Internal control defect collaborative prediction system based on federal causal forest and security aggregation

Abstract

The application discloses an internal control defect collaborative prediction system based on federal causal forest and safety aggregation, relates to the technical field of data intelligence, and comprises a data standard module, an internal control running state data of each participant is collected from an internal control environment, standardized processing is performed, and a standardized data set is obtained; a causal modeling module, based on the standardized data set, a causal forest structure is constructed locally at each participant, and iterative optimization is performed on causal path selection, and a local causal model is obtained; a path feedback module, according to prediction representation information, a corresponding causal path of each participant is fed back to indicate adjustment, and a global causal forest model is optimized. The application introduces a consistency driving causal path dynamic selection mechanism based on a time index in the causal inference process, so that the selection of the causal path can maintain continuity and consistency with the evolution of the internal control running state.

Description

Technical Field

[0001] This invention relates to the field of data intelligence technology, and in particular to a collaborative prediction system for internal control defects based on federated causal forests and security aggregation. Background Technology

[0002] With the increasing complexity of corporate governance structures and the high degree of informatization in business processes, internal control operational status data exhibits characteristics of multiple sources, strong temporal sequence, and strong coupling. In recent years, data-driven methods have been gradually applied to the field of internal control risk identification and defect prediction. By modeling and analyzing control parameters, operational status, and their temporal evolution relationships, potential internal control defects can be detected in advance. In this process, causal inference models have attracted attention because they can characterize the causal relationships between variables, and federated learning frameworks have also been gradually introduced into multi-agent collaborative analysis scenarios to meet the needs of cross-organizational data collaborative modeling.

[0003] Existing technologies for predicting internal control deficiencies typically focus on statistical correlation or static rule inference, making it difficult to simultaneously consider both time-adaptive modeling of causal paths and data privacy protection requirements in multi-participant environments. Especially when the internal control operation status evolves continuously over time, existing methods often struggle to achieve stable aggregation and consistent inference of causal path effects without exposing local data or model details, thus limiting the reliability and interpretability of collaborative prediction results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a collaborative prediction system for internal control defects based on federated causal forests and secure aggregation to solve the problem of difficulty in collaboratively predicting internal control defects when internal control data of multiple participants cannot be shared.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a collaborative prediction system for internal control deficiencies based on federated causal forests and secure aggregation. It includes: a data standardization module, where each participant collects internal control operational status data from the internal control environment and performs standardization processing to obtain a standardized dataset; a causal modeling module, which constructs a causal forest structure locally on each participant's site based on the standardized dataset and iteratively optimizes the selection of causal paths to obtain a local causal model; a homomorphic encryption module, which uses Paillier homomorphic encryption to encrypt the local causal models to obtain an encrypted causal model; a privacy aggregation module, which uploads the encrypted causal model to a central aggregation node and uses a differential privacy aggregation method to perform privacy-controlled collaborative aggregation calculations on the encrypted causal models of multiple participants to form a global causal forest model; a causal inference module, which performs causal inference calculations on newly accessed internal control operational status data based on the global causal forest model to obtain predictive representation information of internal control deficiencies; and a path feedback module, which adjusts the causal paths corresponding to the feedback from each participant based on the predictive representation information to optimize the global causal forest model.

[0008] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the internal control operation status data includes control parameter values, control parameter changes, operation status identifiers, and corresponding time indexes.

[0009] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the specific steps for obtaining the standardized dataset are as follows:

[0010] The internal control operation status data is linked and organized according to the time index to form the original data sequence arranged in chronological order;

[0011] The values ​​and changes of control parameters in the original data sequence are subjected to dimensional unification and numerical normalization, and the operation status identifiers are subjected to unified mapping to form a standardized data sequence.

[0012] The standardized data sequence is aligned and missing items are handled by time index, and then encapsulated to form a standardized dataset.

[0013] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the specific steps for constructing the causal forest structure are as follows:

[0014] Training samples are organized by time index based on a standardized dataset, and the training samples are divided into covariates, processing variables and target variables to form a training sample set.

[0015] Perform residual orthogonalization preprocessing on the training sample set to remove the influence of covariates on the treatment and target variables, resulting in a preprocessed sample set.

[0016] A causal tree structure is generated by dividing the sample space layer by layer based on the preprocessed sample set, and a local treatment effect representation is formed at the last-level node. The local treatment effect representations are then integrated to construct a causal forest structure.

[0017] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the specific steps for obtaining the local causal model are as follows:

[0018] Based on the causal forest structure, the splitting condition sequence from the root node to the leaf node of each causal tree is extracted, and each set of splitting condition sequences is associated with and stored with the local processing effect representation of its corresponding leaf node to form a candidate causal path library.

[0019] The causal forest structure, the candidate causal path library, and metadata recording the split points of each causal tree on the training set are encapsulated together to obtain a local causal model.

