ERP financial data sharing safety management system based on AI and privacy calculation

Abstract

The invention relates to the technical field of machine learning and privacy computing, and particularly discloses an ERP financial data sharing security management system based on AI and privacy computing. The system comprises a data grading desensitization module, a distributed privacy calculation engine, a cross-chain audit evidence storage module and a dynamic compliance verification module, and through closed-loop cooperation of grading desensitization, secret state calculation, double-chain evidence storage and real-time verification, whole-process traceability, tampering prevention and dynamic compliance guarantee of financial data in the sharing process are achieved. And the data security and auditing transparency are improved.

Description

Technical Field

[0001] This invention belongs to the field of machine learning and privacy computing technology, specifically relating to an ERP financial data sharing and security management system based on AI and privacy computing. Background Technology

[0002] Enterprise Resource Planning (ERP) systems, as the core platform for modern enterprise operations and management, carry critical business data such as finance, supply chain, and human resources. Data security and compliance have become fundamental requirements for enterprise digital transformation. With the accelerating marketization of data elements, how to achieve cross-organizational data value transfer while ensuring data privacy has become a crucial issue in enterprise digital construction.

[0003] Privacy-preserving computation-based data sharing technology utilizes cryptography and trusted execution environments to achieve data fusion computation without leaving the original data domain, providing a technical path for the secure flow of financial data. Existing solutions typically interface the privacy-preserving computation module as an independent component with the ERP system, forming an add-on architecture.

[0004] The existing add-on integration model of privacy-preserving computation technology with ERP systems faces compatibility challenges when adapting to the domestic IT innovation environment, and it is difficult to achieve deep collaboration with the financial compliance management functions built into the ERP system, such as electronic voucher archiving and data ownership confirmation. The reasons for insufficient compatibility are: isolated architectural design, with add-on modules deployed independently and externally, not integrated into the ERP system kernel, and lacking underlying adaptation design with the ERP's financial data model and business process logic. Especially in the application scenarios of state-owned enterprises, the usage trajectory of financial data processed by privacy computation is difficult to trace, failing to meet the compliance requirements of the "Electronic Voucher Accounting Data Standard" for auditability throughout the entire financial lifecycle, resulting in audit blind spots and compliance risks in the data sharing process. Therefore, there is an urgent need to build a secure sharing system that deeply integrates privacy computation with ERP financial data management capabilities. Summary of the Invention

[0005] This invention aims to provide an ERP financial data sharing and security management system based on AI and privacy computing, in order to resolve the technical contradictions between existing plug-in privacy computing architecture and ERP system in the context of information technology innovation, the inability to achieve full lifecycle auditability of financial data, and the difficulty in meeting the compliance requirements of electronic voucher accounting data standards.

[0006] This invention provides an ERP financial data sharing and security management system based on AI and privacy computing. The system is deeply embedded in the ERP system kernel and constructs a closed-loop collaborative system that includes four major modules: data hierarchical desensitization, distributed privacy computing, cross-chain audit evidence storage, and dynamic compliance verification. This system achieves full-process coverage from data preprocessing and privacy computing to result auditing. A dual-chain structure is used to record the entire process of privacy computing, including operation traces and data usage records, ensuring traceability and immutability in the sharing and transfer of financial data. The dynamic compliance verification module performs real-time automated verification of the output results based on the electronic voucher accounting data standard. Through the design of a domestically produced hardware security layer and a layered computing architecture, it is compatible with the domestic IT innovation environment; it can be directly embedded into the kernel of a self-developed domestic IT innovation version ERP product, reuse the domestic hardware resources of the ERP, reduce deployment costs, and link with the ERP electronic voucher management module and financial risk control module. The privacy calculation results directly trigger the ERP voucher archiving and risk control early warning process, solving the problems of secure sharing and compliance of financial data.

[0007] Preferably, the data classification and desensitization module establishes a persistent connection with the ERP financial database through a standard database connection interface. Internally, it includes a data classification unit, an algorithm execution unit, and a data encapsulation unit. The data classification unit reads a preset classification strategy configuration file, performs sensitive semantic analysis and attribute determination on the financial data fields through an AI semantic recognition model, and classifies the data levels based on business attributes. The algorithm execution unit jointly calls the desensitization algorithm according to the data level and sensitive detail tags. The data encapsulation unit packages the processed data into structured data packets with metadata identifiers and transmits them to the distributed privacy computing engine through a secure protocol.

[0008] Preferably, the AI ​​semantic recognition model has a dynamic learning mechanism. When a financial field is added to the ERP system, incremental training is performed on a small number of manually labeled samples to quickly adapt to the new field without retraining the full amount of data. The data classification unit sets a confidence verification threshold and performs automatic classification, AI suggestion + manual review, or pauses automatic classification and supplements samples for retraining based on the confidence result.

