A privacy computing service method and system based on federated learning

Abstract

The application provides a privacy computing service method and system based on federated learning. The method identifies the sensitive field type of the original data of the federated learning participant, divides the privacy level, and generates desensitized data through local masking and superposition of dynamic noise. Based on the complexity of the encryption algorithm, the data is divided into data blocks in proportion to the number of operation units for parallel encryption, transmitted to the collaboration node through the encrypted channel to obtain the aggregation result, and the time consumption and privacy intensity value are recorded synchronously. The data block segmentation number is dynamically adjusted by calculating the time consumption difference value, and the correlation parameters are generated by cross-validation with the privacy intensity value. When the privacy intensity is insufficient, the dynamic noise of the corresponding level is improved, and the data block segmentation is optimized synchronously, forming a dynamic linkage mechanism of noise enhancement and block encryption. Through the dynamic mutual feedback mechanism of difference value and privacy intensity, the application realizes the real-time balance of data desensitization intensity and encryption calculation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of privacy computing services, in particular to a privacy computing service method and system based on federated learning. BACKGROUND

[0002] In the cross-institutional multi-party joint modeling scenario, the participants need to achieve joint modeling under the premise of protecting data privacy, while meeting the needs of high security and low communication overhead. Specifically, the data of the participating institutions needs to be strictly stored locally, and the transmission of raw data is prohibited to avoid violating privacy regulations; at the same time, the model training needs to maintain efficient convergence in the heterogeneous data distribution scenario, and the parameter transmission volume needs to be greatly compressed in the communication process to reduce the network bandwidth pressure and avoid wasting computing resources due to the high complexity of encryption algorithms.

[0003] The current mainstream solution is a federated learning secure aggregation protocol based on homomorphic encryption and model compression. This solution encrypts local model parameters through homomorphic encryption technology to ensure that the gradients or parameters uploaded by the participants complete aggregation operations in the ciphertext state, thereby preventing servers or third parties from obtaining raw data privacy through reverse deduction; at the same time, model parameter compression technology is combined to perform lightweight processing on the communication content, and the compression rate is dynamically adjusted to balance model accuracy and transmission efficiency. In addition, some solutions introduce differential privacy mechanisms to add controllable noise to the gradient during local model training to resist attacks based on parameter reverse data features, and adjust the noise intensity through a privacy budget to balance model availability. SUMMARY

[0004] The present application provides a privacy computing service method and system based on federated learning to solve the problem of imbalance between privacy protection and computing efficiency in federated learning in the prior art.

[0005] In a first aspect, the present application provides a privacy computing service method based on federated learning, comprising:

[0006] Obtaining raw data of a federated learning participant, identifying the type of sensitive fields in the raw data, dividing the sensitive fields into different privacy levels according to the type, and performing local masking on the content of the sensitive fields and superimposing dynamic noise on the locally masked area according to the privacy level to generate desensitized data;

[0007] Dividing the desensitized data into multiple data blocks through a preset encryption algorithm complexity, and distributing the data blocks to independent operation units for parallel encryption processing, the number of operation units being proportional to the encryption algorithm complexity to form an operation unit distribution strategy;

[0008] transmit the encrypted data block to the federated learning collaboration node through an encrypted transmission channel, and obtain an aggregation calculation result fed back by the federated learning collaboration node, and synchronize a time consumption of the encryption operation and a privacy intensity value corresponding to the aggregation calculation result;

[0009] calculate a difference value between the time consumption of the encryption operation and a preset threshold value, and adjust the segmentation number of the desensitized data block according to the difference value to match the operation unit allocation strategy, and cross-verify the difference value and the privacy intensity value to generate an association parameter of the operation unit allocation strategy and the privacy intensity value;

[0010] based on the association parameter, when the privacy intensity value is lower than a preset intensity, enhancing the dynamic noise corresponding to the privacy level and synchronously adjusting the segmentation number of the desensitized data block, so that the desensitized data block after the dynamic noise is enhanced is re-segmented and encrypted, to form a dynamic linkage mechanism.

[0011] Optionally, the original data of the federated learning participant is obtained, the type of the sensitive field in the original data is identified, the sensitive field is divided into different privacy levels according to the type, and the content of the sensitive field is locally shielded and superimposed with dynamic noise in the locally shielded area according to the privacy level, to generate desensitized data, including:

[0012] The original data provided by the federated learning participant is obtained, and the fields of the original data are matched with a preset sensitive field type list one by one, the preset sensitive field type list includes the types of identity and location trajectory, and the fields that match successfully are marked as the type of the sensitive field.

[0013] The privacy risk weight table associated with the type of the sensitive field is obtained, and the fields containing the identity are divided into a first privacy level according to the privacy risk weight table, and the fields containing the location trajectory are divided into a second privacy level according to the density value calculated according to the number of trajectory points in the location trajectory when the density value exceeds a trajectory point threshold value.

[0014] The first three characters of the fields in the first privacy level are retained, and the subsequent characters are replaced with shielding symbols, and dynamic noise is superimposed on the shielding symbols one by one according to the positions of the first three characters.

[0015] The segmented shielding operation is performed on the trajectory field in the second privacy level, the segmented shielding operation splits the continuous trajectory points into segments according to the time interval, and shields the trajectory points at the end of each segment, while obtaining the fluctuation range of the numerical value of the trajectory points, and superimposes dynamic noise matching the fluctuation range in the shielded area.

[0016] The sensitive field after shielding and superimposing dynamic noise is reorganized according to the format of the original data to generate desensitized data.

[0017] Optionally, the de-sensitized data is divided into a plurality of data blocks by a preset encryption algorithm complexity, and the data blocks are distributed to independent operation units for parallel encryption processing, the number of the operation units being proportional to the encryption algorithm complexity, to form an operation unit distribution strategy, comprising:

[0018] The number of divisions of the de-sensitized data is determined according to a preset encryption algorithm complexity, the encryption algorithm complexity being defined by a hierarchical depth of encryption operation and a length of data block;

[0019] The de-sensitized data is cut into sequentially marked data blocks according to a number of divisions corresponding to a current encryption algorithm complexity level, the length of the data blocks being inversely proportional to the encryption algorithm complexity;

[0020] The data blocks are distributed to independent operation units according to the number of divisions, the operation units binding one data block and triggering encryption processing, the total number of the operation units being strictly consistent with the number of divisions, and the processing capacity of the operation units matching the length of the data blocks;

[0021] The real-time rate at which the operation units process the data blocks is monitored, and if the real-time rate is lower than a preset reference rate corresponding to the current encryption algorithm complexity, the data blocks are cut into sub-data blocks according to a preset splitting ratio, and new operation units are distributed to synchronously process the sub-data blocks, to form the operation unit distribution strategy.

[0022] Optionally, the real-time rate at which the operation units process the data blocks is monitored, and if the real-time rate is lower than a preset reference rate corresponding to the current encryption algorithm complexity, the data blocks are cut into sub-data blocks according to a preset splitting ratio, and new operation units are distributed to synchronously process the sub-data blocks, to form the operation unit distribution strategy, comprising:

[0023] The real-time rate at which the operation units process the data blocks is monitored, and the real-time rate is compared with a preset reference rate corresponding to the current encryption algorithm complexity on a unit-by-unit basis, to generate a rate difference state mark;

[0024] When it is detected that the rate difference state mark is lower than the preset reference rate, a splitting operation of the data blocks is triggered, the splitting operation cutting the current data blocks into a plurality of sub-data blocks according to a preset splitting ratio, the length of each sub-data block being inversely proportional to the splitting ratio;

[0025] The new operation units are distributed to the sub-data blocks to be processed, and each sub-data block is bound to an independent operation unit for encryption processing, to form the operation unit distribution strategy.

[0026] Optionally, the encrypted data block is transmitted to the federated learning collaboration node through an encrypted transmission channel, and an aggregation calculation result fed back by the federated learning collaboration node is obtained, a time consumption of the encryption operation corresponding to the aggregation calculation result is recorded, and a privacy intensity value is recorded, comprising:

[0027] A sequence identifier and a channel key are attached to the encrypted data block, the encrypted data block is sent to the federated learning collaboration node in batches according to the sequence identifier through the encrypted transmission channel, and the start and end transmission timestamps of each batch of data blocks are recorded when sending to record the time consumption of the encryption operation;

[0028] The continuity of the sequence identifier and the validity of the channel key of the encrypted data block are verified at the receiving end of the federated learning collaboration node, and the data block that passes the verification is subjected to aggregation calculation to generate an aggregation calculation result;

[0029] The privacy intensity value corresponding to the aggregation calculation result is calculated based on the encryption characteristics of the data block.

