Cross-department enterprise data security collaborative modeling method based on federated learning

Abstract

The application discloses a cross-department enterprise data security collaborative modeling method based on federal learning, relates to the technical field of data modeling, and comprises the following steps: S1: each department node uniformly processes the department subject identifier, generates a round temporary identifier in combination with a current training round random factor, and then sends the round temporary identifier to a security alignment agent node; the security alignment agent node completes cross-department subject alignment of the current training round according to the round temporary identifier, maps the alignment result into a plurality of temporary alignment groups, and triggers security replacement processing for the temporary alignment group with a subject quantity lower than a security lower limit; S2: each department node generates a plurality of contribution slices according to a preset slicing rule for the local features of each subject in the department, and the application reduces the continuous association risk of the same subject in the multi-round cross-department training through the round temporary mapping and temporary alignment grouping mechanism.

Description

Technical Field

[0001] This invention relates to the field of data modeling technology, specifically to a cross-departmental collaborative modeling method for enterprise data security based on federated learning. Background Technology

[0002] Federated learning enables multiple departments within an enterprise to conduct collaborative modeling without directly exchanging raw data. For AI tasks such as customer risk control, employee management, supply chain assessment, and business analysis, the same entity often retains transaction information, behavioral information, performance information, service information, and tag information in multiple departments. Each department completes joint modeling through local training and centralized coordination, thus forming an application model of cross-departmental enterprise data security collaborative modeling.

[0003] In cross-departmental collaborative modeling, a unique scenario exists: the same entity appears in multiple departments simultaneously, with significant differences in entity coverage between departments. Some entities have significantly fewer records in certain departments, and the training process requires multiple rounds of collaborative iteration. In this scenario, if the same entity consistently corresponds to stable alignment relationships and intermediate representations across multiple training rounds, the coordinating side may gradually determine which departments jointly hold the entity by following the chain of alignment index—intermediate representation—multi-round update, thereby accumulating and inferring the entity's cross-departmental relationships and sensitive patterns. Therefore, those skilled in the art need a technical solution that can maintain cross-departmental collaborative modeling capabilities, sever stable associative relationships of the same entity across multiple training rounds, and proactively control the repeated exposure of sparse entities. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-departmental enterprise data security collaborative modeling method based on federated learning to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a cross-departmental enterprise data security collaborative modeling method based on federated learning, which is deployed across multiple departmental nodes, a coordinating node, and a security alignment proxy node, wherein one of the multiple departmental nodes is designated as the primary supervisory departmental node; the method includes the following steps: S1: Each department node unifies the identifier of its own subject and generates a temporary identifier for the current training round by combining the random factor of the current training round. The temporary identifier is then sent to the security alignment proxy node. The security alignment proxy node completes the cross-department subject alignment for the current training round according to the temporary identifier and maps the alignment result into multiple temporary alignment groups. For temporary alignment groups with a subject number lower than the safety lower limit, the security alignment proxy node triggers security filling processing. S2: Each department node generates multiple contribution slices according to the preset slicing rules for the local features of each subject in its department, and selects only one contribution slice to participate in training in the current training round. Each department node inputs the currently activated contribution slice into the local encoding network to obtain an intermediate representation of a unified dimension. Subsequently, each department node uses the central representation in the temporary alignment group to mask the intermediate representation of a single subject before uploading it to the coordination node. S3: The coordinating node receives masking representations from multiple department nodes for each temporary alignment group, and performs weighted fusion based on the effective contribution of each department node in the current round to generate a fusion representation for the current training round. The target probability distribution generated based on the fusion representation is only valid within the current training round and is only used to transmit the target probability distribution back to each department node. It is not retained as a stable identity index across rounds. The main supervisory department node participates in training constraints based on the task label and the target probability distribution. S4: Each department node performs local model-guided updates based on the target probability distribution and synchronously receives group density and coverage information returned by the coordination node. Each department node calculates the current subject's exposure based on the number of local subject records. For subjects whose exposure exceeds the first threshold, the system switches them to group summary collaboration mode; for subjects whose exposure exceeds the second threshold, the system switches them to local constraint update mode. S5: After the current training round ends, each department node retains the local model parameters updated by guidance, and updates the subject participation coefficient for the next training round based on the subject exposure. The coordination node and the security alignment proxy node destroy the round temporary identifier, temporary alignment group index, fusion representation, target probability distribution and its association index formed in the current training round. The next training round uses the new training round random factor to re-execute the temporary alignment, thereby blocking the cross-round stable association relationship of the same subject.

[0006] According to the above technical solution, step S1 specifically includes the following sub-steps: S1-1: Each department node corresponds to the main body In training rounds Generate temporary identifier for round ,in as the main body A unified main identity, For training rounds The round random factor, For the shared mapping key pre-stored in each department node, It is a one-way hash function. as the main body In training rounds The function of this formula is to enable the same subject to generate different temporary identifiers in different training rounds, thereby cutting off cross-round stable indexes from the source. S1-2: Security alignment proxy nodes construct temporary alignment groups based on round temporary identifiers. ,in as the main body In training rounds The corresponding temporary alignment group number, For training rounds The total number of temporary alignment groups; the function of this formula is to upgrade the alignment result that was originally at the single-subject granularity to the group granularity result, so that the coordination node subsequently receives the group-level collaborative relationship instead of the stable single-subject relationship; S1-3: Security Alignment Proxy Node Statistics for Each Temporary Alignment Group In training rounds initial number of subjects And calculate the number of padding records. ,in To preset a safety lower limit, this formula is used to safely fill in low-density temporary aligned groups so that the statistical characteristics within the group reach a predetermined safety density; the fill-in records are extracted from the non-intersection record pool of this department according to the same field structure, and only participate in the calculation of the central representation within the group, and do not participate in the calculation of the supervision loss. S1-4: The security alignment agent node only returns the temporary alignment group sequence number and group density statistics to the coordination node, and does not return the unified subject identifier and round temporary identifier; each department node only saves the local mapping table of its own department's unified subject identifier, round temporary identifier, and temporary alignment group sequence number for subsequent subject-level exposure write-back.