[0009] Preferably, the distributed privacy computing engine is deployed on a trusted computing cluster and adopts a layered computing architecture, which includes a computing scheduling layer, a cryptographic protocol layer and a hardware security layer from top to bottom. The computing scheduling layer parses the multi-party secure computing task graph in the encrypted data packet through a task parser. The resource manager allocates subtasks to the trusted execution environment of the node according to the load and latency. The cryptographic protocol layer provides cryptographic primitive implementations to ensure that the original data does not appear in plaintext form. The hardware security layer is built on domestic cryptographic chips to provide hardware-level protection for key management and cryptographic operations.

[0010] Preferably, the cross-chain audit and evidence storage module includes an audit blockchain and an evidence storage blockchain, both of which adopt a permissioned blockchain architecture. The audit blockchain records metadata of privacy computing tasks, while the evidence storage blockchain stores hash digests of data usage trajectories and integrates an AI abnormal behavior monitoring model to analyze the temporal characteristics and correlations of trajectories in real time, identify suspicious patterns, and perform graded early warning processing.

[0011] Preferably, the audit blockchain and the evidence storage blockchain are interconnected through a cross-chain protocol, and the latest Merkle root hash value is written to each other at regular intervals of a certain number of blocks to achieve consistency of record time sequence and immutability; when the AI ​​abnormal behavior monitoring model triggers an early warning, the system generates an early warning log and calculates its hash value, and writes it to the two blockchains simultaneously through the cross-chain protocol to achieve three-way hash binding of the early warning log, abnormal trajectory record and audit task record.

[0012] Preferably, the AI ​​abnormal behavior monitoring model adopts a dynamic optimization mechanism of offline iteration and online fine-tuning. During offline iteration, the early warning logs and manual review results are summarized monthly, the model parameters are updated incrementally and verified by performance indicators, and during online fine-tuning, the administrator can trigger the model by annotating a small number of samples, so as to quickly adapt to new risks.

[0013] Preferably, the dynamic compliance verification module is embedded in the ERP business process controller. The verification process is divided into two stages: structural verification and logical verification. The structural verification retrieves the electronic voucher accounting data standard rule library and performs field-by-field matching verification on the structure of the output data fields. The logical verification adopts a dual mechanism of preset rule verification and AI deep learning verification to capture explicit logical relationships and implicit compliance risks, and performs hierarchical processing on the verification results, linking the cross-chain audit and evidence storage module to record relevant information.

[0014] Preferably, the dynamic compliance verification module has an AI-driven feedback adjustment mechanism. When the same type of data is detected to continuously trigger compliance warnings, the historical warning data is encapsulated as features and input into the AI ​​intelligent decision-making model. The model outputs strategy adjustment suggestions and confidence scores. Based on the scores, a strategy update signal is generated and sent to the data classification and desensitization module. The strategies of each participating node are updated synchronously and the ledger is recorded to the audit blockchain.

[0015] Preferably, in a heterogeneous computing environment, the distributed privacy computing engine's computing scheduling layer introduces an intelligent resource allocation algorithm, integrates an AI-integrated performance prediction model to output the expected execution time and execution stability score of the task, and the resource manager uses a time-stability weighted scoring mechanism to select the optimal node combination; the cryptographic protocol layer adopts an optimized share generation algorithm and an AI adaptive compression algorithm; the hardware security layer expands the compatibility of domestic cryptographic chips and adds a dynamic measurement mechanism; the cross-chain audit and evidence storage module adopts a hierarchical Merkle tree structure and a memory cache pool to improve retrieval efficiency; and the data hierarchical desensitization module supports custom classification rules and dynamic level adjustment.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By deeply embedding privacy computing capabilities into the kernel of the ERP system, a closed-loop collaborative system of four major modules—data hierarchical desensitization, distributed privacy computing, cross-chain audit and evidence storage, and dynamic compliance verification—was constructed, achieving full-process coverage from data preprocessing and privacy computing to result auditing. 2. The system uses a dual-chain structure to record the operation trajectory and data usage traces of the entire privacy computing process, ensuring the traceability and immutability of every piece of financial data in the shared circulation, and completely eliminating audit blind spots; 3. The dynamic compliance verification module performs real-time automated verification of the output results based on the electronic voucher accounting data standard, effectively ensuring the compliance of shared data in terms of format, logic and business rules; 4. Through the design of a domestically produced hardware security layer and a layered computing architecture, the system has good compatibility with the domestic IT innovation environment and can meet the requirements for secure sharing of financial data in scenarios with high standards of compliance such as state-owned enterprises and central enterprises. 5. This system can be directly embedded into the kernel of the company's self-developed IT-innovation version of ERP products without the need for additional independent server deployment. It can reuse the existing domestic hardware resources of the ERP, reducing the deployment cost for enterprises. At the same time, it can be linked with the ERP's electronic voucher management module and financial risk control module. The privacy calculation results can directly trigger the ERP's voucher archiving and risk control early warning processes, solving the core issues of data security sharing and financial compliance for state-owned and central enterprises. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the ERP financial data sharing and security management system based on AI and privacy computing proposed in this invention; Figure 2 This is a flowchart of the hierarchical processing logic of the data hierarchical desensitization module in this invention; Figure 3 This is a schematic diagram of the layered computing architecture of the distributed privacy computing engine in this invention; Figure 4 This is a flowchart of the two-stage process of structural verification and logical verification of the dynamic compliance verification module in this invention. Detailed Implementation