[0020] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the specific steps for obtaining the encrypted causal model are as follows:

[0021] Based on the local causal model, the causal forest structure, candidate causal path library and path decision records are read to form a local causal model representation to be encrypted;

[0022] Paillier homomorphic encryption is performed on the local causal model representation, and the encrypted local causal model representation is encapsulated to form an encrypted causal model.

[0023] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the specific process for forming the global causal forest model is as follows:

[0024] At the central aggregation node, encrypted causal models are received, and consistent alignment is performed on the number of causal trees, hierarchical order, and candidate causal path indexes according to the causal forest structure to form a set of encrypted causal models.

[0025] Based on the set of encrypted causal models, the local processing effect representations under candidate causal paths are subjected to collaborative aggregation processing within the encrypted domain to form a path-level global processing effect representation.

[0026] Based on the global treatment effect characterization, the path effect among different participants is corrected for consistency, and the path effect that deviates from the overall distribution range is weakened, thus obtaining the path treatment effect characterization.

[0027] The path processing effect is represented back to the corresponding causal forest structure position, forming a global causal forest model while keeping the splitting condition sequence unchanged.

[0028] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and secure aggregation described in this invention, the central aggregation node refers to a logical processing entity that performs unified aggregation calculation on the encrypted causal models uploaded by multiple participants without decrypting the encrypted causal models.

[0029] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the specific process for obtaining the predictive characterization information of internal control defects is as follows:

[0030] Based on the global causal forest model, the newly accessed internal control operation status data is read item by item in the order of time index, and the internal control operation status data is input into the corresponding splitting condition sequence in the global causal forest model;

[0031] The causal trees in the global causal forest are traversed layer by layer along the splitting condition sequence to locate the corresponding candidate causal paths and read the associated path processing effect representations.

[0032] The path processing effect representations output by multiple causal trees under the same time index are summarized to form the predictive representation information under the time index.

[0033] As a preferred embodiment of the internal control defect collaborative prediction system based on federated causal forest and security aggregation described in this invention, the optimization of the global causal forest model is specifically carried out as follows:

[0034] Based on the predicted representation information, the path treatment effect representations of the corresponding candidate causal paths are read, and the path treatment effect representations that continuously deviate in the time series are selected as the update objects.

[0035] In the global causal forest model, the candidate causal path position corresponding to the updated object is located. While keeping the splitting condition sequence unchanged, the path processing effect representation is replaced and written back to the corresponding final node position to optimize the global causal forest model.

[0036] The beneficial effects of this invention are as follows: by introducing a time-index-based consistency-driven dynamic selection mechanism for causal paths in the causal inference process, the selection of causal paths can maintain continuity and consistency with the evolution of the internal control operation status; without changing the causal forest structure and the splitting condition sequence, a path decision record is formed based on the consistency between the direction of change of the processing variable and the representation of the path processing effect, providing a stable and clear causal basis for model feedback. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of a collaborative prediction system for internal control defects based on federated causal forests and security aggregation.

[0039] Figure 2 A flowchart for data standard processing.

[0040] Figure 3 A flowchart for modeling local causality.

[0041] Figure 4 A flowchart for privacy aggregation feedback. Detailed Implementation

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0045] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides an internal control defect collaborative prediction system based on federated causal forest and security aggregation, comprising the following modules:

[0046] The data standard module allows each participant to collect internal control operation status data from the internal control environment and perform standardization processing to obtain a standardized dataset.

[0047] The internal control operation status data is linked and organized according to the time index to form the original data sequence arranged in chronological order.

[0048] Specifically, by reading the time index field carried in the internal control operation status data, time matching and association are performed on the internal control operation status data from different sources or different records. Data items with the same or continuous time index are combined and sorted according to the order of the time index to form a raw data sequence arranged continuously in time order.

[0049] It should be noted that the internal control operation status data includes control parameter values, control parameter changes, operation status identifiers, and corresponding time indexes.

[0050] The values ​​and changes of control parameters in the original data sequence are subjected to dimensional unification and numerical normalization, and the operation status identifiers are subjected to unified mapping processing to form a standardized data sequence.

[0051] Specifically, the range and variation of different control parameters in the original data sequence are transformed and scaled, and each control parameter is mapped to a comparable numerical range. At the same time, the discrete operating status identifiers are uniformly mapped according to the preset state correspondence, and the control parameter values, control parameter variations and operating status identifiers are represented in the same numerical system to form a standardized data sequence.

[0052] It should be noted that the preset status correspondence is set by performing category statistics and semantic grouping on historical internal control operation status data, mapping each type of operation status identifier to a discrete integer code or an ordered level code. The value range is determined to be a finite integer interval based on the number of operation status categories. For example, when there are The value range is set to the class state. .

[0053] The standardized data sequence is aligned and missing items are handled by time index, and then encapsulated to form a standardized dataset.

