribution scores reflect the entity's actual contribution in a real-time data access environment and serve as a key basis for adjusting user permissions. During this adjustment process, dynamic data characteristics at the data layer are comprehensively considered, such as the real-time proportion of K-category masked data and changes in user access patterns, as well as the permissions of other users at the user layer. This ensures that permission adjustments meet the requirements for secure data masking access and reflect each entity's actual contribution to data sharing and access. Specifically, entities with higher dynamic contribution scores are granted higher user permissions, allowing them to access more or more sensitive data. Entities with lower dynamic contribution scores have their user permissions reduced, restricting their access to sensitive data. This permission adjustment mechanism based on dynamic contribution scores not only improves the refinement of data masking access but also ensures data access security. For example, if a subject provides high-quality desensitized data in real-time data access and is frequently accessed by users with high privilege levels, their dynamic contribution may be significantly improved, thereby obtaining higher user privileges and participating more comprehensively in data sharing and access. Other users who do not participate in federated learning, on the other hand, do not contribute data or model training to the system, so their privileges remain unchanged, which to a certain extent also protects the stability and security of the system. Finally, based on the adjusted user privilege set, when a user accesses the trusted data space, the data desensitization agent performs data desensitization processing and display based on the access user privileges, ensuring the secure sharing and effective use of data.

[0024] In a preferred embodiment, obtaining a data desensitization processing agent based on federated learning training includes: obtaining multiple sample desensitization data sets of multiple subjects performing federated learning; using the multiple sample desensitization data sets to train the data desensitization processing agent based on federated learning.

[0025] Specifically, the data desensitization agent collects multiple sample desensitized datasets from multiple participants in federated learning. These sample desensitized datasets are obtained by processing each participant's original dataset based on specific desensitization rules. The goal is to protect data privacy while preserving the data's key characteristics and value. During this process, each participant performs local training using its own sample desensitized dataset, then sends model updates to a central server for aggregation to generate a global data desensitization agent. This agent learns how to effectively desensitize data while protecting data privacy.

[0026] In a preferred embodiment, multiple sample desensitized data sets of multiple subjects performing federated learning are obtained, including: collecting multiple sample original data sets provided by multiple subjects; obtaining P-type user authority levels and K-type desensitized data, and desensitizing the K-type desensitized data in the multiple sample original data sets according to the P-type user authority levels to obtain multiple sample desensitized data sets; combining the multiple sample original data sets and the multiple sample desensitized data sets to obtain multiple sample desensitized data sets.

[0027] Furthermore, we can first collect multiple sample original data sets provided by multiple subjects. These data sets contain original information that has not been desensitized and may cover various sensitive fields. Then, clarify the P-type user authority level and the K-type desensitized data type. Among them, the P-type user authority level represents the user's level of authority to access the data, such as level 1 to 5. The higher the authority level, the higher the sensitivity and accuracy of the data the user can access. K-type desensitized data refers to the sensitive information categories in the data that need to be desensitized, such as age, height, income, etc.

[0028] Based on the above classification, for each sample original data set, differentiated desensitization processing is implemented on K-type desensitized data according to different P-type user authority levels. For example, for level 1 authority users, age data may be fuzzified in the form of "X±50%" to generate highly generalized desensitized data to fully protect data privacy; while for level 5 authority users, a desensitization strategy of "X±10%" may be adopted to retain more data details to meet the data accuracy requirements of high-level users.

[0029] Through this process, multiple corresponding sample desensitized datasets are generated for each sample original dataset. These datasets protect data privacy while taking into account the data access needs of users with different permissions. Finally, the sample original dataset and the sample desensitized dataset are combined to form multiple sample desensitized datasets, providing comprehensive data support for the subsequent training of the federated learning model. For example, in a medical data sharing scenario, medical institutions of different levels (corresponding to different P-type user permission levels) may need to access medical records with different degrees of desensitization (corresponding to K-type desensitized data). The above process can ensure that each medical institution obtains appropriately desensitized medical record data in accordance with its own permissions, which not only ensures data security but also promotes the reasonable sharing and use of medical data.

