A Real-Time Intelligent Authentication Method and System for Bank Transaction Records

By establishing a parallel detection model for image legality verification and anti-PS authentication, and an intelligent analysis model for transaction images, combined with account-level three-layer convergent large language model analysis and external data cross-validation, multiple problems in bank transaction verification are solved, and an efficient and automated multi-level verification process is achieved.

CN123118233APending Publication Date: 2026-11-17BANK OF SHANGHAI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611619275.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing bank transaction verification technologies lack pre-verification of image legitimacy and parallel detection of anti-PS, have a single verification dimension, lack transaction mode-level anti-counterfeiting, lack a layered aggregation design for large-scale transaction analysis, lack a unified framework for integrating external data sources, lack a hierarchical design for judgment results, and fail to form reusable entity information from image recognition results.

Method used

A parallel detection model for image legality verification and anti-PS authentication is established, along with a streaming image intelligent analysis model, an account-granular three-layer convergent large language model analysis and an external data cross-validation model. Parallel detection is performed using a large visual language model and image algorithms, and a three-level judgment is made by combining multi-source external data.

Benefits of technology

It enables rapid legality and anti-counterfeiting detection of image files, supports minute-level processing of multi-account transaction records, provides multi-level judgment results, improves the automation and accuracy of authentication, and supports large-scale enterprise-level transaction analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118233A_ABST
    Figure CN123118233A_ABST
Patent Text Reader

Abstract

This invention relates to the field of bank statement authentication technology, and particularly to a real-time intelligent bank statement authentication method and system. The method and system comprise three core subsystems and an end-to-end processing pipeline, achieving full automation of the entire process: image uploading and archiving, parallel detection of statement validity and anti-PS measures, intelligent analysis of statement images, aggregation of account-level and external data-based small model judgment results for LLM comprehensive analysis, and output of three-level judgment results. This invention constructs reusable statement entities, linked to internal customer entities, which can then serve as an important basis for determining the risk events of that customer in other business scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of bank statement authentication technology, and in particular to a real-time intelligent bank statement authentication method and system. Background Technology

[0002] The existing technology has the following shortcomings: 1. Lack of pre-verification of image legitimacy and parallel detection mechanism against Photoshop manipulation: Existing OCR pipeline extraction schemes allow any image file to enter the extraction pipeline without using a visual language model to verify whether the image content is indeed a bank statement before extraction, and without introducing image algorithms to detect Photoshop manipulation of the image. This results in the wrong extraction of non-pipeline images and forged images, wasting computing resources and potentially introducing misleading data.

[0003] 2. Limited verification dimensions and lack of comprehensive cross-validation: Most existing solutions only focus on one dimension, such as numerical verification (balance continuity), single-transaction anomaly detection (machine learning classification), or external data comparison. They lack the ability to conduct comprehensive verification from multiple layers, including image legality, image tampering detection, transaction patterns, and multi-source external data (business registration, credit information, spouse information, etc.).

[0004] 3. Lack of anti-counterfeiting measures at the transaction pattern level: Existing technologies rely heavily on single-transaction numerical anomaly detection and image tampering detection in terms of counterfeiting detection, but do not rely on the semantic understanding capabilities of large language models to identify batch counterfeit transactions from the macro-dimensional perspective of "transaction patterns"—such as typical counterfeiting patterns like quick in and out or large in and out.

[0005] 4. Large-scale pipeline analysis lacks hierarchical aggregation design: Existing solutions mostly adopt flat single-layer analysis and do not organize the analysis process according to the hierarchical structure of "account granularity → chunk sharding → account granularity aggregation → business granularity (multi-account) aggregation". When multiple accounts and multiple customers are involved, the results cannot be effectively isolated and progressively summarized.

[0006] 5. Lack of a unified framework for integrating external data sources: Judging compliance red lines requires comprehensive comparison of multiple external data sources such as business registration information, credit data, spouse information, regulatory blacklists, and industry access. Existing solutions are independent and have not formed a unified hierarchical verification closed loop of "small model prediction + LLM analysis + comprehensive external data verification".

[0007] 6. Lack of hierarchical design in judgment results: Existing solutions mostly output binary judgments (pass / fail), which cannot distinguish intermediate states such as "confirmed violation", "suspected fraud pending verification" and "normal operation", and cannot output the basis for analysis and judgment, which is not conducive to targeted intervention in business decision-making and manual review.

[0008] 7. The image recognition results do not form reusable entity information, and the flow entity information cannot be used as risk events for internal customer entities to provide decision support for other business scenarios.

[0009] Therefore, it is necessary to provide a real-time intelligent authentication method and system for bank statements to achieve fully automated authentication of bank statements, including "image uploading and archiving, statement legality verification + anti-PS parallel detection, intelligent analysis of statement images, aggregation of account-level and external data small model judgment results into LLM comprehensive analysis, and output of three-level judgment results". Summary of the Invention

[0010] The purpose of this invention is to provide a real-time intelligent authentication method and system for bank statements, which can achieve fully automated authentication of bank statements through "image uploading and archiving, statement legality verification + anti-PS parallel detection, intelligent parsing of statement images, aggregation of account-level and external data small model judgment results into LLM comprehensive analysis, and three-level judgment result output".

[0011] To address the problems existing in the prior art, this invention provides a real-time intelligent authentication method for bank transaction records, comprising the following steps: S1: Establish a parallel detection model for image legality verification and anti-PS authentication to generate streaming images that pass both legality verification and anti-PS detection; S2: Establish a transaction flow image intelligent analysis model. Transaction flow images that have passed legality verification and anti-PS detection enter the analysis pipeline of the analysis model. First, the PDF format image file is rendered page by page into high-resolution image frames. Then, transaction detail data is extracted from the high-resolution image frames. After analysis, a transaction flow entity with the account involved in a single transaction as the finest granularity and its related identifiers and attributes is pre-constructed and summarized and associated with the customer entity by transaction number. The structured transaction flow detail data and transaction flow summary indicators are combined with the transaction flow entity as instance information of the corresponding customer's current business and stored in the Doris database for the intelligent query interface to retrieve by entity name and entity identifier. S3: Establish a three-layer convergent big language model analysis and external data cross-validation model at the account granularity. Embed external data query, business information query, and credit data query into the Function Call tool for big language model rule execution. This allows the judgment of each rule to complete evidence collection and conclusion generation in a single big language model call. The judgment result is then verified by external data evidence through the cross-validation engine, and finally outputs a three-level judgment: pass, doubtful, or fail.

[0012] Optionally, in the real-time intelligent authentication method for bank statements, the method for generating statement images that pass legality verification and anti-PS detection is as follows: An image legality verification module is established based on a page-based differentiation strategy. A large visual language model is used, and the system prompt guides the model to perform two binary judgments on each sampled image: whether the content of the sampled image is a transaction statement from a bank or payment platform, and whether the text and transaction data in the sampled image are clear and identifiable. If both judgments are correct, the sampled image passes the verification. A parallel image algorithm-based detection module for preventing Photoshop manipulation (PS) attacks is established. The detection methods are as follows: First dimension, error level analysis: detecting differences in compression error levels across different regions of the image to identify abnormal regions introduced through pasting or splicing; Second dimension, noise pattern analysis: analyzing the consistency of image noise distribution to detect inconsistent noise regions introduced by local editing; Third dimension, metadata integrity verification: parsing software editing records and modification timestamps in the image's EXIF / IPTC metadata to detect the presence of editing traces from image processing software; Fourth dimension, tampering edge feature detection: using a deep learning convolutional network to analyze pixel-level tampering edge features to identify splicing and smearing boundary traces. The anti-PS detection module weights and fuses the detection results from the four dimensions, outputting a comprehensive tampering confidence score. When the score exceeds a preset threshold, it is marked as suspected tampering. A parallel linkage decision-making mechanism is established, in which legality verification and anti-PS detection are executed in parallel, reducing the end-to-end latency from the serial T1+T2 to (max(T1,T2)) under the parallel scheme. Here, T1 is the time consumed by legality verification and T2 is the time consumed by anti-PS detection. After each is completed independently, the results are merged by the linkage decision-maker. The merged decision is as follows: if both legality verification and anti-PS detection pass, the image file enters the parsing stage in S2; otherwise, the overall business process of the account corresponding to the current image file is immediately interrupted, and an interruption reason and detailed evidence report are generated and returned to the business front end. The subsequent processing pipeline of the account corresponding to the current image file will no longer be executed.

[0013] Optionally, in the real-time intelligent authentication method for bank statements, the intelligent analysis model for bank statement images includes the following modules: Dual-path parallel multimodal visual language large model extraction pipeline module: Two visual language models are concurrently invoked using asyncio coroutines. The first model extracts transaction details: guided by a prompt, it sequentially identifies header rows, data rows, and total rows from the image, outputting structured JSON information. The second model extracts metadata: guided by another prompt, it extracts structured information from non-table areas of the image, outputting a JSON object. Both models share the same batch of image input but use different prompt templates, preventing mutual blocking and maximizing parallel throughput. Field mapping and balance continuity verification module: Field mapping engine: Maps individual column names from various banks to standard fields, outputting standardized comma-separated values; Balance continuity validator: Verifies balance item by item as follows: i =balance i-1 ±amount i Among them, balance i This is the account balance. i-1 This is the previous account balance, ±amount i For transactions involving lending or borrowing; Module for storing parsed results in the database: After the dual-path visual language model is extracted, the structured transaction details data after field mapping and balance verification is written into the Doris database. The structured transaction details data includes a transaction details table and an account metadata table. The transaction summary metrics returned by the external data platform are also written into the Doris metrics table; Intelligent data query interface: By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity, along with their associated identifiers and attributes, and summarizing and associating them with customer entities by transaction number; the structured transaction details data and transaction summary indicators are combined with the transaction entities as instance information of the corresponding customer's current business and stored in the Doris database, so that the intelligent query interface can retrieve them by entity name and entity identifier. Each module evolves and expands independently.