[0018] This embodiment details the technical implementation of an ERP financial data sharing and security management system based on AI and privacy computing. Please refer to the appendix. Figure 1 To be continued Figure 4 This system is deeply integrated into the software kernel of the Enterprise Resource Planning (ERP) system, forming a closed-loop security system covering four core aspects: data preprocessing, encrypted computation, audit evidence storage, and compliance verification. The entire system engages in bidirectional data interaction and command communication with the ERP system's financial database, business process controller, and a trusted computing cluster composed of multiple participating computing nodes.

[0019] The system begins with the data classification and anonymization module. This module establishes a persistent connection with the financial database of the Enterprise Resource Planning (ERP) system via a standard database connection interface. Internally, it comprises three sub-components: a data classification unit, an algorithm execution unit, and a data encapsulation unit. The data classification unit first reads a preset data classification strategy configuration file, then scans the raw financial data in the ERP system's financial database line by line. It first uses an AI semantic recognition model to perform sensitive semantic analysis and attribute determination on the fields, and then combines this with the business attributes of the data fields to classify them into three distinct levels: Level 1 (Core Financial Data): The AI ​​semantic recognition model extracts the semantics of fund association and unique identifier in the field to determine that it has high compliance risk and is directly related to core assets, and finally incorporates it into the core financial data, covering the accounting voucher amount field, bank account number field and taxpayer identification number field. Level 2 (Business Support Data): AI models identify fields with business-related semantics but no direct financial attributes. They determine compliance risks and support business decision-making, and include them in business support data, including supplier name field, product name field, and transaction date field. Level 3 (Statistical Derivative Data): The AI ​​model identifies the semantics of the aggregated fields and determines that they have low compliance risk and no association with individual / core assets, and includes them in the statistical derivative data. This mainly involves the field of departmental statistical summary value and the field of monthly transaction frequency.

[0020] The AI ​​semantic recognition model features a dynamic learning mechanism: when a new financial field is added to the ERP system, the model can quickly adapt to the new field through a small number of manually labeled samples and incremental training. Administrators only need to upload the business semantic description and compliance tags for the new field in the system backend, and the model automatically performs incremental fine-tuning based on historical training parameters, without needing to retrain the entire dataset. After fine-tuning, the model's sensitivity judgment accuracy for the new field must reach ≥98% before it can be put into use. Simultaneously, the data classification unit sets a confidence level verification threshold: the AI ​​model outputs a sensitivity judgment confidence level in the 0-1 range for each field. When the confidence level is ≥0.95, the level is directly classified according to the model result; when 0.8 ≤ confidence level < 0.95, an AI suggestion + manual review process is triggered; when the confidence level < 0.8, automatic classification is paused, and the administrator is prompted to supplement labeled samples and retrain the model.

[0021] After classification is completed, the algorithm execution unit jointly invokes the de-identification algorithm based on the data level and the sensitive detail tags output by the AI ​​semantic recognition model: Level 1 core financial data: The AI ​​model outputs sensitive detail labels, the algorithm execution unit loads the key pair of the homomorphic encryption algorithm, the bank account number adopts the ring learning error problem encryption scheme + 2048-bit key, the amount field adopts partial homomorphic encryption + precision preservation mode, and generates the corresponding ciphertext field. The original plaintext field is immediately cleared to zero in memory. Level 2 Business Auxiliary Data: The AI ​​model outputs noise perturbation intensity to suggest labels, the algorithm execution unit initializes the differential privacy engine, sets the Gaussian noise standard deviation σ=0.08 for high repetition fields, sets σ=0.12 for time series fields, and superimposes a noise perturbation that conforms to the distribution for each target field value; Level 3 Statistical Derivative Data: The AI ​​model outputs a k-value to suggest labels, and the algorithm execution unit performs k-anonymization processing. Through generalization and suppression techniques, it is ensured that each record is indistinguishable from at least another k-1 records on the quasi-identifier set.

[0022] Finally, the data encapsulation unit packages the processed data into structured data packets. Each data packet is appended with metadata identifiers including data level, de-identification algorithm identifier, valid timestamp, and AI-determined confidence level. These packets are then transmitted to the task queue of the distributed privacy computing engine via the Secure Sockets Layer protocol.

[0023] Please refer to the attached document. Figure 4 The distributed privacy computing engine is deployed in a trusted computing cluster consisting of at least three participating nodes. The engine adopts a clear layered computing architecture, which includes a computing scheduling layer, a cryptographic protocol layer, and a hardware security layer from top to bottom.