[0030] Optionally, the difference value between the time consumption of the encryption operation and the preset threshold is calculated, and the number of partitions of the desensitized data block is adjusted according to the difference value to match the operation unit allocation strategy, and the difference value and the privacy intensity value are cross-verified to generate an association parameter of the operation unit allocation strategy and the privacy intensity value, comprising:

[0031] The time consumption of the encryption operation is compared with the preset threshold point by point to calculate the percentage value of the time consumption of the encryption operation and the preset threshold as the difference value;

[0032] The deviation direction of the difference value is extracted, and the number of partitions of the desensitized data block is adjusted according to the deviation direction and the percentage value. If the difference value deviates positively, the number of partitions is increased in proportion to match the upper limit of the number of units of the operation unit allocation strategy. If the difference value deviates negatively, the number of partitions is reduced in proportion to release the redundant operation units of the operation unit allocation strategy;

[0033] The privacy intensity value is extracted, and the difference value of each data block is bound with the corresponding privacy intensity value one by one to generate a verification pair of the difference value and the privacy intensity value;

[0034] The correlation between the difference value and the privacy intensity value in the verification pair is analyzed. When the difference value deviates positively and the privacy intensity value is lower than the preset intensity, the number of partitions is increased and the positive correlation parameter of the privacy intensity value is increased. When the difference value and the privacy intensity value both meet the preset standard, a balance correlation parameter maintaining the current number of partitions and the privacy intensity is generated;

[0035] The positive correlation parameter and the balance correlation parameter are integrated to generate an association parameter.

[0036] Optionally, based on the association parameter, when the privacy intensity value is lower than the preset intensity, the dynamic noise of the corresponding privacy level is enhanced, and the number of partitions of the de-sensitized data block is adjusted synchronously, so that the de-sensitized data block after the dynamic noise enhancement is re-partitioned and encrypted to form a dynamic linkage mechanism, comprising:

[0037] checking whether the privacy intensity value marked in the association parameter is lower than the preset intensity, and when it is detected that the privacy intensity value is lower than the preset intensity, enhancing the dynamic noise of the field belonging to the corresponding privacy level in the de-sensitized data block to generate the de-sensitized data block after the noise enhancement;

[0038] calculating the length change of the data block after the dynamic noise is enhanced, and adjusting the number of partitions of the de-sensitized data block according to the length change according to the preset partitioning rule, to ensure that the length of the data block is inversely proportional to the number of partitions;

[0039] redistributing the de-sensitized data block after the number of partitions is adjusted to an independent operation unit for encryption processing, and each operation unit binds the corresponding data block according to the updated number of partitions and performs encryption processing to form a dynamic linkage mechanism.

[0040] In a second aspect, the present application provides a privacy computing service system based on federated learning, comprising:

[0041] an identification module configured to obtain original data of a federated learning participant, identify the type of a sensitive field in the original data, divide the sensitive field into different privacy levels according to the type, and perform local masking on the content of the sensitive field and superimpose dynamic noise on the locally masked area to generate de-sensitized data;

[0042] a distribution module configured to divide the de-sensitized data into a plurality of data blocks through a preset encryption algorithm complexity, and distribute the data blocks to independent operation units for parallel encryption processing, wherein the number of operation units is proportional to the encryption algorithm complexity, to form an operation unit distribution strategy;

[0043] a transmission module configured to transmit the encrypted data blocks to a federated learning collaboration node through an encrypted transmission channel, and obtain an aggregation calculation result fed back by the federated learning collaboration node, and synchronously record the time consumption of the encryption operation and the privacy intensity value corresponding to the aggregation calculation result;

[0044] a calculation module configured to calculate the difference value between the time consumption of the encryption operation and a preset threshold, adjust the number of partitions of the de-sensitized data block to match the operation unit distribution strategy according to the difference value, and cross-verify the difference value and the privacy intensity value to generate an association parameter of the operation unit distribution strategy and the privacy intensity value;

[0045] The forming module is configured to, based on the association parameter, when a privacy intensity value is lower than a preset intensity, increase dynamic noise corresponding to a privacy level, synchronously adjust a segmentation number of the desensitized data block, and re-segment and encrypt the desensitized data block after the dynamic noise is enhanced, so as to form a dynamic linkage mechanism.

[0046] In a third aspect, an embodiment of the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component, so as to realize the method for privacy computing service based on federated learning according to the first aspect.

[0047] In a fourth aspect, an embodiment of the present application provides a computer storage medium, which stores a computer program; when the computer program is executed by a computer, the method for privacy computing service based on federated learning according to the first aspect is realized.

[0048] In the technical solution of the present application, through the multi-layer protection mechanism of sensitive field hierarchical desensitization, elastic data encryption segmentation, dynamic resource scheduling and encrypted transmission verification, an adaptive framework of privacy intensity quantitative evaluation and dynamic linkage of encryption strategy is constructed. Based on the data block segmentation adjustment and noise enhancement compensation technology driven by the difference value, the conflict between privacy protection and model utility in the traditional static desensitization mode is broken through, and the privacy leakage risk is accurately prevented while the data availability is ensured. The cross verification mechanism of encryption operation efficiency and privacy intensity forms a dynamic balance of real-time feedback optimization, solves the problem that security and computing resource utilization rate are difficult to be considered in federated learning multi-node cooperation, and provides full-link closed-loop protection and elastic expansion capability for cross-domain data security sharing.

[0049] Further, through the preset sensitive field accurate identification and privacy level dynamic division technology, a differentiated management system of core privacy elements such as identity and location trajectory is constructed. By using the double strategies of pre-sequence character reservation masking and trajectory segmentation masking, the sensitive information is accurately hidden while the partial data availability is ensured. Combined with the dynamic noise superposition technology, the noise intensity is matched according to the masking area characteristics and the trajectory fluctuation range, which effectively solves the inherent contradiction between data distortion and privacy leakage in the traditional desensitization method. Through the sensitive field format reorganization technology, the desensitized data is compatible with the federated learning framework, and a new protection paradigm with privacy intensity and model utility is provided for multi-party data security cooperation.

[0050] These and other aspects of the application will become more apparent in the following description of embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to make the technical scheme of the present application or the prior art clearer, the accompanying drawings needed in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative effort based on these drawings.

[0052] Figure 1 A flow chart of a privacy computing service method based on federated learning provided by the present application is shown;

[0053] Figure 2 A scenario diagram of a privacy computing service method based on federated learning provided by the present application is shown;

[0054] Figure 3 A scenario diagram of a privacy computing service method based on federated learning provided by the present application is shown;

[0055] Figure 4 A structural schematic diagram of a privacy computing service system based on federated learning provided by the present application is shown;

[0056] Figure 5 A structural schematic diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0057] In order to make the technical scheme of the present application or the prior art clearer, the accompanying drawings needed in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative effort based on these drawings.

[0058] In some of the processes described in the specification and the claims of the present application and the above-mentioned accompanying drawings, a plurality of operations appearing in a specific order are included, but it should be clearly understood that these operations can be executed or in parallel without the order appearing in the text, and the serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and these operations can be executed in sequence or in parallel. It should be noted that the "first", "second", etc. in this text are used to distinguish different messages, devices, modules, etc., and do not represent the order, nor do "first" and "second" represent different types.

[0059] The research found that the current cross-agency federated learning security aggregation framework, although it realizes preliminary privacy protection and communication optimization through homomorphic encryption and parameter compression technology, has the core bottleneck that the static privacy protection mechanism and the adaptability of dynamic computing resource demand are insufficient, and the quantitative basis for the coordination and regulation of privacy strength and communication efficiency is lacking. On the one hand, although homomorphic encryption can guarantee the security of data aggregation in the ciphertext state, its inherent computational complexity and communication overhead grow exponentially (especially in the case of high-dimensional model parameters), leading to an exacerbation of the power imbalance problem between heterogeneous devices; on the other hand, the existing scheme uses a globally unified differential privacy noise injection strategy, which is difficult to adapt to the differentiated protection needs of sensitive fields with multiple privacy levels - for example, if the same noise intensity is used for diagnosis records (high privacy level) and demographic information (low privacy level) in medical data, the model accuracy may be reduced or the privacy protection may be insufficient due to excessive perturbation. In addition, although model compression technology can reduce the communication load of a single communication, the fixed compression rate setting cannot dynamically adapt to the changes in the correlation between encryption time and privacy strength of different data blocks, which can easily cause loss of key parameters or reduce convergence stability.