[0014] Optionally, in the real-time intelligent authentication method for bank statements, the account-granular three-layer convergent large language model analysis and external data cross-validation model includes the following modules: First layer: Chunk-level rule determination, including Function Call external data query; For each account's transaction data, the transaction records are divided into K chunks according to the preset chunk size, where K is a natural number. Each chunk independently executes the authentication and falsification rules as follows: the rules are constructed in combination with the chunked Cartesian product task matrix, the dependencies between the rules are resolved by the directed acyclic graph topology sorter to determine the execution order, and semaphores are used to control the batch parallel large language model calls. In each leaf task, the large language model performs semantic analysis on the transaction records within the chunk according to the corresponding rules, and combines the external data query results returned by the Function Call to output the judgment result, reasoning basis, external data evidence reference and confidence score of the current chunk; Second layer: Account granularity summary; Once all rule judgment results for all chunks under an account are completed, account-level aggregation is initiated, and a user operation authenticity profile is constructed. The aggregation logic is as follows: Authentication Rules: Positive rules. Hitting a rule increases the weight of the real business profile. Based on the judgment results of all chunks in the current account under each authentication rule, if any chunk hits an authentication rule, the weight of the real profile at the granular level of the account will be increased. False detection rule: Negative rule. If a false detection rule is hit, the weight of the false business profile will be increased. Based on the results of the false detection rules of all chunks, if any chunk hits the false detection rule, the weight of the false business profile of that account will be increased. Red Line Rule: The bottom line rule. If any chunk hits the red line rule, the current account will be marked as failing. During the aggregation process, the results of the small model's authenticity verification, fake product detection, and red line rule judgment serve as the reference baseline for the large language model's judgment. Combining the rule results of the small and large models, if the red line rule is not hit, the result is determined according to the rule weight: if the user profile is more inclined towards fraudulent operations, the account's granularity result is questionable; if the user profile is more inclined towards genuine operations, the account's granularity result is approved. Third layer: Business granularity summary; Once all accounts associated with a transaction have completed account-level aggregation, transaction-level aggregation is initiated. The aggregation logic follows a three-level decision rule: Failure: If any related account triggers the red line rule, the entire current business is deemed to fail. Questionable: If no account violates the red line rules, but any related account is identified as having a user profile weight biased towards fraudulent operations, then the current business as a whole is judged to be questionable. If the user profile weights of all associated accounts are biased towards genuine business operations and no red line rules are triggered, then the current business as a whole is judged to be approved.

[0015] Optionally, in the real-time intelligent authentication method for bank statements, the large language model call supports dual-mode switching, namely thinking mode and non-thinking mode. Thinking mode is used for streaming interaction scenarios, and non-thinking mode is used for synchronous batch execution scenarios.

[0016] This invention also provides a real-time intelligent authentication system for bank statements, comprising: A parallel detection model for image legitimacy verification and anti-PS authentication is configured to generate streaming images that pass both legitimacy verification and anti-PS detection. The intelligent parsing model for transaction flow images is configured to process images that have passed legality verification and anti-PS detection. First, the PDF image file is rendered page by page into high-resolution image frames. Then, transaction details are extracted from these high-resolution image frames. After parsing, a pre-constructed transaction entity model, with the account involved in each transaction as the finest granularity, along with its associated identifiers and attributes, is built and aggregated by transaction number and associated with the customer entity. The structured transaction details and aggregated transaction indicators are then combined with the transaction entity as instance information for the corresponding customer's current business and stored in the Doris database for the intelligent query interface to retrieve by entity name and identifier. The system integrates a three-layered, account-granular large language model analysis with an external data cross-validation model. It is configured to embed external data queries, business information queries, and credit data queries into a Function Call tool for executing large language model rules. This allows each rule's judgment to complete evidence collection and conclusion generation in a single large language model call. The cross-validation engine then verifies the judgment results with external data evidence, ultimately outputting a three-level judgment: pass, questionable, or fail.

[0017] Optionally, in the real-time intelligent bank statement authentication system, the parallel detection model for image legality verification and anti-PS authentication generates transaction images that pass both legality verification and anti-PS detection in the following way: An image legality verification module is established based on a page-based differentiation strategy. A large visual language model is used, and the system prompt guides the model to perform two binary judgments on each sampled image: whether the content of the sampled image is a transaction statement from a bank or payment platform, and whether the text and transaction data in the sampled image are clear and identifiable. If both judgments are correct, the sampled image passes the verification. A parallel image algorithm-based detection module for preventing Photoshop manipulation (PS) attacks is established. The detection methods are as follows: First dimension, error level analysis: detecting differences in compression error levels across different regions of the image to identify abnormal regions introduced through pasting or splicing; Second dimension, noise pattern analysis: analyzing the consistency of image noise distribution to detect inconsistent noise regions introduced by local editing; Third dimension, metadata integrity verification: parsing software editing records and modification timestamps in the image's EXIF / IPTC metadata to detect the presence of editing traces from image processing software; Fourth dimension, tampering edge feature detection: using a deep learning convolutional network to analyze pixel-level tampering edge features to identify splicing and smearing boundary traces. The anti-PS detection module weights and fuses the detection results from the four dimensions, outputting a comprehensive tampering confidence score. When the score exceeds a preset threshold, it is marked as suspected tampering. A parallel linkage decision-making mechanism is established, in which legality verification and anti-PS detection are executed in parallel, reducing the end-to-end latency from the serial T1+T2 to (max(T1,T2)) under the parallel scheme. Here, T1 is the time consumed by legality verification and T2 is the time consumed by anti-PS detection. After each is completed independently, the results are merged by the linkage decision-maker. The merged decision is as follows: if both legality verification and anti-PS detection pass, the image file enters the parsing stage in S2; otherwise, the overall business process of the account corresponding to the current image file is immediately interrupted, and an interruption reason and detailed evidence report are generated and returned to the business front end. The subsequent processing pipeline of the account corresponding to the current image file will no longer be executed.

[0018] Optionally, in the real-time intelligent bank statement authentication system, the intelligent analysis model for bank statement images includes the following modules: Dual-path parallel multimodal visual language large model extraction pipeline module: Two visual language models are concurrently invoked using asyncio coroutines. The first model extracts transaction details: guided by a prompt, it sequentially identifies header rows, data rows, and total rows from the image, outputting structured JSON information. The second model extracts metadata: guided by another prompt, it extracts structured information from non-table areas of the image, outputting a JSON object. Both models share the same batch of image input but use different prompt templates, preventing mutual blocking and maximizing parallel throughput. Field mapping and balance continuity verification module: Field mapping engine: Maps individual column names from various banks to standard fields, outputting standardized comma-separated values; Balance continuity validator: Verifies balance item by item as follows: i =balance i-1 ±amount i Among them, balance i This is the account balance. i-1 This is the previous account balance, ±amount i For transactions involving lending or borrowing; Module for storing parsed results in the database: After the dual-path visual language model is extracted, the structured transaction details data after field mapping and balance verification is written into the Doris database. The structured transaction details data includes a transaction details table and an account metadata table. The transaction summary metrics returned by the external data platform are also written into the Doris metrics table; Intelligent data query interface: By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity, along with their associated identifiers and attributes, and summarizing and associating them with customer entities by transaction number; the structured transaction details data and transaction summary indicators are combined with the transaction entities as instance information of the corresponding customer's current business and stored in the Doris database, so that the intelligent query interface can retrieve them by entity name and entity identifier. Each module evolves and expands independently.

[0019] Optionally, in the real-time intelligent verification system for bank statements, The account-granular three-layer convergent large language model analysis and external data cross-validation model includes the following modules: First layer: Chunk-level rule determination, including Function Call external data query; For each account's transaction data, the transaction records are divided into K chunks according to the preset chunk size, where K is a natural number. Each chunk independently executes the authentication and falsification rules as follows: the rules are constructed in combination with the chunked Cartesian product task matrix, the dependencies between the rules are resolved by the directed acyclic graph topology sorter to determine the execution order, and semaphores are used to control the batch parallel large language model calls. In each leaf task, the large language model performs semantic analysis on the transaction records within the chunk according to the corresponding rules, and combines the external data query results returned by the Function Call to output the judgment result, reasoning basis, external data evidence reference and confidence score of the current chunk; Second layer: Account granularity summary; Once all rule judgment results for all chunks under an account are completed, account-level aggregation is initiated, and a user operation authenticity profile is constructed. The aggregation logic is as follows: Authentication Rules: Positive rules. Hitting a rule increases the weight of the real business profile. Based on the judgment results of all chunks in the current account under each authentication rule, if any chunk hits an authentication rule, the weight of the real profile at the granular level of the account will be increased. False detection rule: Negative rule. If a false detection rule is hit, the weight of the false business profile will be increased. Based on the results of the false detection rules of all chunks, if any chunk hits the false detection rule, the weight of the false business profile of that account will be increased. Red Line Rule: The bottom line rule. If any chunk hits the red line rule, the current account will be marked as failing. During the aggregation process, the results of the small model's authenticity verification, fake product detection, and red line rule judgment serve as the reference baseline for the large language model's judgment. Combining the rule results of the small and large models, if the red line rule is not hit, the result is determined according to the rule weight: if the user profile is more inclined towards fraudulent operations, the account's granularity result is questionable; if the user profile is more inclined towards genuine operations, the account's granularity result is approved. Third layer: Business granularity summary; Once all accounts associated with a transaction have completed account-level aggregation, transaction-level aggregation is initiated. The aggregation logic follows a three-level decision rule: Failure: If any related account triggers the red line rule, the entire current business is deemed to fail. Questionable: If no account violates the red line rules, but any related account is identified as having a user profile weight biased towards fraudulent operations, then the current business as a whole is judged to be questionable. If the user profile weights of all associated accounts are biased towards genuine business operations and no red line rules are triggered, then the current business as a whole is judged to be approved.