[0024] The computation scheduling layer runs a task parser and a resource manager. The task parser receives encrypted data packets from the data hierarchical de-identification module and parses the embedded multi-party secure computation task graph. The task graph describes computational dependencies in the form of a directed acyclic graph, where nodes represent computational operations and edges represent data flows. The resource manager dynamically monitors the computational load and network latency of each participating node and allocates computational subtasks to the trusted execution environment of the target node according to the least latency priority strategy.

[0025] The cryptographic protocol layer provides implementations of two cryptographic primitives. The first is a secret-sharing protocol, employing the Shamir secret-sharing scheme, which divides each encrypted data into multiple shares and distributes them to different participants. The threshold for the secret-sharing protocol is set at 2 / 3, meaning at least two participants are required to reconstruct the secret. The second is a homomorphic encryption algorithm, employing a fully homomorphic encryption scheme, which supports direct addition and multiplication operations on the ciphertext.

[0026] The cryptographic protocol layer coordinates all participants during the computation process, and realizes the exchange and computation of encrypted data through a secure multi-party computation protocol, ensuring that the original data does not appear in plaintext form at any node.

[0027] The hardware security layer is built on domestically produced cryptographic chips, and each participating node is equipped with a physical security module.

[0028] The physical security module internally establishes an isolated trusted execution environment, providing hardware-level protection for key management and cryptographic operations at the cryptographic protocol layer. All cryptographic operations are performed within the trusted execution environment, and key materials are never exposed to ordinary operating system memory.

[0029] After completing the computation, the distributed privacy computing engine generates an encrypted output of the result. Simultaneously, the engine's internal event monitor captures key operational events of the computation. Event data includes the task identifier, a list of participating nodes, the computation algorithm type, and the task execution time window accurate to the microsecond level. This event data is formatted into event messages and sent to the cross-chain audit and evidence storage module via a high-throughput event bus.

[0030] The cross-chain audit and evidence storage module is designed with a dual-chain structure, including an audit blockchain and an evidence storage blockchain. The audit blockchain adopts a permissioned chain architecture, allowing only authorized participating nodes to act as ledger nodes, specifically for recording metadata of privacy computing tasks. Each block contains multiple transaction records, and each transaction corresponds to the event data of a privacy computing task. The transaction data structure contains five fixed fields: field 1 is the identity identifier of the task initiator; field 2 is the list of participating nodes; field 3 is the computation algorithm type; field 4 is the task start timestamp; and field 5 is the task end timestamp.

[0031] The evidence storage blockchain also adopts a permissioned blockchain architecture and has dual functions: trajectory hash storage and AI anomaly detection. On the one hand, it focuses on storing hash digests of data usage trajectories, with each trajectory record containing four elements. Element 1: Identity of the data access subject; Element 2: Access time accurate to milliseconds; Element 3: Calculation operation type, such as query calculation or statistical calculation; Element 4: The data fingerprint of the output result, calculated using the SHA-256 algorithm; On the other hand, it integrates an AI abnormal behavior monitoring model based on LSTM (Long Short-Term Memory Network). The model learns from 100,000+ historical data trajectory samples and can analyze the temporal characteristics and correlations of trajectories in real time, identifying suspicious patterns not covered by preset rules.

[0032] The real-time analysis process of the AI ​​abnormal behavior monitoring model consists of three steps: Real-time input of trajectory data: When the distributed privacy computing engine generates trajectory data, the trajectory data is first synchronized to the trajectory preprocessing unit of the evidence storage blockchain through the event bus. The unit converts the unstructured trajectory information into vector features that the model can recognize. Anomaly score calculation: The preprocessed feature vector is input into the AI ​​model in real time. The model extracts the anomaly patterns of the access time series through the temporal convolutional layer, focuses on the association conflicts between the access subject, data category, and node permission through the attention mechanism, and finally outputs the anomaly score in the range of 0-1. Warning Triggering and Tiered Handling: When the anomaly score is ≥0.85, the model automatically triggers a high-risk warning, immediately pushing a warning notification containing details of the anomaly trajectory and analysis of suspicious points to the financial security management platform of the ERP system, while suspending the subsequent data access permissions of the access subject; when the anomaly score is ≤0.7 and <0.85, a medium-risk warning is triggered, and a notification is only pushed to the administrator's workbench; when the anomaly score is <0.7, it is determined to be a normal trajectory, and only the hash digest is recorded to the evidence storage blockchain.

[0033] The two chains are interconnected via a cross-chain protocol, and a new hash anchoring mechanism has been added between AI alert logs and records from both chains. Basic hash anchoring: Every 100 blocks, the audit blockchain writes the Merkle root hash value of the latest block into a special transaction of the evidence storage blockchain, and the evidence storage blockchain simultaneously writes the latest Merkle root hash value into the audit blockchain, ensuring the temporal consistency and immutability of the audit record and the evidence storage record. AI Warning Log Anchoring: When an AI model triggers a high / medium risk warning, the system automatically generates an AI warning log and calculates the SHA-256 hash value of the log. Subsequently, the hash value is simultaneously written to the abnormal event block of the audit blockchain and the warning association block of the evidence storage blockchain through a cross-chain protocol, realizing the three-way hash binding of AI warning log, abnormal trajectory record, and audit task record. If any record is tampered with, the change in its hash value will cause the cross-chain verification to fail.