[0030] In a preferred embodiment, according to the proportion of the number of K-type desensitized data in each sample desensitized data set and the proportion of the number of desensitized data with P-type user permissions, multiple static contribution degrees of each subject are allocated, including: extracting the total number of K-type desensitized data in all sample desensitized data sets to obtain the total number of K data; extracting the total number of data desensitized according to P-type user permissions in all sample desensitized data sets to obtain the total number of P permissions; in each sample desensitized data set, extracting the number of K-type desensitized data and the number of data desensitized according to P-type user permissions to obtain A data quantity set and a permission quantity set for each subject, wherein each data quantity set includes K data quantities, and each permission quantity set includes P permission quantities; respectively calculating the ratio of the K data quantities of each subject to the total number of K data quantities to obtain K data quantity ratios, and calculating the average to obtain the data contribution of each subject; respectively calculating the ratio of the P permission quantities of each subject to the total number of P permissions to obtain P permission quantity ratios, and calculating the average to obtain the permission contribution of each subject; based on the data contribution and permission contribution of each subject, calculate and obtain multiple static contributions of each subject.

[0031] In detail, in order to quantify the contribution of each subject in the training process of the data desensitization agent, it is necessary to assign a static contribution to each subject based on the proportion of K-type desensitized data in each sample desensitized data set and the proportion of desensitized data with P-type user permissions. The specific process is as follows: the total number of K-type desensitized data is extracted from all sample desensitized data sets to obtain the total number of K data. These total numbers reflect the distribution scale of each type of desensitized data in the overall data set; at the same time, for P-type user permissions, the total number of desensitized data under each permission level is counted separately, that is, the total number of permissions is obtained. In other words, the total amount of each type of desensitized data under each permission is counted separately according to P-type permissions. Since each sample desensitized data is processed according to P-type permissions during training, it can be understood as the total amount statistics classified by permission in each type of desensitized data.

[0032] Next, for each subject, the number of K-type desensitized data and the number of data desensitized according to P-type user permissions are extracted from the sample desensitized data set provided by the subject, forming a data quantity set and a permission quantity set for each subject. Each data quantity set contains K data quantities, corresponding to the specific number of K-type desensitized data in the subject's data set; each permission quantity set contains P permission quantities, corresponding to the number of desensitized data classified by P-type permissions in the subject's data set.

[0033] Next, the ratio of each subject's K data items to the total number of corresponding K data items is calculated, resulting in the K data volume ratios. The mean of these ratios is then calculated as the subject's data contribution, a metric that measures the subject's relative contribution to providing each type of masked data. Similarly, the ratio of each subject's P permissions to the total number of corresponding P permissions is calculated, resulting in the P permission quantity ratios. The mean of these ratios is then calculated to obtain the permission contribution, a metric that reflects the subject's contribution to providing masked data at different permission levels. Finally, by comprehensively considering the data contribution and permission contribution of multiple subjects, a static contribution is obtained for each subject through weighted calculation or averaging (the specific calculation method can be set based on actual needs or an appropriate aggregation strategy). For example, in the scenario of building a disease prediction model using cross-hospital federated learning, the amount and permission distribution of masked patient data (such as K types of masked data such as age and blood pressure) provided by different hospitals vary. Through this process, each hospital's data contribution and permission support during model training can be accurately quantified, providing a basis for subsequent dynamic adjustment of contribution and permission allocation.

[0034] In a preferred embodiment, a real-time trusted access data set and an access user set in a trusted data space are obtained, and each static contribution degree is dynamically adjusted to obtain multiple dynamic contribution degrees, including: obtaining a real-time trusted access data set in the trusted data space, and obtaining an access user set that records access in the trusted data space, wherein the access user set includes the multiple subjects; and dynamically adjusting and calculating the multiple static contribution degrees based on the trusted access data set and the access user set to obtain multiple dynamic contribution degrees.

[0035] Optionally, to ensure that contribution evaluation is compatible with dynamic changes in data access, the static contribution needs to be dynamically adjusted. Specifically, a trusted access dataset is obtained from the trusted data space in real time. This dataset covers the anonymized data records that are currently being accessed or to be accessed. Its data type, access frequency and other characteristics reflect the real-time needs and patterns of data access. At the same time, the set of access users that records access behavior in the trusted data space is obtained. This set includes not only ordinary access users, but also multiple entities that previously participated in the static contribution allocation, and records key information such as the time when the user accessed the data and the level of authority.

