A bank transaction flow authenticity detection system and method based on intelligent analysis

Abstract

The application discloses a bank transaction flow authenticity detection system and method based on intelligent analysis, and belongs to the field of financial data security. The system comprises a data preprocessing and structured extraction module, a high-level intelligent feature engineering module, a multi-modal integrated learning module and an explainability analysis module. The detection method extracts structured data in the flow file through high-precision OCR technology and completes cleaning and standardization. A multi-dimensional discrimination feature space is constructed, which covers macro statistical and distribution characteristics, micro timing and behavior characteristics, and local anomaly and combination characteristics. A multi-modal integrated architecture of LSTM / Transformer timing model and gradient boosting tree statistical model is adopted, and a real probability value is output through a stacking generalization strategy. An explainability analysis result is generated in combination with SHAP values. The application solves the problems of low efficiency and weak discrimination ability for high-simulation forgery of traditional detection methods, realizes automatic and high-precision detection of transaction flow authenticity, and effectively reduces the operating cost and fraud risk of financial institutions.

Description

Technical Field

[0001] This invention relates to the field of financial data security technology, and in particular to a system and method for detecting the authenticity of bank transaction records based on intelligent analysis. Background Technology

[0002] Bank transaction records, as core documents proving a company's or individual's financial status, transaction capabilities, and creditworthiness, are widely used in high-risk business scenarios such as credit approval, financial auditing, and business cooperation. Their authenticity is a key element of the financial risk control system.

[0003] However, existing technologies have limitations in verifying the authenticity of bank transaction records: 1. Inefficiency and Dependence on Human Resource Management: Traditional auditing processes heavily rely on human experience for format comparison, seal verification, or simple internal cross-validation. This method is extremely inefficient and costly, and its accuracy is easily limited by the auditor's subjective experience and fatigue when faced with massive amounts of heterogeneous (different banks, different formats) transaction documents.

[0004] 2. Lack of in-depth behavioral identification capability against highly realistic forgeries: Modern forgery methods can accurately simulate the format and basic elements of bank transaction records. The core flaw of existing detection systems lies in their lack of in-depth analysis capabilities of the "behavioral patterns" of transaction data, only able to verify surface elements. Professional forgers meticulously fabricate transaction sequences to meet basic verification standards on the surface, but the deep statistical distribution, temporal dependencies, or account behavior habits of forged transactions differ fundamentally from those of genuine accounts. Traditional rule-based verification and manual review mechanisms cannot capture these advanced features reflecting forgery traces, making highly realistic forged transactions difficult to identify effectively, posing a significant fraud risk to financial institutions.

[0005] Therefore, there is an urgent need for a new type of detection system and method that can perform in-depth and intelligent analysis from multiple dimensions such as data behavior, statistical distribution, and time series patterns. Summary of the Invention

[0006] The purpose of this invention is to address the problem of insufficient authenticity of existing bank transaction records by proposing a bank transaction record authenticity detection system and method based on intelligent analysis.

[0007] To achieve the above objectives, the present invention employs the following technology: a bank transaction record authenticity detection system based on intelligent analysis, comprising: The data preprocessing and structured extraction module is used to receive the transaction files to be inspected, use OCR technology to extract structured data such as date, amount, payee and payee, and summary from the transaction records, and perform format unification and missing value handling on the extracted data, while standardizing and vectorizing the summary text. The advanced intelligent feature engineering module is used to construct a multi-dimensional discriminative feature space, extracting macro-statistical and distributional features, micro-temporal and behavioral features, and local anomalies and combination features. Among them, macro-statistical and distributional features include transaction concentration indicators, balance fluctuation and retention indicators, and transaction type information entropy; micro-temporal and behavioral features include higher-order moments of timestamps, periodic patterns, transaction graph centrality, and correlation measures; local anomalies and combination features include local anomaly factors (LOF) and highly discriminative features constructed by feature cross-construction. The multimodal ensemble learning module includes a base classifier training unit and an ensemble fusion judgment unit. The base classifier training unit uses LSTM or Transformer models for temporal dependency modeling and gradient boosting trees for non-temporal statistical modeling. The ensemble fusion judgment unit adopts a stacked generalization strategy, using the output of the base classifier as the input of the meta classifier, learning the optimal combination weights and outputting the final true probability value. The interpretability analysis module is used to generate a list of core anomaly features and their corresponding contributions, providing a quantitative basis for judgment.

[0008] As a further description of the above technical solution: the transaction concentration index is the Gini coefficient of the account transaction amount; the balance volatility and retention index is the ratio of the standard deviation to the mean of the account's daily balance; and the transaction type information entropy is expressed by the formula... The calculation yielded the result.

[0009] As a further description of the above technical solution: the higher-order moments of the timestamp include skewness and kurtosis; the periodic pattern is detected by decomposing the transaction sequence using Fourier transform; and the transaction graph centrality index is used to quantify the degree of correlation between the account to be detected and the abnormal trading counterparty.

[0010] As a further description of the above technical solution: the high-discrimination features constructed by the feature cross include the proportion of large transactions occurring outside of working hours and the ratio of transaction frequency to balance change.