[0034] In addition, the AI ​​abnormal behavior monitoring model has a dynamic optimization mechanism of offline iteration + online fine-tuning: Offline iteration: The system automatically summarizes the AI ​​warning logs and manual review results of the month, adds the newly added abnormal trajectory samples to the training dataset, updates the model parameters by incremental training, and verifies the performance by the dual indicators of "historical abnormal trajectory recall rate ≥95%" and "false positive rate ≤3%". Only after the standards are met can the online model be replaced. Online fine-tuning: When a new abnormal pattern is detected, the administrator can trigger online fine-tuning by annotating a small number of samples (≥30). The model only updates the relevant network layers, and the fine-tuning cycle is ≤1 hour, which can quickly adapt to new risks.

[0035] After the distributed privacy computing engine outputs its results, the dynamic compliance verification module begins its work. This module is embedded in the business process controller of the enterprise resource planning system, and its verification process is clearly divided into two stages: structural verification and logical verification.

[0036] The structure verification phase first retrieves the electronic voucher accounting data standard rule base, which is stored in the form of relational database tables. It defines the specifications of all fields in the standard voucher template. The module matches the field structure of the privacy calculation output data with the template in the rule base field by field, verifying three aspects: aspect 1 is the existence check of required fields; aspect 2 is the compliance check of field format; and aspect 3 is the consistency check of data type.

[0037] During the logic verification phase, the module employs a dual mechanism of pre-defined rule verification and AI deep learning verification: on the one hand, it verifies explicit logical relationships between data based on a pre-defined business rule base; on the other hand, it uses a finance-specific AI deep learning model, based on a graph neural network (GNN) + Transformer hybrid architecture, adapted to the correlation characteristics of financial data, to identify implicit logical risks. This model learns from over 150,000 financial data samples from the past three years, including 100,000 compliant samples, 30,000 explicitly non-compliant samples, and 20,000 fuzzy samples marked with implicit risks. The marking dimensions include voucher correlation, periodic logic, and business scenario matching, automatically capturing implicit compliance risks not covered by pre-defined rules. The final verification scope includes: Explicit rule verification: Debit and credit balance verification, verified by the formula total debits = total credits; Account level consistency verification, verified by account code prefix matching; Voucher sequence continuity verification, verified by continuous voucher numbers without jumps. AI-based hidden risk verification: abnormal voucher combinations, cross-cycle logic conflicts, and mismatched business scenarios.

[0038] The specific execution process of logical verification consists of 3 steps: Explicit rule validation initialization: The module first loads the preset business rule library and performs explicit logical validation on the data field by field and voucher by voucher; if it fails, it directly marks the explicit violation and generates a rule violation report.

[0039] AI implicit risk verification trigger: The module triggers the AI ​​model only after passing the rule verification. Data feature extraction: Convert financial data into a voucher-account-amount relationship diagram and extract time-dimensional features; Risk prediction: The GNN layer captures the risk associated with vouchers, the Transformer layer captures the risk of time series data, and outputs the "confidence level of implicit risk". Risk classification: Confidence level ≥ 0.8 (highly suspicious), 0.6 ≤ confidence level < 0.8 (lowly suspicious), confidence level < 0.6 (no risk).

[0040] Verification Result Integration: The module integrates the two types of results into a unified compliance report, marking them as passed, explicit violation, and implicit risk (high / low suspicious).

[0041] The module employs a tiered processing mechanism for the logical verification results and collaborates with the cross-chain audit and evidence storage module. Handling explicit violations: Send an interrupt signal, generate a violation report, and write the report's hash value to the audit blockchain anomaly log; Handling of high-suspicious hidden risks: Generate a high-risk warning report, push it to the administrator for review, and write the review result and the report hash value into the evidence storage blockchain; Handling low-suspicion hidden risks: Generate risk warning logs, attach tags that require attention, and write the log hash value to the audit blockchain; Risk-free processing: Generate a compliance label, and write the compliance result into the normal log of the audit blockchain.

[0042] In addition, the AI ​​deep learning validation model has an offline training + online adaptive optimization mechanism: Offline batch training: AI results are summarized quarterly and manually reviewed to generate new samples for batch training. The model is updated after being verified by "recall rate ≥ 92% and false positive rate ≤ 4%". Online adaptive fine-tuning: When adding new business scenarios, the model is fine-tuned using 500 new scenario data points to adapt to new risks; Manual intervention optimization: Administrators label new risk cases, triggering incremental fine-tuning, realizing the transformation of human experience into model capabilities.