[0020] Optionally, in the real-time intelligent authentication system for bank transactions, the large language model call supports dual-mode switching: a thinking mode and a non-thinking mode. The thinking mode is used for streaming interaction scenarios, while the non-thinking mode is used for synchronous batch execution scenarios.

[0021] Compared with the prior art, the present invention has the following advantages: (1) Parallel pre-filtering for image validity verification and anti-PS detection: Before extraction, two parallel checkpoints are used: the validity verification VLM and the image algorithm anti-PS detection. The "dual-path parallel + one-vote veto" mechanism is adopted to quickly intercept non-stream images, blurred images and PS-tampered images, reducing the latency from serial T1+T2 to (max(T1,T2)) to prevent invalid data from entering the subsequent pipeline. The Fail-open fault tolerance strategy ensures that the failure of a single model call does not block the overall process.

[0022] (2) End-to-end automation: From image uploading and archiving, legality verification and anti-PS parallel detection, extraction to final authentication and three-level judgment, the entire process is completed automatically in a unified pipeline without manual intervention. It supports enterprise-level large-scale multi-account transaction processing (thousands of records) at the minute level.

[0023] (3) Multimodal high-precision extraction: The dual-path parallel VLM is used to extract transaction details and metadata simultaneously, and the identification deviation is automatically corrected through balance continuity verification.

[0024] (4) Transaction pattern-level anti-counterfeiting authentication: By using LLM to perform two-way analysis of transaction flow for authenticity and counterfeiting rules, the characteristics of counterfeit flow (quick in and quick out, large in and large out, etc.) are identified at the transaction pattern level, breaking through the limitation of traditional solutions that only perform numerical verification. The differentiated Prompt design realizes a two-dimensional evaluation of transaction authenticity.

[0025] (5) Account-level three-layer aggregation architecture: Taking the account as the smallest analysis granularity, a three-layer progressive aggregation architecture of "Chunk-level rule judgment, account-level aggregation, and business-level (multi-account) aggregation" is designed to effectively isolate the judgment differences between multiple accounts and clearly and progressively output the final judgment result.

[0026] (6) LLM rules embed external data query (Function Call): External data, business information, credit data, spouse information and other external data queries are used as the Function Call embedded tool for LLM rule execution, so that each rule can complete evidence collection and conclusion generation in a single LLM call, avoiding the traditional two-step mode of "analyze first and then query".

[0027] (7) Three-level judgment output mechanism: The three-level judgment logic driven by the unique red line and the false judgment: the red line rule trigger is "not passed", the false judgment rule trigger is "doubtful", and all passes are "passed"; providing clear and operable judgment results for business decisions.

[0028] (8) Massive parallel processing: Through a three-layer parallel mechanism of rule-sharded Cartesian product task matrix + DAG topology sorting + semaphore concurrent control, it supports simultaneous analysis of thousands of transactions and dozens of rules from multiple accounts.

[0029] (9) Explainable judgment results: Each rule judgment is accompanied by the reasoning basis generated by LLM, the small model prediction reference and the authoritative evidence reference of external data sources. The anti-PS rule provides additional confidence score to meet the financial regulatory requirements for auditability and explainability.

[0030] (10) High availability and fault tolerance design: The legality verification stage adopts the fail-open strategy and exponential backoff retry mechanism; LLM supports dual-mode switching between thinking mode and non-thinking mode, taking into account both real-time interaction and batch processing; the image platform allocates a unique file link to ensure traceability of the entire link.

[0031] (11) Construct reusable transaction entities and associate them with customer entities within the bank. These entities can then serve as important evidence for risk events in other business scenarios. Attached Figure Description

[0032] Figure 1 This is a flowchart of a method provided in an embodiment of the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. The advantages and features of the present invention will become clearer from the following description. It should be noted that the drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0034] In the following, if the methods described herein include a series of steps, the order of these steps presented herein is not necessarily the only order in which these steps can be performed, and some of the steps described may be omitted and / or some other steps not described herein may be added to the method.

[0035] To address the problems existing in the prior art, this invention provides a real-time intelligent authentication method for bank transaction records. Before structured data extraction, it employs two parallel pre-screening mechanisms: image algorithm-based anti-PS detection and Visual Language Model (VLM) legitimacy verification, to filter out non-compliant user transaction records. At the transaction data level, it achieves anti-counterfeiting authentication through LLM-driven transaction pattern analysis. Furthermore, it constructs a three-layer converged analysis architecture of "chunk → account → business" with the account as the smallest analytical granularity. This architecture integrates multi-source external data, credit data, spouse information, and authentication results from several small model rules for cross-validation of red-line rules, counterfeit rules, and authentication rules. Finally, it outputs a three-level judgment result: pass / questionable / fail.

[0036] like Figure 1 As shown in Table 1, the steps of this method are as follows: Table 1:

[0037] Specifically, S1: Establish a parallel detection model for image legality verification and anti-PS authentication to generate streaming images that pass both legality verification and anti-PS detection; (1) Scene description: In actual business operations such as bank loan approval, the format and content of bank statement files submitted by customers vary widely. The following problems may exist: submission of scanned copies of ID cards, bank account information pages, receipts, or other non-transactional details; blurry screenshots with unrecognizable text; submission of fake bank statement images altered or spliced ​​using image processing software such as Photoshop; and submission of bank statement files with a time span not meeting the minimum six-month requirement. If these problematic images directly enter the subsequent extraction and analysis pipeline, it will not only waste computational resources but may also lead to incorrect authentication conclusions. Therefore, this model sets up two parallel checkpoints at the entry point before extraction: image legality verification (verifying whether the content is a genuine bank statement and is compliant, clear, and identifiable) and anti-Photoshop forgery detection (detecting whether the image has been tampered with or spliced ​​using image algorithms). These two checkpoints execute in parallel and jointly determine whether to allow the image to enter the subsequent analysis pipeline.

[0038] (2) Technical introduction: Dataset / Input Description: The input data received by this model is the image files to be authenticated uploaded by the customer through the image platform. The image files to be authenticated are shown in Table 2. Table 2:

[0039] Specifically, Establish an image upload and archiving module. The front end uploads sequential image files at the account level. The image platform assigns a unique file link and archives it, outputting image file references with unique links.

[0040] S11: Establish an image validity verification module based on a page-number differentiation strategy. Single-page file: The page will be directly sent to the dedicated VLM for validity verification. Multi-page PDF files: Employ an intelligent sampling strategy—extracting only the first page (usually account summary information), the last page (transaction history / summary information), and the middle pages (representing the main transaction details) evenly distributed according to the total number of pages. The default maximum number of samples is 15 pages. Each sampled image is converted into an image frame using the PyMuPDF rendering engine at a preset DPI (default 200).

[0041] Furthermore, using the Visual Language Model (VLM), the System Prompt guides the model to perform two binary judgments on each sampled image (or single page): determining whether the content of the sampled image is a transaction statement from a bank or payment platform, and determining whether the text and transaction data in the sampled image are clear and identifiable. If both judgments are correct, the sampled image passes the verification. Preferably, a veto + early exit control mechanism is adopted: each image is verified in the sampling order. Once an image is determined to be "no" (not sequential or blurry), the verification of subsequent images and the entire extraction process are immediately terminated, the file is skipped and the next task is processed to avoid invalid calculations.

[0042] Fail-open fault tolerance mechanism: To prevent legitimate images from being mistakenly blocked due to occasional failures of the VLM model service (network timeout, service overload, etc.), this step adopts an exponential backoff retry strategy (maximum of 3 retries, with intervals of 1s, 2s, and 4s respectively) and a fail-open fault tolerance strategy—blocking is only performed when the VLM explicitly returns a negative judgment. If all retries fail (timeout / abnormality), the image is automatically allowed to proceed to the next stage, and alarm logs are recorded for operation and maintenance to investigate.

[0043] S12: A parallel detection module based on image algorithms is established to prevent Photoshop manipulation and forgery detection. This module runs in parallel with the legitimacy verification process. It uses a dedicated image algorithm to detect tampering in streaming image files, determining whether the images have been edited, spliced, blurred, or artificially generated using image processing tools such as Photoshop. The detection method is as follows: The first dimension, Error Level Analysis (ELA): detects the differences in compression error levels in different regions of an image and identifies abnormal regions introduced by pasting and splicing. The second dimension is noise pattern analysis: by analyzing the consistency of image noise distribution, it detects regions of inconsistency in noise introduced by local editing. The third dimension is metadata integrity verification: parsing the software editing records and modification timestamps in the EXIF / IPTC metadata of the image to detect whether there are any editing traces from image processing software; The fourth dimension is the detection of tampering edge features: using deep learning convolutional networks to analyze the pixel-level tampering edge features of images and identify the boundary traces of splicing and smearing. The anti-PS detection module performs weighted fusion of the detection results from four dimensions and outputs a comprehensive tampering confidence score (0-1). When the score exceeds a preset threshold, it is marked as suspected tampering. S13: Establish a parallel linkage decision-making mechanism, where legality verification and anti-PS detection are executed in parallel, reducing the end-to-end latency from the serial T1+T2 to (max(T1,T2)) under the parallel scheme. Here, T1 is the time consumed by legality verification and T2 is the time consumed by anti-PS detection. After each is completed independently, the results are merged by the linkage decision-maker. The merged decision is as follows: if both legality verification and anti-PS detection pass, the image file enters the parsing stage in S2; otherwise, the overall business process of the account corresponding to the current image file is immediately interrupted, and an interruption reason and detailed evidence report are generated and returned to the business front end. The subsequent processing pipeline of the account corresponding to the current image file will no longer be executed.