[0060] To address the above challenges, the present application proposes a federated learning dynamic encryption and noise linkage regulation method for multiple privacy levels, which is innovative in that it constructs a closed-loop regulation system of "privacy grading-parallel encryption-feedback correction" through the synergistic optimization of sensitive field grading masking and dynamic allocation of encryption computing power. Specifically, based on the division of sensitive field types, differential de-identification data is generated by combining local masking and dynamic noise superposition; further, the de-identified data is divided into data blocks for parallel processing according to the encryption algorithm complexity, and the computing power resources are dynamically allocated based on the number of operation units, realizing decoupling control of encryption efficiency and privacy strength. By real-time monitoring of the cross-validation indicators of encryption time and privacy strength, the number of data block segmentation and the noise injection strength are dynamically adjusted, forming a multi-objective balance mechanism of "computing power consumption-privacy protection-communication load". This method breaks through the static privacy-efficiency trade-off dilemma of traditional federated learning: on the one hand, the local masking technique realizes fine-grained information hiding for high-privacy-level fields, reducing the negative impact of global noise injection on model accuracy; on the other hand, the parallel encryption strategy based on data block segmentation significantly reduces the computational delay of homomorphic encryption by dynamically adjusting the segmentation granularity to match the computing power distribution of heterogeneous devices. More importantly, by modeling the correlation parameters of encryption time and privacy strength, a linkage feedback mechanism of noise strength and segmentation number is established, enabling the system to adaptively optimize privacy protection strength and communication efficiency according to real-time performance indicators, realizing a technological leap from "fixed compression rate" to "dynamic resource adaptation" and from "global noise perturbation" to "multi-level masking enhancement", providing a solution for cross-agency joint modeling that balances high security, low communication overhead, and high model availability.

[0061] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0062] Figure 1 A flowchart of a privacy computing service method based on federated learning is provided for the embodiments of the present application, as shown in Figure 1 The method comprises the following steps.

[0063] 101. Obtain the original data of a federated learning participant, identify the types of sensitive fields in the original data, divide the sensitive fields into different privacy levels according to the types, and perform local masking on the contents of the sensitive fields and superimpose dynamic noise on the masked area according to the privacy levels to generate desensitized data.

[0064] Optionally, step 101 can specifically include the following steps.

[0065] 1011. Obtain the original data provided by the federated learning participant, match the fields of the original data with a preset sensitive field type list one by one, the preset sensitive field type list contains the types of identity and location trajectory, and mark the matched fields as the types of sensitive fields;

[0066] 1012. Obtain the privacy risk weight table associated with the types of the sensitive fields, and divide the fields containing identity into the first privacy level according to the privacy risk weight table, and divide the fields containing location trajectory into the second privacy level according to the density value calculated according to the number of track points in the location trajectory when the density value exceeds the track point threshold;

[0067] 1013. Keep the first three characters of the fields in the first privacy level and replace the subsequent characters with masking symbols, and superimpose dynamic noise on the masking symbols one by one according to the positions of the first three characters;

[0068] 1014. Perform a segmented masking operation on the track fields in the second privacy level, the segmented masking operation splits the continuous track points into segments according to time intervals, and masks the track points at the end of each segment, while obtaining the fluctuation range of the values of the track points, and superimposes dynamic noise matching the fluctuation range on the masked area;

[0069] 1015. Reorganize the sensitive fields with completed masking and superimposed dynamic noise according to the format of the original data to generate desensitized data.

[0070] In the above scheme, the federated learning participant is a data holder or individual participating in collaborative machine learning. The raw data is a set of unprocessed user information. The type of sensitive field is a classification category of information in the data related to personal privacy. The preset sensitive field type list is a list of predefined sensitive data types such as identity and location trajectory. The identity is an information element that can directly or indirectly identify the identity of an individual. The location trajectory is time series data recording the moving path of an individual. The privacy risk weight table is a table of quantitative indicators established according to the degree of privacy leakage risk associated with the type of sensitive field. The first privacy level is the high-risk privacy protection level corresponding to the identity field. The second privacy level is the medium-risk privacy protection level of the location trajectory field according to the trajectory point density. The trajectory point threshold is a critical value for determining whether the location trajectory density is too high. The masking symbol is a placeholder marker used to replace the original characters of the sensitive field. The dynamic noise is random disturbance data superimposed on the masking area and the value changes over time or context. The segmented masking operation is a processing process of segmenting continuous trajectory points by time interval and masking part of the data. The section is a continuous subsequence of trajectory data divided by time interval. The fluctuation range is the variation amplitude interval of the trajectory point value in the time dimension. The reorganization is the operation of reorganizing the desensitized field according to the original data structure. The desensitized data is a privacy-protected data set after masking and noise processing.

[0071] In the embodiments of the present application, first, through step 1011, the system obtains the raw data provided by the federated learning participant, and matches the fields of the raw data with the preset sensitive field type list (including identity, location trajectory, etc.) one by one through a field matching algorithm (such as regular expression or keyword matching). The fields that match are marked as the type of sensitive field, for example, the fields with field names "identity card number" or "GPS trajectory" are classified as identity or location trajectory type respectively.

[0072] Subsequently, through step 1012, according to the type of the marked sensitive field, the pre-defined privacy risk weight table (including risk level mapping rules of different types of fields) is queried, and the identity field is directly classified into the first privacy level. For the location trajectory field, the number of trajectory points (such as the number of trajectory points in a unit of time or space) is counted through a density calculation algorithm, and compared with the preset trajectory point threshold. If the density value exceeds the threshold, it is classified into the second privacy level, ensuring that the privacy protection strength of high-density trajectory data is higher.

[0073] Next, by step 1013, a partial masking operation is performed on the field of the first privacy level: the first three characters of the field content (such as the first three administrative code of the ID number) are retained, and the subsequent characters are replaced with a fixed number of masking symbols (such as "*"). In the position of the masking symbol, dynamic noise is superimposed by a dynamic noise generation algorithm (such as a random disturbance based on a hash function or a differential privacy noise mechanism), ensuring that the masked data is irreversible and retains some statistical characteristics. The amplitude of the noise is dynamically adjusted according to the field type and privacy level, for example, low-amplitude random character replacement is used for identity fields.

[0074] Then, by step 1014, a segmented masking operation is performed on the position trajectory field of the second privacy level: the continuous trajectory points are divided into multiple segments by a time window division algorithm (such as fixed interval or dynamic clustering). The trajectory points at the end of each segment (such as the last few points of each time window) are masked, and the masking range is adaptively determined according to the distribution density of the trajectory points. At the same time, by statistically analyzing the numerical fluctuation range of the trajectory points (such as the latitude and longitude offset or the speed change range), dynamic noise (such as Laplace mechanism-based noise or Gaussian noise) matching the fluctuation range is superimposed in the masking area, ensuring that the noise amplitude is consistent with the change trend of the original data, and avoiding the destruction of the data distribution characteristics.

[0075] Finally, by step 1015, the sensitive fields that have completed masking and noise superposition are reorganized according to the original data structure format. The original content of the non-sensitive fields is retained during the reorganization process, and only the processed sensitive fields are replaced to generate desensitized data that meets the federal learning data specification, ensuring the balance between data usability and privacy protection.