[0007] According to the above technical solution, step S2 specifically includes the following sub-steps: S2-1: Department Node The main body The local feature vector is divided into Each contribution slice is determined, and the current department node is identified. In training rounds Activation slice number ,in Representing the subject At the department node The first Each contribution slice For departmental nodes The total number of contribution slices is calculated by the formula, which aims to ensure that the complete local features of each subject are not exposed all at once in the same training round, but rather participate in collaborative training in rounds. S2-2: Department Node Perform local encoding on the currently active contribution slice ,in For departmental nodes The local encoder, as the main body At the department node and training rounds The unified dimensional intermediate representation; the function of this formula is to transform the heterogeneous features of different departments into a unified dimensional representation that can be fused. S2-3: Department Node In temporary alignment group Internal computation group center representation ,in For departmental nodes In training rounds The middle belongs to the temporary alignment group The main body set, For the main body set The number of elements, As a group center representation, the function of this formula is to construct a group-level statistical representation, providing a reference benchmark for masking single-subject representations; S2-4: Department Node Generate masking representations based on group center representations. ,in as the main body At the department node and training rounds The masking representation, as the main body In training rounds The temporary alignment group number to which it falls. Temporarily align group numbers The corresponding temporary alignment group is in the department node. and training rounds The grouping center representation is calculated according to method S2-3. For training rounds The subject representation retention coefficient, which ranges from 0 to 1, is used to mix the single subject representation with group-level statistical representation, so that the single-round upload result does not have a complete subject identifiable representation.

[0008] According to the above technical solution, step S3 specifically includes the following sub-steps: S3-1: Coordinating node statistics on the performance of each department node in the training rounds. Number of effective masking representations within And calculate the departmental weighting coefficient. ,in The total number of department nodes participating in the current training round. For departmental nodes In training rounds The departmental weighting coefficient; the function of this formula is to determine the fusion weight based on the effective contribution of the departmental node in the current round; S3-2: The coordinating node performs weighted fusion of the masking representations uploaded by each department node within the same temporary alignment group to generate a fused representation. ,in as the main body In training rounds The fusion representation; the function of this formula is to form a joint knowledge representation that can be used for collaborative reasoning without restoring the original data and without retaining the stable subject index; S3-3: The coordination node generates the target probability distribution based on the fused representation. ,in as the main body In training rounds The target probability distribution, For training rounds The temperature smoothing coefficient; the function of this formula is to transform the fused representation into the target probability distribution, reduce the sharp recognition features in the single round output, and facilitate the execution of guided updates by each department node; S3-4: The coordinating node sends the target probability distribution to each department node, which can only be used within the current training round; neither the security alignment agent node nor the coordinating node may associate the fused representation with the temporary alignment state of the next training round.

[0009] According to the above technical solution, step S4 specifically includes the following sub-steps: S4-1: Department Node The local model is updated based on the target probability distribution, and the department nodes are calculated according to the following formula. In training rounds Joint guidance loss ,in The entity held by the main oversight department node Task tags, For departmental nodes The local model for the subject In training rounds The predicted distribution of the output, For training rounds The weight of the supervision loss, The cross-entropy function, The Kullback-Leibler divergence function is used to simultaneously preserve the constraints of the main supervision task and the target probability distribution constraints generated by the fused representation, so as to reduce the model performance loss caused by safety masking. S4-2: Department nodes calculate the subject based on grouping density, subject sparsity, and cross-department coverage. In training rounds Exposure increment: ,in as the main body In training rounds Belonging to temporary alignment group The initial number of subjects in this training round, , , To expose the incremental weighting coefficients, as the main body The number of local records within the preset statistical time window. For training rounds Internal Participation Entities The number of departmental nodes in collaborative training. This represents the total number of department nodes participating in the current training round. The purpose of this formula is to transform the judgment rule that the more sparse the grouping, the more sparse the subject, and the more departments covered, the higher the exposure risk, into an executable quantitative control quantity. S4-3: Department Node Calculation Entity In training rounds Cumulative exposure: ,in The exposure memory coefficient is between 0 and 1. The purpose of this formula is to accumulate the exposure risk of the current round with the exposure risk of the historical rounds, forming a basis for cross-round risk control. S4-4: When the cumulative exposure meets the requirements At that time, the department node will be the main body Switch to group summary collaboration mode, in which only group-level summary representations are uploaded and not subject-level masking representations; when the cumulative exposure meets the requirements... At that time, the department node will be the main body Switch to local constraint update mode; in this mode, the main body... Stop participating in cross-departmental guidance and integration; complete the update for this round solely based on the target probability distribution already received in the current round. Subsequent rounds will continue training only based on the updated local model parameters. The first exposure threshold, This is the second exposure threshold.

[0010] According to the above technical solution, step S5 specifically includes the following sub-steps: S5-1: Department Node In training rounds The local model parameters are updated according to the following formula. , For training rounds learning rate, This formula is used to jointly guide the gradient of the loss with respect to the local model parameters; its purpose is to solidify the joint knowledge obtained in the current training round into the local models of each department node. S5-2: Departmental nodes target the main body Calculate the subject participation coefficient for the next training round. ,in This is the minimum participation coefficient; the purpose of this formula is to proactively reduce the frequency of a subject's continued participation in cross-departmental collaboration as the subject's cumulative exposure increases, thereby suppressing repeated exposure. S5-3: Securely aligned proxy nodes destroy the set of temporary identifiers for the current training round after the training round ends. And the corresponding cross-departmental alignment relationships; the coordinating node destroys the temporary alignment group index set for this round. fusion representation set and the set of target probability distributions Each department node retains only local model parameters and the cumulative exposure of the subject. S5-4: In the next training round, the coordinating node generates a new round random factor, and each department node regenerates the round temporary identifier, re-executes the temporary alignment, and re-selects the currently activated contribution slice, so that cross-round collaboration continues only at the model parameter level, and not at the subject alignment relationship level.