[0044] S2: Establish a transaction flow image intelligent analysis model. Transaction flow images that have passed legality verification and anti-PS detection enter the analysis pipeline of the model. First, the PDF format image file is rendered page by page into high-resolution image frames (where the high resolution range is 180–220 DPI, with a good balance between clarity and performance). Then, transaction details (debit amount, credit amount, balance, transaction summary, counterparty account, counterparty name, etc.) and unstructured metadata (account name, account number, currency, transaction start and end dates, etc.) are extracted from the high-resolution image frames. After analysis, transaction flow entities with the accounts involved in a single transaction as the finest granularity and their related identifiers and attributes are pre-constructed and summarized and associated with customer entities according to the transaction number. The structured transaction flow details and transaction flow summary indicators are combined with the transaction flow entities as instance information of the corresponding customer's current business and stored in the Doris database for the intelligent query interface to retrieve by entity name and entity identifier. Further details regarding the dataset / input: Image files that pass both S1 verification and anti-PS detection are listed in Table 3. Table 3:

[0045] The analysis is as follows: (1) Dual-path parallel multimodal visual language large model extraction pipeline module: The two visual language models are invoked concurrently using asyncio coroutines, as follows: The first approach involves extracting transaction details using a visual language model: Guided by a prompt, the visual language model sequentially identifies the header row, data row, and total row from the image, outputting structured JSON information. The second approach involves extracting metadata from a large visual language model: another set of prompts guides the large visual language model to extract structured information (account name, account number, currency, transaction start and end dates, bank name, etc.) from non-table areas in the image, and outputs a JSON object. Two visual language large models call the same batch of image inputs but use different Prompt templates, without blocking each other, maximizing parallel throughput; (2) Field mapping and balance continuity verification module: Field mapping engine: Maps individual column names from various banks ("posting date", "Transaction Date", etc.) to standard fields, outputting standardized comma-separated values; Balance continuity validator: Verification is performed item by item as follows: balance i =balance i-1 ±amount i Among them, balancei This is the account balance. i-1 This is the previous account balance, ±amount i To determine the transaction amount carrying the lending / borrowing direction, derive the theoretical amount sign (+amount / -amount) and compare it with the current lending / borrowing identifier. If they do not match, invert the lending / borrowing identifier and check again; that is, if the balance is achieved... i Not equal to balance i-1 +amount i At that time, check the balance. i Does it equal balance? i-1 -amount i If balance i It is not the same as balance. i-1 -amount i At the same time, the reversed amount data is retained and stored in the database, and a case of discontinuous balance is marked as the basis for judgment rules in subsequent analysis stages.

[0046] (3) Module for storing parsed results in the database: After the dual-path visual language model is extracted, the structured transaction details data (standardized field records for each transaction) after field mapping and balance verification is written into the Doris database. The structured transaction details data includes a transaction details table and an account metadata table. The transaction summary metrics returned by the external data platform (total number of transactions, total transaction amount, percentage of integer transactions, number of counterparties, etc.) are also written into the Doris metric table; Doris's high-performance columnar storage and MPP parallel query capabilities provide basic support for the rapid retrieval of massive transaction data in downstream analysis modules.

[0047] (4) Intelligent Questioning Interface: By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity, along with their associated identifiers and attributes, and summarizing and associating them with customer entities by transaction number; the structured transaction details data and transaction summary indicators are combined with the transaction entities as instance information of the corresponding customer's current business and stored in the Doris database, so that the intelligent query interface can retrieve them by entity name and entity identifier. After the parsed results are stored in the database, the downstream analysis module retrieves the real-time parsed structured transaction details and indicator table data from the Doris database on demand via the intelligent query interface. The analysis scheduling engine determines the required data conditions (account dimension, transaction time range, field combination, statistical caliber, etc.) according to rules and initiates a data retrieval request. The interface returns a standardized result set, and the analysis module does not need to be aware of the underlying storage and query implementation details. This interface decouples the parsing module from the analysis module, allowing each module to evolve and expand independently.

[0048] The application results are as follows: After introducing the dual-path parallel VLM extraction and Doris database parsing pipeline, the dual-path VLM extraction of 1200 transactions can save about 40%-50% of the extraction time; structured transaction details and indicator data are uniformly stored in Doris and can be obtained on demand through the intelligent query interface. The parsing module and the analysis module are completely decoupled, and each module can evolve and expand independently.

[0049] S3: After the parsing phase, the structured transaction details and summary indicators have been entered into the Doris database. Since a single transaction may involve multiple accounts (such as the business owner's personal account, spouse's account, and company account), the transaction records of each account need to be independently verified and analyzed before being aggregated upwards to form a comprehensive judgment across business dimensions. Existing solutions mostly employ a flat, single-layer analysis, which cannot effectively isolate the judgment differences between multiple accounts. Furthermore, compliance red-line rules (such as counterparty business irregularities, involvement in negative lists or regulatory penalties, and credit irregularities) suffer from insufficient timeliness if relying solely on LLM "internal knowledge."

[0050] Therefore, it is necessary to establish a three-layer convergent big language model analysis and external data cross-validation model at the account granularity. External data query, business information query, and credit data query are embedded into the FunctionCall tool for big language model rule execution. This allows the determination of each rule to complete evidence collection and conclusion generation in a single big language model call. The judgment results are then verified by external data evidence through a cross-validation engine, and finally outputs a three-level judgment: pass, doubtful, or fail.

[0051] Further details regarding the dataset and input: During the analysis phase, the structured transaction details and metrics data are obtained from Doris via the intelligent query interface, and the rule list is obtained via the rule engine interface. External data sources are queried via Function Call. See Table 4 for a description of the dataset and inputs. Table 4:

[0052] Furthermore, the account-granular three-layer convergent large language model analysis and external data cross-validation model includes the following modules: First layer: Chunk-level rule determination, including Function Call external data query; For each account's transaction data, the transaction records are divided into K chunks according to a preset chunk size (e.g., preset chunk size S = 500 records / chunk), where K is a natural number. Each chunk independently executes the authentication and deception rules as follows: Construct a rule-based Cartesian product task matrix (M authentication rules × K chunks = M × K independent LLM analysis tasks), resolve the dependencies between rules using a directed acyclic graph topology sorter to determine the execution order, and use semaphores to control the batch parallel large language model invocation; External data query during rule execution (Function Call): When performing LLM semantic analysis, each rule can call an external data interface through the Function Call mechanism to obtain authoritative information as a basis for judgment, specifically including: External data query: Search the enterprise registration database (operating status, registered capital, business scope), regulatory penalty record database (administrative penalties, dishonest persons subject to enforcement) and industry access negative list by the name of the counterparty; Business registration information inquiry: Verify whether the business registration status of the counterparty is normal and valid; Credit data inquiry: Obtain credit records, debt information, and credit status of customers and related parties; Beneficial owner lookup: Identify the controlling shareholders and actual controllers of counterparties to uncover hidden related-party transactions; Spouse-related information inquiry: Obtain the spouse's identity and related account information to identify hidden connections in fund transfers; Small model authentication prediction results acquisition: The prediction results of calling the dedicated small model service are used as the reference baseline for LLM determination.

[0053] All external data queries mentioned above are invoked as embedded tools within the LLM rule execution. The query results are directly integrated into the Prompt context of the rule, and the LLM then provides a judgment conclusion by comprehensively considering transaction data and authoritative external data. This design avoids the "two-step" model where external data queries are independent of rule execution, allowing evidence collection and conclusion generation for each rule to be completed in a single LLM call. Preferably, a piecewise Cartesian product task matrix DAG topological sorting method is used to resolve rule dependencies.

[0054] In each leaf task, the LLM performs semantic analysis on the transaction records within the batch chunk according to the corresponding rules (the authenticity rule uses a "forward evidence search" type prompt, and the false detection rule uses a "reverse anomaly identification" type prompt). Combined with the external data query results returned by the Function Call, it outputs the chunk's judgment result (pass / questionable / fail), reasoning basis, external data evidence citation, and confidence score. The false detection rule additionally outputs a confidence score (0-1) and a specific list of suspicious transactions.

[0055] LLM calls support dual-mode switching: thinking mode (enable_thinking=True, for streaming interactive scenarios) and non-thinking mode (enable_thinking=False, for synchronous batch execution scenarios).

[0056] Second layer: Account granularity summary; Once all rule judgment results for all chunks under an account are completed, account-level aggregation is initiated, and a user operation authenticity profile is constructed. The aggregation logic is as follows: Authentication Rules: Positive rules. Hitting a rule increases the weight of the real business profile. Based on the judgment results of all chunks in the current account under each authentication rule, if any chunk hits an authentication rule, the weight of the real profile at the granular level of the account will be increased. False detection rule: Negative rule. If a false detection rule is hit, the weight of the false business profile will be increased. Based on the results of the false detection rules of all chunks, if any chunk hits the false detection rule, the weight of the false business profile of that account will be increased. Red Line Rule: The bottom line rule. If any chunk hits the red line rule, the current account will be marked as failing. During the aggregation process, the results of the small model's authenticity verification, fake product detection, and red line rule judgment serve as the reference baseline for the large language model's judgment. Combining the rule results of the small and large models, if the red line rule is not hit, the result is determined according to the rule weight: if the user profile is more inclined towards fraudulent operations, the account's granularity result is questionable; if the user profile is more inclined towards genuine operations, the account's granularity result is approved. The third layer: summarizing business granularity to form a three-level judgment result; Once all accounts associated with a transaction have completed account-level aggregation, transaction-level aggregation is initiated. The aggregation logic follows a three-level decision rule: Failure: If any related account triggers the red line rule, the entire current business is deemed to fail. Questionable: If no account violates the red line rules, but any related account is identified as having a user profile weight biased towards fraudulent operations, then the current business as a whole is judged to be questionable. If the user profile weights of all associated accounts are biased towards genuine business operations and no red line rules are triggered, then the current business as a whole is judged to be approved.

[0057] LLM uses the third-layer business granularity summary as the basis for analysis and stores it in Doris database for auditing and traceability.

[0058] Finally, a results output layer is established to output the judgment conclusions and evidence reports back to the business front end.