[0076] In practical applications, in the cross-hospital electronic medical record joint modeling scenario, the original data uploaded by the federal learning participant (a top three hospital) contains patient ID and department trajectory (step 1011). The system matches the preset sensitive field type list, identifies the patient ID as an identity type, and the department time sequence as a location trajectory type (step 1011). According to the privacy risk weight table, the patient ID is classified into the first privacy level, and the department trajectory is classified into the second privacy level because the single-day cross-department frequency exceeds the trajectory point threshold (step 1012). Perform local masking on the patient ID field: keep the "PAT" prefix and replace the subsequent characters with "", and superimpose dynamic noise in the masked area, such as replacing the original field "PAT2023-045" with "PAT#k7%9" (step 1013). For the department trajectory, mask the last department name by segmenting every hour, and superimpose noise according to the historical trajectory fluctuation range - for example, replace the last segment of "Cardiology Department → Radiology Department → Laboratory Department" with "Cardiology Department → Pharmacy Department" after masking (step 1014). When reorganizing the desensitized data, the patient ID and the noisy department trajectory are packaged in the original JSON format (step 1015) for use in the federal learning model training liver disease prediction task. After this processing, the local data of each hospital in participating in joint modeling protects patient privacy and maintains the effectiveness of cross-institutional feature alignment, achieving a balance between privacy and model performance.

[0077] The overall scheme of step 101 realizes a multi-level privacy protection and dynamic noise adaptive data desensitization mechanism. Through sensitive field type identification and privacy level division technology, a differentiated management system based on core sensitive elements such as identity and location trajectory is constructed, and a double protection strategy of local masking and dynamic noise superposition is adopted to realize privacy intensity grading control on the basis of preserving part of the data usability. For fields of different privacy levels, implement fine processing such as pre-sequence character masking and trajectory segment masking, and combine the association matching of dynamic noise range and trajectory fluctuation characteristics, effectively solving the problem that data usability and privacy protection are difficult to balance in traditional desensitization methods, laying a foundation for the safe sharing of multi-source heterogeneous data in the federal learning scenario.

[0078] 102, the desensitized data is divided into multiple data blocks by a preset encryption algorithm complexity, and the data blocks are distributed to independent operation units for parallel encryption processing, the number of operation units is proportional to the encryption algorithm complexity, to form an operation unit distribution strategy.

[0079] Optionally, step 102 can specifically include the following steps:

[0080] 1021, determine the number of divisions of the desensitized data according to a preset encryption algorithm complexity, the encryption algorithm complexity being defined by the hierarchical depth of encryption operation and the length of data block;

[0081] 1022. split the de-identified data into sequentially labeled data blocks according to the number of splits corresponding to the current encryption algorithm complexity level, the length of the data blocks being inversely proportional to the encryption algorithm complexity;

[0082] 1023. distribute the data blocks to independent computing units according to the number of splits, the computing units being bound to one data block and triggering encryption processing, the total number of computing units strictly corresponding to the number of splits, and the processing capacity of the computing units matching the length of the data blocks;

[0083] 1024. monitor the real-time rate at which the computing units process the data blocks, and if the real-time rate is lower than a preset reference rate corresponding to the current encryption algorithm complexity, split the data blocks into sub-data blocks according to a preset splitting ratio, and allocate new computing units to process the sub-data blocks synchronously, to form a computing unit allocation strategy.

[0084] In the above scheme, the encryption algorithm complexity is an index that measures the difficulty of encryption processing. The number of splits is the number of data blocks determined according to the encryption algorithm complexity. The hierarchical depth is the number of recursive or iterative processing steps in the encryption algorithm. The length of the data block is the number of bytes of a single data segment. The sequentially labeled data block is an encryption processing unit numbered in the original data order. The independent computing unit is a computing resource entity that performs encryption tasks. The processing capacity is the number of data bytes processed per second by the computing unit. The real-time rate is the actual speed value of the computing unit processing the data block. The preset reference rate is the minimum processing speed requirement corresponding to the encryption algorithm complexity. The splitting ratio is the ratio of the number of sub-blocks to the number of original blocks when the data block is split twice. The sub-data block is a smaller data processing unit generated by the second splitting of the data block. The computing unit allocation strategy is a processing scheme that dynamically adjusts the matching relationship between the computing unit and the data block.

[0085] In the embodiments of the present application, first, through step 1021, the system determines the number of splits of the de-identified data according to the preset encryption algorithm complexity (defined by the hierarchical depth of encryption operation and the length of the data block). The complexity evaluation module analyzes the computing resources required for encryption, such as higher complexity corresponding to multi-layer encryption or long key algorithm, to generate a number of splits proportional to the complexity (such as more data blocks when the complexity is high).

[0086] Subsequently, through step 1022, according to the number of splits corresponding to the current complexity level, the de-identified data is split into multiple sequentially labeled data blocks using a data fragmentation algorithm. The fragmentation rule is that the higher the complexity, the shorter the length of the single data block, for example, using fixed length fragmentation or dynamic adaptive fragmentation to ensure that the computing load of each data block matches the processing capacity of the computing unit.

[0087] Next, through step 1023, independent operation units (such as containerized instances or threads) are created based on the number of divisions, each unit binds a data block and triggers encryption processing. The resource scheduling module ensures that the total number of operation units strictly matches the number of divisions through a load balancing algorithm, and the computing capacity (such as CPU core number, memory) of the unit adapts to the length of the data block. For example, longer data blocks are allocated to high-spec operation units, and shorter data blocks are allocated to low-spec units, maximizing parallel efficiency.

[0088] Finally, through step 1024, the real-time rate at which each operation unit processes the data block is monitored in real time. If the rate is lower than the preset benchmark rate corresponding to the current complexity (such as the threshold of the number of bytes processed per second), the dynamic adjustment module further divides the original data block into sub-data blocks according to the preset splitting ratio (such as halving), and automatically expands new operation units to process the sub-data blocks. During the adjustment process, the task scheduler ensures that the new units are one-to-one bound with the sub-data blocks through a dynamic resource allocation algorithm (such as elastic scaling), and finally forms an operation unit allocation strategy that is dynamically optimized according to real-time performance, ensuring efficient completion of the encryption task.

[0089] In actual application, in the joint training scene of credit scoring models among banks, the system performs encryption preprocessing on the desensitized customer occupation and income interval data (output after step 101 processing). According to the encryption algorithm complexity of AES-256 algorithm (step 1021), the desensitized data is divided into 8 sequentially labeled data blocks, each data block length adapts to the complexity level - the occupation field is divided into fine-grained blocks due to short characters, and the income interval is divided into coarse-grained blocks due to wide numerical range (step 1022). Assign 8 independent operation units to bind data blocks (step 1023), among which the operation unit corresponding to the occupation field block starts the cache mechanism to match the short character processing requirement. During operation, it is found that the rate of processing the income interval block by a certain operation unit is lower than the benchmark value (step 1024), triggering the dynamic splitting strategy: the data block is split into 4 sub-data blocks according to the horizontal splitting ratio, and 4 new operation units are activated for parallel processing. After adjustment, the total number of operation units increases to 12, the length of the sub-data block adapts to the processing capacity of the unit, and the overall rate of the encryption task recovers to the preset threshold. This allocation strategy ensures that the occupation and income data meet the real-time requirements when cross-bank joint modeling, while maintaining the encryption strength and privacy security boundary.

[0090] The overall scheme in step 102 above achieves elastic data encryption and dynamic resource scheduling optimization based on algorithm complexity. Through a dynamic mapping mechanism between encryption algorithm complexity and the number of data block segments, a collaborative matching model of data block length, number of computing units, and processing capacity is constructed. Real-time rate monitoring and secondary data block splitting techniques are employed to dynamically adjust the execution efficiency of encryption tasks, overcoming the processing latency bottleneck caused by sudden changes in algorithm complexity under a fixed resource allocation model. The strong binding strategy between computing units and data blocks, combined with an elastic expansion mechanism, improves resource utilization for large-scale data parallel processing while ensuring encryption security, providing high throughput support for encrypted communication in federated learning.

[0091] Step 1024 may specifically include the following processes: real-time monitoring of the processing rate of the data block by the computing unit, and comparing the real-time rate with a preset benchmark rate bound to the complexity of the current encryption algorithm on a unit-by-unit basis to generate a rate difference status marker; when the rate difference status marker is detected to be lower than the preset benchmark rate, a data block splitting operation is triggered, the splitting operation divides the current data block into multiple sub-data blocks according to a preset splitting ratio, and the length of each sub-data block is inversely proportional to the splitting ratio; a sub-data block to be processed is allocated to the newly added computing unit, and each sub-data block is bound to an independent computing unit for encryption processing to form a computing unit allocation strategy.

[0092] 103. Transmit the encrypted data block to the federated learning collaboration node through the encrypted transmission channel, obtain the aggregation calculation result fed back by the federated learning collaboration node, and synchronously record the time consumed by the encryption operation and the privacy strength value corresponding to the aggregation calculation result.

