Data circulation method and system based on privacy protection

Abstract

The invention provides a data circulation method and system based on privacy protection, and the method comprises the steps: carrying out the secret sharing of a local data set through a plurality of data owners, and generating share data, and enabling the share data to be distributed in the plurality of data owners; the plurality of data owners compare the field mode, the sample range and the feature dimension of the share data in an encryption state based on a secure multi-party computing protocol to obtain a comparison result; a plurality of data owners execute selected transverse federation or longitudinal federation merging operation, and share data is processed by using homomorphic encryption to obtain a merging result; a plurality of data owners negotiate a unified time axis reference in an encrypted state through a secure multi-party computing protocol, and calculate the offset of a data set of each party relative to the unified time axis reference by using timestamps of share data; and the plurality of data owners perform time sequence alignment on the share data according to the offset to obtain the share data after time sequence alignment and generate a joint data set.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a data circulation method and system based on privacy protection. Background Technology

[0002] In the field of data collaboration and sharing, cross-organizational data processing is considered a key pillar for driving technological innovation and enhancing business value. Efficient data integration and analysis can significantly optimize decision-making and resource allocation. However, this field faces a profound contradiction between privacy protection and data utilization. How to achieve efficient collaboration of multi-party data without disclosing sensitive information has become a pressing challenge. Currently, although some methods attempt to protect data privacy through encryption or anonymization, these solutions often fall short in complex scenarios. Especially when multiple data holders need to merge data, existing methods struggle to address issues such as inconsistent data formats and timelines, leading to either sacrificing data integrity or requiring significant time and resources for manual adjustments during collaboration—resulting in inefficiency and a high risk of errors.

[0003] A deeper challenge lies in coordinating the core element of time consistency while protecting privacy during data merging. Differences in the recording times of data from various parties—for example, one party might record data hourly while another aggregates it daily, or there might be timestamp offsets—directly impact the accuracy and usability of the merged data. For instance, in financial transaction analysis, if data from multiple parties cannot be aligned to the same point in time, it can lead to misjudgments of transaction trends, thus affecting decision-making. Further, in cross-organizational collaboration, one bank's transaction data might record minute-by-minute fluctuations, while another institution's payment data is aggregated daily. Without securely comparing and adjusting these time differences in an encrypted state, it's difficult to establish a unified time benchmark, significantly diminishing the value of the data. Summary of the Invention

[0004] This invention provides a privacy-preserving data circulation method and system, which aims to ensure data privacy through encryption technology while automating and intelligently merging multi-dimensional data alignment and significantly improving the security and efficiency of distributed data collaboration.

[0005] In a first aspect, the present invention provides a privacy-preserving data circulation method, which mainly includes: Multiple data holders secretly share their local datasets to generate share data, which is distributed among them. These data holders then compare the field patterns, sample ranges, and feature dimensions of the share data in an encrypted state using a secure multi-party computation protocol, obtaining comparison results. Based on the comparison results, the data holders determine the specific parts of the data to be merged and select horizontal or vertical federation. Horizontal federation merges data based on sample ranges, while vertical federation merges data based on feature dimensions. The data holders execute the selected horizontal or vertical federation merge operation, using homomorphic encryption to process the share data and obtain the merged result. The data holders then negotiate a unified timeline benchmark in an encrypted state using the secure multi-party computation protocol, calculating the offset of each data set relative to the unified timeline benchmark using the timestamps of the share data. Finally, the data holders perform time-series alignment of the share data based on the offsets, obtaining time-series aligned share data and generating a joint dataset.

[0006] Furthermore, the multiple data holders compare the field patterns, sample ranges, and feature dimensions of the share data in an encrypted state based on a secure multi-party computation protocol to obtain comparison results. This includes: the multiple data holders encrypting the share data using homomorphic encryption to generate encrypted share data; the multiple data holders calculating the similarity of field patterns, the overlap of sample ranges, and the complementarity of feature dimensions in the encrypted share data using the secure multi-party computation protocol; generating the comparison results based on the similarity, overlap, and complementarity, wherein the comparison results indicate alignable fields, overlapping sample ranges, and complementary feature dimensions; and the multiple data holders extracting alignable fields from the comparison results for field mapping, overlapping sample ranges for sample alignment, and complementary feature dimensions for feature supplementation.

[0007] Furthermore, the multiple data holders determine the specific part of the data merging based on the comparison results and select horizontal federation or vertical federation, including: the multiple data holders obtaining the coverage and feature richness of each party's dataset; calculating the merging utility value based on the coverage and feature richness; if the overlapping sample range exceeds a preset range, then selecting horizontal federation to merge the sample range; if the complementary feature dimension exceeds a preset dimension, then selecting vertical federation to merge the feature dimension; the multiple data holders determine that the specific part of the data merging is a combination of the alignable fields, the overlapping sample range, and the complementary feature dimension.

[0008] Furthermore, the plurality of data holders calculate a combined utility value based on the coverage and the feature richness, including: the plurality of data holders aggregating the coverage of each party through a secure multi-party computation protocol to obtain a total coverage; the plurality of data holders calculating the ratio of the total coverage to the coverage of each party's dataset as coverage utility; the plurality of data holders aggregating the feature richness of each party to obtain a total feature richness; the plurality of data holders calculating the ratio of the total feature richness to the feature richness of each party's dataset as richness utility; and the plurality of data holders adding the coverage utility and the richness utility to obtain the combined utility value.

[0009] Furthermore, the multiple data holders perform a selected horizontal federation or vertical federation merging operation, and use homomorphic encryption to process the share data to obtain the merging result, including: if the horizontal federation is selected, the multiple data holders perform joint aggregation on the share data within the overlapping sample range; if the vertical federation is selected, the multiple data holders perform feature concatenation on the share data within the complementary feature dimensions; the multiple data holders apply homomorphic encryption to calculate intermediate results during the joint aggregation or feature concatenation process; the multiple data holders recover the merging result from the intermediate results, and the merging result represents the structure of the combined dataset.

[0010] Furthermore, the multiple data holders negotiate a unified timeline benchmark in an encrypted state through a secure multi-party computation protocol, and calculate the offset of each party's dataset relative to the unified timeline benchmark using the timestamps of the share data. This includes: the multiple data holders secretly sharing the timestamps of the share data to generate timestamp shares; the multiple data holders calculating the minimum value of each party's timestamp shares using the secure multi-party computation protocol as the unified timeline benchmark; the multiple data holders using homomorphic encryption to calculate the difference between each party's timestamp shares and the unified timeline benchmark as the offset; and the multiple data holders verifying the consistency of the offset and distributing it to each party.

[0011] Furthermore, the plurality of data holders use homomorphic encryption to calculate the difference between the timestamp share of each party and the unified timeline benchmark as the offset, including: the plurality of data holders homomorphically encrypting the timestamp shares of each party to obtain encrypted timestamp shares; the plurality of data holders calculating the encrypted timestamp shares minus the encrypted value of the unified timeline benchmark; the plurality of data holders decrypting the calculation result to obtain the offset of each party; and the plurality of data holders adjusting the timestamp position of the share data according to the offset.

[0012] Furthermore, the plurality of data holders perform time-series alignment of the share data according to the offset to obtain time-series aligned share data and generate a joint dataset, including: the plurality of data holders applying the offset to the share data of each party to perform timestamp shifting; the plurality of data holders verifying the consistency of the shifted timestamps on the unified time axis reference; the plurality of data holders inputting the time-series aligned share data into the merge operation; and the plurality of data holders outputting the joint dataset from the merge operation.

[0013] Secondly, the present invention provides a privacy-protected data circulation system, which mainly includes: A sharing generation module is used for multiple data holders to secretly share their local datasets to generate share data, which is distributed among the multiple data holders. An encrypted comparison module is used by the multiple data holders to compare the field patterns, sample ranges, and feature dimensions of the share data in an encrypted state based on a secure multi-party computation protocol, and to obtain the comparison results. The part determination module is used by the multiple data holders to determine the specific part of the data to be merged based on the comparison results, and to select horizontal federation or vertical federation. The horizontal federation merges samples based on the range, and the vertical federation merges features based on the dimension. The merge operation module is used by the multiple data holders to perform a selected horizontal or vertical federation merge operation, and to process the share data using homomorphic encryption to obtain the merge result. The negotiation and unification module is used by the multiple data holders to negotiate and unify the time axis benchmark in an encrypted state through a secure multi-party computation protocol, and to calculate the offset of each party's dataset relative to the unified time axis benchmark using the timestamp of the share data. The time alignment module is used by the multiple data holders to perform time alignment of the share data according to the offset, so as to obtain the time-aligned share data and generate a joint dataset.