[0054] Specifically, using the time index as a unified alignment benchmark, consistency checks are performed on the data items corresponding to each time point in the standardized data sequence. Data items with time misalignment are realigned to the corresponding time index, and missing data items are filled in or marked according to the data relationship of adjacent time indices. This is done by inserting placeholders at the corresponding time index positions and recording the missing status, or by using existing data items under adjacent time indices to fill in the missing data. The aligned data items are then encapsulated in chronological order into a data set with a unified structure, forming a standardized dataset.

[0055] Standardized datasets refer to a collection of data formed by aligning internal control operation status data with time indexes, unifying units, normalizing values, mapping status, and handling missing items, and then encapsulating the data in a unified structure and time order.

[0056] The causal modeling module, based on a standardized dataset, constructs a causal forest structure locally for each participant and iteratively optimizes the selection of causal paths to obtain a local causal model.

[0057] Training samples are organized by time index based on a standardized dataset, and the training samples are divided into covariates, processing variables and target variables to form a training sample set.

[0058] Specifically, each internal control operation status data point in the standardized dataset is read according to the time index. The control parameter values, control parameter changes, and operation status identifiers under the same time index are combined and organized to form a training sample record sequence with time order constraints. In the training sample record sequence, the control parameter values ​​and control parameter changes fields describing the background conditions of the internal control operation status data are labeled as covariates, the operation status identifiers that characterize triggering relationships or intervention factors are labeled as processing variables, and the fields that characterize whether internal control defects occur or their level are labeled as target variables. Each training sample record is serialized into a training sample containing covariates, processing variables, and target variables, and then summarized and encapsulated according to the time index order to form a training sample set.

[0059] Perform residual orthogonalization preprocessing on the training sample set to remove the influence of covariates on the treatment and target variables, resulting in a preprocessed sample set.

[0060] Specifically, covariates, treatment variables, and target variables are read in the training sample set in sample order. Covariates are used as explanatory inputs, and least squares fitting calculations are performed on the treatment variables and target variables respectively. The least squares fitting calculations obtain the fitting representation of the covariates on the treatment variables and target variables by establishing the linear correspondence between the covariates and the treatment variables, as well as the linear correspondence between the covariates and the target variables.

[0061] Treatment variable residuals are generated by comparing the differences between the treatment variable and its fitted representation, and target variable residuals are generated by comparing the differences between the target variable and its fitted representation. This eliminates the linear influence of covariates on the treatment and target variables. The treatment and target variables in the training sample set are then replaced with their corresponding treatment and target variable residuals, which are then encapsulated together with the original covariates to form a preprocessed sample set.

[0062] A causal tree structure is generated by dividing the sample space layer by layer based on the preprocessed sample set, and a local treatment effect representation is formed at the last-level node. The local treatment effect representations are then integrated to construct a causal forest structure.

[0063] Specifically, the preprocessed sample set reads the covariates, processing variable residuals, and target variable residuals in the order of samples. Then, the sample space partitioning operation is performed layer by layer on the preprocessed sample set. By traversing the candidate split variables of the covariates and performing group comparison on the candidate split points of each candidate split variable, the dispersion of the target variable residuals in each group after splitting is statistically analyzed and the reduction of dispersion is compared to obtain the optimal split variable and the optimal split point, forming a causal tree structure.

[0064] In the final node of the causal tree structure, the residuals of the treatment variables and the residuals of the target variables covered by the final node are summarized, and linear fitting is performed on the residuals of the treatment variables and the residuals of the target variables to output the local treatment effect representation corresponding to the final node. Multiple causal tree structures are repeatedly generated and the local treatment effect representations output by the final nodes of each causal tree structure are integrated and summarized to construct a causal forest structure.

[0065] Among them, the causal tree structure refers to the tree-like hierarchical structure formed by dividing the sample space layer by layer based on the preprocessed sample set; the sample space refers to the overall feature domain composed of the range of multi-dimensional feature values ​​formed by all training samples in the preprocessed sample set in each covariate value dimension.

[0066] The local treatment effect characterization refers to the intensity and direction of the causal influence of the treatment variable on the target variable in a local sample subspace, based on the fitting relationship between the residuals of the treatment variable and the residuals of the target variable within a certain final node of the causal tree structure.

[0067] Based on the causal forest structure, the splitting condition sequence from the root node to the leaf node of each causal tree is extracted, and each set of splitting condition sequences is associated with and stored with the local processing effect representation of its corresponding leaf node to form a candidate causal path library.