[0036] The trusted access dataset and access user set are the core inputs to the dynamic adjustment process. The former reflects the real-time dynamics of data access, while the latter reflects the distribution and characteristics of user access behavior. Based on this real-time data, dynamic adjustment calculations are performed for multiple static contribution metrics. Specifically, by analyzing the real-time access popularity of various types of desensitized data in the trusted access dataset, the access activity, and permission usage of different users within the access user set, each entity's actual contribution to data sharing and access is reassessed. For example, if the desensitized data provided by a certain entity has been frequently accessed recently, and the proportion of access to their data by high-privileged users within the access user set has significantly increased, the entity's contribution may increase accordingly during the dynamic adjustment process. Conversely, if the access popularity of a certain entity's data decreases, or if its data is primarily accessed by low-privileged users, its contribution may decrease. Through this series of calculations and analyses based on real-time data, multiple dynamic contribution metrics are ultimately obtained, accurately quantifying the contribution of each entity to the dynamic changes in data access. This provides a more timely and accurate basis for subsequent permission allocation and data desensitization. For example, in the financial data sharing scenario, the access popularity of desensitized customer data provided by different financial institutions (corresponding to multiple entities) will change with factors such as the market environment and business needs. Through the above-mentioned dynamic adjustment process, the actual contribution of each institution in data sharing can be reflected in a timely manner, ensuring that data access rights match the contribution level, and ensuring the security and efficiency of data sharing.

[0037] In a preferred embodiment, each static contribution is dynamically adjusted and calculated based on the trusted access data set and the access user set to obtain multiple dynamic contributions, including: extracting the total number of K real-time data of K types of desensitized data in the trusted access data set, and extracting the number of P real-time permissions of P types of user permissions in the access user set; calculating the change amplitude of the K total number of real-time data and the K total number of data to obtain K data volume change amplitudes, adjusting the K data volume ratios of each subject to obtain K adjusted data volume ratios, and calculating the average to obtain the adjusted data contribution of each subject; calculating the ratio of each real-time permission quantity to the average of P real-time permission quantities to obtain P permission coefficients, adjusting the P permission quantity ratios of each subject to obtain P adjusted permission quantity ratios, and calculating the average to obtain the adjusted permission contribution of each subject; according to the multiple adjusted data contribution degrees and multiple adjusted permission contribution degrees of each subject, multiple dynamic contribution degrees of multiple subjects are calculated.

[0038] For example, in a federated learning-based trusted data dynamic masked access mechanism, to adapt to dynamic changes in data access, the static contribution level needs to be dynamically adjusted based on the trusted access dataset and the access user set. First, the total number of K real-time data points for K types of masked data is extracted from the trusted access dataset. These numbers reflect the actual distribution scale of each type of masked data during the current access cycle. Simultaneously, the number of P real-time permissions for P types of user permissions is extracted from the access user set. These numbers reflect the degree of participation of users of different permission levels in data access.

[0039] The total amount of real-time data and the number of real-time permissions constitute the core data foundation for dynamic adjustments. Subsequently, for K categories of masked data, the change in the total amount of real-time data for each category compared to the previously calculated K total amounts is calculated, yielding the K data volume change magnitudes. For example, if the total amount of real-time data for a particular masked data category increases by 10%, an adjustment factor of 110% is multiplied by the original data volume ratio for the corresponding subject to reflect the impact of increased data access popularity on the contribution. Conversely, if the data volume decreases by 10%, a 90% adjustment factor is used. This adjustment calculation yields K adjusted data volume ratios for each subject, and the average is calculated to yield the adjusted data contribution, which quantifies the subject's actual contribution to the change in data access volume. Furthermore, for P categories of user permissions, the ratio of the number of real-time permissions for each permission level to the average of the P real-time permissions is calculated, yielding the P permission coefficients. For example, if the number of real-time permissions for a permission level is 10% higher than the average, its permission coefficient is 1.1; if it is 10% lower than the average, its permission coefficient is 0.9. Subsequently, these authority coefficients are multiplied by the original authority quantity ratios of each subject at the corresponding authority level to obtain P adjusted authority quantity ratios, and the average is calculated to obtain the adjusted authority contribution, which reflects the contribution of the subject in the changes in user access behavior at different authority levels. Finally, the adjusted data contribution and adjusted authority contribution of multiple subjects are comprehensively considered, and multiple dynamic contributions are obtained by weighted summation or calculation of the mean (the specific method can set weights or select appropriate aggregation strategies according to actual needs). For example, suppose that the ratio of the amount of data provided by a subject in the original static contribution calculation is 20% in Class A desensitized data and 30% in Class B desensitized data. At the same time, the ratio of the authority quantity of the subject's data among level 1 permission users is 25%, and the ratio of the authority quantity among level 2 permission users is 35%. If the current access volume of Class A desensitized data increases by 10%, the access volume of Class B desensitized data decreases by 5%, and the number of users with level 1 permissions is 10% higher than the average, and the number of users with level 2 permissions is 5% lower than the average, then the adjusted data contribution of the subject is (1.1×20%+0.95×30%) / 2, and the adjusted permission contribution is (1.1×25%+0.95×35%) / 2. The final dynamic contribution is calculated based on the weighted or average of the two, realizing dynamic optimization of the static contribution and ensuring that the contribution assessment matches the dynamic changes in data access.