[0059] Application effect: Taking a corporate loan approval scenario (2 linked accounts, totaling 1200 transactions) as an example: After the transaction log is parsed, it is split by account: Account A: 800 records / 2 chunks (500 records per chunk), Account B: 400 records / 1 chunk; Each chunk of each account executes rule determination independently, and external data is queried synchronously via Function Call during rule execution; with 8 concurrent processes, all chunk-level analysis is completed in approximately 15 seconds. After the account granularity is summarized, the business granularity is summarized and all account results are combined to determine the final business result. The three-level judgment logic clearly defines pass / questionable / fail, meeting business decision-making needs.

[0060] Key indicators for detecting false positives, such as: Integer transaction percentage detection: ; Among them, D score The score (indicator value) is the percentage of integer transactions, N. total To count the total number of transactions within the specified interval, amount i The transaction amount of the i-th transaction, where 1 [...] represents an indicator function; the value is 1 if the condition within the parentheses is true, and 0 if it is false; amount i mod100 represents taking the i-th amount modulo 100, which checks if the amount is a multiple of 100 (modulo = 0 means the amount is a multiple of 100, e.g., 1000, 25000); when D score When the value is greater than 0.7, a "round in and round out" rule alarm is triggered. That is, if the percentage of transactions with a value of 100 yuan or more in the total number of transactions exceeds 70% in the statistical period, an alarm will be triggered.

[0061] Fast in / out detection: For large transactions (amount > θ) large θ large A threshold value for large transactions (i.e., a preset critical value; only transactions exceeding this value are considered large transactions, e.g., 200,000) is calculated. Δt = t out -tin , where t in The time of the transaction (the time when funds are transferred into the account), t out The time of the transfer-out transaction is Δt, where Δt is the time difference between the transfer-in and transfer-out transactions. When Δt ≤ 24 hours and When marked as fast in and fast out, where amount in The amount of funds transferred into the account (incoming funds), amount out This refers to the amount of funds transferred out of the account (outflow amount). A large sum of money is transferred in and then withdrawn within 24 hours; and the difference between the outflow and inflow amounts is very small (less than 10% of the inflow amount), which is considered a quick inflow and quick outflow.

[0062] This invention also provides a real-time intelligent authentication system for bank statements, comprising: 1. A parallel detection model for image legitimacy verification and anti-PS authentication is configured to generate streaming images that pass both legitimacy verification and anti-PS detection. (1) Scene description: In actual business operations such as bank loan approval, the format and content of bank statement files submitted by customers vary widely. The following problems may exist: submission of scanned copies of ID cards, bank account information pages, receipts, or other non-transactional details; blurry screenshots with unrecognizable text; submission of fake bank statement images altered or spliced ​​using image processing software such as Photoshop; and submission of bank statement files with a time span not meeting the minimum six-month requirement. If these problematic images directly enter the subsequent extraction and analysis pipeline, it will not only waste computational resources but may also lead to incorrect authentication conclusions. Therefore, this model sets up two parallel checkpoints at the entry point before extraction: image legality verification (verifying whether the content is a genuine bank statement and is compliant, clear, and identifiable) and anti-Photoshop forgery detection (detecting whether the image has been tampered with or spliced ​​using image algorithms). These two checkpoints execute in parallel and jointly determine whether to allow the image to enter the subsequent analysis pipeline.

[0063] (2) Technical introduction: Dataset / Input Description: The input data received by this model is the image files to be authenticated uploaded by the customer through the image platform. The image files to be authenticated are shown in Table 5. Table 5:

[0064] Specifically, The image upload and archiving module is configured to allow the front end to upload sequential image files at the account level. The image platform assigns a unique file link and archives it, outputting image file references with unique links.

[0065] 1) Establish an image legality verification module based on a page-number differentiation strategy. Single-page file: The page will be directly sent to the dedicated VLM for validity verification. Multi-page PDF files: Employ an intelligent sampling strategy—extracting only the first page (usually account summary information), the last page (transaction history / summary information), and the middle pages (representing the main transaction details) evenly distributed according to the total number of pages. The default maximum number of samples is 15 pages. Each sampled image is converted into an image frame using the PyMuPDF rendering engine at a preset DPI (default 200).

[0066] Furthermore, using the Visual Language Model (VLM), the System Prompt guides the model to perform two binary judgments on each sampled image (or single page): determining whether the content of the sampled image is a transaction statement from a bank or payment platform, and determining whether the text and transaction data in the sampled image are clear and identifiable. If both judgments are correct, the sampled image passes the verification. Preferably, a veto + early exit control mechanism is adopted: each image is verified in the sampling order. Once an image is determined to be "no" (not sequential or blurry), the verification of subsequent images and the entire extraction process are immediately terminated, the file is skipped and the next task is processed to avoid invalid calculations.

[0067] Fail-open fault tolerance mechanism: To prevent legitimate images from being mistakenly blocked due to occasional failures of the VLM model service (network timeout, service overload, etc.), this subsystem adopts an exponential backoff retry strategy (maximum of 3 retries, with intervals of 1s, 2s, and 4s respectively) and a fail-open fault tolerance strategy—blocking only when the VLM explicitly returns a negative judgment; if all retries fail (timeout / abnormality), it is automatically allowed to proceed to the next stage, and alarm logs are recorded for operation and maintenance to investigate.

[0068] 2) A parallel image algorithm-based detection module is established to prevent Photoshop manipulation and forgery detection. This module runs in parallel with the legitimacy verification process. A dedicated image algorithm is used to detect tampering in the streaming image files, determining whether the images have been edited, spliced, blurred, or artificially generated using image processing tools such as Photoshop. The detection method is as follows: The first dimension, Error Level Analysis (ELA): detects the differences in compression error levels in different regions of an image and identifies abnormal regions introduced by pasting and splicing. The second dimension is noise pattern analysis: by analyzing the consistency of image noise distribution, it detects regions of inconsistency in noise introduced by local editing. The third dimension is metadata integrity verification: parsing the software editing records and modification timestamps in the EXIF / IPTC metadata of the image to detect whether there are any editing traces from image processing software; The fourth dimension is the detection of tampering edge features: using deep learning convolutional networks to analyze the pixel-level tampering edge features of images and identify the boundary traces of splicing and smearing. The anti-PS detection module performs weighted fusion of the detection results from four dimensions and outputs a comprehensive tampering confidence score (0-1). When the score exceeds a preset threshold, it is marked as suspected tampering. 3) Establish a parallel linkage decision-making mechanism, in which legality verification and anti-PS detection are executed in parallel, reducing the end-to-end latency from serial T1+T2 to (max(T1,T2)) under the parallel scheme. Here, T1 is the time consumed by legality verification and T2 is the time consumed by anti-PS detection. After each is completed independently, the results are merged by the linkage decision-maker. The merged decision is as follows: if both legality verification and anti-PS detection pass, the image file enters the parsing stage in S2; otherwise, the overall business process of the account corresponding to the current image file is immediately interrupted, and the interruption reason and detailed evidence report are generated and returned to the business front end. The subsequent processing pipeline of the account corresponding to the current image file will no longer be executed.

[0069] II. The intelligent parsing model for transaction flow images is configured to enter the parsing pipeline of the parsing model after passing legality verification and anti-PS detection. First, the PDF format image file is rendered page by page into high-resolution image frames (where the high resolution range is 180-220 DPI, with a good balance between clarity and performance). Then, transaction details (debit amount, credit amount, balance, transaction summary, counterparty account, counterparty name, etc.) and unstructured metadata (account name, account number, currency, transaction start and end dates, etc.) are extracted from the high-resolution image frames. After parsing, the transaction flow entities, with the accounts involved in a single transaction as the finest granularity, and their related identifiers and attributes are pre-constructed and summarized and associated with the customer entity by transaction number. The structured transaction flow details and transaction flow summary indicators are combined with the transaction flow entities as instance information of the corresponding customer's current business and stored in the Doris database for the intelligent query interface to retrieve by entity name and entity identifier. Further details regarding the dataset / input: Image files that pass both the image legitimacy verification and anti-PS detection parallel detection model verification and anti-PS detection are listed in Table 6. Table 6:

[0070] The analysis is as follows: (1) Dual-path parallel multimodal visual language large model extraction pipeline module, the method is as follows: The first approach involves extracting transaction details using a visual language model: Guided by a prompt, the visual language model sequentially identifies the header row, data row, and total row from the image, outputting structured JSON information. The second approach involves extracting metadata from a large visual language model: another set of prompts guides the large visual language model to extract structured information (account name, account number, currency, transaction start and end dates, bank name, etc.) from non-table areas in the image, and outputs a JSON object. Two visual language large models call the same batch of image inputs but use different Prompt templates, without blocking each other, maximizing parallel throughput; (2) Field mapping and balance continuity verification module: Field mapping engine: Maps individual column names from various banks ("posting date", "Transaction Date", etc.) to standard fields, outputting standardized comma-separated values; Balance continuity validator: Verification is performed item by item as follows: balance i =balance i-1 ±amount i Among them, balance i This is the account balance. i-1 This is the previous account balance, ±amount i To carry the transaction amount with lending and borrowing directions; derive the theoretical amount symbol (+amount / -amount) and compare it with the current lending and borrowing identifier; if they do not match, invert the lending and borrowing identifier and check again; (3) Module for storing parsed results in the database: After the dual-path visual language model is extracted, the structured transaction details data (standardized field records for each transaction) after field mapping and balance verification is written into the Doris database. The structured transaction details data includes a transaction details table and an account metadata table. The transaction summary metrics returned by the external data platform (total number of transactions, total transaction amount, percentage of integer transactions, number of counterparties, etc.) are also written into the Doris metric table; Doris's high-performance columnar storage and MPP parallel query capabilities provide basic support for the rapid retrieval of massive transaction data in downstream analysis modules.