[0068] Specifically, based on the causal forest structure, each causal tree in the causal forest structure is read sequentially. Starting from the root node, each branch is traversed layer by layer to the leaf node. During the traversal, the covariate value constraints, branch pointing relationships, and hierarchical position information corresponding to each node are extracted layer by layer. The splitting conditions that appear consecutively on the same path are concatenated and organized in the order from the root node to the leaf node to form a set of splitting condition sequences. According to the splitting condition sequence, the leaf node corresponding to the termination position of the splitting condition sequence is located. The local processing effect representation associated with the leaf node is read and stored in association with the local processing effect representation according to the one-to-one correspondence of the same path. At the same time, the causal tree number and path order information are recorded to distinguish different paths in different causal trees. The traversal extraction, sequential concatenation, leaf node location, and association storage operations are repeated for all causal trees in the causal forest structure to form a candidate causal path library composed of multiple sets of splitting condition sequences and corresponding local processing effect representations.

[0069] A superior approach, compared to traditional decision tree algorithms, divides the sample space layer by layer; by closely associating and storing the splitting condition sequence of each causal tree with the local processing effect representation of its corresponding leaf node, a candidate causal path library is formed, ensuring the integration of splitting conditions and effect representations. Through path order arrangement and concatenation of splitting condition sequences, interpretable causal path inference is provided.

[0070] The correlation between leaf nodes and local treatment effects is represented by the following formula:

[0071] ;

[0072] In the formula, Indicates the first The path effect association of trees, This represents the corresponding splitting condition sequence. This represents the local treatment effect of the final-level node. Indicates the first The number of layers in a cause-and-effect tree.

[0073] The path effect storage of all causal trees is represented by the formula:

[0074] ;

[0075] In the formula, Indicates multiple sets of splitting condition sequences and corresponding local treatment effect characterization The candidate causal path library jointly constitutes This represents the total number of trees in a causal forest.

[0076] Leaf nodes and local processing effects characterize the association formula and the path effect storage formula of all causal trees. These are standardized data, and their dimensions have been unified. These are split points in a tree structure, and are usually data that has undergone dimensional processing, or simply standardized binary conditions. It is the output value of a specific path, which will be expressed in some form of score or probability (such as a dimensionless value or a quantified effect value), and therefore, the dimensions are uniform.

[0077] The causal forest structure, the candidate causal path library, and metadata recording the split points of each causal tree on the training set are encapsulated together to obtain a local causal model.

[0078] Specifically, the training samples are read one by one according to the time index order of the training set based on the candidate causal path library, and the covariate value of each training sample is matched layer by layer with the splitting condition sequence in the candidate causal path library to determine the candidate causal path identifier corresponding to the training sample. The time index identifier is associated with the candidate causal path identifier and written to form a path decision record.

[0079] Following a unified encapsulation order, the causal forest structure, candidate causal path library, path decision records, and metadata recording the split point information of each causal tree on the training set are read sequentially. The causal tree number, hierarchical order, leaf node position, and local treatment effect representation in the causal forest structure are structurally organized. The split condition sequence, path identifier, and corresponding local treatment effect representation in the candidate causal path library are path mapped and organized. The time index identifier and candidate causal path identifier in the path decision records are time-series indexed and organized. The causal tree number, split variable identifier, split point value, hierarchical number, and node position in the metadata recording the split point information of each causal tree on the training set are indexed and organized.

[0080] The causal forest structure, candidate causal path library, path decision record, and metadata recording the split point information of each causal tree on the training set are combined and encapsulated in a fixed order. Consistency verification is completed by checking the correspondence between causal tree numbers, path identifiers, and split point indexes item by item to obtain the local causal model.

[0081] It should be noted that the splitting condition is composed of the covariate value satisfaction relationship and the splitting branch pointing relationship. The covariate value satisfaction relationship comes from the layer-by-layer recording of the covariate value range or judgment condition during the construction of the causal tree, and the splitting branch pointing relationship comes from the path record of the sample being assigned to different branches under the corresponding splitting condition.

[0082] Path decision records refer to time-series records formed by associating the candidate causal path identifiers matched for each training sample with their corresponding time index identifiers and storing them continuously according to the time index order of the training set.

[0083] The homomorphic encryption module uses Paillier homomorphic encryption to encrypt the local causal model, resulting in an encrypted causal model.

[0084] Based on the local causal model, the causal forest structure, candidate causal path library, and path decision records are read to form a local causal model representation to be encrypted.

[0085] Specifically, based on the local causal model, the splitting condition sequence and hierarchical organization information of each causal tree in the causal forest structure are read sequentially. The splitting condition sequence and the binding relationship of the local treatment effect representation of each candidate causal path in the candidate causal path library are read. The causal path selection information arranged by time index in the path decision record is read. The causal forest structure is used as structural description information, the candidate causal path library is used as path mapping information, and the path decision record is used as path usage information under time index. These are then uniformly organized and sequentially arranged to form the local causal model representation to be encrypted.

[0086] Paillier homomorphic encryption is performed on the local causal model representation, and the encrypted local causal model representation is encapsulated to form an encrypted causal model.