[0043] The dynamic compliance verification module also features an AI-driven feedback adjustment mechanism. When the module detects that the same type of data continuously triggers compliance warnings during multiple verifications, it first encapsulates the historical warning data into structured input features and inputs them into a pre-trained AI intelligent decision-making model for in-depth analysis.

[0044] This AI-powered intelligent decision-making model is built upon the Gradient Boosting Tree (GBDT) algorithm. The training data covers three core data categories from the past 12 months: historical compliance warning logs, historical desensitization strategy adjustment records, and a database of financial data sensitivity levels. The model training optimizes the rate of compliance warning decline for this type of data within three months of desensitization strategy adjustments. Parameters are optimized using 5-fold cross-validation (tree depth set to 8, learning rate 0.1, iterations 100). The final output includes two core results: strategy adjustment suggestions and a confidence score. The module generates a strategy update signal with a confidence level indicator only when the model outputs a confidence score ≥ 0.8. If the score is between 0.6 and 0.8, it must first be pushed to the financial security administrator's workbench for manual approval before generating a signal. If the score is < 0.6, only a warning log is recorded without triggering adjustments.

[0045] The signal is sent to the data classification and desensitization module via the dynamic policy update interface. Upon receiving the policy update signal with a confidence level indicator, the data classification and desensitization module automatically invokes its local policy adjustment engine and adjusts the processing policy for the corresponding data category according to the AI ​​model suggestions in the signal.

[0046] The updated de-identification policy is synchronized to the local policy library of all participating nodes via the cluster P2P broadcast protocol. During the synchronization process, each node performs SHA-256 hash verification on the new policy to ensure that the policy transmission has not been tampered with. After synchronization is complete, each node returns a policy effectiveness confirmation signal. The data hierarchical de-identification module also generates a policy update ledger and synchronously writes this ledger to the audit blockchain of the cross-chain audit and evidence storage module, forming a closed-loop chain of "early warning - AI analysis - policy adjustment - synchronous evidence storage" to ensure that the policy adjustment process is traceable and auditable.

[0047] Before the distributed privacy computing engine outputs its final results, it generates a ciphertext digest of the computation. The digest consists of two parts. Part 1 is the hash value of the output data, calculated using the secure hash algorithm 256. Part 2 is a set of digital signatures from the computation participants, each signing the hash value using their private key. This digest is then sent to the cross-chain audit and evidence storage module, which binds the digest to the corresponding audit record, forming a complete data lineage chain, and persistently stores it on the evidence storage blockchain.

[0048] If the dynamic compliance verification module detects a serious violation at any stage of the verification process, such as detecting tampering with voucher amounts or invalid account codes, the module will immediately send a high-priority interrupt signal to the business process controller of the enterprise resource planning system. Upon receiving the signal, the business process controller will terminate all data processing flows for the current batch, preventing abnormal data from entering any subsequent business modules. Simultaneously, the dynamic compliance verification module will generate a detailed compliance anomaly report, which includes a sample of the violating data, the triggered rule entries, and recommended handling measures. This report will be logged in a dedicated anomaly log on the audit blockchain for subsequent review by auditors.

[0049] Please refer to the appendix for the overall system's collaborative working principle. Figure 2 The data flow begins in the enterprise resource planning system's financial database, sequentially undergoing hierarchical processing by the data classification and anonymization module, encrypted computation by the distributed privacy computing engine, full-process recording by the cross-chain audit and evidence storage module, and compliance review by the dynamic compliance verification module. The final output satisfies both data sharing requirements and full compliance with security and compliance standards. All modules are loosely coupled via an event bus and message queue, ensuring high availability and scalability of the system.

[0050] This embodiment also provides another implementation of an ERP financial data sharing and security management system based on AI and privacy computing, focusing on its adaptive optimization mechanism in a heterogeneous computing environment. Please refer to the appendix. Figure 1 The overall system architecture remains unchanged, but an intelligent resource allocation algorithm is introduced into the computing scheduling layer of the distributed privacy computing engine.