[0093] Optionally, step 103 may specifically include the following steps:

[0094] 1031. Attach a sequence identifier and a channel key to the encrypted data blocks, and send the encrypted data blocks to the federated learning collaboration nodes in batches according to the sequence identifier through the encrypted transmission channel. Record the start and end timestamps of each batch of data blocks during transmission to record the time consumed by the encryption operation.

[0095] 1032. At the receiving end of the federated learning collaboration node, verify the continuity of the sequential identifier of the encrypted data block and the validity of the channel key, and perform aggregation calculation on the verified data block to generate the aggregation calculation result;

[0096] 1033. Calculate the privacy strength value corresponding to the aggregated calculation result based on the encryption features of the data block.

[0097] In the above scheme, the encrypted transmission channel is a secure communication link that uses encryption protocols to protect the data transmission process. The federated learning collaboration node is a server or terminal device that performs multi-party data aggregation calculations. The aggregation calculation result is the joint output value of multiple participants' data processed by the federated learning model. The encryption operation time is the total length of time from the start of data block transmission to the completion of encryption processing. The privacy intensity value is a quantitative indicator that measures the degree of privacy protection of the aggregation calculation result. The sequence identifier is a unique number that marks the transmission sequence of the data block. The channel key is a key parameter used to verify the identity of both parties in the encrypted transmission channel. The start transmission timestamp is the system clock record value when the data block starts transmission. The end timestamp is the system clock record value when the data block completes transmission. The receiving end is a module in the federated learning collaboration node responsible for receiving and verifying the data block. The encryption feature is a set of mathematical properties formed by the combination of the data block encryption algorithm and the key parameter.

[0098] In the embodiments of the present application, first, by step 1031, the system attaches a unique sequence identifier (such as an incremental serial number or a hash value) and a channel key (such as a session key based on the TLS protocol) to the encrypted data block, ensuring that the sequence of the data block is traceable and the channel is encrypted and secure during transmission. The data block is sent to the federated learning collaboration node in batches according to the sequence identifier through the encrypted transmission channel (such as SSL / TLS or IPSec tunnel), and the start transmission timestamp and the end transmission timestamp of each batch of data block are recorded at the time of sending through the time synchronization protocol to calculate the total encryption operation time (such as the end timestamp minus the start timestamp).

[0099] Subsequently, by step 1032, the federated learning collaboration node verifies whether the sequence identifier of the data block is continuous (such as checking whether the serial number is increasing) through the sequence verification algorithm at the receiving end, and confirms the validity of the channel key through the key verification algorithm (such as digital signature or hash verification). The data block that passes the verification is input into the aggregation calculation module (such as the FedAvg algorithm) to perform weighted average or gradient fusion of multi-node data, generating an aggregation calculation result. If the sequence or key verification fails, a data retransmission mechanism is triggered.

[0100] Finally, by step 1033, the privacy intensity value is calculated based on the encryption features (such as encryption algorithm type, key length, noise intensity) of the data block through the privacy intensity evaluation algorithm. For example, the differential privacy budget (ε value) or the security level of homomorphic encryption is used to quantify the degree of privacy protection, ensuring that the risk of privacy leakage of the aggregation result is controllable. This value is used to optimize the encryption strategy or adjust the privacy protection intensity of the federated learning model in the future.

[0101] In practical applications, in the federated learning scenario of cross-e-commerce platform user behavior analysis, platform A attaches the encrypted user browsing duration and commodity click sequence data block (step 102 output) to the incremental order identifier and the dynamically generated channel key (step 1031), and sends them to the federated collaboration node in order through the TLS encrypted transmission channel. After verifying that the data block order identifier is continuous and the channel key is valid (step 1032), the collaboration node performs weighted average aggregation calculation on the multi-platform data to generate a global user behavior feature vector. The system calculates the privacy strength value of the aggregation result based on the homomorphic property of the encrypted data block (step 1033), and confirms that the model update parameters do not leak individual browsing preferences. The transmission time consumption of the synchronously recorded commodity click sequence data block is prolonged due to the high encryption level, triggering the subsequent optimization of the channel key rotation frequency to balance security and timeliness. This process ensures that when e-commerce platforms jointly optimize the recommendation model, user sensitive behavior data meets the privacy strength threshold in the entire process of encrypted transmission and aggregation calculation, while supporting dynamic optimization of model iteration efficiency.

[0102] The overall scheme of step 103 realizes an end-to-end encrypted transmission and privacy strength quantitative evaluation closed-loop verification system. Through the composite verification mechanism of sequential identifier and channel key, a dual protection of encrypted data block transmission integrity and channel security is constructed, combined with transmission timestamp recording and aggregation calculation result feedback, forming a full-link monitoring of encryption time consumption and privacy strength. The privacy strength value calculation model based on encrypted feature analysis maps the encryption quality at the data block level to a quantifiable privacy protection indicator, solving the problem of objective evaluation of data security status in the federated learning scenario, and providing a verifiable privacy protection benchmark for cross-node collaboration.

[0103] 104, calculate the difference value between the time consumption of the encryption operation and the preset threshold, and adjust the segmentation number of the desensitized data block according to the difference value to match the operation unit allocation strategy, and cross-verify the difference value and the privacy strength value to generate the correlation parameter of the operation unit allocation strategy and the privacy strength value.

[0104] Optionally, step 104 can specifically include the following steps:

[0105] 1041, compare the time consumption of the encryption operation with the preset threshold point by point to calculate the percentage value of the time consumption of the encryption operation and the preset threshold as the difference value;

[0106] 1042, extract the deviation direction of the difference value, and adjust the segmentation number of the desensitized data block according to the deviation direction and the percentage value. If the difference value deviates positively, the segmentation number is increased in proportion to match the upper limit of the number of operation units of the operation unit allocation strategy, and if the difference value deviates negatively, the segmentation number is reduced in proportion to release the redundant operation units of the operation unit allocation strategy.

[0107] 1043、extracting the privacy intensity value and binding the difference value of each data block with the corresponding privacy intensity value one by one to generate a verification pair of difference value and privacy intensity value;

[0108] 1044、analyzing the correlation between the difference value and the privacy intensity value in the verification pair, increasing the number of partitions and increasing the positive correlation parameter of the privacy intensity value when the difference value is positively deviated and the privacy intensity value is lower than the preset intensity, and generating a balance correlation parameter of maintaining the current number of partitions and privacy intensity when the difference value and the privacy intensity value both meet the preset standard;

[0109] 1045、integrating the positive correlation parameter and the balance correlation parameter to generate a correlation parameter.

[0110] In the above scheme, the preset threshold is a pre-set upper or lower limit reference value of the encryption processing time consumption. The difference value is a percentage difference value of the encryption operation time consumption and the preset threshold. The deviation direction is the positive or negative change trend of the difference value relative to the preset threshold. The upper limit of the number of units of the operation unit allocation strategy is the maximum number of parallel processing units allowed by the operation unit allocation strategy. The redundant operation unit is an idle operation unit that exceeds the current encryption task requirement. The verification pair is a verification data group formed by binding the difference value and the privacy intensity value. The correlation is the statistical correlation degree between the difference value and the privacy intensity value. The preset intensity is the minimum protection standard threshold that the privacy intensity value needs to meet. The positive correlation parameter is a positive influence coefficient for adjusting the privacy intensity value when the difference value is positively deviated. The balance correlation parameter is a stable parameter for maintaining the current strategy when the difference value and the privacy intensity value both meet the standard. The correlation parameter is a dynamic adjustment coefficient that integrates the positive correlation parameter and the balance correlation parameter.

[0111] In the embodiments of the present application, first, through step 1041, the system compares the time consumption of the encryption operation with the preset reference threshold (such as the maximum allowed processing time) through real-time monitoring module at each time point, and calculates the percentage value of the time consumption exceeding or being lower than the threshold as the difference value. The comparison algorithm uses sliding window technology to dynamically intercept the time consumption data of each time period and the threshold to calculate the percentage difference value, for example, if the time consumption can exceed the threshold by 20%, the difference value can be +20%.