[0087] Specifically, the structural description information, path mapping information, and time index association information in the local causal model representation are read item by item in a predetermined order, and the quantifiable parameters are organized into an encryptable parameter sequence. Paillier public key encryption is performed on each of the encryptable parameter sequences to map each numerical parameter to a corresponding ciphertext representation. The splitting condition sequence, covariate value satisfaction relationship, and splitting branch pointing relationship in the causal forest structure are kept unchanged in terms of structural identifier form and are aligned and labeled using a unified structural benchmark order. The ciphertext representation is backfilled and encapsulated according to the original causal forest structure position and organized together with the structural identifier information that has not been encrypted to form an encrypted causal model.

[0088] The predetermined order refers to the fixed order in which the encryptable parameters in the local causal model representation are uniformly arranged and read item by item according to the causal tree numbering order, the hierarchical numbering order, the candidate causal path index order, and the time index order.

[0089] The privacy aggregation module uploads the encrypted causal model to the central aggregation node and uses the differential privacy aggregation method to perform privacy-controlled collaborative aggregation computation on the encrypted causal models of multiple participants to form a global causal forest model.

[0090] The encrypted causal model is received at the central aggregation node, and consistent alignment is performed on the number of causal trees, hierarchical order, and candidate causal path indexes according to the causal forest structure to form a set of encrypted causal models.

[0091] Specifically, at the central aggregation node, encrypted causal models sent by multiple participants are received sequentially. The number and hierarchical order of causal trees in the causal forest structure and the index of candidate causal paths in the candidate causal path library are read from each encrypted causal model. Using the baseline order of the causal forest structure as an alignment reference, the number of causal trees recorded in each encrypted causal model is compared with the number of trees in the unified causal forest structure baseline to obtain the unified structure baseline.

[0092] Read the hierarchical numbers in each encrypted causal model in order of causal tree numbering, and rearrange the hierarchical records in order of hierarchical numbering within each causal tree to generate a hierarchical organization sequence; split the candidate causal path index into causal tree number, hierarchical number, and final node number, and relocate them according to their corresponding positions in the unified index mapping relationship to obtain the candidate causal path index relationship; summarize and encapsulate the unified structural benchmark, hierarchical organization sequence, and candidate causal path index relationship in order of participant identification to form an encrypted causal model set.

[0093] It should be noted that the central aggregation node refers to the logical processing entity that performs unified aggregation calculations on the encrypted causal models uploaded by multiple participants without decrypting the encrypted causal models.

[0094] Before each participant constructs the causal forest structure, a unified causal forest structure template is pre-issued by the central aggregation node. Each participant only calculates the local processing effect representation of the final-level node within the framework of the splitting condition sequence defined by the causal forest structure template.

[0095] Based on the set of encrypted causal models, the local processing effect representations under candidate causal paths are subjected to collaborative aggregation processing within the encrypted domain to form a path-level global processing effect representation.

[0096] Specifically, based on the encrypted causal model set, the encrypted causal models are read one by one in the order of the participant identifiers. According to the candidate causal path index relationship, the encrypted local processing effect representations corresponding to each participant under the same candidate causal path are located. The encrypted local processing effect representations under the same candidate causal path are aligned and organized to form a path ciphertext group. Homomorphic summation aggregation operation in the encrypted domain is performed on the encrypted local processing effect representations in the path ciphertext group. Ciphertext addition is performed on each ciphertext representation in the path ciphertext group to generate an aggregated ciphertext representation. The aggregated ciphertext representation is backfilled and encapsulated according to the candidate causal path index relationship to form a global processing effect representation.

[0097] Among them, the global processing effect characterization refers to the path-level overall causal influence characterization formed by homomorphically aggregating the encrypted local processing effect characterizations of each participant under the same candidate causal path index relationship.

[0098] Based on the global treatment effect characterization, the path effect among different participants is corrected for consistency, weakening the path effect that deviates from the overall distribution range, thus obtaining the path treatment effect characterization.

[0099] Specifically, based on the global processing effect representation, the corresponding global processing effect representation is read one by one according to the candidate causal path index relationship. Simultaneously, the corresponding encrypted local processing effect representation under the same candidate causal path is read from the encrypted causal model set according to the participant identifier. Homomorphic difference operation is performed on the encrypted local processing effect representation and the global processing effect representation in the encrypted domain to generate path effect difference ciphertext. The path effect difference ciphertext is submitted to the threshold Paillier joint decryption mechanism, where multiple participants holding private key shares jointly participate in threshold decryption, and only the difference value of the interval comparison is output.

[0100] Based on the path effect difference value and the preset difference allowable range, an interval comparison calculation is performed. When the difference value falls within the preset difference allowable range, it is marked as no correction is needed. When the difference value exceeds the preset difference allowable range, it is marked as correction is needed. Corresponding judgment information is generated and returned to the central aggregation node. The central aggregation node performs a homomorphic proportional adjustment operation on the corresponding encrypted local processing effect representation in the encrypted domain according to the judgment information to generate the corrected encrypted path effect representation. The corrected encrypted path effect representation is backfilled into the corresponding position according to the candidate causal path index relationship and encapsulated to form the path processing effect representation.