[0014] Furthermore, the encryption comparison module includes: An encryption processing unit is used by the multiple data holders to encrypt the share data using homomorphic encryption to generate encrypted share data; The similarity calculation unit is used by the multiple data holders to calculate the similarity of field patterns, the overlap of sample ranges, and the complementarity of feature dimensions in the encrypted share data through a secure multi-party computation protocol. The comparison result generation module is used to generate the comparison result based on the similarity, overlap and complementarity, wherein the comparison result indicates the alignable fields, the range of overlapping samples and the dimension of complementary features; The alignment field extraction unit is used by the multiple data holders to extract alignable fields from the comparison results for field mapping, overlapping sample ranges for sample alignment, and complementary feature dimensions for feature supplementation.

[0015] The technical solutions provided by the embodiments of the present invention have the following beneficial effects: This invention discloses a distributed data collaborative processing method based on secure multi-party computation and federated learning. It addresses the unique business scenario of data merging and temporal alignment among multiple data holders while protecting privacy, particularly the challenge of securely and efficiently determining the merging points and ensuring temporal consistency when data field patterns, sample ranges, and feature dimensions are heterogeneous. This invention generates share data through secret sharing, utilizes homomorphic encryption and secure multi-party computation protocols, compares data characteristics under encrypted conditions, intelligently selects horizontal or vertical federated merging methods, negotiates a unified time axis benchmark, and calculates offsets for temporal alignment, ultimately generating a joint dataset. The core innovation of this invention lies in automating and intelligently performing multi-dimensional data alignment and merging while ensuring data privacy through encryption technology. This significantly improves the security and efficiency of distributed data collaboration, providing reliable technical support for cross-organizational data sharing and joint modeling. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a privacy-preserving data flow method according to the present invention.

[0017] Figure 2 This is another schematic diagram of a privacy-preserving data circulation method according to the present invention.

[0018] Figure 3 This is another schematic diagram of a privacy-preserving data circulation method according to the present invention.

[0019] Figure 4 This is another schematic diagram of a privacy-preserving data circulation method according to the present invention.

[0020] Figure 5 This is another schematic diagram of a privacy-preserving data circulation method according to the present invention.

[0021] Figure 6 This is another schematic diagram of a privacy-preserving data circulation method according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0023] like Figures 1-6The privacy-protected data circulation method provided in this embodiment may specifically include the following steps: S1, multiple data holders secretly share their local datasets to generate share data, which is used for merging operations in subsequent secure multi-party computation protocols.

[0024] Each data holder segments its local dataset, dividing the original data into multiple data fragments and encrypting each fragment to form encrypted data shares for data protection during subsequent collaborative computation. These encrypted data shares are distributed to participating nodes, ensuring that each node receives only a partial share to prevent any single node from holding the complete data, thus protecting data privacy. Based on their encrypted data shares, each node performs a preliminary data merging operation using a secure multi-party computation protocol, maintaining the data's encryption throughout the computation to generate an intermediate merge result. The intermediate merge result undergoes further verification to determine if it meets preset integrity conditions. If it does, the data merging is complete, enabling collaborative data operations among multiple data holders while protecting privacy.

[0025] In one embodiment, each data holder first treats the local dataset as the original data set and performs uniform segmentation on it, dividing the original data set into several data segments. The purpose of this segmentation is to disperse data risks and make it less likely for subsequent operations to leak complete information.

[0026] For example, when distributing encrypted data shares to the participating nodes, it is necessary to ensure that the distribution mechanism uses a random allocation method. For instance, in joint analysis of financial data, a bank's transaction record share is distributed to multiple server nodes, with each node receiving only a portion of the share. This distribution has the beneficial effect of avoiding centralized data leakage because if a node is attacked, only limited information is exposed. During the distribution process, a share allocation log is also recorded for subsequent tracking. This helps improve the robustness of the overall system and connects with the aforementioned encrypted shares, ensuring privacy protection throughout the entire process.

[0027] In one possible implementation, when each node performs an initial data merging operation based on the acquired encrypted data share, it coordinates through a secure multi-party computation protocol. For example, in supply chain data collaboration, the supplier's inventory share and the retailer's sales share are added in an encrypted state to generate an intermediate merged result. The purpose of maintaining the encrypted state of the data in this operation is to prevent the intermediate result from being reverse engineered, thereby improving the security of the computation process.

[0028] Specifically, the protocol involves bit-by-bit comparison and aggregation of shares without decrypting the original values. This connects with the privacy protections in the aforementioned distribution phase, ensuring that the merging operation does not introduce new risks. Further verification of the intermediate merging results checks whether they meet preset integrity conditions. For example, in IoT data fusion, the sum of the merged sensor readings is checked to see if it matches the expected range. If it does, the data merging is complete. This verification process effectively verifies accuracy because it detects potential calculation errors or tampering. It directly connects with the results of the aforementioned merging operation, enabling multiple data holders to jointly operate their data under privacy protection, thus supporting the reliable application of the secret sharing mechanism as a whole.

[0029] S2, the multiple data holders compare the field patterns, sample ranges and feature dimensions of their respective share data in an encrypted state based on a secure multi-party computation protocol, and the comparison is used to identify alignable mergeable parts.

[0030] In an encrypted state, the multiple data holders compare the patterns of their respective data fields using a secure multi-party computation protocol to obtain similarity information on the structural distribution and naming rules of the fields, generating a preliminary field matching result set. For this preliminary result set, the multiple data holders further compare the overlap of sample ranges in an encrypted state, extracting the intersection of sample identifiers to generate a sample overlap interval record. For this sample overlap interval record, the multiple data holders continue to compare the complementarity of feature dimensions in an encrypted state, determining the differences in feature coverage among the data holders, forming a feature dimension comparison list. Based on the feature dimension comparison list, the multiple data holders comprehensively determine alignable merging parts according to the preliminary field matching result set and the sample overlap interval record, generating a final merging part mapping table for subsequent part alignment in data fusion scenarios.

[0031] In one embodiment, multiple data holders compare data field patterns in an encrypted state using a secure multi-party computation protocol. This protocol allows parties to jointly compute results without exposing the original data. For example, in a financial risk control scenario, one bank holds customer credit score fields, while another holds transaction history fields. They transmit field structure distribution information, such as whether the field type is numeric or string, and naming rules, such as the similarity between credit_score and credit_rating, through the protocol to generate a preliminary set of field matching results. This ensures privacy protection while identifying potential alignment points, which is beneficial for subsequent merging efficiency. Specifically, for the preliminary set of field matching results, the degree of overlap of sample ranges is further compared. For example, assuming one party's data has customer IDs from 1 to 1000, and the other party's from 500 to 1500, they calculate the intersection, such as the sample identifiers from 500 to 1000, in an encrypted state, thereby extracting the intersection part to generate a sample overlap interval record. This method helps avoid computational waste caused by invalid samples and improves the accuracy of data federation, because the overlap interval directly supports the possibility of horizontal federation.

[0032] In one embodiment, for records of overlapping sample intervals, the complementarity of feature dimensions is further compared. For example, one side has age and income characteristics, while the other side has occupation and education level. Differences, such as complementary occupational dimensions, are identified under encrypted conditions, forming a feature dimension comparison list. This reveals opportunities for vertical federation, enriching the overall feature space rather than creating redundancy, thus providing a more comprehensive risk assessment basis in the risk control model. Specifically, for the feature dimension comparison list, the merging points that can be aligned are comprehensively determined based on the preliminary result set of field matching and records of overlapping sample intervals. For example, if the fields are similar and the samples overlap sufficiently, it is determined to be a horizontal merge; otherwise, it is a vertical merge. A final merge point mapping table is generated for point alignment in data federation scenarios. This comprehensive judgment ensures the optimal merging strategy, which helps minimize privacy leakage risks and optimize computational resources.

[0033] S3, In response to the comparison result, the multiple data holders determine the specific part of the data merging and select horizontal federation or vertical federation, wherein the horizontal federation is used for sample merging and the vertical federation is used for feature merging.