[0051] At the computation scheduling layer, the task parser, in addition to parsing the computation task graph, also integrates an AI-integrated performance prediction model. This model, trained using historical execution logs and node state data, can simultaneously output the "expected execution time of the computation task" and an execution stability score (stability reflects the risk of fluctuations in task execution time). Its core employs a hybrid architecture of "XGBoost + Lightweight Multilayer Perceptron (MLP)," adapting to the linear / non-linear correlation features between computation tasks and node states. Algorithm division of labor: Gradient Boosting Tree (XGBoost): responsible for capturing the correlation between linear features (such as the number of CPU cores in a node, available memory size, static network bandwidth and other stable hardware parameters) and execution time, and iteratively optimizing the time prediction error through multiple decision trees; Lightweight MLP (2 hidden layers, ReLU activation function): responsible for capturing the correlation between nonlinear features (such as node load fluctuation, network latency standard deviation, and the matching degree between task data volume and node computing power) and execution stability, and outputting the fluctuation coefficient of task execution time; Core computational logic: Expected execution time The base time prediction value output by XGBoost Combined with the dynamic correction coefficients of the MLP output (Based on real-time node load fluctuation calculation), finally ; Execution stability score : Fluctuation coefficient output by MLP (The ratio of the standard deviation to the mean of task execution time), expressed by the formula Conversion; Model training data includes three core data categories (over 200,000 samples) from the past six months: node status data (CPU utilization fluctuations, peak / trough memory usage), task feature data (data volume, number of computational subtasks), and label data (actual execution time, execution duration fluctuation coefficient). The training and validation mechanism of the AI ​​ensemble model focuses on the dual objectives of "time prediction accuracy" and "stability prediction reliability." Loss function: Joint loss function is used. ,in This is the root mean square error (RMSE). The mean absolute error (MAE) is used. Training optimization: 5-fold cross-validation was used to split the training / test set (8:2), and grid search was used to optimize hyperparameters; Performance verification metrics: Time prediction error rate ≤ 8% on the test set, stability score accuracy ≥ 90%. When allocating computational subtasks, the resource manager uses a "time-stability weighted scoring" mechanism to select the optimal node combination. Node combination candidate pool generation: Filter nodes with "immediate load ≤ 70%" and "stability score ≥ 6 points" to generate 10-15 potential combinations; Weighted score calculation: ,in Sort by expected execution time Assess stability score; Optimal combination selection: selection The highest-scoring combination will be prioritized if the scores are the same, based on the node with a "historical task success rate ≥ 99%".

[0052] In addition, the AI-integrated performance prediction model has an incremental online update mechanism: Triggering conditions: Triggered when the time prediction error rate of 100 consecutive tasks is greater than 12% or when a new node / task type is added; Update process: Collect 10,000 new samples for incremental training, replace the online model after verification, otherwise roll back.

[0053] At the cryptographic protocol layer, a share generation algorithm based on Fast Fourier Transform is used to generate secret shared shares, reducing the computational complexity of share generation from quadratic to linear logarithmic level. At the same time, an AI adaptive compression algorithm is introduced to intelligently select the optimal lossless compression scheme according to the data characteristics of the secret shares, and accurately compress the secret shares before transmission, further reducing network bandwidth consumption.

[0054] At the hardware security layer, this embodiment extends compatibility with various domestically produced cryptographic chips. The hardware abstraction layer encapsulates the application programming interfaces of different chip manufacturers, providing a unified key management and cryptographic operation interface. The creation process of the trusted execution environment adds a dynamic measurement mechanism, performing integrity verification on the loaded code during environment initialization to ensure that the runtime environment has not been tampered with.

[0055] In this embodiment, the cross-chain audit and evidence storage module enhances data retrieval efficiency. Both the audit blockchain and the evidence storage blockchain adopt a hierarchical Merkle tree structure, storing transaction data in shards according to time windows. During queries, relevant data shards can be quickly located using time ranges, reducing the overhead of full-chain scanning. Simultaneously, blockchain nodes maintain a memory cache pool, caching the index data of the most recent 1000 blocks to accelerate the response speed of frequent queries.

[0056] In this embodiment, the dynamic compliance verification module incorporates machine learning-assisted anomaly detection. The module collects historical compliance verification results (including compliance data, explicitly non-compliant data, and suspected risk data) and trains an AI anomaly detection model that integrates isolated forests and lightweight convolutional neural networks. This model is used to accurately identify potential compliance risk patterns (including known rule violations and unknown latent risks). When new output data enters the verification process, in addition to performing fixed rule verifications, an anomaly score is calculated using the AI ​​anomaly detection model. If the anomaly score exceeds a threshold of 0.75, a risk warning log is generated for manual review, even if no specific rule is triggered.

[0057] In this embodiment, the data classification and anonymization module provides more granular configuration of anonymization strategies. System administrators can customize data classification rules through a graphical interface, supporting field matching based on regular expressions and dynamic level adjustment based on business scenarios. For example, it can be set to automatically upgrade a record to level 1 processing when the transaction amount exceeds 1 million yuan.

[0058] Configuration information for all system modules is centrally stored in the configuration management center, supporting hot updates without requiring a service restart. Each module receives configuration change notifications through a publish-subscribe mechanism, ensuring consistency across the entire system. Monitoring agents are deployed within each module to collect performance metrics and error logs in real time, displaying the system's health status through a unified monitoring dashboard.

[0059] 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.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI and privacy computing-based ERP financial data sharing security management system, characterized in that, The system deeply embeds the ERP system kernel, builds a closed-loop collaborative system containing four modules of data hierarchical desensitization, distributed privacy calculation, cross-chain audit evidence storage, and dynamic compliance verification, and realizes full-process coverage from data preprocessing, secret calculation to result audit; The operation track and data usage trace of the whole process of privacy calculation are recorded by adopting a double-chain structure, ensuring the traceability and non-tamperability in the sharing and circulation of financial data; The dynamic compliance verification module automatically verifies the output results in real time based on the electronic voucher accounting data standard; It is compatible with the national security environment through the design of domestic hardware security layer and layered calculation architecture; It can be directly embedded into the kernel of self-developed ERP products, reuse ERP domestic hardware resources, reduce deployment costs, and link with ERP electronic voucher management module and financial risk control module, so that the privacy calculation results directly trigger the voucher archiving and risk warning processes of ERP, solving the problems of financial data security sharing and compliance.