[0011] According to the above technical solution, the system used in this method includes: The alignment mapping module is used to convert the same subject held by each department node into a temporary alignment identifier that is only valid in the current round within each training round, and further map the temporary alignment identifier to a group-level index so that cross-department joint processing is based on the granularity of temporary groups rather than the granularity of stable subjects. The training guidance module is used to divide the main features of each department node into multiple contribution slices according to preset rules. In the current round, only one contribution slice is activated for local encoding and intra-group masking. The coordination node generates a fusion representation and target probability distribution that are only valid for the current round based on the masking representation uploaded by each department node, and then transmits the target probability distribution back to each department node. The control recycling module is used to calculate the subject exposure based on the density of the alignment group to which the subject belongs, the sparsity of the subject's local records, and the cross-departmental coverage of the subject in the current round. When the subject exposure reaches a preset threshold, the collaboration mode is switched, and after the round ends, only the local model parameters are retained while the fusion representation, target probability distribution and its associated index formed in the current training round are destroyed.

[0012] According to the above technical solution, the alignment mapping module includes: an identifier standardization and round temporary identifier generation module, which is used to convert the subject identifiers within each department node into a unified identifier format, and generate round temporary identifiers in combination with the random factor of the current training round, so that the same subject corresponds to different temporary identifiers in different training rounds; and an alignment group construction and security fill module, which is used by the security alignment proxy node to complete the cross-department subject alignment according to the round temporary identifier, and map the alignment result to a temporary alignment group. When the number of subjects in the temporary alignment group is lower than the safety lower limit, a fill record is generated. The fill record is only used for statistical masking within the group and not for the calculation of supervised loss. The training guidance module includes: a feature slicing and unified encoding module, used to divide the local features of each subject at each department node into multiple contribution slices, and to encode the activated contribution slices in a unified dimension in the current training round to form an intermediate representation that can be fused across departments; and a fusion representation generation and target probability distribution back transmission module, used by the coordination node to perform weighted fusion of the intermediate representations uploaded by each department node within the same temporary alignment group, to generate a fusion representation that is only valid in the current round, and to generate a target probability distribution based on the fusion representation, and back transmission the target probability distribution to each department node to guide each department node to update its local model; The control and recycling module includes: an exposure calculation and threshold switching module, used to calculate the subject exposure by combining the temporary alignment group density, the number of local records of the subject, and the number of cross-department coverage of the subject, and to switch the subject to the group summary collaboration mode when the subject exposure reaches the first threshold, and to switch the subject to the local constraint update mode when the subject exposure reaches the second threshold; and a temporary state destruction and participation coefficient update module, used to destroy the temporary identifier of the round, the temporary alignment group index, the fusion representation, the target probability distribution and its associated index after each training round, and to update the subject participation coefficient of the next training round according to the subject exposure.

[0013] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention transforms the long-term stable mapping of the same subject in cross-departmental collaborative training into a temporary mapping that is only valid within the current training round, and transforms the original joint processing oriented towards a single subject into joint processing oriented towards temporary alignment groups, thereby reducing the possibility of coordination nodes continuously identifying and tracking the cross-departmental relationships of the same subject during multiple training rounds; at the same time, this invention divides the local features of each subject into contribution slices that are activated sequentially according to the training round, and uses the group center representation of the temporary alignment group to which the subject belongs to perform hybrid masking on the single-round upload results, thereby reducing the direct identifiability of the intermediate representation of a single subject.

[0014] Based on this, the present invention generates a target probability distribution by a coordinating node according to the fusion representation, and then transmits the target probability distribution back to each department node to guide each department node to update its local model. This reduces the risk of sensitive subject exposure while maintaining cross-departmental collaborative modeling capabilities. Furthermore, the present invention implements hierarchical control for sparse subjects, low-density grouped subjects, and cross-departmental covered subjects through exposure calculation and threshold switching mechanisms. This balances the effects of cross-departmental collaborative training, subject privacy protection, and model update stability, thus completely solving the aforementioned technical problems. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall modular structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a technical solution: a cross-departmental enterprise data security collaborative modeling method based on federated learning. This method is deployed across multiple departmental nodes, a coordinating node, and a security alignment proxy node, with one of the departmental nodes designated as the primary supervisory departmental node. The method includes the following steps: S1: Each department node unifies the identifier of its own subject and generates a temporary identifier for the current training round by combining the random factor of the current training round. The temporary identifier is then sent to the security alignment proxy node. The security alignment proxy node completes the cross-department subject alignment for the current training round according to the temporary identifier and maps the alignment result into multiple temporary alignment groups. For temporary alignment groups with a subject number lower than the safety lower limit, the security alignment proxy node triggers security filling processing. S2: Each department node generates multiple contribution slices according to the preset slicing rules for the local features of each subject in its department, and selects only one contribution slice to participate in training in the current training round. Each department node inputs the currently activated contribution slice into the local encoding network to obtain an intermediate representation of a unified dimension. Subsequently, each department node uses the central representation in the temporary alignment group to mask the intermediate representation of a single subject before uploading it to the coordination node. S3: The coordinating node receives masking representations from multiple department nodes for each temporary alignment group, and performs weighted fusion based on the effective contribution of each department node in the current round to generate a fused representation for the current training round. The target probability distribution generated based on the fused representation is only valid within the current training round and is only used to transmit the target probability distribution back to each department node. It is not retained as a stable identity index across rounds. The main supervisory department node participates in training constraints based on the task label and the target probability distribution. The target probability distribution is not the original task label, but a category probability distribution formed by the coordinating nodes based on the fused representation, used to provide the collaborative update direction for the current training round to each department node.