[0034] For datasets from multiple data holders, field descriptions and sample distribution information are obtained from each party. Through encrypted comparison, the semantic similarity between fields and the overlap of sample identifiers are determined, forming preliminary field correspondences and sample overlap range records. From these records, the degree of matching between fields and the proportion of sample overlap are extracted. It is determined whether data holders with high field matching are suitable for feature merging. If the matching degree meets a preset threshold, they are recorded as feature merging candidate groups. For the sample overlap range records, the overlap ratio of sample identifiers is analyzed to see if it reaches a preset threshold. If the overlap ratio is high, the relevant data holders are recorded as sample merging candidate groups, and the feature merging candidate groups and sample merging candidate groups are categorized separately. Based on these candidate groups, the specific parts of the data merging are determined. If the categorization results favor feature merging, a vertical federation approach is selected for feature dimension integration; if they favor sample merging, a horizontal federation approach is selected to expand the sample range, ultimately achieving precise location of the data merging part.

[0035] For example, in financial risk control scenarios, multiple banks act as data holders, each holding customer credit data. They obtain field descriptions such as age and income, as well as sample distribution information such as customer ID ranges from each party. They calculate the semantic similarity between fields through encrypted comparison, that is, assess whether the field names and types are similar, and the overlap of sample identifiers, that is, check the proportion of the same customer IDs, forming a preliminary field correspondence, such as the correspondence of the age field, and a record of the overlap range of samples, such as the percentage of overlapping customers. This ensures the identification of mergeable parts under privacy protection, which is beneficial for subsequent efficient data integration and avoids the waste of resources caused by blind merging.

[0036] Extract matching degree, such as similarity score, and overlap ratio, from the records of field correspondence and sample overlap range. Determine whether the data holder with a high degree of field matching is suitable for feature merging. That is, if multiple fields, such as income and loan records, can be aligned, they are considered complementary. If the matching degree meets a preset threshold, such as 0.8 or above, the record is a candidate group for feature merging. This process is beneficial for accurately screening complementary data and improving the accuracy of model training, because it can integrate more dimensional information without introducing irrelevant noise.

[0037] In one embodiment, the overlap ratio of the overlapping samples is analyzed to determine if it reaches a preset threshold, such as 60%. If the ratio is high, it indicates that the samples are expandable. The relevant data holders are recorded as sample merging candidate groups, and the feature merging candidate groups and sample merging candidate groups are categorized separately. This helps distinguish merging types and avoids calculation errors caused by confusion. A high overlap ratio supports sample expansion to increase training volume, while the categorization process ensures clear subsequent selection logic, thereby optimizing overall data utilization efficiency. For example, based on the feature merging candidate groups and sample merging candidate groups, the specific parts to be merged are determined, i.e., the intra-group tendency is evaluated. If feature merging is preferred, a vertical federated integration dimension is selected, such as adding new fields; if sample merging is preferred, a horizontal federated expansion range is selected, such as adding customer samples. Ultimately, precise positioning is achieved. This helps minimize privacy risks because only necessary parts are merged, improving model robustness in cross-institutional collaboration. Simultaneously, this selection mechanism supports intelligent training of the risk control model, avoiding the leakage risks caused by invalid data participation.

[0038] S4, the multiple data holders perform the selected federated merge operation, using homomorphic encryption to ensure the correctness of the calculation results during the merge process.

[0039] Multiple data holders perform additive secret sharing on each column attribute value of their local dataset, dividing the original attribute value into multiple secret shares, which are distributed to all participants, ensuring that each participant holds a corresponding share for all records. After receiving all secret shares, each participant performs homomorphic addition on its secret share for the same attribute column, obtaining an encrypted intermediate sum for that attribute across the entire dataset. This intermediate value remains encrypted within the homomorphic encryption domain. Each participant then performs homomorphic multiplication on the encrypted intermediate sum to calculate the product terms required for the federated merge operation, resulting in the complete encrypted aggregation result. All participants jointly hold a share of the encrypted aggregation result, which is directly used as the final output of the federated merge operation in its encrypted state for correct calculations within the subsequent homomorphic encryption domain.

[0040] For example, when multiple data holders process a local dataset, they first perform addition and secret sharing for each column of attribute values. This sharing method divides the original attribute values ​​into multiple secret shares and distributes them, so that each participant holds the corresponding share of all records. The purpose of this is to ensure data privacy, because a single participant cannot recover the original value from the share, thereby protecting sensitive information in joint computation. The beneficial effect is to improve the security of data circulation.

[0041] In one possible implementation, after each participant receives all the secret shares, they perform homomorphic addition on the secret shares of the same column attribute to obtain the encrypted summation intermediate value of that column attribute on the entire dataset. This intermediate value is kept within the homomorphic encryption domain. This operation allows direct addition in the encrypted state without decryption, which has the advantage of ensuring the correctness and efficiency of the calculation.

[0042] In one possible implementation, all participants jointly hold a share of the encrypted aggregation result, which is directly used as the final output of the federated merge operation for correct computation within the subsequent homomorphic cryptographic domain. This facilitates further joint analysis and maintains end-to-end privacy throughout the process. For example, in cross-institutional research, this output can be directly used for model training without revealing the source data. For instance, in a specific implementation of additive secret sharing, assuming one party has the value 10, it can be divided into two shares, 3 and 7, distributed to the two parties. Each party only knows a portion and not the whole picture, thus preventing data theft.

[0043] In one possible implementation, the homomorphic addition operation involves adding the encrypted shares, such as adding two encrypted shares [3] and [7] together

[10] , while maintaining encryption. The advantage of this is that it supports secure merging. For example, the homomorphic multiplication operation is extended to product calculation, such as multiplying the sum by a coefficient to obtain an aggregate value. The advantage of this is that it enhances computational flexibility.

[0044] In one possible implementation, a shared mechanism of shareholding allows participants to collaboratively decrypt the final result only when needed, achieving a balance between privacy and availability. For example, the chain from sharing to aggregation ensures privacy continuity, resulting in more reliable overall data flow.

[0045] S5, during the merging process, the multiple data holders negotiate a unified timeline benchmark through a secure multi-party computation protocol. The unified timeline benchmark is used to calculate the offset of each party's dataset relative to the benchmark.

[0046] During the merging process, the multiple data holders encrypt the timestamp information of their respective datasets using a secure multi-party computation protocol. Key time point features are extracted from the encrypted timestamp data of each party for subsequent negotiation to determine a unified timeline benchmark. Based on the extracted key time point features, the data holders jointly calculate a unified reference time point in an encrypted state. This reference time point serves as the unified timeline benchmark. By comparing the differences between each party's dataset and this benchmark, a preliminary basis for calculating the time offset is established. Starting from this unified timeline benchmark, each data holder uses the encrypted time point features to calculate the time offset of their respective dataset relative to the benchmark, ensuring that the offset is calculated under privacy protection and generating corresponding adjustment parameters. Based on the generated adjustment parameters, each data holder applies the parameters to adjust the time dimension of their respective datasets, ensuring that the merged dataset is consistent with the unified timeline benchmark in the time dimension, thus completing the time alignment process.

[0047] For example, during the merging process, the multiple data holders encrypt the timestamp information of their respective datasets through a secure multi-party computation protocol. The secure multi-party computation protocol refers to the specific implementation method in which each party inputs its own private data and then jointly executes the computation function without disclosing the input in plaintext. Key time point features, such as the earliest and latest record times of the datasets, are extracted from the encrypted timestamp data of each party for subsequent negotiation to determine the unified timeline benchmark.

[0048] For the extracted key time point features, all data holders jointly calculate a unified reference time point under encrypted conditions. Specifically, a secure multi-party computation protocol is used to calculate the median or average value among the features of each party as the reference time point, i.e., the unified time axis benchmark. By comparing the differences between each party's dataset and this benchmark, a preliminary basis for calculating the time offset is formed.

[0049] Starting from the unified timeline benchmark, each data holder uses encrypted time point features to calculate the time offset of their own dataset relative to the benchmark. The offset is defined as the difference between the average time point of each dataset and the benchmark, ensuring that the offset is calculated under privacy protection and that corresponding adjustment parameters are generated.

[0050] For the generated adjustment parameters, each data holder applies the parameters to adjust the time dimension of their own dataset, that is, to add or subtract the corresponding offset for each timestamp, to ensure that the merged dataset is consistent with the unified time axis benchmark in the time dimension and complete the time alignment process.

[0051] S6, the multiple data holders perform time-series alignment of the share data according to the calculated offset, and the time-series alignment ensures that the merged dataset maintains consistency in the time dimension.