[0101] The preset allowable range of differences is set by calculating the central tendency and discrete scale of the historical statistical distribution of the path effect differences of the participants under the same candidate causal path. Specifically, the mean and standard deviation of the historical samples are statistically analyzed, and a standard deviation interval of a fixed multiple is used as the allowable range of differences. The historical statistical distribution is formed by the central aggregation node continuously collecting the path effect difference descriptions of each participant according to the candidate causal path index during multiple aggregation processes and accumulating them over time. The value range is limited to a finite numerical range that covers the normal fluctuation range and excludes abnormal offsets.

[0102] A better approach is to perform a one-time aggregation process on the local processing effect representations corresponding to the same causal path in federated or distributed scenarios. Instead, based on the global processing effect representation, the global processing effect representation is used as a reference for path effect consistency. The encrypted local processing effect representations carrying path effects are conditionally adjusted. While maintaining the structural integrity, abnormal offsets are suppressed and the stability of global path effects is improved.

[0103] The path processing effect is represented back to the corresponding causal forest structure position, forming a global causal forest model while keeping the splitting condition sequence unchanged.

[0104] Specifically, the path processing effect representation is read one by one according to the candidate causal path index relationship. The causal tree number, level number, and final node number that are consistent with the candidate causal path index relationship are located in the causal forest structure. The path processing effect representation is written into the storage location of the local processing effect representation associated with the final node. During the writing process, the splitting condition sequence content and level order recorded in the causal forest structure are kept unchanged. Only the local processing effect representation associated with the final node is replaced and backfilled. The mapping verification information between the candidate causal path index relationship and the causal forest structure position is recorded for each backfilling operation. The backfilled causal forest structure is encapsulated to form a global causal forest model.

[0105] Specifically, a training sample set containing covariates, treatment variables, and target variables is formed by organizing a standardized dataset according to time index. Residual orthogonalization is performed on the training sample set to remove the linear influence of covariates on treatment and target variables. Multiple causal tree structures are generated by traversing candidate split variables and candidate split points to divide the sample space layer by layer on the preprocessed sample set. Linear fitting is performed on the residuals of treatment variables and target variables of the covered samples in the final node of each causal tree to estimate the local treatment effect representation. The local treatment effect representations of multiple causal trees are integrated and summarized to train the causal forest model.

[0106] The causal inference module, based on the global causal forest model, performs causal inference calculations on newly accessed internal control operation status data to obtain predictive characterization information of internal control defects.

[0107] Based on the global causal forest model, newly accessed internal control operation status data are read item by item in time index order, and the internal control operation status data is input into the corresponding splitting condition sequence in the global causal forest model.

[0108] Specifically, the newly accessed internal control operation status data is read item by item according to the time index order, and the covariate values ​​consistent with the splitting condition sequence in the causal forest structure are extracted from each internal control operation status data. According to the hierarchical order of the splitting condition sequence, the covariate value constraint relationship specified in each splitting condition is read layer by layer, and the corresponding covariate values ​​in the internal control operation status data are compared with the constraint relationship item by item to determine the splitting branch pointing relationship, and the internal control operation status data is mapped to the position of the corresponding splitting condition sequence in the global causal forest model.

[0109] Specifically, when it is necessary to predict defects in newly accessed internal control operation status data, the newly accessed internal control operation status data is input into the data standard module. The data standard module transforms the data into a standardized dataset according to the same processing flow as the training data (including sorting by time index, unit normalization, status identifier mapping, missing value handling, etc.).

[0110] The data source for internal control operation status data and the data source for newly connected internal control operation status data belong to the same unified internal control environment monitoring system and follow the same collection specifications and definitions. For example, whether in the training phase or the prediction phase, the collected internal control operation status data includes data in four core dimensions: "control parameter values, control parameter changes, operation status identifiers, and corresponding time indices".

[0111] The system traverses each causal tree in the global causal forest layer by layer along the splitting condition sequence, locates the corresponding candidate causal paths, and reads the associated path processing effect representations.

[0112] Specifically, based on the determined splitting branch pointing relationship, the system traverses the splitting condition sequence layer by layer to the final node position in each causal tree in the global causal forest. The final node number is matched with the causal tree number and the level number to form a candidate causal path index relationship. The associated path processing effect representation is then read from the position pointed to by the candidate causal path index relationship.

[0113] The path processing effect representations output by multiple causal trees under the same time index are summarized to form the predictive representation information under the time index.