[0018] S4: Each department node performs local model-guided updates based on the target probability distribution and synchronously receives group density and coverage information returned by the coordination node. Each department node calculates the current subject's exposure based on the number of local subject records. For subjects with exposure exceeding the first threshold, the system switches them to group summary collaboration mode; for subjects with exposure exceeding the second threshold, the system switches them to local constraint update mode. Coverage information is used to characterize how many department nodes jointly include the current subject in collaborative processing in the current training round.

[0019] S5: After the current training round ends, each department node retains the local model parameters updated by guidance, and updates the subject participation coefficient for the next training round based on the subject exposure. The coordination node and the security alignment proxy node destroy the round temporary identifier, temporary alignment group index, fusion representation, target probability distribution and its association index formed in the current training round. The next training round uses the new training round random factor to re-execute the temporary alignment, thereby blocking the cross-round stable association relationship of the same subject. S1 specifically includes the following sub-steps: S1-1: Each department node corresponds to the main body In training rounds Generate temporary identifier for round ,in as the main body A unified main identity, For training rounds The round random factor, For the shared mapping key pre-stored in each department node, It is a one-way hash function. as the main body In training rounds The function of this formula is to enable the same subject to generate different temporary identifiers in different training rounds, thereby cutting off cross-round stable indexes from the source. S1-2: Security alignment proxy nodes construct temporary alignment groups based on round temporary identifiers. ,in as the main body In training rounds The corresponding temporary alignment group number, For training rounds The total number of temporary alignment groups; the function of this formula is to upgrade the alignment result that was originally at the single-subject granularity to the group granularity result, so that the coordination node subsequently receives the group-level collaborative relationship instead of the stable single-subject relationship; S1-3: Security Alignment Proxy Node Statistics for Each Temporary Alignment Group In training rounds initial number of subjects And calculate the number of padding records. ,in To preset a safety lower limit, this formula is used to fill in the low-density temporary aligned groups with safety, so that the statistical characteristics within the group reach the predetermined safety density. The filler records are extracted from the non-intersecting record pool of this department according to the same field structure, and only participate in the calculation of the central characterization within the group, but do not participate in the calculation of the supervision loss. S1-4: The security alignment proxy node only returns the temporary alignment group sequence number and group density statistics to the coordination node, and does not return the unified subject identifier and round temporary identifier; each department node only saves the local mapping table of its own department's unified subject identifier, round temporary identifier, and temporary alignment group sequence number for subsequent subject-level exposure write-back; The key to this step is not simply changing the subject identifier, but transforming the subject correspondence, which might remain unchanged across multiple training rounds, into a temporary correspondence that only exists within the current training round. Furthermore, it elevates the objects of subsequent joint processing from individual subjects to temporary alignment groups. If those skilled in the art were to use conventional methods, they would typically reuse stable subject alignment results across multiple training rounds. While this approach is simple to implement and facilitates cumulative training across rounds, the coordinating node is more likely to gradually identify the presence of the same subject across multiple departments along a fixed correspondence.

[0020] In contrast, this scheme uses round-based temporary mapping and temporary alignment grouping to invalidate the original correspondences after each training round. New temporary correspondences are then established in subsequent rounds, fundamentally weakening the persistent correlation of the same subject across multiple training rounds. This step in this scheme provides the prerequisite for subsequent group-level fusion, group-level masking, and exposure control, ensuring that the entire scheme avoids establishing long-term stable single-subject joint processing paths from the outset. Its originality lies in not adding protective measures to existing stable alignment relationships, but directly changing the basic organizational method of cross-departmental joint training, transforming stable subject correspondences into round-based temporary group correspondences, thus reducing the risk of persistent identification from the source.

[0021] S2 specifically includes the following sub-steps: S2-1: Department Node The main body The local feature vector is divided into Each contribution slice is determined, and the current department node is identified. In training rounds Activation slice number ,in Representing the subject At the department node The first Each contribution slice For departmental nodes The total number of contribution slices is calculated by the formula, which aims to ensure that the complete local features of each subject are not exposed all at once in the same training round, but rather participate in collaborative training in rounds. S2-2: Department Node Perform local encoding on the currently active contribution slice ,in For departmental nodes The local encoder, as the main body At the department node and training rounds The unified dimensional intermediate representation; the function of this formula is to transform the heterogeneous features of different departments into a unified dimensional representation that can be fused. S2-3: Department Node In temporary alignment group Internal computation group center representation ,in For departmental nodes In training rounds The middle belongs to the temporary alignment group The main body set, For the main body set The number of elements, As a group center representation, the function of this formula is to construct a group-level statistical representation, providing a reference benchmark for masking single-subject representations; S2-4: Department Node Generate masking representations based on group center representations. ,in as the main body At the department node and training rounds The masking representation, as the main body In training rounds The temporary alignment group number to which it falls. Temporarily align group numbers The corresponding temporary alignment group is in the department node. and training rounds The group centrality representation is calculated according to method S2-3. For training rounds The main subject representation retention coefficient, with a value range between 0 and 1, is used to mix the single subject representation with group-level statistical representation, so that the single-round upload result does not have a complete identifiable subject representation. The key to this step is to avoid allowing the complete local features of a single subject to participate in cross-departmental joint processing all at once in the same training round. Instead, the features are first split into multiple contribution slices, then activated step by step according to the round. At the same time, the group center representation of the temporary alignment group to which the subject belongs is used to perform a blending mask on the single-round upload results. If the conventional approach is adopted, those skilled in the art would usually directly encode the complete local features currently available for the subject and participate in joint modeling. While this is beneficial for obtaining more comprehensive information at once, it is also easy for the single-round upload results to retain strong traces of single-subject features. Once combined with the cross-departmental fusion process, it is easier to form a recognizable subject representation.