[0071] (4) Intelligent Questioning Interface: By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity, along with their associated identifiers and attributes, and summarizing and associating them with customer entities by transaction number; the structured transaction details data and transaction summary indicators are combined with the transaction entities as instance information of the corresponding customer's current business and stored in the Doris database, so that the intelligent query interface can retrieve them by entity name and entity identifier. After the parsed results are stored in the database, the downstream analysis module retrieves the real-time parsed structured transaction details and indicator table data from the Doris database on demand via the intelligent query interface. The analysis scheduling engine determines the required data conditions (account dimension, transaction time range, field combination, statistical caliber, etc.) according to rules and initiates a data retrieval request. The interface returns a standardized result set, and the analysis module does not need to be aware of the underlying storage and query implementation details. This interface decouples the parsing module from the analysis module, allowing each module to evolve and expand independently.

[0072] The application results are as follows: After introducing the dual-path parallel VLM extraction and Doris database parsing pipeline, the dual-path VLM extraction of 1200 transactions can save about 40%-50% of the extraction time; structured transaction details and indicator data are uniformly stored in Doris and can be obtained on demand through the intelligent query interface. The parsing module and the analysis module are completely decoupled, and each module can evolve and expand independently.

[0073] Third, after the analysis phase is completed, the structured transaction details and summary indicators have been entered into the Doris database. Since a single transaction may involve multiple accounts (such as the business owner's personal account, spouse's account, and corporate account), it is necessary to conduct independent verification analysis on the transaction records of each account and then aggregate them upwards to form a comprehensive judgment based on the business dimension. Existing solutions mostly adopt a flat, single-layer analysis, which cannot effectively isolate the judgment differences between multiple accounts. Furthermore, compliance red-line rules (such as abnormal business registration of counterparties, involvement in negative lists or regulatory penalties, and abnormal credit records) suffer from insufficient timeliness if relying solely on LLM "internal knowledge."

[0074] Therefore, it is necessary to establish a three-layer convergent big language model analysis and external data cross-validation model at the account granularity. External data query, business information query, and credit data query are embedded into the FunctionCall tool for big language model rule execution. This allows the determination of each rule to complete evidence collection and conclusion generation in a single big language model call. The judgment results are then verified by external data evidence through a cross-validation engine, and finally outputs a three-level judgment: pass, doubtful, or fail.

[0075] Further details regarding the dataset and input: During the analysis phase, the structured transaction details and metrics data are obtained from Doris via the intelligent query interface, and a list of rules is obtained via the rule engine interface. External data sources are queried via Function Call. See Table 7 for a description of the dataset and inputs. Table 7:

[0076] Furthermore, the account-granular three-layer convergent large language model analysis and external data cross-validation model includes the following modules: First layer: Chunk-level rule determination, including Function Call external data query; For each account's transaction data, the transaction records are divided into K chunks according to a preset chunk size (e.g., preset chunk size S = 500 records / chunk), where K is a natural number. Each chunk independently executes the authentication and deception rules as follows: Construct a rule-based Cartesian product task matrix (M authentication rules × K chunks = M × K independent LLM analysis tasks), resolve the dependencies between rules using a directed acyclic graph topology sorter to determine the execution order, and use semaphores to control the batch parallel large language model invocation; External data query during rule execution (Function Call): When performing LLM semantic analysis, each rule can call an external data interface through the Function Call mechanism to obtain authoritative information as a basis for judgment, specifically including: External data query: Search the enterprise registration database (operating status, registered capital, business scope), regulatory penalty record database (administrative penalties, dishonest persons subject to enforcement) and industry access negative list by the name of the counterparty; Business registration information inquiry: Verify whether the business registration status of the counterparty is normal and valid; Credit data inquiry: Obtain credit records, debt information, and credit status of customers and related parties; Beneficial owner lookup: Identify the controlling shareholders and actual controllers of counterparties to uncover hidden related-party transactions; Spouse-related information inquiry: Obtain the spouse's identity and related account information to identify hidden connections in fund transfers; Small model authentication prediction results acquisition: The prediction results of calling the dedicated small model service are used as the reference baseline for LLM determination.

[0077] All external data queries mentioned above are invoked as embedded tools within the LLM rule execution. The query results are directly integrated into the Prompt context of the rule, and the LLM then provides a judgment conclusion by comprehensively considering transaction data and authoritative external data. This design avoids the "two-step" model where external data queries are independent of rule execution, allowing evidence collection and conclusion generation for each rule to be completed in a single LLM call. Preferably, a piecewise Cartesian product task matrix DAG topological sorting method is used to resolve rule dependencies.

[0078] In each leaf task, the LLM performs semantic analysis on the transaction records within the batch chunk according to the corresponding rules (the authenticity rule uses a "forward evidence search" type prompt, and the false detection rule uses a "reverse anomaly identification" type prompt). Combined with the external data query results returned by the Function Call, it outputs the chunk's judgment result (pass / questionable / fail), reasoning basis, external data evidence citation, and confidence score. The false detection rule additionally outputs a confidence score (0-1) and a specific list of suspicious transactions.

[0079] LLM calls support dual-mode switching: thinking mode (enable_thinking=True, for streaming interactive scenarios) and non-thinking mode (enable_thinking=False, for synchronous batch execution scenarios).

[0080] Second layer: Account granularity summary; Once all rule judgment results for all chunks under an account are completed, account-level aggregation is initiated, and a user operation authenticity profile is constructed. The aggregation logic is as follows: Authentication Rules: Positive rules. Hitting a rule increases the weight of the real business profile. Based on the judgment results of all chunks in the current account under each authentication rule, if any chunk hits an authentication rule, the weight of the real profile at the granular level of the account will be increased. False detection rule: Negative rule. If a false detection rule is hit, the weight of the false business profile will be increased. Based on the results of the false detection rules of all chunks, if any chunk hits the false detection rule, the weight of the false business profile of that account will be increased. Red Line Rule: The bottom line rule. If any chunk hits the red line rule, the current account will be marked as failing. During the aggregation process, the results of the small model's authenticity verification, fake product detection, and red line rule judgment serve as the reference baseline for the large language model's judgment. Combining the rule results of the small and large models, if the red line rule is not hit, the result is determined according to the rule weight: if the user profile is more inclined towards fraudulent operations, the account's granularity result is questionable; if the user profile is more inclined towards genuine operations, the account's granularity result is approved. The third layer: summarizing business granularity to form a three-level judgment result; Once all accounts associated with a transaction have completed account-level aggregation, transaction-level aggregation is initiated. The aggregation logic follows a three-level decision rule: Failure: If any related account triggers the red line rule, the entire current business is deemed to fail. Questionable: If no account violates the red line rules, but any related account is identified as having a user profile weight biased towards fraudulent operations, then the current business as a whole is judged to be questionable. If the user profile weights of all associated accounts are biased towards genuine business operations and no red line rules are triggered, then the current business as a whole is judged to be approved.

[0081] LLM uses the third-layer business granularity summary as the basis for analysis and stores it in Doris database for auditing and traceability.

[0082] Finally, a results output layer is established to output the judgment conclusions and evidence reports back to the business front end.

[0083] Example 1: Corporate Loan Approval Scenario (Multiple Accounts): Scenario description: A corporate client applied for a business loan and submitted bank transaction records for the past 6 months (PDF format, 14 pages, approximately 800 records from the corporate account and approximately 400 records from the legal representative's personal account, totaling 1200 transaction records from 2 accounts).

[0084] Processing procedure: (1) Image upload and archiving: The client uploads the corporate account transaction records and the legal representative's personal account transaction records to the image platform according to the account granularity. The platform assigns a unique file link and archives it.

[0085] (2) Module 1: Perform legality verification and anti-PS detection on the transaction records of the two accounts in parallel. The corporate account PDF (14 pages) uses intelligent sampling—extracting the first page, page 7, and last page—and all legality verifications (VLM) pass. The legal representative's personal account transaction record (single-page image) passes VLM verification directly. At the same time, the image algorithm anti-PS detection is performed on both files separately—ELA, noise, metadata, and edge features—and no tampering traces are found. The anti-PS score for both is 0.02 (far below the threshold of 0.5). The linkage decision-maker determines that both tests pass and proceeds to parsing.

[0086] (3) Module 2 – Parsing and Database Storage: Transaction details (a total of 1200 records) and metadata of two accounts are extracted in parallel using dual-path VLM. The balance continuity verification deviation rate is 2%, which is satisfactory. By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity and their associated identifiers and attributes, the transaction entities are summarized and associated with the customer entity by transaction number. The structured transaction details data and the summarized transaction-derived indicators are combined with the transaction entities as instance information of the customer for that transaction and stored in the Doris database, so that downstream analysis modules can obtain them by entity name and instance identifier through the intelligent query interface.

[0087] (4) Module 3 – Analysis: In the analysis phase, structured transaction details and indicator data are obtained from Doris through the intelligent query interface. Analysis is performed independently by account – Account A (800 corporate accounts) is divided into 2 chunks, and Account B (400 individual accounts) is divided into 1 chunk. Authentication rules: 8 rules × 3 chunks = 24 tasks; False detection rules: 5 rules × 3 chunks = 15 tasks; Red line rules: 2 rules × 3 chunks = 6 tasks. The process is completed in approximately 45 seconds with 8 concurrent connections. Results: Account A passed authentication; the false detection rule detected 3 quick-in / quick-out transactions (confidence level 0.72); Account B passed authentication. External data fusion: The small model's authentication prediction was consistent; business registration information was normal (registered capital of 5 million yuan, still in operation); credit information was normal; and there were no abnormal funds associated with the spouse's account.

[0088] (5) Account granularity summary: Account A - Fake judgment rule triggered → Account level suspicious; Account B - All passed.

[0089] (6) Summary of business granularity: Account A is judged as a false trigger, no red line trigger → business level judgment is "doubtful".

[0090] Result: The entire process took about 140 seconds (compared to about 2 hours for traditional manual review), and the identification was successful. The final judgment after three levels of evaluation was "questionable".

[0091] Example 2: Personal Consumer Loan Approval Scenario (Single Account): Scenario description: An individual customer applies for a consumer loan and submits a screenshot of their mobile banking app (a single long image containing approximately 200 transaction records), involving only a single account.