[0052] The multiple data holders first acquire the timestamp information from their respective datasets, extracting the collection time of each data record to form an initial time series set. For these initial time series sets, the parties share necessary time range boundaries using encryption, ensuring a common time base interval is determined without exposing specific data content. After determining the common time base interval, the multiple data holders use this interval to calculate the offset of their respective datasets relative to the base interval. For the calculated offsets, the parties encrypt the offsets using a security protocol, forming an encrypted offset set for subsequent time series adjustment. For the encrypted offset set, the multiple data holders adjust their respective share data according to the offsets in the time dimension, generating an adjusted share data set. During the adjustment process, it is ensured that each party's data records are aligned with the base interval in the time dimension. From the adjusted share data set, the multiple data holders merge the data, ensuring that the merged dataset maintains consistency in the time dimension, achieving the goal of time series alignment.

[0053] In one possible implementation, multiple data holders obtain timestamp information from their respective datasets, extract the collection time of each data record, and form an initial time series set. The purpose of this approach is to provide basic data support for subsequent alignment.

[0054] Specifically, all parties share necessary time range boundaries via encryption to ensure a common time base interval is determined without exposing specific data content. This boundary-sharing process is achieved through a secure multi-party computation protocol, where boundaries such as the earliest and latest time points are transmitted encrypted. This serves to negotiate and unify the benchmark, preventing the leakage of sensitive patterns. The resulting technical effect is enhanced data security and alignment accuracy. For example, after determining the common time base interval, multiple data holders use this interval to calculate the offset of their respective datasets relative to the benchmark interval. This calculation involves comparing the differences between their local time series and the benchmark. For instance, if the benchmark interval starts from a certain standard starting point, the deviation of the local dataset's starting point from it is used as the offset. This quantifies time inconsistencies. The calculated offsets are then encrypted by all parties using a secure protocol, forming an encrypted offset set for subsequent time series adjustments. This encryption uses homomorphic encryption, allowing computation in an encrypted state. Its purpose is to protect the offsets from direct reading, thus maintaining collaborative computation while preserving privacy.

[0055] In one possible implementation, for the encrypted offset set, multiple data holders adjust their respective share data along the time dimension according to the offset, generating an adjusted share data set. This adjustment process is achieved by adding or subtracting the offset from the timestamp of each data record. For example, if a record's timestamp lags behind the baseline, it is shifted forward by the corresponding amount. This achieves data synchronization on the timeline. During the adjustment process, it is ensured that the data records of each party are aligned with the baseline interval in the time dimension. This assurance is accomplished by verifying whether the adjusted sequence falls within the baseline range, thus eliminating time-series deviations. The resulting technical effect is improved usability and consistency of the merged data.

[0056] For example, from the adjusted share dataset, multiple data holders merge the data to ensure that the merged dataset remains consistent over time, achieving the goal of time-series alignment. This merging process involves combining the adjusted shares through a secret sharing protocol.

[0057] S21, the multiple data holders compare the field patterns, sample ranges and feature dimensions of their respective share data in an encrypted state based on a secure multi-party computation protocol. The comparison is used to identify alignable mergeable parts, including identifying overlapping fields and complementary feature dimensions.

[0058] In an encrypted state, the multiple data holders compare the patterns of their respective data fields using a secure multi-party computation protocol to obtain matching information on field names and types, forming a preliminary field correspondence table for subsequent determination of alignable merging points. Based on this preliminary table, the data holders further compare the degree of overlap in sample ranges, matching encrypted sample identifiers to determine the distribution range of overlapping samples, generating a sample overlap distribution record to provide a basis for selecting merging points. Building upon this record, the data holders compare the complementarity of feature dimensions, identifying the feature coverage and missing parts of each party, forming a feature complementarity list to determine whether to merge samples or features. Based on the feature complementarity list and the sample overlap distribution record, the data holders determine the final merging point using preset threshold judgment rules, ensuring that the comparison results accurately identify alignable fields and dimensions, completing the identification of merging points in data collaboration scenarios.

[0059] In one possible implementation, the multiple data holders compare the patterns of their respective data fields in an encrypted state using a secure multi-party computation protocol. For example, in a financial risk control scenario, one bank holds the customer credit score field, while another holds the transaction history field. By encrypting and comparing field names such as "credit_score" and "transaction_date" and their types such as numeric and date, the matching results are obtained to form a preliminary field correspondence table. This approach ensures the identification of common fields while protecting privacy, which is beneficial for subsequent merging to avoid the risk of data leakage. It also supports the improvement of joint training efficiency from multiple perspectives, as the matching results directly affect the accuracy of data alignment.

[0060] It should be noted that this comparison process involves the exchange of encrypted pattern summaries by all parties without exposing the original data, thus maintaining trust in cross-institutional collaboration. Further comparison of the overlap of sample ranges is performed on the initial field correspondence table. For example, assuming the data samples from two institutions are identified by encrypted user IDs, the overlap is found by matching, such as the distribution range of 1000 user samples that are jointly covered, such as the overlapping part from ID 1 to ID 500. This generates a sample overlap distribution record. This provides a basis for selecting the merging points and is beneficial for optimizing the choice of federated learning type, because a high degree of overlap tends to favor feature merging, while a low degree of overlap favors sample merging, supporting each other from multiple directions, such as combining field matching to avoid invalid computation.

[0061] In one embodiment, the complementarity of feature dimensions is compared based on the overlapping distribution records of samples. For example, one side of the data may have income characteristics but lack education level, while the other side may have the opposite. By identifying the completeness of coverage, such as income data, and the missing parts, such as gaps in education data, a list of feature complementarities is formed. This can be used to determine whether to merge samples or features, which is beneficial for intelligently selecting the optimal merging method because the list clearly shows the complementary points and considers multiple aspects, such as combining with overlapping records to ensure comprehensiveness. Specifically, this comparison process is achieved through encrypted dimensional vector comparison, without the need for plaintext transmission, thereby enhancing the value of data utilization while maintaining privacy. For example, in e-commerce data scenarios, the list shows that one side has browsing history while the other side has purchasing preferences, supporting vertical federated decision-making to enrich feature dimensions. For example, based on the feature complementarity list and the sample overlap distribution record, the final merging part is determined by a preset threshold judgment rule. For example, if the overlap sample ratio exceeds 50% and there are more than 3 complementary features, it is judged as vertical federation; otherwise, it is horizontal. This ensures that the comparison results accurately identify the fields and dimensions that can be aligned, which is beneficial for completing the identification of merging parts in data federation scenarios. This is because the rule is based on quantitative thresholds such as ratio thresholds to avoid subjective bias and supports the integration of lists and records from multiple directions to improve the reliability of decision-making.

[0062] It should be noted that this judgment rule is a predefined set of conditions, such as the interval coverage of overlapping distribution records combined with the complementary count of the list, to form a logical chain. In practical applications, such as when telecommunications data is combined, this process helps to accurately align user behavior fields and location dimensions, and optimize the training efficiency of risk control models.

[0063] S22, the multiple data holders intelligently select the optimal merging part based on the coverage and feature richness of each party's dataset, and the selection determines whether to perform horizontal federation or vertical federation.

[0064] For different datasets held by multiple data holders, the field distribution and sample coverage information of each party are first obtained. By comparing the field names and data ranges of each dataset, potential correspondences between fields and overlapping areas of samples are determined, and a data table recording the field matching degree and sample overlap ratio is created. From the data table of field matching degree and sample overlap ratio, the field consistency and sample intersection size between each dataset are extracted. For fields with high consistency and sample ranges with large intersection, the complementarity of feature dimensions and the breadth of sample coverage are judged to form a preliminary evaluation result for the merging part. Based on the preliminary evaluation result, the richness of feature dimensions and the comprehensiveness of sample coverage are further analyzed. If the feature dimensions are highly complementary, the direction of feature merging is favored; if the sample coverage is broad, the direction of sample merging is favored, generating a priority ranking list of merging parts. From the priority ranking list, the merging part with the highest priority is selected to determine whether feature merging or sample merging will be performed in the end, ensuring that the optimal merging part is intelligently selected based on coverage and feature richness during the data union process.

[0065] For different datasets held by multiple data holders, the field distribution and sample coverage information of each party are first obtained. By comparing the field names and data ranges of each dataset, potential correspondences between fields and overlapping areas of samples are determined, and a data table recording the degree of field matching and the proportion of sample overlap is created. Specifically, the degree of field matching is obtained by comparing the string similarity of field names and the compatibility of data types. For example, in one embodiment, for the field names "Customer ID" and "User Identifier" in two datasets, if the similarity exceeds a preset threshold, they are considered to be potential correspondences. At the same time, the data ranges, such as numerical intervals, are checked to confirm the overlapping areas of samples. The advantage of this approach is that it can initially identify mergeable parts and avoid computational waste caused by blind comparison.