[0022] This scheme combines stepwise exposure with group dilution: the former avoids carrying too much subject-specific information in a single round, while the latter prevents the uploaded result from being too close to the original representation of a single subject. Without directly weakening the collaborative training structure, it reduces the recognizability of intermediate results in a single round for a single subject, while preserving sufficient statistical validity for the generation of subsequent fused representations and target probability distributions. Instead of simply adding noise or masking the uploaded representation, it combines round-robin slicing with group statistical masking, exposing only local information in each round, and this local information is further diluted by the group center, thus simultaneously balancing modeling usability and recognition suppression effectiveness.

[0023] S3 specifically includes the following sub-steps: S3-1: Coordinating node statistics on the performance of each department node in the training rounds. Number of effective masking representations within And calculate the departmental weighting coefficient. ,in The total number of department nodes participating in the current training round. For departmental nodes In training rounds The departmental weighting coefficient; the function of this formula is to determine the fusion weight based on the effective contribution of the departmental node in the current round; S3-2: The coordinating node performs weighted fusion of the masking representations uploaded by each department node within the same temporary alignment group to generate a fused representation. ,in as the main body In training rounds The fusion representation; the function of this formula is to form a joint knowledge representation that can be used for collaborative reasoning without restoring the original data and without retaining the stable subject index; S3-3: The coordination node generates the target probability distribution based on the fused representation. ,in as the main body In training rounds The target probability distribution, For training rounds The temperature smoothing coefficient; the function of this formula is to transform the fused representation into the target probability distribution, reduce the sharp recognition features in the single round output, and facilitate the execution of guided updates by each department node; S3-4: The coordinating node sends the target probability distribution to each department node, which can only be used within the current training round; neither the security alignment agent node nor the coordinating node may associate the fused representation with the temporary alignment state of the next training round. The purpose of this step is to reorganize the effective cross-departmental information retained after the aforementioned temporary grouping and masking processes into guiding information that can be used by the local models of each department, thereby preventing the models of each department from being completely isolated from each other due to the introduction of the protection mechanism. If only the aforementioned temporary mapping, slicing, and masking processes are performed without the subsequent guiding update step, although the nodes of each department will achieve a better main protection effect, the cross-departmental collaborative modeling capability will be significantly reduced, and the overall solution will be unable to balance security and modeling effectiveness.

[0024] This scheme further transforms the fused representation into a target probability distribution and feeds it back to each department node. This allows each department node to absorb collaborative information from other departments without restoring the original cross-departmental correspondences. The unique value of this step lies in the fact that it does not restore the traditional stable joint training path, but rather retains a collaborative information feedback method sufficient to support model updates, even after the subject's identifiability has been weakened. This allows the entire scheme to not only prevent exposure but also continue training.

[0025] S4 specifically includes the following sub-steps: S4-1: Department Node The local model is updated based on the target probability distribution, and the department nodes are calculated according to the following formula. In training rounds Joint guidance loss ,in The entity held by the main oversight department node Task tags, For departmental nodes The local model for the subject In training rounds The predicted distribution of the output, For training rounds The weight of the supervision loss, The cross-entropy function, The Kullback-Leibler divergence function is used to simultaneously preserve the constraints of the main supervision task and the target probability distribution constraints generated by the fused representation, so as to reduce the model performance loss caused by safety masking. S4-2: Department nodes calculate the subject based on grouping density, subject sparsity, and cross-department coverage. In training rounds Exposure increment: ,in as the main body In training rounds Belonging to temporary alignment group The initial number of subjects in this training round, , , To expose the incremental weighting coefficients, as the main body The number of local records within the preset statistical time window. For training rounds Internal Participation Entities The number of departmental nodes in collaborative training. This represents the total number of department nodes participating in the current training round. The purpose of this formula is to transform the judgment rule that the more sparse the grouping, the more sparse the subject, and the more departments covered, the higher the exposure risk, into an executable quantitative control quantity. S4-3: Department Node Calculation Entity In training rounds Cumulative exposure: ,in The exposure memory coefficient is between 0 and 1. The purpose of this formula is to accumulate the exposure risk of the current round with the exposure risk of the historical rounds, forming a basis for cross-round risk control. S4-4: When the cumulative exposure meets the requirements At that time, the department node will be the main body Switch to group summary collaboration mode, in which only group-level summary representations are uploaded and not subject-level masking representations; when the cumulative exposure meets the requirements... At that time, the department node will be the main body Switch to local constraint update mode; in this mode, the main body... Stop participating in cross-departmental guidance and integration; complete the update for this round solely based on the target probability distribution already received in the current round. Subsequent rounds will continue training only based on the updated local model parameters. The first exposure threshold, This is the second exposure threshold; that is, in this mode, the system only retains statistical information at the group level for collaborative processing, and no longer retains uploaded results at the single-subject level. In this mode, subjects no longer participate in the subsequent cross-departmental fusion process, but only continue to complete subsequent training within their own departmental nodes.

[0026] The key to this step is to move away from assuming all subjects will consistently use the same cross-departmental collaboration method throughout the training process. Instead, it differentiates the exposure risk of subjects based on the density of their group, the sparsity of their own records, and their cross-departmental coverage, and gradually reduces their participation intensity as the risk increases. Conventionally, those skilled in the art would typically use the same protection strategy for all subjects, either maintaining full collaboration or uniformly reducing the granularity of uploaded information. Both approaches have significant limitations: the former can easily expose a small number of high-risk subjects continuously across multiple training rounds, while the latter can cause unnecessary losses to the overall collaborative modeling capability.