[0040] In a preferred embodiment, the user rights of the multiple subjects are adjusted according to the multiple dynamic contribution degrees to obtain an access user permission set. When a user accesses the trusted data space, the data desensitization processing agent performs data desensitization processing and display according to the access user rights, including: calculating the ratio of the dynamic contribution degree of each subject to the minimum dynamic contribution degree to obtain a permission adjustment coefficient; using the permission adjustment coefficient to adjust and round the user permission level of each subject to obtain adjusted user rights; combining the adjusted user rights of each subject and the user rights of other users to obtain an access user permission set; according to the access user permission set, when a user accesses the trusted data space, the access user rights and access data are input into the data desensitization processing agent to perform data desensitization processing to obtain desensitized data for display.

[0041] Specifically, in order to adapt to the dynamic contribution of each subject in data sharing, it is necessary to adaptively adjust the subject user rights based on multiple dynamic contribution degrees. First, calculate the ratio of each subject's dynamic contribution to the minimum dynamic contribution among all subjects to generate multiple permission adjustment coefficients. Among them, the permission adjustment coefficient is a key indicator to measure the extent of the subject's permission adjustment, and its value reflects the relative advantage of the subject compared to the lowest contributor. For example, if the dynamic contribution of a subject is 0.8 and the minimum dynamic contribution is 0.4, the permission adjustment coefficient of the subject is 2, indicating that its contribution is twice that of the lowest contributor, and it should obtain a higher permission increase.

[0042] These permission adjustment coefficients are then used to recalculate the user permission levels of multiple entities and rounded to obtain multiple adjusted user permissions. This step aims to differentially enhance data access permissions based on the entity's dynamic contribution, ensuring that high-contributing entities have access to more accurate and comprehensive masked data to support their business needs or research objectives. Specifically, if a entity's original permission level is 2 and its permission adjustment coefficient is 1.5, the adjusted permission level may be increased to Level 3 (this is determined by combining the rounding rule). For example, suppose there are three entities, A, B, and C, with dynamic contributions of 0.6, 0.8, and 1.0, respectively, and the minimum dynamic contribution is 0.6. The permission adjustment coefficient for entity A is 1, for entity B is 1.33, and for entity C is 1.67. If the rounding rule is round up and the original permission level is 2, then after the adjustment, entity A's permission level may remain at Level 2, entity B's permission level may be increased to Level 3, and entity C's permission level may be increased to Level 3. (If the rounding rule is stricter, such as rounding up, entity C may be increased to Level 4, depending on the rounding threshold.)

[0043] The adjusted user permissions are then combined with those of other users who do not participate in federated learning but have fixed permissions to form a set of user permissions. This set comprehensively covers all users who may access the trusted data space and their corresponding permissions. Finally, when a user accesses the trusted data space, the user's permission information and the data to be accessed are input into the data desensitization agent. The agent performs differentiated desensitization on the data based on the user permission set. For example, a high-privilege user may access desensitized data with an accuracy of "X ± 10%", while a low-privilege user may only access data with an accuracy of "X ± 50%". The desensitized data is then displayed to the user, ensuring data privacy and security while meeting the access needs of users with different permissions.