[0114] Specifically, under the same time index, the path treatment effect representations output by each causal tree in the global causal forest model are aggregated. The path treatment effect representations are uniformly organized according to the causal tree numbering order, and consistent alignment is performed based on the causal path identifiers corresponding to each path treatment effect representation. Multiple path treatment effect representations under the same time index are mapped to a unified effect scale space. Normalized scale integration calculation is performed on the mapped path treatment effect representations to generate the comprehensive effect value corresponding to the time index.

[0115] The comprehensive effect value is input into the defect risk mapping function that has been trained in advance using historical labeled samples, and the corresponding internal control defect occurrence probability value is output. Then, according to the preset probability interval division rules, the internal control defect occurrence probability value is converted into the corresponding internal control defect level identifier, thereby forming internal control defect prediction information that can be directly output.

[0116] It should be noted that the preset probability interval division rules are set by statistically analyzing the predicted probability distribution of historically marked internal control defect samples, using quantile division or selecting the threshold corresponding to the optimal Youden index based on the ROC curve as the dividing point, and dividing the probability interval into multiple continuous segments according to the number of risk levels.

[0117] Internal control deficiencies refer to situations in the internal control environment that fail to meet standards or safety requirements, which may manifest as anomalies, operational errors, or inadequate risk management.

[0118] The path feedback module, based on the predicted representation information, instructs and adjusts the causal paths corresponding to the feedback from each participant, thereby optimizing the global causal forest model.

[0119] Based on the predicted representation information, the path treatment effect representations of the corresponding candidate causal paths are read, and the path treatment effect representations that continuously deviate in the time series are selected as the update objects.

[0120] Specifically, based on the predicted representation information, the corresponding candidate causal path identifiers are read one by one in the order of time index, and the associated path treatment effect representations are located accordingly. The changes of the same path treatment effect representation under adjacent time indexes are continuously compared along the time series direction to determine whether the path treatment effect representation shows a stable deviation in multiple consecutive time indices. The path treatment effect representations that show a stable deviation are marked as update objects.

[0121] Among them, the stable deviation state refers to the continuous deviation state in which the same path processing effect is characterized by consistent direction and continuous amplitude changes at multiple consecutive time indices relative to the existing historical level.

[0122] The difference between the current path processing effect representation and its historical rolling average is used as the deviation. A continuous time window length N and a deviation magnitude threshold δ are set. When the absolute value of the deviation of the path is greater than δ and the deviation direction is consistent for N consecutive time indices, it is determined to be in a stable deviation state and is selected as an update object.

[0123] In the global causal forest model, the candidate causal path position corresponding to the updated object is located. While keeping the splitting condition sequence unchanged, the path processing effect representation is replaced and written back to the corresponding final node position to optimize the global causal forest model.

[0124] Specifically, in the global causal forest model, the candidate causal path identifiers corresponding to the updated object are read one by one according to the candidate causal path index relationship. The position of the last-level node is located in the causal forest structure based on the causal tree number, level number and last-level node number obtained by splitting the candidate causal path identifier. At the located last-level node position, the content of the splitting condition sequence and the hierarchical order remain unchanged. Only the original path processing effect representation of the local processing effect representation associated with the last-level node is replaced with the last-level node position corresponding to the updated object, thereby optimizing the global causal forest model.