[0027] This scheme categorizes subjects into different risk levels through exposure calculation and threshold switching. High-risk subjects are gradually switched to either a group summary collaboration mode or a local constraint update mode. This concentrates protection measures on subjects that are truly vulnerable to exposure, while most normal subjects maintain good collaborative training capabilities. The previously established temporary grouping and masking mechanisms are further transformed into a dynamically adjustable risk control closed loop, ensuring the scheme is not a one-time protection but continuously adjusts during training. Subject risk assessment is embedded in the training process itself, dynamically changing the collaboration mode based on the real-time exposure status of different subjects during training, thus balancing security and usability.

[0028] S5 specifically includes the following sub-steps: S5-1: Department Node In training rounds The local model parameters are updated according to the following formula. , For training rounds learning rate, This formula is used to jointly guide the gradient of the loss with respect to the local model parameters; its purpose is to solidify the joint knowledge obtained in the current training round into the local models of each department node. S5-2: Departmental nodes target the main body Calculate the subject participation coefficient for the next training round. ,in This is the minimum participation coefficient; the purpose of this formula is to proactively reduce the frequency of a subject's continued participation in cross-departmental collaboration as the subject's cumulative exposure increases, thereby suppressing repeated exposure. S5-3: Securely aligned proxy nodes destroy the set of temporary identifiers for the current training round after the training round ends. And the corresponding cross-departmental alignment relationships; the coordinating node destroys the temporary alignment group index set for this round. fusion representation set and the set of target probability distributions Each department node retains only local model parameters and the cumulative exposure of the subject. S5-4: In the next training round, the coordinating node generates a new round random factor, and each department node regenerates the round temporary identifier, re-executes the temporary alignment, and re-selects the currently activated contribution slice, so that cross-round collaboration continues only at the model parameter level, and not at the subject alignment relationship level. The system used in this method includes: The alignment mapping module is used to convert the same subject held by each department node into a temporary alignment identifier that is only valid in the current round within each training round, and further map the temporary alignment identifier to a group-level index so that cross-department joint processing is based on the granularity of temporary groups rather than the granularity of stable subjects. The training guidance module is used to divide the main features of each department node into multiple contribution slices according to preset rules. In the current round, only one contribution slice is activated for local encoding and intra-group masking. The coordination node generates a fusion representation and target probability distribution that are only valid for the current round based on the masking representation uploaded by each department node, and then transmits the target probability distribution back to each department node. The control recycling module is used to calculate the subject exposure based on the density of the alignment group to which the subject is located, the sparsity of the subject's local records, and the cross-departmental coverage of the subject in the current round. When the subject exposure reaches a preset threshold, the collaboration mode is switched, and after the round ends, only the local model parameters are retained while the fusion representation, target probability distribution and its associated index formed in the current training round are destroyed. The alignment mapping module includes: an identifier standardization and round temporary identifier generation module, which converts the subject identifiers within each department node into a unified identifier format and generates round temporary identifiers by combining the random factor of the current training round, so that the same subject corresponds to different temporary identifiers in different training rounds; and an alignment group construction and safety fill module, which is used by the safety alignment proxy node to complete cross-department subject alignment according to the round temporary identifiers and map the alignment results to temporary alignment groups. When the number of subjects in the temporary alignment group is lower than the safety lower limit, a fill record is generated. The fill record is only used for statistical masking within the group and not for supervised loss calculation. The training guidance module includes: a feature slicing and unified encoding module, which divides the local features of each subject at each department node into multiple contribution slices, and performs unified dimension encoding on the activated contribution slices in the current training round to form an intermediate representation that can be fused across departments; and a fusion representation generation and target probability distribution backhaul module, which is used by the coordination node to perform weighted fusion of the intermediate representations uploaded by each department node within the same temporary alignment group, generate a fusion representation that is only valid in the current round, generate a target probability distribution based on the fusion representation, and backhaul the target probability distribution to each department node to guide each department node to update its local model; The control and recycling module includes: an exposure calculation and threshold switching module, which calculates the subject exposure by combining the temporary alignment group density, the number of local records of the subject, and the number of cross-department coverage of the subject, and switches the subject to the group summary collaboration mode when the subject exposure reaches the first threshold, and switches the subject to the local constraint update mode when the subject exposure reaches the second threshold; and a temporary state destruction and participation coefficient update module, which destroys the temporary identifier of the round, the temporary alignment group index, the fusion representation, the target probability distribution and its associated index after each training round, and updates the subject participation coefficient for the next training round according to the subject exposure.

[0029] Overall, the technical concept of this invention is not a single-point repair of a local link in the existing cross-departmental collaborative training process, but rather a reconstruction of the organization of joint training, addressing the specific problem of the same subject being easily and continuously associated in multiple rounds of cross-departmental training. First, through round-based temporary mapping and temporary alignment grouping, cross-departmental collaboration is no longer based on long-term stable subject correspondences. Second, through contribution slice rotation and group center representation masking, the uploaded results of a single round no longer directly correspond to a complete single-subject representation. Third, by fusing representations to generate a target probability distribution and guiding the local model updates of each department, the cross-departmental collaborative training capability is preserved while weakening subject identifiability. Finally, through exposure calculation and threshold switching mechanisms, the protection strength can be dynamically adjusted according to the subject's risk status. In other words, this invention does not simply use common unified encryption, unified noise addition, or unified downsampling methods to process all subjects, but uses temporary association, local participation, group dilution, and dynamic degradation as a coordinated overall mechanism, thus giving the invention strong feasibility, specificity, and stability.