2. The system of claim 1, wherein, The data hierarchical desensitization module establishes a persistent connection with the ERP financial database through a standard database connection interface, and internally contains a data classification unit, an algorithm execution unit, and a data packaging unit. The data classification unit reads the preset classification strategy configuration file, performs sensitive semantic analysis and attribute determination on the financial data fields through an AI semantic recognition model, and divides the data levels according to the business attributes. The algorithm execution unit jointly calls the desensitization algorithm according to the data level and sensitive detail label, and the data packaging unit packs the processed data into structured data packets with metadata identifiers and transmits them to the distributed privacy calculation engine through a secure protocol.

3. The system of claim 2, wherein, The AI semantic recognition model has a dynamic learning mechanism. When new financial fields are added to the ERP system, a small amount of samples are manually labeled for incremental training to quickly adapt to new fields without retraining the full data. The data classification unit sets a confidence threshold, and according to the confidence result, it performs automatic classification, AI suggestion + manual review, or suspends automatic classification and supplements samples for retraining.

4. The system of claim 1, wherein, The distributed privacy calculation engine is deployed in a trusted computing cluster, adopts a layered calculation architecture, and includes a calculation scheduling layer, a cryptographic protocol layer, and a hardware security layer from top to bottom. The calculation scheduling layer parses the multi-party secure computation task graph in the secret data packet through a task parser, and the resource manager allocates subtasks to node trusted execution environments according to the load and delay. The cryptographic protocol layer provides cryptographic primitives to ensure that raw data does not appear in plaintext form. The hardware security layer is based on domestic cryptographic chips to provide hardware-level protection for key management and cryptographic operations.

5. The system of claim 1, wherein, The cross-chain audit evidence storage module includes an audit blockchain and an evidence blockchain, both of which adopt a permissioned chain architecture. The audit blockchain records privacy calculation task metadata, and the evidence blockchain stores the hash digest of data usage traces and integrates an AI abnormal behavior monitoring model to analyze the timing characteristics and correlation of traces in real time, identify suspicious patterns, and perform hierarchical warning processing.

6. The system of claim 5, wherein, The audit blockchain and the evidence blockchain are interconnected through a cross-chain protocol, and each writes the latest Merkle root hash value every certain number of blocks to achieve consistency and non-tamperability in recording time sequence; when the AI abnormal behavior monitoring model triggers an early warning, the system generates a warning log and calculates its hash value, which is written into both blockchains through the cross-chain protocol, realizing the three-party hash binding of the warning log, abnormal trajectory record and audit task record.

7. The system of claim 5, wherein, The AI abnormal behavior monitoring model adopts a dynamic optimization mechanism of offline iteration and online fine-tuning. During offline iteration, the system aggregates the early warning logs and manual review results every month, incrementally trains and updates the model parameters, and verifies the performance indicators. During online fine-tuning, a small number of samples labeled by administrators can trigger a quick adaptation to new risks.

8. The system of claim 1, wherein, The dynamic compliance verification module is embedded in the ERP business process controller. The verification process is divided into two stages: structure verification and logic verification. The structure verification retrieves the electronic voucher accounting data standard rule library and performs field-by-field matching verification on the output data field structure. The logic verification adopts a dual mechanism of preset rule verification and AI deep learning verification to capture explicit logical relationships and implicit compliance risks, and processes the verification results in stages, linking the cross-chain audit and evidence module to record relevant information.

9. The system of claim 8, wherein, The dynamic compliance verification module has an AI-driven feedback regulation mechanism. When the same type of data continuously triggers compliance warnings, the historical warning data is packaged as feature input into the AI intelligent decision model. The model outputs strategy adjustment suggestions and confidence scores, generates strategy update signals based on the scores, and sends them to the data classification desensitization module. The strategy is updated simultaneously in each participating node and recorded in the audit blockchain.

10. The system of claim 1, wherein, In a heterogeneous computing environment, the computing scheduling layer of the distributed privacy computing engine introduces an intelligent resource allocation algorithm, integrates an AI integrated performance prediction model to output the expected execution time and execution stability score of the task, and the resource manager selects the optimal node combination using a time-stability weighted scoring mechanism. The cryptographic protocol layer uses an optimized share generation algorithm and an AI adaptive compression algorithm. The hardware security layer expands the compatibility of domestic cryptographic chips and adds a dynamic measurement mechanism. The cross-chain audit and evidence module uses a hierarchical Merkle tree structure and memory cache pool to improve retrieval efficiency. The data classification desensitization module supports custom classification rules and dynamic level adjustment.

Images