[0066] From the data table of field matching degree and sample overlap ratio, extract the field consistency and sample intersection scale between each dataset. For fields with high consistency and sample ranges with large intersection, determine the complementarity of feature dimensions and the breadth of sample coverage to form a preliminary evaluation result of the merged part.

[0067] In one possible implementation, field consistency refers to the reliability of field correspondence, which is obtained by calculating the average of the matching degree. The sample intersection size is quantified based on the size of the overlapping area. For example, for a financial dataset, if one side has a credit score field while the other side does not, it is judged to be highly complementary. The beneficial effect of doing so is to highlight the complementary value between data and ensure that subsequent selections are more accurate.

[0068] Based on the preliminary evaluation results, further analysis is conducted on the richness of feature dimensions and the comprehensiveness of sample coverage. If the feature dimensions are highly complementary, the direction of feature merging is preferred; if the sample coverage is broad, the direction of sample merging is preferred. A priority ranking list of merging parts is generated.

[0069] For example, the richness of feature dimensions is evaluated by counting the number of unique fields, while the comprehensiveness of sample coverage is examined by the ratio of overlapping to non-overlapping samples. For instance, in risk control scenarios, if one side's data covers more unique features, such as transaction history, then feature merging is prioritized. This optimizes data utilization efficiency and reduces privacy risks. From the priority ranking list, the highest priority merging part is selected to determine the final method of feature merging or sample merging, ensuring that the optimal merging part is intelligently selected based on coverage and feature richness during data federation. It can be understood that the priority ranking list is arranged according to a complementary and breadth-weighted score, selecting the highest priority, such as performing vertical federation during feature merging. This enables intelligent decision-making and improves the security and effectiveness of joint training.

[0070] S23, the horizontal federated merge operation combines datasets for overlapping sample ranges, and the vertical federated merge operation aggregates features for complementary feature dimensions.

[0071] Encrypted dataset description information is obtained from each participant. This description information includes sample identifiers and metadata for feature fields. Overlapping sample ranges and complementary feature dimensions are identified in the encrypted state using a secure comparison protocol, forming a preliminary joint range mapping table for subsequent merging decisions. For overlapping sample ranges in the joint range mapping table, a sample alignment operation is performed to associate sample data with the same identifier, forming a unified sample set. Simultaneously, the data coverage of each participant in this set is recorded, providing a basis for horizontal merging. For complementary feature dimensions in the joint range mapping table, a feature field matching operation is performed to aggregate feature fields provided by different participants, forming an extended feature set and ensuring the non-redundancy of feature fields, supporting vertical merging. Based on the sample set and the extended feature set, a horizontally merged data view and a vertically merged feature view are constructed. The horizontally merged view is used for merging datasets with overlapping sample ranges, and the vertically merged view is used for feature aggregation of complementary feature dimensions, completing the final merging operation.

[0072] For example, in cross-institutional data collaboration scenarios, the process of obtaining encrypted dataset description information from each participant involves each institution first hashing and encrypting the sample identifiers and feature fields of its own dataset to form metadata packets. These metadata packets are transmitted to the coordinating node through a secure channel, thereby ensuring that data patterns are compared without exposing the original content. The secure comparison protocol specifically refers to using homomorphic encryption methods, allowing each participant to calculate the intersection of sample identifiers and the difference of feature fields within the encrypted domain. For example, it calculates the number of overlapping samples and the list of complementary features using additive homomorphic attributes, ultimately generating a joint range mapping table. This table records the identifier range of shared samples and the dimensional list of unique features. This approach allows subsequent decisions to be based on accurate matching information, avoiding privacy risks caused by blind merging.

[0073] In one embodiment, when performing sample alignment operations on overlapping sample ranges in the joint range mapping table, the intersection portion identified in the table is first extracted. Then, each participant exchanges the index of encrypted samples using a key-sharing protocol, associating matching samples into a unified set. Simultaneously, the source coverage of each sample is marked; for example, participant A covers the sample's credit record while participant B covers the transaction history. This helps to efficiently integrate different attributes of the same sample during horizontal joint operations, improving the sample completeness of model training. Specifically, when the feature field matching operation aggregates complementary feature dimensions, it checks the list of non-overlapping fields in the table. Through semantic similarity calculations, such as using cosine similarity to compare field description vectors, similar fields are merged and redundancy is removed. For example, the "age" field of one institution and the "year of birth" field of another are aggregated into a single age feature. This expands the richness of the feature set, providing a more comprehensive input dimension for vertical joint operations and avoiding computational redundancy caused by feature duplication.

[0074] In one embodiment, the process of constructing a data view based on a sample set and an extended feature set includes creating a horizontal joint view that horizontally concatenates data rows from the unified sample set according to their source, while the vertical joint view vertically appends the extended feature set to the samples. This ensures that the merging operation precisely targets overlapping samples and complementary features, optimizing dataset merging and feature aggregation, and improving the accuracy and privacy protection of financial risk control models. For example, from multiple perspectives, this view construction can also support dynamic adjustments, such as prioritizing data from participants with high coverage when sample coverage is uneven, which further enhances the efficiency and security of the merging. Specifically, in practical applications, if the range of overlapping samples is small but complementary features are abundant, the aggregation of the vertical view can significantly improve the model's generalization ability, while the merging of the horizontal view can amplify the training scale when the sample size is sufficient. These mutually supportive directions together ensure the effectiveness of federated learning and avoid the limitations caused by data silos.

[0075] In one embodiment, the formation of the initial joint range mapping table can also serve as a feedback mechanism to guide the participants to optimize the metadata encryption method, such as adjusting the hash function to improve the comparison accuracy. This is closely linked to the subsequent alignment and matching operations, forming a closed-loop optimization and improving the robustness of the overall merging operation.

[0076] S31, the multiple data holders compare the field patterns, sample ranges and feature dimensions of their respective share data in an encrypted state, including using homomorphic encryption to calculate the pattern similarity of the share data of each party.

[0077] Each data holder splits its share of data into multiple shares via secret sharing, including the field names and value distributions, and submits them to a secure multi-party computation node. Under encrypted conditions, the shares of all field names are compared bit-by-bit to determine the consistency share of each field's existence across all parties' data. These consistency shares are then accumulated using homomorphic encryption to obtain the frequency share of each field across all data holders. This frequency share is then compared with the total number of participants using homomorphic encryption to determine the encryption identifier of the global common field set. For each field in the global common field set, the statistical histogram share of that field is extracted from each party's share data. The sum of squared Euclidean distances across all histogram shares is calculated under homomorphic encryption to obtain an encrypted value representing the similarity between the sample range and feature dimensions. The encrypted similarity values ​​of all fields are then accumulated using homomorphic encryption to obtain the overall share data's pattern similarity encryption result, which is directly used to determine whether the data from each party meets the merging criteria.

[0078] For example, in a multi-party data collaboration scenario, each data holder first applies a secret sharing protocol to the field names and value distribution of their respective data, splitting this information into multiple independent shares and uploading them to a secure computing node. This splitting ensures that the original data is not directly exposed, thereby maintaining privacy. The purpose of this approach is to allow all parties to participate in the computation without disclosing the specific content.

[0079] In one possible implementation, these consistent shares are accumulated using homomorphic encryption. This accumulation process allows for direct addition within the encrypted domain without decryption. For example, the presence identifier share of each field is accumulated to obtain an encrypted share representing the frequency of occurrence. Then, this share is homomorphically compared with the encrypted form of the total number of participants to determine whether all parties possess the field. This forms an encrypted identifier for the global set of common fields. This accumulation and comparison filters out reliable sets of shared fields, reducing redundancy in subsequent calculations, ensuring that only consistent parts are processed, thereby optimizing resource utilization and reducing the risk of privacy breaches.

[0080] Specifically, for a given set of global common fields, the statistical histogram share of each field is extracted from the share data of each party. The histogram here represents the discrete representation of the value distribution, such as the distribution buckets of the age field. Then, under homomorphic encryption, the sum of squared Euclidean distances between these histogram shares is calculated. The calculation process involves summing the squared encryption differences of the corresponding bucket values ​​to obtain the similarity encryption value. This characterizes the degree of matching between the sample range and the feature dimension. The purpose of this calculation is to quantify the similarity of the data distribution, and its beneficial effect is to help detect potential inconsistencies.