[0030] This invention does not apply a fixed-strength collaborative protection method to all subjects. Instead, it implements graded control based on the actual exposure risk of the subjects during the training process. Therefore, it can protect high-risk subjects while reducing unnecessary weakening of the collaborative training ability of low-risk subjects, thereby improving the practicality of the overall solution.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0032] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cross-departmental enterprise data security collaborative modeling method based on federated learning, characterized by: The method is deployed across multiple department nodes, a coordination node, and a security alignment proxy node, with one of the department nodes designated as the primary oversight department node; the method includes the following steps: S1: Each department node unifies the identifier of its own subject and generates a temporary identifier for the current training round by combining the random factor of the current training round. The temporary identifier is then sent to the security alignment proxy node. The security alignment proxy node completes the cross-department subject alignment for the current training round according to the temporary identifier and maps the alignment result into multiple temporary alignment groups. For temporary alignment groups with a subject number lower than the safety lower limit, the security alignment proxy node triggers security filling processing. S2: Each department node generates multiple contribution slices according to the preset slicing rules for the local features of each subject in its department, and selects only one contribution slice to participate in training in the current training round. Each department node inputs the currently activated contribution slice into the local encoding network to obtain an intermediate representation of a unified dimension. Subsequently, each department node uses the central representation in the temporary alignment group to mask the intermediate representation of a single subject before uploading it to the coordination node. S3: The coordinating node receives masking representations from multiple department nodes for each temporary alignment group, and performs weighted fusion based on the effective contribution of each department node in the current round to generate a fusion representation for the current training round. The target probability distribution generated based on the fusion representation is only valid within the current training round and is only used to transmit the target probability distribution back to each department node. It is not retained as a stable identity index across rounds. The main supervisory department node participates in training constraints based on the task label and the target probability distribution. S4: Each department node performs local model-guided updates based on the target probability distribution and synchronously receives group density and coverage information returned by the coordination node. Each department node calculates the current subject's exposure based on the number of local subject records. For subjects whose exposure exceeds the first threshold, the system switches them to group summary collaboration mode; for subjects whose exposure exceeds the second threshold, the system switches them to local constraint update mode. S5: After the current training round ends, each department node retains the local model parameters updated by guidance, and updates the subject participation coefficient for the next training round based on the subject exposure. The coordination node and the security alignment proxy node destroy the round temporary identifier, temporary alignment group index, fusion representation, target probability distribution and its association index formed in the current training round. The next training round uses the new training round random factor to re-execute the temporary alignment, thereby blocking the cross-round stable association relationship of the same subject.

2. The cross-departmental enterprise data security collaborative modeling method based on federated learning according to claim 1, characterized in that: S1 specifically includes the following sub-steps: S1-1: Each department node corresponds to the main body In training rounds Generate temporary identifier for round ,in as the main body A unified main identity, For training rounds The round random factor, For the shared mapping key pre-stored in each department node, It is a one-way hash function. as the main body In training rounds Temporary round identifier; S1-2: Security alignment proxy nodes construct temporary alignment groups based on round temporary identifiers. ,in as the main body In training rounds The corresponding temporary alignment group number, For training rounds The total number of temporary alignment groups; S1-3: Security Alignment Proxy Node Statistics for Each Temporary Alignment Group In training rounds initial number of subjects And calculate the number of padding records. ,in To preset a safety lower limit, this formula is used to safely fill in low-density temporary aligned groups so that the statistical characteristics within the group reach a predetermined safety density; the fill-in records are extracted from the non-intersection record pool of this department according to the same field structure, and only participate in the calculation of the central representation within the group, and do not participate in the calculation of the supervision loss. S1-4: The security alignment proxy node only returns the temporary alignment group sequence number and group density statistics to the coordinating node, and does not return the unified subject identifier and round temporary identifier; Each department node only stores a local mapping table of its own unified subject identifier, round temporary identifier, and temporary alignment group number, which is used for subsequent subject-level exposure write-back.

3. The cross-departmental enterprise data security collaborative modeling method based on federated learning according to claim 2, characterized in that: S2 specifically includes the following sub-steps: S2-1: Department Node The main body The local feature vector is divided into Each contribution slice is determined, and the current department node is identified. In training rounds Activation slice number ,in Representing the subject At the department node The first Each contribution slice For departmental nodes The total number of contributing slices; S2-2: Department Node Perform local encoding on the currently active contribution slice ,in For departmental nodes The local encoder, as the main body At the department node and training rounds A unified dimensional intermediate representation; S2-3: Department Node In temporary alignment group Internal computation group center representation ,in For departmental nodes In training rounds The middle belongs to the temporary alignment group The main body set, For the main body set The number of elements, As a group center representation; S2-4: Department Node Generate masking representations based on group center representations. ,in as the main body At the department node and training rounds The masking representation, as the main body In training rounds The temporary alignment group number to which it falls. Temporarily align group numbers The corresponding temporary alignment group is in the department node. and training rounds The grouping center representation is calculated according to method S2-3. For training rounds The main characterization retention coefficient ranges from 0 to 1.

4. The cross-departmental enterprise data security collaborative modeling method based on federated learning according to claim 3, characterized in that: S3 specifically includes the following sub-steps: S3-1: Coordinating node statistics on the performance of each department node in the training rounds. Number of effective masking representations within And calculate the departmental weighting coefficient. ,in The total number of department nodes participating in the current training round. For departmental nodes In training rounds Departmental weighting coefficients; S3-2: The coordinating node performs weighted fusion of the masking representations uploaded by each department node within the same temporary alignment group to generate a fused representation. ,in as the main body In training rounds Fusion representation; S3-3: The coordination node generates the target probability distribution based on the fused representation. ,in as the main body In training rounds The target probability distribution, For training rounds Temperature smoothness coefficient; S3-4: The coordinating node sends the target probability distribution to each department node, which can only be used within the current training round; neither the security alignment agent node nor the coordinating node may associate the fused representation with the temporary alignment state of the next training round.