[0125] In summary, this invention achieves this by introducing a time-indexed consistency-driven dynamic causal path selection mechanism during the causal inference process, ensuring the continuity and consistency of causal path selection as the internal control operation status evolves; and by forming path decision records based on the consistency between the direction of change of the processing variables and the representation of the path processing effect without changing the causal forest structure and the splitting condition sequence, thus providing a stable and clear causal basis for model feedback.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An internal control defect collaborative prediction system based on federated causal forest and security aggregation, characterized in that: include, The data standard module allows each participant to collect internal control operation status data from the internal control environment. The internal control operation status data includes control parameter values, control parameter changes, operation status identifiers, and corresponding time indices. The data is then standardized to obtain a standardized dataset. The causal modeling module, based on a standardized dataset, constructs a causal forest structure locally for each participant and iteratively optimizes the selection of causal paths to obtain a local causal model. The specific steps are as follows: Training samples are organized by time index based on standardized datasets and divided into covariates, processing variables and target variables to form a training sample set. The control parameter values ​​and control parameter changes that describe the background conditions of internal control operation status data are labeled as covariates, the operation status identifiers that characterize triggering relationships or intervention factors are labeled as processing variables, and the fields that characterize whether internal control defects occur or the level of internal control defects are labeled as target variables. Perform residual orthogonalization preprocessing on the training sample set to remove the influence of covariates on the treatment and target variables, resulting in a preprocessed sample set. Based on the preprocessed sample set, the sample space is divided layer by layer to generate a causal tree structure, and a local treatment effect representation is formed at the last-level node. The local treatment effect representation refers to the strength and direction of the causal influence of the treatment variable on the target variable in the local sample subspace, which is characterized by the fitting relationship between the residuals of the treatment variable and the residuals of the target variable covered by the node in the last-level node of the causal tree structure. The local treatment effect representation is integrated to construct a causal forest structure. Based on the causal forest structure, the splitting condition sequence from the root node to the leaf node of each causal tree is extracted. The splitting conditions that appear consecutively on the same path are linked together in the order from the root node to the leaf node to form a set of splitting condition sequences. The splitting conditions are composed of the covariate value satisfaction relationship and the splitting branch pointing relationship. Each set of splitting condition sequences is associated with and stored with the local processing effect representation of its corresponding leaf node to form a candidate causal path library. The causal forest structure, the candidate causal path library, and metadata recording the split point information of each causal tree on the training sample set are jointly encapsulated to obtain a local causal model. The homomorphic encryption module uses Paillier homomorphic encryption to perform encryption processing on the local causal model, resulting in an encrypted causal model. The privacy aggregation module uploads the encrypted causal model to the central aggregation node. It then uses a differential privacy aggregation method to perform privacy-controlled collaborative aggregation computation on the encrypted causal models of multiple participants, forming a global causal forest model. The specific process is as follows: At the central aggregation node, encrypted causal models are received, and consistent alignment is performed on the number of causal trees, hierarchical order, and candidate causal path indexes according to the causal forest structure to form a set of encrypted causal models. Based on the set of encrypted causal models, the local processing effect representations under candidate causal paths are subjected to collaborative aggregation processing within the encrypted domain to form a path-level global processing effect representation. Based on the global treatment effect characterization, the global treatment effect characterization is used as a path effect consistency reference to correct the consistency of path effects among different participants, weaken the path effects that deviate from the overall distribution range, and obtain the path treatment effect characterization. The path processing effect is written back to the corresponding causal forest structure position to form a global causal forest model while keeping the splitting condition sequence unchanged. The causal inference module, based on the global causal forest model, performs causal inference calculations on newly accessed internal control operation status data to obtain predictive characterization information of internal control defects. The path feedback module, without changing the causal forest structure and the sequence of splitting conditions, adjusts the causal paths corresponding to the feedback from each participant based on the predicted representation information, thereby optimizing the global causal forest model.

2. The internal control defect collaborative prediction system based on federated causal forest and security aggregation as described in claim 1, characterized in that: The standardized dataset is obtained through the following steps: The internal control operation status data is linked and organized according to the time index to form a raw data sequence arranged in chronological order; The values ​​and changes of control parameters in the original data sequence are subjected to dimensional unification and numerical normalization, and the operation status identifiers are subjected to unified mapping to form a standardized data sequence. The standardized data sequence is aligned and missing items are handled by time index, and then encapsulated to form a standardized dataset.

3. The internal control defect collaborative prediction system based on federated causal forest and security aggregation as described in claim 1, characterized in that: The specific steps to obtain the encrypted causal model are as follows: Based on the local causal model, the causal forest structure, candidate causal path library and path decision records are read to form a local causal model representation to be encrypted. Paillier homomorphic encryption is performed on the local causal model representation. The encrypted local causal model representation is then encapsulated to form an encrypted causal model. In this process, training samples are read one by one according to the time index of the training set based on the candidate causal path library. The covariate value of each training sample is matched layer by layer with the splitting condition sequence in the candidate causal path library to determine the candidate causal path identifier corresponding to the training sample. The time index identifier is associated with the candidate causal path identifier and written to form a path decision record.

4. The internal control defect collaborative prediction system based on federated causal forest and security aggregation as described in claim 1, characterized in that: The central aggregation node refers to a logical processing entity that performs unified aggregation calculations on encrypted causal models uploaded by multiple participants without decrypting the encrypted causal models.

5. The internal control defect collaborative prediction system based on federated causal forest and security aggregation as described in claim 1, characterized in that: The specific process for obtaining the predictive characterization information of internal control defects is as follows: Based on the global causal forest model, the newly accessed internal control operation status data is read item by item in the order of time index, and the internal control operation status data is input into the corresponding splitting condition sequence in the global causal forest model; The causal trees in the global causal forest are traversed layer by layer along the splitting condition sequence to locate the corresponding candidate causal paths and read the associated path processing effect representations. The path processing effect representations output by multiple causal trees under the same time index are summarized to form the predictive representation information under the time index.

6. The internal control defect collaborative prediction system based on federated causal forest and security aggregation as described in claim 5, characterized in that: The specific process for optimizing the global causal forest model is as follows: Based on the predicted representation information, the path treatment effect representations of the corresponding candidate causal paths are read, and the path treatment effect representations that continuously deviate in the time series are selected as the update objects. In the global causal forest model, the candidate causal path position corresponding to the updated object is located. While keeping the splitting condition sequence unchanged, the path processing effect representation is replaced and written back to the corresponding final node position to optimize the global causal forest model.

Images