[0081] In one possible implementation, these encrypted similarity values ​​for all fields are further homomorphically encrypted and accumulated. The accumulated result yields an encrypted overall pattern similarity, which is directly applied to determine data merging conditions. For example, if the similarity is higher than a preset threshold, the data is considered to have a basis for merging. This overall accumulation provides a comprehensive evaluation, avoiding a single field dominating the decision. Its beneficial effect is to enhance the robustness of the judgment, support multi-party data comparison under privacy protection, and achieve the goal of encrypted comparison of field patterns, sample range, and feature dimensions.

[0082] S32, the alignment result indicates that the fields that can be aligned are used for field mapping, the overlapping sample ranges are used for sample alignment, and the complementary feature dimensions are used for feature supplementation.

[0083] Encrypted dataset description information is obtained from each participating party. This information includes the basic structure of field names, sample identifiers, and feature dimensions. Field names from each dataset are compared in an encrypted state using a secure multi-party computation protocol to determine an alignable set of fields for subsequent field mapping operations. For this alignable set of fields, the sample identifier ranges of each dataset are further compared to identify overlapping sample intervals. These overlapping intervals are used for sample alignment, ensuring that information from different datasets on the same sample can be correlated. Based on sample alignment, the distribution of feature dimensions across each dataset is analyzed, and a complementary set of feature dimensions is extracted. These complementary dimensions are used for feature supplementation to enrich the overall data's feature representation capabilities. A joint data view is constructed based on the alignable set of fields, overlapping sample intervals, and complementary feature dimension set. This view guides the subsequent data merging direction, ensuring coordinated operations of field mapping, sample alignment, and feature supplementation while protecting privacy.

[0084] For example, in a financial risk control scenario, encrypted dataset description information is obtained from various banking institutions. For instance, a bank's dataset description might include customer ID as a sample identifier, credit score as a field name, and income level as a feature dimension. A secure multi-party computation protocol is used to compare field names in an encrypted state. This ensures privacy is not compromised while identifying a set of alignable fields, such as credit scores. The beneficial effect is that subsequent field mapping operations are more efficient, avoiding errors caused by manual matching. In one possible implementation, the secure multi-party computation protocol involves each participating party exchanging encrypted field hash values ​​and then jointly calculating the intersection to determine the set of alignable fields. This helps achieve data pattern comparison without exposing the original data because this protocol uses homomorphic encryption technology to allow computation in encrypted domains, resulting in technical benefits such as reduced privacy risks and improved joint efficiency.

[0085] In one possible implementation, for the aforementioned set of alignable fields, the sample identifier ranges are further compared. For example, if the customer ID ranges of two banks are 1 to 1000 and 500 to 1500 respectively, overlapping sample intervals between 500 and 1000 are identified through encrypted comparison. This ensures sample alignment, allowing information from the same customer to be linked. The beneficial effects are avoiding data redundancy and improving model accuracy. Specifically, the comparison process uses a secure set intersection protocol. Each party inputs an encrypted identifier set and outputs a shared overlapping interval. This connects to the aforementioned field comparison because samples can only be effectively aligned on the basis of field alignment. The reason is that the overlapping interval directly supports subsequent merging, avoiding additional privacy risks caused by irrelevant samples participating in the calculation. The technical effects include optimizing resource utilization and enhancing data consistency.

[0086] For example, based on sample alignment, the distribution of feature dimensions can be analyzed. For instance, one institution may have an income level dimension while another has a consumption habit dimension. By extracting a set of complementary feature dimensions through encrypted statistics, the overall data expression capability can be enriched, and the beneficial effect is to improve the prediction accuracy of the risk control model.

[0087] In one possible implementation, the analysis process involves all parties computing encrypted dimension vectors and exchanging statistical summaries, such as dimension uniqueness indices, and then identifying non-overlapping portions as complementary sets. This is related to the aforementioned sample alignment, because only aligned samples can accurately assess dimensional complementarity. The rationale is that complementary dimension additions can fill gaps in a single dataset, bringing technical benefits such as improved feature coverage and reduced risk of overfitting.

[0088] In one possible implementation, a joint data view is constructed based on a set of alignable fields, overlapping sample intervals, and complementary feature dimensions. For example, the view maps credit scoring fields, aligns overlapping customer IDs, and supplements consumption habit dimensions. This guides the direction of data merging, ensuring that field mapping, sample alignment, and feature supplementation operate collaboratively while protecting privacy. A beneficial effect is the intelligent selection of horizontal or vertical federation, avoiding ineffective computation. Specifically, the construction process uses multi-party secure aggregation to generate view metadata, which connects to the aforementioned extraction steps because these sets form the basis of the view. The rationale is that the view provides a blueprint for merging. Technical benefits include minimizing privacy exposure and maximizing data utility, supporting efficient model training in financial risk control.

[0089] S41, the multiple data holders negotiate a unified timeline benchmark through a secure multi-party computation protocol, the negotiation being based on the secret sharing of timestamps of each party's share of data.

[0090] The multiple data holders first fragment their timestamp data into pieces. These fragmented timestamp pieces are then distributed to each participant via a secret sharing mechanism, ensuring that each party holds only a portion of the timestamp information and cannot reconstruct the complete timestamp data independently, thus generating distributed shared timestamp pieces. For these distributed shared timestamp pieces, each participant performs encrypted comparison operations on the pieces using a secure multi-party computation protocol to determine the relative order between the pieces, forming a preliminary timestamp order distribution map for subsequent benchmark negotiation. From this preliminary timestamp order distribution map, the participants further negotiate and determine a unified benchmark time point. Based on the order relationship in the distribution map, they obtain the offset of each timestamp piece from the benchmark time point through encrypted calculation, generating an encrypted offset set. For this encrypted offset set, each participant applies the offsets to adjust the timestamps in their respective datasets, completing time alignment operations in an encrypted state to ensure that the final dataset is consistent with the unified timeline benchmark in the time dimension.

[0091] In one possible implementation, the process of multiple data holders sharding the timestamp data in their respective datasets involves splitting the timestamp sequence into multiple smaller segments.

[0092] For example, for distributed shared timestamp segments, the specific implementation method of the comparison operation in the encrypted state by each participating party through a secure multi-party computation protocol is to use additive homomorphic encryption to compare the size of the segment values ​​without decrypting the data. For example, one party holds a share of segment A and the other party holds a share of segment B. They jointly calculate the encrypted result of A minus B and check the symbol. In this way, the relative order relationship is determined, forming a preliminary timestamp order distribution map. The beneficial effect is that a global order view can be established without exposing the actual time value, supporting subsequent negotiation without revealing the specific time pattern.

[0093] In one possible implementation, the process of negotiating a unified reference time point from the initial timestamp order distribution map includes each participant voting or averaging to calculate an intermediate point, for example, by calculating the average of all segments using cryptographic summation as the reference. Then, based on the order relationship in the distribution map, cryptographic calculation methods such as Paillier homomorphic encryption are used to obtain the difference between each segment and the reference, generating an encrypted set of offsets. The beneficial effect is to ensure the privacy and security of offset calculation and prevent any party from inferring the precise timestamps of other parties.

[0094] For example, the specific method for applying the encrypted set of offsets to the timestamp adjustment of each participant's dataset is to add the corresponding offset to their own data locally while maintaining the encryption state to complete the time alignment operation.

[0095] In one possible implementation, the overall application of this offset calculation mechanism in the secure merging of multi-party data can also be extended to handling differences in data update frequency. For example, when one party updates its data frequently while the other updates it more slowly, offset alignment ensures that the timeline is consistent after merging. The beneficial effect is to improve the accuracy and usability of the joint dataset and avoid analytical bias caused by time inconsistency.

[0096] S42, the multiple data holders use homomorphic encryption to calculate the offset of each party's dataset relative to a unified time axis reference, the offset maintaining the privacy of each party's data acquisition time pattern.

[0097] The multiple data holders first acquire the timestamp information of their respective datasets and protect these timestamps using homomorphic encryption to form an encrypted timestamp set. This encrypted timestamp set is shared with the participating parties without exposing the original data. For this encrypted timestamp set, each data holder compares it to a pre-established unified timeline benchmark, calculating an initial time deviation value between their dataset and the benchmark. This initial time deviation value is stored in encrypted form to ensure that each party cannot directly access the time information of other holders. Starting from the initial time deviation value, each data holder adjusts the deviation value through a secure negotiation mechanism to generate final offset data. This offset data is used for subsequent data alignment operations while maintaining the privacy of each party's data acquisition time pattern. For this final offset data, each data holder applies it to adjust the time dimension of their respective datasets, ensuring the consistency of the merged datasets on the timeline, while using encryption to prevent the leakage of specific acquisition time patterns, thus achieving the offset calculation target relative to the unified timeline benchmark.