[0044] The method for dynamic desensitized access to trusted data based on federated learning provided by the embodiments of the present invention has at least the following technical effects:

[0045] 1. By integrating sample desensitized datasets from multiple subjects for federated learning, a data desensitization processing agent with cross-subject data understanding capabilities is trained. This breaks through the limitations of traditional desensitization methods in data silos and generalization capabilities, enabling the agent to accurately adapt to the data characteristics and desensitization requirements of different subjects. While ensuring data privacy, it maximizes data availability and provides a safe and efficient solution for cross-institutional data collaboration.

[0046] 2. A static contribution allocation mechanism has been established based on the proportion of desensitized data types and the distribution of user permissions. This mechanism further combines real-time access data and user behavior within the trusted data space to dynamically adjust the contribution of each entity. By quantifying the linkage between contribution and permissions, a differentiated and adaptive allocation of user permissions is achieved. For example, high-contributing entities can obtain more granular data access rights, while low-contributing entities are restricted to basic permissions. This creates a positive incentive loop of "contribution-permission" in data sharing, effectively improving the fairness and sustainability of the data sharing ecosystem.

[0047] 3. Based on the access user's permission set, the data desensitization agent can analyze user permissions and the sensitivity of the accessed data in real time, dynamically generating differentiated desensitization strategies. For example, a high-privilege user may access desensitized data that "preserves numerical distribution patterns but hides specific values," while a low-privilege user only obtains summary information with "fuzzy interval ranges." This fine-grained permission-aware desensitization mechanism not only avoids the loss of business value caused by "one-size-fits-all" desensitization, but also prevents the risk of data misuse through precise permission control, providing technical support for secure access and compliant sharing in trusted data spaces.

[0048] Example 2:

[0049] like Figure 2As shown, based on the same inventive concept as the method for dynamic desensitized access to trusted data based on federated learning provided in Example 1, an embodiment of the present invention also provides a dynamic desensitized access system for trusted data based on federated learning, the system comprising:

[0050] The access activation module 11 is used to obtain a data desensitization processing agent based on federated learning training, wherein the data desensitization processing agent is obtained by performing federated learning on multiple sample desensitization data sets of multiple subjects.

[0051] The weight allocation module 12 is used to allocate multiple static contribution degrees of each subject according to the proportion of the number of K-type desensitized data in each sample desensitized data set and the proportion of the number of desensitized data with P-type user permissions, where K and P are positive integers.

[0052] The dynamic adjustment module 13 is used to obtain the real-time trusted access data set and access user set in the trusted data space, and dynamically adjust each static contribution to obtain multiple dynamic contributions.

[0053] The result display module 14 is used to adjust the user rights of the multiple subjects according to the multiple dynamic contribution degrees, obtain the access user permission set, and when the user accesses the trusted data space, the data desensitization processing agent performs data desensitization processing and display according to the access user permissions.

[0054] Furthermore, the access activation module 11 is further configured to perform the following steps:

[0055] Acquire multiple sample desensitized data sets of multiple subjects performing federated learning; and train a data desensitization processing agent based on federated learning using the multiple sample desensitized data sets.

[0056] Furthermore, the access activation module 11 is further configured to perform the following steps:

[0057] Collect multiple sample original data sets provided by multiple subjects; obtain P-type user authority levels and K-type desensitized data, and desensitize the K-type desensitized data in the multiple sample original data sets according to the P-type user authority levels to obtain multiple sample desensitized data sets; combine the multiple sample original data sets and the multiple sample desensitized data sets to obtain multiple sample desensitized data sets.

[0058] Furthermore, the weight allocation module 12 is further configured to perform the following steps:

[0059] Extract the total number of K types of desensitized data in all sample desensitized data sets to obtain the total number of K data; extract the total number of data desensitized according to P types of user permissions in all sample desensitized data sets to obtain the total number of P permissions; in each sample desensitized data set, extract the number of K types of desensitized data and the number of data desensitized according to P types of user permissions to obtain multiple data quantity sets and multiple permission quantity sets for each subject, wherein each data quantity set includes K data quantities, and each permission quantity set includes P permission quantities; calculate the ratio of the K data quantities of each subject to the K total data quantities to obtain K data quantity ratios, and calculate the average to obtain the data contribution of each subject; calculate the ratio of the P permission quantities of each subject to the P total permission quantities to obtain P permission quantity ratios, and calculate the average to obtain the permission contribution of each subject; according to the multiple data contribution degrees and multiple permission contribution degrees of each subject, calculate and obtain multiple static contribution degrees of each subject.