[0092] Processing procedure: (1) Module 1: The image is directly sent to the legality verification VLM and is judged to be smooth and clear; the anti-PS detection is executed in parallel, and ELA / noise / metadata are all normal, with an anti-PS score of 0.01. The linkage decision is passed.

[0093] (2) Module 2 - Parsing and Database Storage: Dual-path VLM extracts 200 transaction records and metadata. By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity and their related identifiers and attributes, the transaction entities are summarized and associated with the customer entity by transaction number. The structured transaction details and the summarized transaction-derived indicators are combined with the transaction entities as instance information of the customer for that transaction and stored in the Doris database, so that downstream analysis modules can obtain them by entity name and instance identifier through the intelligent query interface.

[0094] (3) Module 3 – Analysis and Summary: During the analysis phase, 200 transaction details are obtained through the intelligent query interface, with 500 transactions per chunk and 200 transactions per chunk. 10 rules are loaded (5 for verification + 3 for fraud detection + 2 for red lines), 10 rules × 1 chunk = 10 tasks, completed in approximately 3 seconds with 8 concurrent processes. Verification results: No fixed operating expenses (normal personal transaction history), dispersed trading partners (characteristics of daily consumption), small number of transactions account for 67%. All fraud detections passed. Account-level / business-level summary (single account, consistent results at both levels): All verifications passed, no fraud detections or no red line triggers → judged as "passed".

[0095] Result: The entire process took approximately 40 seconds, and the final judgment after three levels of evaluation was "pass".

[0096] The comparison table of technical effects is shown in Table 8: Table 8:

[0097] Compared with the prior art, the present invention has the following advantages: (1) Parallel pre-filtering for image legality verification and anti-PS: Before extraction, two parallel checkpoints are used: the legality verification VLM and the image algorithm anti-PS detection. The "dual-path parallel + one-vote veto" mechanism is adopted to quickly intercept non-flowing images, blurred images and PS-tampered images to prevent invalid data from entering the subsequent pipeline. The Fail-open fault tolerance strategy ensures that the failure of a single model call does not block the overall process.

[0098] (2) End-to-end automation: From image uploading and archiving, legality verification and anti-PS parallel detection, the end-to-end extraction latency is reduced from serial T1+T2 to (max(T1,T2)). The entire process from extraction to final authentication is completed automatically in a unified pipeline without manual intervention. It supports minute-level processing of enterprise-level large-scale multi-account transaction records (thousands of records).

[0099] (3) Multimodal high-precision extraction: The dual-path parallel VLM is used to extract transaction details and metadata simultaneously, and the identification deviation is automatically corrected through balance continuity verification.

[0100] (4) Transaction pattern-level anti-counterfeiting authentication: By using LLM to perform two-way analysis of transaction flow for authenticity and counterfeiting rules, the characteristics of counterfeit flow (quick in and quick out, large in and large out, etc.) are identified at the transaction pattern level, breaking through the limitation of traditional solutions that only perform numerical verification. The differentiated Prompt design realizes a two-dimensional evaluation of transaction authenticity.

[0101] (5) Account-level three-layer aggregation architecture: Taking the account as the smallest analysis granularity, a three-layer progressive aggregation architecture of "Chunk-level rule judgment, account-level aggregation, and business-level (multi-account) aggregation" is designed to effectively isolate the judgment differences between multiple accounts and clearly and progressively output the final judgment result.

[0102] (6) LLM rules embed external data query (Function Call): External data, business information, credit data, spouse information and other external data queries are used as the Function Call embedded tool for LLM rule execution, so that each rule can complete evidence collection and conclusion generation in a single LLM call, avoiding the traditional two-step mode of "analyze first and then query".

[0103] (7) Three-level judgment output mechanism: The three-level judgment logic driven by the unique red line and the false judgment: the red line rule trigger is "not passed", the false judgment rule trigger is "doubtful", and all passes are "passed"; providing clear and operable judgment results for business decisions.

[0104] (8) Massive parallel processing: Through a three-layer parallel mechanism of rule-sharded Cartesian product task matrix + DAG topology sorting + semaphore concurrent control, it supports simultaneous analysis of thousands of transactions and dozens of rules from multiple accounts.

[0105] (9) Explainable judgment results: Each rule judgment is accompanied by the reasoning basis generated by LLM, the small model prediction reference and the authoritative evidence reference of external data sources. The anti-PS rule provides additional confidence score to meet the financial regulatory requirements for auditability and explainability.

[0106] (10) High availability and fault tolerance design: The legality verification stage adopts the fail-open strategy and exponential backoff retry mechanism; LLM supports dual-mode switching between thinking mode and non-thinking mode, taking into account both real-time interaction and batch processing; the image platform allocates a unique file link to ensure traceability of the entire link.

[0107] (11) Construct reusable transaction entities and associate them with customer entities within the bank. These entities can then serve as important evidence for risk events in other business scenarios.

[0108] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A real-time intelligent authentication method for bank statements, characterized in that, Includes the following steps: S1: Establish a parallel detection model for image legality verification and anti-PS authentication to generate streaming images that pass both legality verification and anti-PS detection; S2: Establish a transaction flow image intelligent analysis model. Transaction flow images that have passed legality verification and anti-PS detection enter the analysis pipeline of the analysis model. First, the PDF format image file is rendered page by page into high-resolution image frames. Then, transaction detail data is extracted from the high-resolution image frames. After analysis, a transaction flow entity with the account involved in a single transaction as the finest granularity and its related identifiers and attributes is pre-constructed and summarized and associated with the customer entity by transaction number. The structured transaction flow detail data and transaction flow summary indicators are combined with the transaction flow entity as instance information of the corresponding customer's current business and stored in the Doris database for the intelligent query interface to retrieve by entity name and entity identifier. S3: Establish a three-layer convergent big language model analysis and external data cross-validation model at the account granularity. Embed external data query, business information query, and credit data query into the Function Call tool for big language model rule execution. This allows the judgment of each rule to complete evidence collection and conclusion generation in a single big language model call. The judgment result is then verified by external data evidence through the cross-validation engine, and finally outputs a three-level judgment: pass, doubtful, or fail.

2. The real-time intelligent authentication method for bank statements as described in claim 1, characterized in that, The following methods are used to generate streaming images that pass both legality verification and anti-Photoshop detection: An image legality verification module is established based on a page-based differentiation strategy. A large visual language model is used, and the SystemPrompt guides the model to perform two binary judgments on each sampled image: whether the content of the sampled image is a transaction statement from a bank or payment platform, and whether the text and transaction data in the sampled image are clear and identifiable. If both judgments are correct, the sampled image passes the verification. A parallel image algorithm-based detection module for preventing Photoshop manipulation (PS) attacks is established. The detection methods are as follows: First dimension, error level analysis: detecting differences in compression error levels across different regions of the image to identify abnormal regions introduced through pasting or splicing; Second dimension, noise pattern analysis: analyzing the consistency of image noise distribution to detect inconsistent noise regions introduced by local editing; Third dimension, metadata integrity verification: parsing software editing records and modification timestamps in the image's EXIF / IPTC metadata to detect the presence of editing traces from image processing software; Fourth dimension, tampering edge feature detection: using a deep learning convolutional network to analyze pixel-level tampering edge features to identify splicing and smearing boundary traces. The anti-PS detection module weights and fuses the detection results from the four dimensions, outputting a comprehensive tampering confidence score. When the score exceeds a preset threshold, it is marked as suspected tampering. A parallel linkage decision-making mechanism is established, in which legality verification and anti-PS detection are executed in parallel, reducing the end-to-end latency from the serial T1+T2 to (max(T1,T2)) under the parallel scheme. Here, T1 is the time consumed by legality verification and T2 is the time consumed by anti-PS detection. After each is completed independently, the results are merged by the linkage decision-maker. The merged decision is as follows: if both legality verification and anti-PS detection pass, the image file enters the parsing stage in S2; otherwise, the overall business process of the account corresponding to the current image file is immediately interrupted, and an interruption reason and detailed evidence report are generated and returned to the business front end. The subsequent processing pipeline of the account corresponding to the current image file will no longer be executed.

3. The real-time intelligent authentication method for bank statements as described in claim 1, characterized in that, The intelligent analysis model for flowing water images includes the following modules: Dual-path parallel multimodal visual language large model extraction pipeline module: The asyncio coroutine is used to concurrently call two visual language models. The method is as follows: the first path is to extract the transaction details visual language model: the Prompt guides the visual language model to identify the header row, data row and total row from the image in sequence, and output structured JSON information. The second path performs metadata extraction for the visual language big model: another set of prompts guides the visual language big model to extract structured information in non-table areas of the image and output a JSON format object; the two visual language big models call the same batch of image inputs but use different prompt templates, without blocking each other, maximizing parallel throughput; Field mapping and balance continuity verification module: Field mapping engine: Maps individual column names from various banks to standard fields, outputting standardized comma-separated values; Balance continuity validator: Verifies balance item by item as follows: i =balance i-1 ±amount i Among them, balance i This is the account balance. i-1 This is the previous account balance, ±amount i For transactions involving lending or borrowing; Module for storing parsed results in the database: After the dual-path visual language model is extracted, the structured transaction details data after field mapping and balance verification is written into the Doris database. The structured transaction details data includes a transaction details table and an account metadata table. The transaction summary metrics returned by the external data platform are also written into the Doris metrics table; Intelligent data query interface: By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity, along with their associated identifiers and attributes, and summarizing and associating them with customer entities by transaction number; the structured transaction details data and transaction summary indicators are combined with the transaction entities as instance information of the corresponding customer's current business and stored in the Doris database, so that the intelligent query interface can retrieve them by entity name and entity identifier. Each module evolves and expands independently.