[0098] For example, in a scenario of secure data merging among multiple parties, multiple data holders obtain the timestamp information of their respective datasets and perform protective processing on these timestamps using homomorphic encryption to form an encrypted set of timestamps. This processing ensures that the original timestamp data is not directly exposed during sharing, thereby maintaining privacy.

[0099] Specifically, homomorphic encryption allows addition or multiplication operations to be performed within the encrypted domain without decryption. For example, if a holder has a specific timestamp, after generating ciphertext using an encryption key, other parties can only operate on the ciphertext and cannot reverse-engineer the original value. This provides privacy protection because even with a shared set, parties cannot know the specific time patterns of others' data collection, helping to prevent the leakage of time-related information.

[0100] In one possible implementation, for the encrypted timestamp set, each data holder compares it against a pre-established unified timeline benchmark to calculate an initial time deviation value between their respective dataset and the benchmark. Here, the unified timeline benchmark refers to a negotiated standard time reference line, such as an axis constructed starting from the earliest timestamp. The calculation process involves subtracting the encrypted timestamp from the benchmark value, performed under homomorphic encryption. For example, if the benchmark is a fixed point, the deviation value is obtained by encrypted subtraction. Storing this deviation value in encrypted form further enhances privacy, ensuring that each party cannot directly obtain the time information of other holders, thereby avoiding the exposure of sensitive patterns in joint research and improving the security of data collaboration.

[0101] It should be noted that, starting from the initial time deviation value, each data holder adjusts the deviation value through a secure negotiation mechanism to generate the final offset data. The secure negotiation mechanism is a multi-party computation-based protocol that allows parties to jointly calculate the output without revealing their private inputs. For example, it can use secret sharing to split and reassemble the deviation value for adjustment. The purpose of this adjustment is to refine the deviation to fit a unified axis. The resulting offset data is used for subsequent alignment operations while maintaining the privacy of each party's data acquisition time pattern, effectively preventing pattern analysis attacks.

[0102] In one possible implementation, for the final offset data, each data holder applies it to adjust the time dimension of their respective datasets, ensuring the consistency of the merged datasets on the timeline. Encryption is used to prevent the leakage of specific data collection time patterns; for example, the offset is added as an encrypted addition factor to the timestamp, achieving alignment without decrypting the original data.

[0103] S51, the multiple data holders perform time-series alignment of share data according to the calculated offset, including adjusting the timestamps of each party's dataset to a unified time axis.

[0104] The multiple data holders first acquire the timestamp information from their respective datasets, extracting the collection time of each data record to form an initial time series set. For these initial time series sets, the parties share necessary time range boundaries using encryption, ensuring a preliminary time alignment range is determined without exposing specific data. Based on this preliminary time alignment range, the multiple data holders use the encrypted time range boundaries to calculate the preliminary offset of their respective datasets relative to a unified time axis. During the calculation, the parties exchange the preliminary offsets in encrypted form to ensure the privacy of the offset data, while simultaneously generating a unified offset benchmark value. Based on the generated unified offset benchmark value, the multiple data holders adjust their respective initial time series sets, shifting the timestamp of each data record according to the offset benchmark value to generate an adjusted time series set. The adjusted time series set is consistent with the unified time axis in the time dimension. Based on the adjusted time series set, the multiple data holders perform a final verification of the share data to ensure that the timestamp of each data record has been accurately aligned to the unified time axis, forming the final time series alignment result for subsequent data merging processing.

[0105] In one embodiment, based on an initial time alignment range, the plurality of data holders calculate an initial offset of their respective datasets relative to a unified time axis using the encrypted time range boundaries. Based on the generated unified offset baseline, the plurality of data holders adjust their respective initial time series sets.

[0106] It should be noted that, based on the adjusted time series set, the multiple data holders perform a final verification of the share data. For example, they compare the time intervals between adjacent records using encryption to see if they meet the expected continuity. If the interval is abnormal, a re-shift is triggered to ensure that the timestamp of each data record is accurately aligned to a unified time axis, forming the final time-series alignment result. This result directly supports subsequent data merging processing. After verification, the adjusted set in the previous step is privacy-secure and time-series consistent, achieving seamless integration of data from multiple parties in the time dimension.

[0107] S52, the time-aligned share data participates in the final merge operation, and the final merge operation generates a joint dataset.

[0108] Encrypted timestamp data shares are obtained from each participant. These shares are then preliminarily compared in an encrypted state to determine the timestamp range boundaries of each share, forming an initial timestamp boundary set. For this initial timestamp boundary set, a comparison protocol in the encrypted state is used to calculate the deviation between each participant's timestamp data and the unified timeline, generating an offset value set for each party. Specific offset data for each party is extracted from the offset value set, and the timestamp shares of each participant are adjusted based on these offset data to form time-aligned share data. The time-aligned share data is then integrated, and a final merging operation is performed in an encrypted environment to generate a joint dataset that conforms to time dimension consistency.

[0109] For example, in a scenario of secure merging of multi-party data, the process of obtaining encrypted timestamp data shares from each participating party can be understood as the starting point of a privacy protection mechanism that ensures that the original information is not exposed during data transmission and processing, thereby protecting the privacy rights of all parties.

[0110] Specifically, when performing an initial comparison of these data shares in an encrypted state, a secure comparison protocol can be used to identify the upper and lower limits of the timestamps. For example, assuming that the timestamps of one data share are from 8:00 AM to 5:00 PM and the other is from 9:00 AM to 6:00 PM, this comparison helps to form an initial set of timestamp boundaries, which is beneficial for subsequent alignment operations, avoids merging errors caused by time inconsistencies, and improves the accuracy of the joint dataset.

[0111] In one embodiment, after determining the timestamp range boundaries of each share of data, the purpose of this set is to provide a reference benchmark for a unified timeline. This improves the efficiency of time alignment because it reduces unnecessary computational overhead and ensures the reliability of the merging process.

[0112] For example, the step of calculating the offset value using a comparison protocol in an encrypted state for an initial set of timestamp boundaries further deepens the privacy computation. This protocol is typically based on homomorphic encryption principles, which allow numerical comparisons to be performed without decryption, thereby generating a set of offset values ​​for each party.

[0113] For example, if one party's timestamp boundary is 10 to 20 and the unified timeline benchmark is 0 to 15, the deviation value may be calculated as positive 5. The beneficial effect of this set of offsets is that it quantifies the time difference, making subsequent adjustments more accurate, avoiding data distortion, and enhancing the temporal consistency of joint research.

[0114] In one embodiment, this calculation process is achieved by comparing boundary values ​​one by one, first comparing the minimum value and then the maximum value, to ensure the accuracy of the deviation value. This approach provides better data compatibility because it supports the seamless integration of multi-source data.

[0115] For example, after extracting the specific offset data for each party from the offset value set, the process of adjusting the timestamp share of each participant is the core step of time alignment. This adjustment is similar to applying addition and subtraction operations within an encrypted domain to form the time-aligned share data. For instance, if the offset is negative 3, then all timestamps are subtracted by 3 to match a unified axis. The advantage of this method is that it achieves standardization of the time dimension without revealing the specific collection mode, thus improving the usability of the merged dataset.

[0116] The adjusted data remains encrypted, which enhances privacy by preventing any party from inferring the other's time patterns. For example, the entire process involves integrating the time-aligned share data and performing a final merge operation in an encrypted environment, generating a joint dataset consistent across the time dimension. This integration can be achieved through secret-sharing protocols, such as adding the adjusted shares or merging them into a complete record, which benefits the production of unbiased analytical results and improves research efficiency.

[0117] In one embodiment, the merging operation ensures that all data are aligned on a unified timeline. This can lead to technical benefits such as improved model training accuracy, as consistent time-series data reduces noise interference and supports the reliability of multi-agency collaboration.