[0060] Furthermore, the dynamic adjustment module 13 is further configured to perform the following steps:

[0061] Obtain a real-time trusted access data set in the trusted data space, and obtain an access user set that records access in the trusted data space, wherein the access user set includes the multiple subjects; based on the trusted access data set and the access user set, dynamically adjust and calculate the multiple static contribution degrees to obtain multiple dynamic contribution degrees.

[0062] Furthermore, the dynamic adjustment module 13 is further configured to perform the following steps:

[0063] Extract the total number of K real-time data of K types of desensitized data in the trusted access data set, and extract the number of P real-time permissions of P types of user permissions in the access user set; calculate the change range of the K total number of real-time data and the K total number of data to obtain K data volume change ranges, adjust the K data volume ratios of each subject to obtain K adjusted data volume ratios, and calculate the average to obtain the adjusted data contribution of each subject; calculate the ratio of each real-time permission number to the average of P real-time permission numbers to obtain P permission coefficients, adjust the P permission number ratios of each subject to obtain P adjusted permission number ratios, and calculate the average to obtain the adjusted permission contribution of each subject; according to the multiple adjusted data contribution degrees and multiple adjusted permission contribution degrees of each subject, calculate and obtain multiple dynamic contribution degrees of multiple subjects.

[0064] Furthermore, the result display module 14 is further configured to perform the following steps:

[0065] Calculate the ratio of each subject's dynamic contribution to the minimum dynamic contribution to obtain multiple authority adjustment coefficients; use the authority adjustment coefficient to adjust and round the user authority level of each subject to obtain multiple adjusted user authorities; combine the adjusted user authorities of each subject and the user authorities of other users to obtain an access user authority set; according to the access user authority set, when the user accesses the trusted data space, the access user authority and access data are input into the data desensitizing processing intelligent body, and data desensitization is performed to obtain desensitized data for display.

[0066] Through the detailed description of the dynamic desensitizing access method for trusted data based on federated learning in the foregoing specification, those skilled in the art can clearly understand the dynamic desensitizing access system for trusted data based on federated learning in this embodiment. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0067] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic desensitized access method for trusted data based on federated learning, characterized by: The method comprises: Obtaining a data desensitization processing agent based on federated learning training, wherein the data desensitization processing agent is obtained by performing federated learning on multiple sample desensitization data sets of multiple subjects; According to the proportion of K-type desensitized data in each sample desensitized data set and the proportion of P-type desensitized data with user permissions, the static contribution of each subject is obtained, where K and P are positive integers. Acquire a real-time trusted access data set and access user set in the trusted data space, dynamically adjust each static contribution, and obtain multiple dynamic contributions; According to the multiple dynamic contribution degrees, the user rights of the multiple subjects are adjusted to obtain an access user permission set. When the user accesses the trusted data space, the data desensitization processing agent performs data desensitization processing and display according to the access user permissions.

2. The method for dynamic desensitized access to trusted data based on federated learning according to claim 1 is characterized in that: Obtain a data desensitization agent trained based on federated learning, including: Obtain multiple sample desensitized datasets from multiple subjects for federated learning; The multiple sample desensitized data sets are used to train a data desensitization processing agent based on federated learning.

3. The method for dynamic desensitized access to trusted data based on federated learning according to claim 2 is characterized in that: Obtain multiple sample desensitized datasets from multiple entities for federated learning, including: Collect multiple sample raw data sets provided by multiple subjects; Obtaining a P-type user authority level and a K-type desensitized data, and performing desensitization processing on the K-type desensitized data in the multiple sample original data sets according to the P-type user authority level, to obtain multiple sample desensitized data sets; The multiple sample original data sets and the multiple sample desensitized data sets are combined to obtain multiple sample desensitized data sets.