4. The real-time intelligent authentication method for bank statements as described in claim 1, characterized in that, The account-granular three-layer convergent large language model analysis and external data cross-validation model includes the following modules: First layer: Chunk-level rule determination, including Function Call external data query; For each account's transaction data, the transaction records are divided into K chunks according to the preset chunk size, where K is a natural number. Each chunk independently executes the authentication and falsification rules as follows: the rules are constructed in combination with the chunked Cartesian product task matrix, the dependencies between the rules are resolved by the directed acyclic graph topology sorter to determine the execution order, and semaphores are used to control the batch parallel large language model calls. In each leaf task, the large language model performs semantic analysis on the transaction records within the chunk according to the corresponding rules, and combines the external data query results returned by the Function Call to output the judgment result, reasoning basis, external data evidence reference and confidence score of the current chunk; Second layer: Account granularity summary; Once all rule judgment results for all chunks under an account are completed, account-level aggregation is initiated, and a user operation authenticity profile is constructed. The aggregation logic is as follows: Authentication Rules: Positive rules. Hitting a rule increases the weight of the real business profile. Based on the judgment results of all chunks in the current account under each authentication rule, if any chunk hits an authentication rule, the weight of the real profile at the granular level of the account will be increased. False detection rule: Negative rule. If a false detection rule is hit, the weight of the false business profile will be increased. Based on the results of the false detection rules of all chunks, if any chunk hits the false detection rule, the weight of the false business profile of that account will be increased. Red Line Rule: The bottom line rule. If any chunk hits the red line rule, the current account will be marked as failing. During the aggregation process, the results of the small model's authenticity verification, fake product detection, and red line rule judgment serve as a reference baseline for the large language model's judgment. When the red line rule is not hit, based on the rule weight, if the user profile is more biased towards fraudulent operations, the account granularity result is considered questionable. When the user profile is more aligned with real business operations, the account granularity result is approved. Third layer: Business granularity summary; Once all accounts associated with a transaction have completed account-level aggregation, transaction-level aggregation is initiated. The aggregation logic follows a three-level decision rule: Failure: If any related account triggers the red line rule, the entire current business is deemed to fail. Questionable: If no account violates the red line rules, but any related account is identified as having a user profile weight biased towards fraudulent operations, then the current business as a whole is judged to be questionable. If the user profile weights of all associated accounts are biased towards genuine business operations and no red line rules are triggered, then the current business as a whole is judged to be approved.

5. The real-time intelligent authentication method for bank statements as described in claim 4, characterized in that, The large language model call supports dual-mode switching: thinking mode and non-thinking mode. Thinking mode is used for streaming interaction scenarios, while non-thinking mode is used for synchronous batch execution scenarios.

6. A real-time intelligent authentication system for bank transaction records, characterized in that, include: A parallel detection model for image legitimacy verification and anti-PS authentication is configured to generate streaming images that pass both legitimacy verification and anti-PS detection. The intelligent parsing model for transaction flow images is configured to process images that have passed legality verification and anti-PS detection. First, the PDF image file is rendered page by page into high-resolution image frames. Then, transaction details are extracted from these high-resolution image frames. After parsing, a pre-constructed transaction entity model, with the account involved in each transaction as the finest granularity, along with its associated identifiers and attributes, is built and aggregated by transaction number and associated with the customer entity. The structured transaction details and aggregated transaction indicators are then combined with the transaction entity as instance information for the corresponding customer's current business and stored in the Doris database for the intelligent query interface to retrieve by entity name and identifier. The system integrates a three-layered, account-granular large language model analysis with an external data cross-validation model. It is configured to embed external data queries, business information queries, and credit data queries into a Function Call tool for executing large language model rules. This allows each rule's judgment to complete evidence collection and conclusion generation in a single large language model call. The cross-validation engine then verifies the judgment results with external data evidence, ultimately outputting a three-level judgment: pass, questionable, or fail.

7. The real-time intelligent authentication system for bank statements as described in claim 6, characterized in that, In the parallel detection model for image legitimacy verification and anti-PS detection, the following method is used to generate streaming images that pass both legitimacy verification and anti-PS detection: An image legality verification module is established based on a page-based differentiation strategy. A large visual language model is used, and the SystemPrompt guides the model to perform two binary judgments on each sampled image: whether the content of the sampled image is a transaction statement from a bank or payment platform, and whether the text and transaction data in the sampled image are clear and identifiable. If both judgments are correct, the sampled image passes the verification. A parallel image algorithm-based detection module for preventing Photoshop manipulation (PS) attacks is established. The detection methods are as follows: First dimension, error level analysis: detecting differences in compression error levels across different regions of the image to identify abnormal regions introduced through pasting or splicing; Second dimension, noise pattern analysis: analyzing the consistency of image noise distribution to detect inconsistent noise regions introduced by local editing; Third dimension, metadata integrity verification: parsing software editing records and modification timestamps in the image's EXIF / IPTC metadata to detect the presence of editing traces from image processing software; Fourth dimension, tampering edge feature detection: using a deep learning convolutional network to analyze pixel-level tampering edge features to identify splicing and smearing boundary traces. The anti-PS detection module weights and fuses the detection results from the four dimensions, outputting a comprehensive tampering confidence score. When the score exceeds a preset threshold, it is marked as suspected tampering. A parallel linkage decision-making mechanism is established, in which legality verification and anti-PS detection are executed in parallel, reducing the end-to-end latency from the serial T1+T2 to (max(T1,T2)) under the parallel scheme. Here, T1 is the time consumed by legality verification and T2 is the time consumed by anti-PS detection. After each is completed independently, the results are merged by the linkage decision-maker. The merged decision is as follows: if both legality verification and anti-PS detection pass, the image file enters the parsing stage in S2; otherwise, the overall business process of the account corresponding to the current image file is immediately interrupted, and an interruption reason and detailed evidence report are generated and returned to the business front end. The subsequent processing pipeline of the account corresponding to the current image file will no longer be executed.

8. The real-time intelligent authentication system for bank statements as described in claim 6, characterized in that, The intelligent analysis model for flowing water images includes the following modules: Dual-path parallel multimodal visual language large model extraction pipeline module: The asyncio coroutine is used to concurrently call two visual language models. The method is as follows: the first path is to extract the transaction details visual language model: the Prompt guides the visual language model to identify the header row, data row and total row from the image in sequence, and output structured JSON information. The second path performs metadata extraction for the visual language big model: another set of prompts guides the visual language big model to extract structured information in non-table areas of the image and output a JSON format object; the two visual language big models call the same batch of image inputs but use different prompt templates, without blocking each other, maximizing parallel throughput; Field mapping and balance continuity verification module: Field mapping engine: Maps individual column names from various banks to standard fields, outputting standardized comma-separated values; Balance continuity validator: Verifies balance item by item as follows: i =balance i-1 ±amount i Among them, balance i This is the account balance. i-1 This is the previous account balance, ±amount i For transactions involving lending or borrowing; Module for storing parsed results in the database: After the dual-path visual language model is extracted, the structured transaction details data after field mapping and balance verification is written into the Doris database. The structured transaction details data includes a transaction details table and an account metadata table. The transaction summary metrics returned by the external data platform are also written into the Doris metrics table; Intelligent data query interface: By pre-constructing transaction entities with the accounts involved in a single transaction as the finest granularity, along with their associated identifiers and attributes, and summarizing and associating them with customer entities by transaction number; the structured transaction details data and transaction summary indicators are combined with the transaction entities as instance information of the corresponding customer's current business and stored in the Doris database, so that the intelligent query interface can retrieve them by entity name and entity identifier. Each module evolves and expands independently.

9. The real-time intelligent authentication system for bank statements as described in claim 6, characterized in that, The account-granular three-layer convergent large language model analysis and external data cross-validation model includes the following modules: First layer: Chunk-level rule determination, including Function Call external data query; For each account's transaction data, the transaction records are divided into K chunks according to the preset chunk size, where K is a natural number. Each chunk independently executes the authentication and falsification rules as follows: the rules are constructed in combination with the chunked Cartesian product task matrix, the dependencies between the rules are resolved by the directed acyclic graph topology sorter to determine the execution order, and semaphores are used to control the batch parallel large language model calls. In each leaf task, the large language model performs semantic analysis on the transaction records within the chunk according to the corresponding rules, and combines the external data query results returned by the Function Call to output the judgment result, reasoning basis, external data evidence reference and confidence score of the current chunk; Second layer: Account granularity summary; Once all rule judgment results for all chunks under an account are completed, account-level aggregation is initiated, and a user operation authenticity profile is constructed. The aggregation logic is as follows: Authentication Rules: Positive rules. Hitting a rule increases the weight of the real business profile. Based on the judgment results of all chunks in the current account under each authentication rule, if any chunk hits an authentication rule, the weight of the real profile at the granular level of the account will be increased. False detection rule: Negative rule. If a false detection rule is hit, the weight of the false business profile will be increased. Based on the results of the false detection rules of all chunks, if any chunk hits the false detection rule, the weight of the false business profile of that account will be increased. Red Line Rule: The bottom line rule. If any chunk hits the red line rule, the current account will be marked as failing. During the aggregation process, the results of the small model's authenticity verification, fake product detection, and red line rule judgment serve as a reference baseline for the large language model's judgment. When the red line rule is not hit, based on the rule weight, if the user profile is more biased towards fraudulent operations, the account granularity result is considered questionable. When the user profile is more aligned with real business operations, the account granularity result is approved. Third layer: Business granularity summary; Once all accounts associated with a transaction have completed account-level aggregation, transaction-level aggregation is initiated. The aggregation logic follows a three-level decision rule: Failure: If any related account triggers the red line rule, the entire current business is deemed to fail. Questionable: If no account violates the red line rules, but any related account is identified as having a user profile weight biased towards fraudulent operations, then the current business as a whole is judged to be questionable. If the user profile weights of all associated accounts are biased towards genuine business operations and no red line rules are triggered, then the current business as a whole is judged to be approved.

10. The real-time intelligent authentication system for bank statements as described in claim 9, characterized in that, The large language model call supports dual-mode switching: thinking mode and non-thinking mode. Thinking mode is used for streaming interaction scenarios, while non-thinking mode is used for synchronous batch execution scenarios.