[0112] Subsequently, through step 1042, the deviation direction of the difference value (positive for time consumption exceeding the limit, negative for time consumption being lower than expected) is extracted, and the number of partitions of the desensitization data block is adjusted according to the deviation direction and the percentage value through the dynamic adjustment module. For example, if the difference value is positively deviated by 30%, the number of partitions is increased by a certain proportion (such as 5% of the number of partitions for every 10% of the difference value), until the upper limit of the number of units of the operation unit allocation strategy is reached; if the difference value is negatively deviated, the number of partitions is reduced by a certain proportion, and the redundant operation units are released to optimize the resource occupation.

[0113] Next, by step 1043, the privacy strength value (such as the differential privacy epsilon value or the encryption security level) of each data block is extracted, and the difference value is associated with the corresponding privacy strength value one by one through the data binding algorithm to generate a verification pair containing the relationship between the two. For example, the difference value of a certain data block can be +15%, and the privacy strength value is medium, then a verification pair (+15%, medium) is generated for subsequent analysis.

[0114] Then, by step 1044, the difference value and the privacy strength value in the verification pair are analyzed for correlation. If the difference value is positively deviated and the privacy strength value is lower than the preset safety standard (such as the privacy strength value not reaching the high level), the number of partitions is increased and the positive correlation parameter of the privacy strength value is increased (such as increasing the privacy strength weight coefficient) through the parameter adjustment algorithm; if both the difference value and the privacy strength value meet the preset standard (such as the time consumption being within the threshold and the privacy strength meeting the standard), a balanced correlation parameter is generated to maintain the stability of the current partition number and privacy strength.

[0115] Finally, by step 1045, the positive correlation parameter (used to optimize performance and privacy) and the balanced correlation parameter (used to maintain a stable state) are integrated into a unified correlation parameter through the parameter fusion algorithm. This parameter will be fed back to the dynamic adjustment module of the system to guide the subsequent optimization of data partitioning, encryption operation and privacy protection, forming a closed-loop adaptive control mechanism.

[0116] In actual application, in the joint risk assessment modeling scenario of insurance institutions, the system detects that the encryption operation time consumption exceeds the preset threshold (step 1041), and the difference value shows positive deviation. Based on the deviation direction (step 1042), the number of partitions of the customer income de-sensitized data block is increased from 12 to 18, releasing more operation units for parallel processing to match the upper limit of the strategy. When verifying the correlation between the difference value and the privacy strength value (step 1043), it is found that the privacy strength value of the newly added partition block is slightly lower than the preset standard due to the decrease in noise superposition density. After the system analyzes the verification pair (step 1044), the positive correlation parameter adjustment is triggered: on the basis of maintaining the number of partitions, the dynamic noise strength of the newly added data block is enhanced to improve the privacy protection level, and the task scheduling logic of the operation unit is optimized. The final generated correlation parameter (step 1045) balances the encryption time consumption and the privacy strength, ensuring that when training the cross-institution premium prediction model, the user income data can meet the real-time encryption transmission requirements, and through the dynamic noise compensation mechanism, it reaches the privacy security baseline of joint modeling, realizing the two-way improvement of risk assessment accuracy and data protection efficiency.

[0117] The overall scheme of step 104 realizes the dynamic balance regulation of encryption efficiency and privacy protection strength. Through the cross-validation mechanism of encryption time-consuming difference value and privacy strength value, a two-way feedback channel of data segmentation number adjustment and privacy parameter optimization is constructed. The difference value deviation direction identification and segmentation proportion dynamic correction technology realizes the elastic expansion of the operation unit allocation strategy, which guarantees the timeliness of the encryption task while avoiding resource redundant consumption. The privacy strength value and processing efficiency associated parameter generation technology integrates encryption performance and privacy protection into a unified evaluation framework, providing dynamic decision basis for the collaborative optimization of security and efficiency in the federated learning scenario.

[0118] 105、based on the associated parameters, when the privacy strength value is lower than the preset strength, the dynamic noise corresponding to the privacy level is enhanced, and the segmentation number of the desensitization data block is adjusted synchronously, so that the desensitization data block after the dynamic noise is enhanced is re-segmented and encrypted to form a dynamic linkage mechanism.

[0119] Optionally, step 105 can specifically include the following steps:

[0120] 1051、check whether the privacy strength value marked in the associated parameters is lower than the preset strength, and when it is detected that the privacy strength value is lower than the preset strength, enhance the dynamic noise in the field belonging to the corresponding privacy level in the desensitization data block to generate a desensitization data block after noise enhancement;

[0121] 1052、calculate the length change of the data block after the dynamic noise is enhanced, and adjust the segmentation number of the desensitization data block according to the length change according to the preset segmentation rule, to ensure that the length of the data block and the segmentation number are inversely proportional;

[0122] 1053、redistribute the desensitization data block after adjusting the segmentation number to an independent operation unit for encryption processing, and each operation unit binds the corresponding data block according to the updated segmentation number and executes encryption processing to form a dynamic linkage mechanism.

[0123] In the above scheme, the dynamic linkage mechanism is an automatic control logic for the cooperative adjustment of the privacy strength value and the data segmentation number. The preset segmentation rule is a block adjustment criterion defined according to the relationship between the data block length and the segmentation number. The desensitization data block after noise enhancement is a privacy protection data set generated by reprocessing after enhancing the dynamic noise strength. The length change is the difference value of the byte number after adjusting the dynamic noise of the data block.

[0124] In the embodiments of the present application, first, the system checks whether the privacy strength value marked in the association parameter is lower than the preset security strength threshold value through step 1051. If it is detected that the privacy strength value is insufficient (for example, due to insufficient data masking or too low noise strength), the dynamic noise enhancement algorithm is used to increase the superimposed strength of the dynamic noise (for example, increase the density or range of random noise) of the field (such as the identity or location trajectory) belonging to the low privacy level in the de-sensitized data block, to generate a noise-enhanced de-sensitized data block, and to ensure that the privacy protection level reaches the preset standard.

[0125] Subsequently, through step 1052, the length change (for example, the increase in the number of bytes due to noise superposition) of the noise-enhanced data block is calculated, the change value is counted by the data block length monitoring module, and the segmentation number is dynamically adjusted according to the preset segmentation rule (such as the inverse relationship between the data block length and the segmentation number). For example, if the data block length becomes larger due to noise increase, the segmentation number is increased in proportion to shorten the length of the individual data block, to ensure the balance between encryption processing efficiency and resource occupation.

[0126] Finally, through step 1053, the adjusted de-sensitized data block is re-distributed to the independent operation unit according to the new segmentation number. The resource scheduling module creates or releases the operation unit through the dynamic expansion or contraction mechanism, to ensure that each unit is bound to a data block and performs encryption processing. During the encryption process, the processing capacity of the operation unit is matched with the data block length in real time (for example, long data blocks are allocated to high-performance units), forming a dynamic linkage mechanism that adjusts the privacy strength and the segmentation number in coordination, to realize the continuous optimization of privacy protection and computing efficiency.

[0127] In actual application, in the federated learning scene of urban traffic flow prediction, the system detects that the privacy strength value of the vehicle type de-sensitized data block is lower than the preset strength (step 1051), triggering the dynamic noise enhancement mechanism: superimposing a random noise mask on the vehicle brand field belonging to the second privacy level, de-sensitizing “new energy-sedan” to “new energy-#-sedan”, and the length of the noise-enhanced data block increases due to character expansion. According to the preset segmentation rule (step 1052), the system adjusts the original segmentation number from 12 blocks to 18 blocks, to ensure that the data block length is inversely proportional to the segmentation number. After adjustment, the newly added 6 independent operation units (step 1053) bind the fine-grained data blocks, and when performing AES encryption, they preferentially process the high-frequency variable area of the noise-enhanced field. Under the dynamic linkage mechanism, the encryption time and the privacy strength are calibrated in real time—when modeling across cities, the vehicle type data not only resists the risk of brand information leakage through noise superposition, but also maintains encryption efficiency by optimizing the segmentation number, so that the traffic flow prediction model continuously absorbs the dynamic change characteristics of multi-source road network data under the premise of protecting the privacy of vehicle owners, to improve the signal light control accuracy during the morning and evening peak periods.