4. The method for dynamic desensitized access to trusted data based on federated learning according to claim 1 is characterized in that: Based on the proportion of K-type desensitized data in each sample desensitized data set and the proportion of P-type desensitized data with user permissions, the static contribution of each subject is allocated, including: Extract the total number of K types of desensitized data in all sample desensitized data sets to obtain the total number of K data; Extract the total amount of data in all sample desensitized data sets that have been desensitized according to P types of user permissions, and obtain the total number of P permissions; In each sample desensitized data set, extract the number of K types of desensitized data and the number of data desensitized according to P types of user permissions, and obtain the data quantity set and permission quantity set of each subject, where each data quantity set includes K data quantities and each permission quantity set includes P permission quantities; Calculate the ratio of the number of K data of each subject to the total number of K data, obtain K data volume ratios, and calculate the average to obtain the data contribution of each subject; Calculate the ratio of each subject's P permissions to the total number of P permissions, obtain the P permission ratios, and calculate the average to obtain the permission contribution of each subject; Based on the data contribution and authority contribution of each subject, the static contribution of each subject is calculated.

5. The method for dynamic desensitized access to trusted data based on federated learning according to claim 4 is characterized in that: Obtain real-time trusted access data sets and access user sets in the trusted data space, dynamically adjust each static contribution, and obtain multiple dynamic contributions, including: Acquire a real-time trusted access data set in the trusted data space, and acquire an access user set that records access in the trusted data space, wherein the access user set includes the multiple subjects; According to the trusted access data set and the access user set, the multiple static contribution degrees are dynamically adjusted and calculated to obtain multiple dynamic contribution degrees.

6. The method for dynamic desensitized access to trusted data based on federated learning according to claim 5 is characterized in that: According to the trusted access data set and the access user set, each static contribution is dynamically adjusted and calculated to obtain multiple dynamic contribution degrees, including: Extracting the total number of K real-time data of K types of desensitized data in the trusted access data set, and extracting the number of P real-time permissions of P types of user permissions in the access user set; Calculate the change range of the K real-time data totals and the K data totals to obtain K data volume change ranges, adjust the K data volume ratios of each subject to obtain K adjusted data volume ratios, and calculate the average to obtain the adjusted data contribution of each subject; Calculate the ratio of each real-time permission to the average of P real-time permissions to obtain P permission coefficients. Adjust the P permission ratios of each subject to obtain P adjusted permission ratios. Calculate the average to obtain the adjusted permission contribution of each subject. Based on the adjustment data contribution and adjustment authority contribution of each subject, multiple dynamic contributions of multiple subjects are calculated.

7. The method for dynamic desensitized access to trusted data based on federated learning according to claim 1 is characterized in that: According to the multiple dynamic contribution degrees, user permissions of the multiple subjects are adjusted to obtain an access user permission set. When the user accesses the trusted data space, the data desensitization processing agent performs data desensitization processing and display according to the access user permissions, including: Calculate the ratio of each subject's dynamic contribution to the minimum dynamic contribution to obtain the authority adjustment coefficient; Using the permission adjustment coefficient, the user permission level of each subject is adjusted and calculated to obtain the adjusted user permission; Combine the adjustment user permissions of each subject with the user permissions of other users to obtain the access user permission set; According to the access user permission set, when a user accesses the trusted data space, the access user permission and access data are input into the data desensitizing processing agent, and data desensitization is performed to obtain desensitized data for display.

8. A dynamic desensitized access system for trusted data based on federated learning, characterized by: A system for implementing the method for dynamic desensitized access to trusted data based on federated learning according to any one of claims 1 to 7, comprising: An access activation module is used to obtain a data desensitization processing agent trained based on federated learning, wherein the data desensitization processing agent is obtained by performing federated learning on multiple sample desensitization data sets of multiple subjects; The weight allocation module is used to allocate multiple static contributions to each subject based on the proportion of K-type desensitized data in each sample desensitized data set and the proportion of P-type desensitized data with user permissions. K and P are positive integers. The dynamic adjustment module is used to obtain the real-time trusted access data set and access user set in the trusted data space, dynamically adjust each static contribution, and obtain multiple dynamic contributions; The result display module is used to adjust the user rights of the multiple subjects according to the multiple dynamic contribution degrees, obtain the access user permission set, and when the user accesses the trusted data space, the data desensitization processing agent performs data desensitization processing and display according to the access user permissions.

Images