[0118] This invention provides a privacy-protected data circulation system, which mainly includes: A sharing generation module is used for multiple data holders to secretly share their local datasets to generate share data, which is distributed among the multiple data holders. An encrypted comparison module is used by the multiple data holders to compare the field patterns, sample ranges, and feature dimensions of the share data in an encrypted state based on a secure multi-party computation protocol, and to obtain the comparison results. The part determination module is used by the multiple data holders to determine the specific part of the data to be merged based on the comparison results, and to select horizontal federation or vertical federation. The horizontal federation merges samples based on the range, and the vertical federation merges features based on the dimension. The merge operation module is used by the multiple data holders to perform a selected horizontal or vertical federation merge operation, and to process the share data using homomorphic encryption to obtain the merge result. The negotiation and unification module is used by the multiple data holders to negotiate and unify the time axis benchmark in an encrypted state through a secure multi-party computation protocol, and to calculate the offset of each party's dataset relative to the unified time axis benchmark using the timestamp of the share data. The time alignment module is used by the multiple data holders to perform time alignment of the share data according to the offset, so as to obtain the time-aligned share data and generate a joint dataset.

[0119] Furthermore, the encryption comparison module includes: An encryption processing unit is used by the multiple data holders to encrypt the share data using homomorphic encryption to generate encrypted share data; The similarity calculation unit is used by the multiple data holders to calculate the similarity of field patterns, the overlap of sample ranges, and the complementarity of feature dimensions in the encrypted share data through a secure multi-party computation protocol. The comparison result generation module is used to generate the comparison result based on the similarity, overlap and complementarity, wherein the comparison result indicates the alignable fields, the range of overlapping samples and the dimension of complementary features; The alignment field extraction unit is used by the multiple data holders to extract alignable fields from the comparison results for field mapping, overlapping sample ranges for sample alignment, and complementary feature dimensions for feature supplementation.

[0120] It should be noted that the modules provided in the embodiments of the present invention have the same implementation principle and technical effects as those in the aforementioned method embodiments. For the sake of brevity, the specific working process of the modules described above can be referred to the corresponding process in the aforementioned method embodiments, and will not be repeated here.

[0121] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A privacy-preserving data circulation method, characterized in that, include: Multiple data holders secretly share their local datasets to generate share data, which is distributed among the multiple data holders. The multiple data holders compare the field patterns, sample ranges, and feature dimensions of the share data in an encrypted state based on a secure multi-party computation protocol to obtain the comparison results; The multiple data holders determine the specific parts of the data to be merged based on the comparison results, and select horizontal federation or vertical federation. Horizontal federation merges data based on the sample range, while vertical federation merges data based on the feature dimension. The multiple data holders perform a selected horizontal or vertical federation merge operation, using homomorphic encryption to process the share data, and obtain the merge result; The multiple data holders negotiate a unified timeline benchmark in an encrypted state through a secure multi-party computation protocol, and use the timestamps of the share data to calculate the offset of each party's dataset relative to the unified timeline benchmark. The multiple data holders perform time-series alignment of the share data according to the offset to obtain time-series aligned share data and generate a joint dataset.

2. The method as described in claim 1, characterized in that, The multiple data holders compare the field patterns, sample ranges, and feature dimensions of the share data in an encrypted state based on a secure multi-party computation protocol to obtain comparison results, including: The multiple data holders use homomorphic encryption to encrypt the share data, generating encrypted share data; The multiple data holders calculate the similarity of field patterns, the overlap of sample ranges, and the complementarity of feature dimensions in the encrypted share data through a secure multi-party computation protocol. The alignment results are generated based on the similarity, overlap, and complementarity, and the alignment results indicate the alignable fields, the range of overlapping samples, and the dimension of complementary features. The multiple data holders extract alignable fields from the comparison results for field mapping, overlapping sample ranges for sample alignment, and complementary feature dimensions for feature supplementation.

3. The method as described in claim 1, characterized in that, The multiple data holders determine the specific parts of the data to be merged based on the comparison results, and select horizontal or vertical federation, including: The multiple data holders obtain the coverage and feature richness of each party's dataset; Calculate the combined utility value based on the coverage and the feature richness; If the overlapping sample range exceeds a preset range, then the horizontal federation is selected to merge the sample ranges. If the complementary feature dimensions exceed the preset dimensions, then the vertical federation is selected to merge the feature dimensions. The multiple data holders determine that the specific part of the data merging is a combination of the alignable fields, overlapping sample ranges, and complementary feature dimensions.

4. The method as described in claim 3, characterized in that, The multiple data holders calculate a combined utility value based on the coverage and the feature richness, including: The multiple data holders aggregate the coverage areas of each party through a secure multi-party computation protocol to obtain the total coverage area; The multiple data holders calculate the ratio of the total coverage to the coverage of each party's dataset as the coverage utility; The total feature richness is obtained by aggregating the feature richness of each of the multiple data holders. The multiple data holders calculate the ratio of the total feature richness to the feature richness of each party's dataset as the richness utility; The multiple data holders add the coverage utility to the enrichment utility to obtain the combined utility value.

5. The method as described in claim 1, characterized in that, The multiple data holders perform a selected horizontal or vertical federation merge operation, using homomorphic encryption to process the share data to obtain the merged result, including: If the horizontal federation is selected, the multiple data holders will jointly aggregate the share data within the overlapping sample range; If the vertical federation is selected, the multiple data holders perform feature splicing on the share data within the complementary feature dimensions; The multiple data holders apply homomorphic encryption to compute intermediate results during the joint aggregation or feature splicing process; The multiple data holders recover the merged result from the intermediate result, and the merged result represents the structure of the combined dataset.

6. The method as described in claim 1, characterized in that, The multiple data holders negotiate a unified timeline benchmark in an encrypted state through a secure multi-party computation protocol, and calculate the offset of each party's dataset relative to the unified timeline benchmark using the timestamps of the share data, including: The multiple data holders secretly share the timestamps of the share data to generate timestamp shares; The multiple data holders calculate the minimum value of each party's timestamp share using a secure multi-party computation protocol as the unified timeline benchmark; The multiple data holders use homomorphic encryption to calculate the difference between each party's timestamp share and the unified timeline benchmark as the offset; The multiple data holders verify the consistency of the offset and distribute it to each party.

7. The method as described in claim 6, characterized in that, The multiple data holders use homomorphic encryption to calculate the difference between their respective timestamp shares and the unified timeline benchmark as the offset, including: The multiple data holders homomorphically encrypt the timestamp shares of each party to obtain encrypted timestamp shares. The multiple data holders calculate the encrypted timestamp share minus the encrypted value of the unified timeline benchmark; The multiple data holders decrypt the calculation results to obtain the offsets provided by each party. The multiple data holders adjust the timestamp position of the share data according to the offset.

8. The method as described in claim 1, characterized in that, The multiple data holders perform time-series alignment of the share data according to the offset, obtain time-series aligned share data, and generate a joint dataset, including: The multiple data holders apply the offset to shift the timestamps of their respective share data; The multiple data holders verify the consistency of the shifted timestamps on the unified timeline reference. The multiple data holders input the time-aligned share data into the merging operation; The multiple data holders output the joint dataset from the merge operation.

9. A privacy-preserving data circulation system, characterized in that, include: A sharing generation module is used for multiple data holders to secretly share their local datasets to generate share data, which is distributed among the multiple data holders. An encrypted comparison module is used by the multiple data holders to compare the field patterns, sample ranges, and feature dimensions of the share data in an encrypted state based on a secure multi-party computation protocol, and to obtain the comparison results. The part determination module is used by the multiple data holders to determine the specific part of the data to be merged based on the comparison results, and to select horizontal federation or vertical federation. The horizontal federation merges samples based on the range, and the vertical federation merges features based on the dimension. The merge operation module is used by the multiple data holders to perform a selected horizontal or vertical federation merge operation, and to process the share data using homomorphic encryption to obtain the merge result. The negotiation and unification module is used by the multiple data holders to negotiate and unify the time axis benchmark in an encrypted state through a secure multi-party computation protocol, and to calculate the offset of each party's dataset relative to the unified time axis benchmark using the timestamp of the share data. The time alignment module is used by the multiple data holders to perform time alignment of the share data according to the offset, so as to obtain the time-aligned share data and generate a joint dataset.

10. The system as described in claim 9, characterized in that, The encryption comparison module includes: An encryption processing unit is used by the multiple data holders to encrypt the share data using homomorphic encryption to generate encrypted share data; The similarity calculation unit is used by the multiple data holders to calculate the similarity of field patterns, the overlap of sample ranges, and the complementarity of feature dimensions in the encrypted share data through a secure multi-party computation protocol. The comparison result generation module is used to generate the comparison result based on the similarity, overlap and complementarity, wherein the comparison result indicates the alignable fields, the range of overlapping samples and the dimension of complementary features; The alignment field extraction unit is used by the multiple data holders to extract alignable fields from the comparison results for field mapping, overlapping sample ranges for sample alignment, and complementary feature dimensions for feature supplementation.

Images