[0128] The overall scheme of step 105 realizes a closed-loop enhancement mechanism of privacy strength self-repair and encryption strategy linkage. Through privacy strength threshold monitoring and dynamic noise enhancement technology, an adaptive compensation system is constructed in the scenario of insufficient privacy protection capability. Based on the dynamic segmentation rule adjustment strategy of data block length change, the resource adaptability of the de-sensitized data after noise enhancement in the encryption processing link is ensured. The real-time redistribution mechanism of the operation unit and the updated data block forms a dynamic balance between privacy protection strength improvement and encryption processing efficiency maintenance, effectively coping with the dynamic changes of privacy leakage risk in the federal learning collaboration process, and realizing the continuous collaborative optimization of security protection capability and system stability.

[0129] The following is a complete embodiment for steps 101-105:

[0130] In the cross-educational institution student behavior analysis scenario, the original data uploaded by the federal learning participants (multiple universities) contains student ID and classroom activity trajectory. The system identifies the student ID as an identity field based on the pre-set sensitive field type list, and the classroom activity trajectory as a location trajectory field, and based on the privacy risk weight table, the student ID is classified into the first privacy level and the high-frequency classroom trajectory is classified into the second privacy level. The first three characters of the student ID field are retained and dynamic noise is superimposed to generate a de-sensitized identifier; the classroom trajectory is segmented by course session and the last position point is masked, and dynamic noise matching the activity frequency is injected in the masked area. According to the encryption algorithm complexity, the de-sensitized data is segmented into multiple data blocks, the number of operation units is proportional to the complexity level, and the data block length is adapted to the processing capacity of the unit. Real-time monitoring found that the speed of a certain operation unit was lower than the baseline value, triggering data block splitting and adding new operation units for parallel processing. After adding the sequence identifier and channel key to the encrypted data block, it is sent to the federal collaboration node through the encrypted transmission channel, and the node verifies the continuity of the identifier and the validity of the key before aggregating to generate global learning behavior features, and synchronously calculates the privacy strength value and records the encryption time. When the encryption time difference value shows a positive deviation and the privacy strength is insufficient, the system increases the noise intensity of the student ID field, dynamically adjusts the number of data block segmentation, and expands the operation unit to balance the processing efficiency. The noise-enhanced data block is re-distributed to the operation unit for encryption, forming a dynamic linkage mechanism. This scheme enables each educational institution to jointly optimize the teaching model by dividing privacy levels, dynamically superimposing noise, and flexibly scheduling operation resources, ensuring that student identity and behavior data meet privacy protection requirements in cross-institution sharing, while supporting model iteration for teaching effectiveness analysis.

[0131] As shown in Figure 2 , Figure 3 illustrates the dynamic collaborative mechanism of privacy protection and computing efficiency in federal learning. Figure 2Focusing on single-scene privacy regulation, a three-layer architecture of "data provider-parallel computing cluster-federated learning node" is presented: the left data source symbolizes multi-field data input through a building icon, the middle parallel cluster completes encrypted computation and gradient aggregation with distributed nodes, and the right federated node receives aggregated results to update the global model. The bottom dynamic monitoring system tracks the privacy strength (such as differential privacy noise level) and computation time in real time, dynamically adjusts the parameters through a feedback loop, and ensures the balance between privacy protection and training efficiency. Figure 3 Extended to a multi-agency collaboration scenario, the left medical and financial data sources are preprocessed through sensitive field anonymization and data segmentation, the middle parallel computing layer integrates encryption processing and data desensitization technology, and the right federated node completes secure aggregation of multiple parties. The dynamic monitoring system focuses on regulating the noise intensity parameter in this scenario, quantitatively analyzes the relationship between privacy leakage risk and model accuracy, and realizes adaptive protection in cross-agency data collaboration. The two figures together build a closed-loop system of "data preprocessing-safe computation-dynamic optimization", which improves the practicality and cross-domain collaboration capability of the federated learning system under the premise of protecting data privacy through layered encryption, gradient perturbation, and real-time monitoring technology.

[0132] Figure 4 A structure diagram of a privacy computing service system based on federated learning is provided for the embodiments of the present application, as shown in Figure 2 The system comprises:

[0133] The identification module 41 is configured to obtain original data of a federated learning participant, identify the type of sensitive fields in the original data, divide the sensitive fields into different privacy levels according to the type, and perform local masking on the content of the sensitive fields and superimpose dynamic noise on the locally masked area to generate desensitized data.

[0134] The allocation module 42 is configured to divide the desensitized data into a plurality of data blocks through a preset encryption algorithm complexity, and allocate the data blocks to independent operation units for parallel encryption processing, the number of operation units being proportional to the encryption algorithm complexity, to form an operation unit allocation strategy.

[0135] The transmission module 43 is configured to transmit the encrypted data blocks to a federated learning collaboration node through an encrypted transmission channel, and obtain aggregated computation results fed back by the federated learning collaboration node, and synchronously record the time consumption of the encrypted operation and the privacy strength value corresponding to the aggregated computation results.

[0136] The calculation module 44 is configured to calculate the difference value between the time consumption of the encrypted operation and a preset threshold, adjust the segmentation number of the desensitized data blocks to match the operation unit allocation strategy according to the difference value, and cross-verify the difference value and the privacy strength value to generate an associated parameter of the operation unit allocation strategy and the privacy strength value.

[0137] The forming module 45 is configured to, based on the association parameter, when the privacy strength value is lower than a preset strength, enhance the dynamic noise corresponding to the privacy level, adjust the segmentation number of the desensitization data block synchronously, and re-segment and encrypt the desensitization data block after the dynamic noise is enhanced, to form a dynamic linkage mechanism.

[0138] Figure 4 The privacy computing service system based on federated learning can perform Figure 1 The privacy computing service method based on federated learning has the implementation principle and technical effects which will not be repeated. For the specific manner in which each module, unit of the privacy computing service system based on federated learning in the above embodiment performs the operation, it has been described in detail in the embodiment related to the method, which will not be described in detail here.

[0139] In one possible design, Figure 4 The privacy computing service system based on federated learning can be implemented as a computing device, such as Figure 5 As shown in the figure, the computing device can include a storage component 51 and a processing component 52.

[0140] The storage component 51 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 52.

[0141] The processing component 52 is configured to perform the above Figure 1 The privacy computing service method based on federated learning.

[0142] The processing component 52 can include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component can also be one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components, for executing the above method.

[0143] The storage component 51 is configured to store various types of data to support the operation of the terminal. The storage component can be realized by any type of volatile or non-volatile storage device or their combination, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0144] Of course, the computing device can also include other components, such as input / output interfaces, display components, communication components, etc.

[0145] The input / output interface provides an interface between the processing component and peripheral interface modules, which can be output devices, input devices, etc.

[0146] The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, etc.

[0147] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform, and the computing device can be a cloud server, and the processing component, the storage component, etc. can be a basic server resource rented or purchased from the cloud computing platform.

[0148] The embodiments of the application also provide a computer storage medium storing a computer program, and the computer program can implement the above-mentioned Figure 1 The embodiment shown in the figure is a privacy computing service method based on federated learning.

[0149] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-mentioned system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0150] The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0151] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software and necessary general hardware platform, and of course, it can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software product, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including a plurality of instructions to make a computer device (which can be a personal computer, server, or network device, etc.) execute the method described in each embodiment or some part of the embodiment.

[0152] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the same; although the present application has been described in detail with reference to the foregoing examples, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A privacy-preserving computing service method based on federated learning, characterized in that, include: The raw data of the participants in the federated learning is obtained, and the types of sensitive fields in the raw data are identified. The sensitive fields are divided into different privacy levels according to the types. The content of the sensitive fields is partially masked according to the privacy level, and dynamic noise is superimposed on the partially masked areas to generate de-identified data. The de-identified data is divided into multiple data blocks by using a preset encryption algorithm complexity, and the data blocks are allocated to independent computing units for parallel encryption processing. The number of computing units is proportional to the encryption algorithm complexity to form a computing unit allocation strategy. The encrypted data block is transmitted to the federated learning collaboration node through the encrypted transmission channel, and the aggregation calculation result fed back by the federated learning collaboration node is obtained. The time consumed by the encryption operation and the privacy strength value corresponding to the aggregation calculation result are recorded simultaneously. The time consumed by the encryption operation is compared point by point with a preset threshold, and the percentage difference between the time consumed by the encryption operation and the preset threshold is calculated as the difference value. Extract the deviation direction of the difference value, and adjust the number of segments of the desensitized data block according to the deviation direction and percentage value. If the difference value is positive, increase the number of segments proportionally to match the upper limit of the number of units in the operation unit allocation strategy. If the difference value is negative, decrease the number of segments proportionally to release redundant operation units in the operation unit allocation strategy. Extract the privacy strength value and bind the difference value of each data block to the corresponding privacy strength value one by one to generate a verification pair between the difference value and the privacy strength value; The correlation between the difference value and the privacy strength value in the verification pair is analyzed. When the difference value deviates positively and the privacy strength value is lower than the preset strength, the number of segments is increased and the positive correlation parameter of the privacy strength value is improved. When both the difference value and the privacy strength value meet the preset standard, a balanced correlation parameter is generated to maintain the current number of segments and the privacy strength. The positive correlation parameters and the balanced correlation parameters are integrated to generate the correlation parameters; Based on the aforementioned correlation parameters, when the privacy strength value is lower than the preset strength, the dynamic noise corresponding to the privacy level is increased and the number of segments of the desensitized data block is adjusted simultaneously, so that the desensitized data block after the dynamic noise enhancement is re-segmented and encrypted to form a dynamic linkage mechanism.