5. The cross-departmental enterprise data security collaborative modeling method based on federated learning according to claim 4, characterized in that: S4 specifically includes the following sub-steps: S4-1: Department Node The local model is updated based on the target probability distribution, and the department nodes are calculated according to the following formula. In training rounds Jointly guided losses ,in The entity held by the main oversight department node Task tags, For departmental nodes The local model for the subject In training rounds The predicted distribution of the output, For training rounds The weight of the supervision loss, The cross-entropy function, The Kullback-Leibler divergence function; S4-2: Department nodes calculate the subject based on grouping density, subject sparsity, and cross-department coverage. In training rounds Exposure increment: ,in as the main body In training rounds Belonging to temporary alignment group Regarding the initial number of subjects in this training round, , , To expose the incremental weighting coefficients, as the main body The number of local records within the preset statistical time window. For training rounds Internal Participation Entities The number of departmental nodes in collaborative training. This represents the total number of department nodes participating in the current training round. S4-3: Department Node Calculation Entity In training rounds Cumulative exposure: ,in The exposure memory coefficient is between 0 and 1; S4-4: When the cumulative exposure meets the requirements At that time, the department node will be the main body Switch to group summary collaboration mode, in which only group-level summary representations are uploaded and not subject-level masking representations; when the cumulative exposure meets the requirements... At that time, the department node will be the main body Switch to local constraint update mode; in this mode, the main body... Stop participating in cross-departmental guidance and integration; complete the update for this round solely based on the target probability distribution already received in the current round. Subsequent rounds will continue training only based on the updated local model parameters. The first exposure threshold, This is the second exposure threshold.

6. The cross-departmental enterprise data security collaborative modeling method based on federated learning according to claim 5, characterized in that: S5 specifically includes the following sub-steps: S5-1: Department Node In training rounds The local model parameters are updated according to the following formula. , For training rounds The learning rate The gradient of the joint guiding loss with respect to the local model parameters; S5-2: Departmental nodes target the main body Calculate the subject participation coefficient for the next training round. ,in The minimum participation coefficient; S5-3: Securely aligned proxy nodes destroy the set of temporary identifiers for the current training round after the training round ends. And the corresponding cross-departmental alignment relationships; the coordinating node destroys the temporary alignment group index set for this round. fusion representation set and the set of target probability distributions Each department node retains only local model parameters and the cumulative exposure of the subject. S5-4: In the next training round, the coordinating node generates a new round random factor, and each department node regenerates the round temporary identifier, re-executes the temporary alignment, and re-selects the currently activated contribution slice, so that cross-round collaboration continues only at the model parameter level, and not at the subject alignment relationship level.

7. The cross-departmental enterprise data security collaborative modeling method based on federated learning according to claim 6, characterized in that: The system used in this method includes: The alignment mapping module is used to convert the same subject held by each department node into a temporary alignment identifier that is only valid in the current round within each training round, and further map the temporary alignment identifier to a group-level index so that cross-department joint processing is based on the granularity of temporary groups rather than the granularity of stable subjects. The training guidance module is used to divide the main features of each department node into multiple contribution slices according to preset rules. In the current round, only one contribution slice is activated for local encoding and intra-group masking. The coordination node generates a fusion representation and target probability distribution that are only valid for the current round based on the masking representation uploaded by each department node, and then transmits the target probability distribution back to each department node. The control recycling module is used to calculate the subject exposure based on the density of the alignment group to which the subject belongs, the sparsity of the subject's local records, and the cross-departmental coverage of the subject in the current round. When the subject exposure reaches a preset threshold, the collaboration mode is switched, and after the round ends, only the local model parameters are retained while the fusion representation, target probability distribution and its associated index formed in the current training round are destroyed.

8. The cross-departmental enterprise data security collaborative modeling method based on federated learning according to claim 7, characterized in that: The alignment mapping module includes: an identifier standardization and round temporary identifier generation module, which converts the subject identifiers within each department node into a unified identifier format and generates round temporary identifiers in combination with the random factor of the current training round, so that the same subject corresponds to different temporary identifiers in different training rounds; and an alignment group construction and security fill module, which is used by the security alignment proxy node to complete cross-department subject alignment according to the round temporary identifiers and map the alignment results to temporary alignment groups. When the number of subjects in the temporary alignment group is lower than the safety lower limit, a fill record is generated. The fill record is only used for statistical masking within the group and not for supervised loss calculation. The training guidance module includes: a feature slicing and unified encoding module, used to divide the local features of each subject at each department node into multiple contribution slices, and to encode the activated contribution slices in a unified dimension in the current training round to form an intermediate representation that can be fused across departments; and a fusion representation generation and target probability distribution back transmission module, used by the coordination node to perform weighted fusion of the intermediate representations uploaded by each department node within the same temporary alignment group, to generate a fusion representation that is only valid in the current round, and to generate a target probability distribution based on the fusion representation, and back transmission the target probability distribution to each department node to guide each department node to update its local model; The control and recycling module includes: an exposure calculation and threshold switching module, used to calculate the subject exposure by combining the temporary alignment group density, the number of local records of the subject, and the number of cross-department coverage of the subject, and to switch the subject to the group summary collaboration mode when the subject exposure reaches the first threshold, and to switch the subject to the local constraint update mode when the subject exposure reaches the second threshold; and a temporary state destruction and participation coefficient update module, used to destroy the temporary identifier of the round, the temporary alignment group index, the fusion representation, the target probability distribution and its associated index after each training round, and to update the subject participation coefficient of the next training round according to the subject exposure.

Images