2. The method according to claim 1, characterized in that, The process involves acquiring raw data from federated learning participants, identifying the types of sensitive fields within the raw data, classifying these sensitive fields into different privacy levels based on their types, and then partially masking the content of these sensitive fields and superimposing dynamic noise into the masked areas to generate de-identified data. Obtain the raw data provided by the participants in the federated learning, and match the fields of the raw data one by one with a preset list of sensitive field types. The preset list of sensitive field types includes the types of identity identifiers and location trajectories. Mark the fields that match successfully as the types of sensitive fields. Obtain the privacy risk weight table associated with the type of the sensitive field, and classify the field containing the identity identifier into the first privacy level according to the privacy risk weight table. Calculate the density value of the field containing the location trajectory based on the number of trajectory points in the location trajectory. When the density value exceeds the trajectory point threshold, classify it into the second privacy level. The first three characters of the field with the first privacy level are retained and the subsequent characters are replaced with masking symbols. Dynamic noise is then added to the masking symbols one by one according to the position of the first three characters. A segmented masking operation is performed on the trajectory field of the second privacy level. The segmented masking operation divides the continuous trajectory points into segments according to time intervals and masks the trajectory points at the end of each segment. At the same time, the fluctuation range of the trajectory point values ​​is obtained, and dynamic noise matching the fluctuation range is superimposed on the masked area. Sensitive fields that have been masked and superimposed with dynamic noise are reorganized according to the format of the original data to generate de-sensitized data.

3. The method according to claim 1, characterized in that, The de-identified data is divided into multiple data blocks based on a preset encryption algorithm complexity, and these data blocks are allocated to independent computing units for parallel encryption processing. The number of computing units is proportional to the encryption algorithm complexity, forming a computing unit allocation strategy, including: The number of segments for the de-identified data is determined according to the preset encryption algorithm complexity, wherein the encryption algorithm complexity is defined by the depth of the encryption operation level and the length of the data block; The de-identified data is divided into sequentially labeled data blocks according to the number of segments corresponding to the complexity level of the current encryption algorithm. The length of the data block is inversely proportional to the complexity of the encryption algorithm. The data blocks are allocated to independent processing units according to the number of segments. Each processing unit is bound to a data block and triggers encryption processing. The total number of processing units is strictly consistent with the number of segments, and the processing capacity of the processing unit is matched with the length of the data block. The real-time rate at which the computing unit processes the data block is monitored. If the real-time rate is lower than the preset baseline rate corresponding to the complexity of the current encryption algorithm, the data block is divided into sub-data blocks according to a preset splitting ratio, and a new computing unit is allocated to process the sub-data blocks synchronously, so as to form a computing unit allocation strategy.

4. The method according to claim 3, characterized in that, The real-time processing rate of the data block by the computing unit is monitored. If the real-time processing rate is lower than the preset baseline rate corresponding to the complexity of the current encryption algorithm, the data block is divided into sub-data blocks according to a preset splitting ratio, and newly added computing units are allocated to process the sub-data blocks synchronously, thus forming a computing unit allocation strategy, including: The real-time rate at which the computing unit processes the data block is monitored, and the real-time rate is compared unit by unit with a preset benchmark rate bound to the complexity of the current encryption algorithm to generate a rate difference status marker. When a rate difference is detected and marked as being lower than a preset baseline rate, a data block splitting operation is triggered. The splitting operation divides the current data block into multiple sub-data blocks according to a preset splitting ratio. The length of each sub-data block is inversely proportional to the splitting ratio. The newly added computing units are allocated sub-data blocks to be processed, and each sub-data block is bound to an independent computing unit for encryption processing to form a computing unit allocation strategy.

5. The method according to claim 1, characterized in that, The encrypted data block is transmitted to the federated learning collaboration node through an encrypted transmission channel, and the aggregation calculation result fed back by the federated learning collaboration node is obtained. The time consumed by the encryption operation and the privacy strength value corresponding to the aggregation calculation result are recorded synchronously, including: A sequence identifier and a channel key are attached to the encrypted data blocks. The encrypted data blocks are sent to the federated learning collaboration nodes in batches according to the sequence identifier through the encrypted transmission channel. The start and end timestamps of each batch of data blocks are recorded during transmission to record the time consumed by the encryption operation. The receiving end of the federated learning collaboration node verifies the continuity of the sequential identifier of the encrypted data block and the validity of the channel key, and performs aggregation calculation on the verified data block to generate the aggregation calculation result. The privacy strength value corresponding to the aggregated calculation result is calculated based on the encryption characteristics of the data block.

6. The method according to claim 1, characterized in that, Based on the aforementioned correlation parameters, when the privacy strength value is lower than a preset strength, the dynamic noise corresponding to the privacy level is increased, and the number of segments of the desensitized data block is adjusted synchronously. This causes the desensitized data block with enhanced dynamic noise to be re-segmented and encrypted, forming a dynamic linkage mechanism, including: Check whether the privacy strength value marked in the associated parameters is lower than the preset strength. When the privacy strength value is detected to be lower than the preset strength, increase the dynamic noise in the fields belonging to the corresponding privacy level in the desensitized data block to generate a noise-enhanced desensitized data block. Calculate the length change of the data block after increasing dynamic noise, and adjust the number of segments of the desensitized data block according to the length change and a preset segmentation rule to ensure that the length of the data block is inversely proportional to the number of segments; The desensitized data blocks, after adjusting the number of segments, are redistributed to independent computing units for encryption. Each computing unit binds the corresponding data block according to the updated number of segments and performs encryption, thus forming a dynamic linkage mechanism.

7. A privacy-preserving computation service system based on federated learning, used in the privacy-preserving computation service method based on federated learning as described in any one of claims 1 to 6, characterized in that, include: The identification module is used to acquire the raw data of the federated learning participants, identify the type of sensitive fields in the raw data, classify the sensitive fields into different privacy levels according to the type, and partially mask the content of the sensitive fields according to the privacy level and superimpose dynamic noise in the partially masked area to generate de-identified data. The allocation module is used to divide the de-identified data into multiple data blocks according to a preset encryption algorithm complexity, and allocate the data blocks to independent computing units for parallel encryption processing. The number of computing units is proportional to the encryption algorithm complexity to form a computing unit allocation strategy. The transmission module is used to transmit the encrypted data block to the federated learning collaboration node through the encrypted transmission channel, and obtain the aggregation calculation result fed back by the federated learning collaboration node, and synchronously record the time consumed by the encryption operation and the privacy strength value corresponding to the aggregation calculation result. The calculation module is used to calculate the difference between the time consumed by the encryption operation and the preset threshold, and adjust the number of segments of the de-identified data block according to the difference to match the operation unit allocation strategy. At the same time, the difference is cross-validated with the privacy strength value to generate the correlation parameter between the operation unit allocation strategy and the privacy strength value. The forming module is used to, based on the associated parameters, increase the dynamic noise corresponding to the privacy level and simultaneously adjust the number of segments of the desensitized data block when the privacy strength value is lower than the preset strength, so that the desensitized data block after the dynamic noise enhancement is re-segmented and encrypted to form a dynamic linkage mechanism.

8. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a privacy computing service method based on federated learning as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, The system contains a computer program that, when executed by a computer, implements a privacy computing service method based on federated learning as described in any one of claims 1 to 6.

Images