[0011] A method for verifying the authenticity of bank transaction records based on intelligent analysis includes the following steps: S1. Data Preprocessing and Structured Extraction: Receive the bank transaction log files to be inspected, use OCR technology to extract the date, amount, payee and payee information and summary information from the transaction records to form structured data, perform format unification and missing value processing on the structured data, and standardize and vectorize the summary text. S2. Advanced Intelligent Feature Engineering and Feature Space Construction: Extract macroscopic statistical and distribution features, microscopic temporal and behavioral features, local anomalies and combination features, and construct a multi-dimensional discriminative feature space; S3. Multimodal Ensemble Learning and Interpretability Detection: Temporal dependency modeling is performed using LSTM / Transformer models, and non-temporal statistical modeling is performed using gradient boosting trees. The outputs of the two models are integrated using a stacked generalization strategy. The authenticity probability value is output through a meta-classifier. When the authenticity probability value is lower than a preset threshold, it is judged as high risk or fake transaction. A list of core abnormal features and their contribution are generated based on interpretability analysis technology to provide quantitative judgment basis.

[0012] As a further description of the above technical solution: In step S2, when extracting macroeconomic statistics and distribution characteristics, the uniformity of transaction amount distribution is measured by calculating the Gini coefficient, the fluctuation of balance is detected by calculating the coefficient of variation, and the transaction type information entropy is calculated to measure transaction diversity.

[0013] As a further description of the above technical solution: In step S2, when extracting micro-time series and behavioral features, the transaction time interval sequence is analyzed and skewness and kurtosis are calculated. Periodic features are detected by Fourier transform, and a transaction graph is constructed and a centrality index is calculated.

[0014] As a further description of the above technical solution: In step S2, when extracting local anomalies and combined features, the local anomaly factor of each transaction is calculated, and a high-discrimination cross feature is constructed through feature cross-conversion.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: Through the synergistic effect of multi-dimensional deep feature engineering and multi-modal ensemble learning models, the recognition performance of highly realistic forged transaction records has been significantly improved, especially in accurately identifying professional forged transaction records with abnormal behavioral patterns. At the same time, the detection process has been fully automated. Relying on high-precision OCR technology and intelligent analysis algorithms, the review time of a single bank transaction record has been greatly reduced from tens of minutes of manual operation to less than a few seconds. This not only improves detection efficiency but also ensures the consistency and stability of detection results in different scenarios. High-precision identification capabilities can accurately intercept risky businesses corresponding to fraudulent transaction records, effectively avoid economic losses such as bad debts and commercial fraud caused by false financial certificates, significantly improve the accuracy and effectiveness of risk management, and the highly automated and intelligent analysis characteristics of the detection process greatly reduce the reliance on high-cost professional manual reviewers, significantly optimize the operating cost structure, and improve business operation efficiency. It has effectively curbed financial crimes involving fraud using fake bank transaction records. By improving the credibility and transparency of transaction data, it helps to establish a healthier business and financial environment and maintain market order. Attached Figure Description

[0016] Figure 1A schematic diagram of the system architecture provided according to an embodiment of the present invention is shown; Figure 2 A schematic diagram of a method flow provided according to an embodiment of the present invention is shown. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Reference Figures 1-2 This embodiment provides a bank transaction record authenticity detection system based on intelligent analysis, including a data preprocessing and structured extraction module, an advanced intelligent feature engineering module, a multimodal ensemble learning module, and an interpretability analysis module, wherein: The data preprocessing and structured extraction module is used to receive the transaction files to be inspected, use OCR technology to extract structured data such as date, amount, payee and payee, and summary from the transaction records, and perform format unification and missing value handling on the extracted data, while standardizing and vectorizing the summary text. The advanced intelligent feature engineering module is used to construct a multi-dimensional discriminative feature space, extracting macro-statistical and distributional features, micro-temporal and behavioral features, and local anomalies and combination features. Among them, macro-statistical and distributional features include transaction concentration indicators, balance fluctuation and retention indicators, and transaction type information entropy; micro-temporal and behavioral features include higher-order moments of timestamps, periodic patterns, transaction graph centrality, and correlation measures; local anomalies and combination features include local anomaly factors (LOF) and highly discriminative features constructed by feature cross-construction. It should be noted that the transaction concentration index is the Gini coefficient of the account's transaction amount, the balance volatility and retention index is the ratio of the standard deviation to the mean of the account's daily balance, and the transaction type information entropy is expressed by the formula... The calculations show that the higher-order moments of the timestamp include skewness and kurtosis. The periodic pattern is detected by decomposing the transaction sequence using Fourier transform. The transaction graph centrality index is used to quantify the degree of association between the account to be detected and the abnormal counterparty. The high-discrimination features constructed by feature cross-construction include the proportion of large transactions occurring outside of working hours and the ratio of transaction frequency to balance change.

[0019] The multimodal ensemble learning module includes a base classifier training unit and an ensemble fusion judgment unit. The base classifier training unit uses LSTM or Transformer models for temporal dependency modeling and gradient boosting trees for non-temporal statistical modeling. The ensemble fusion judgment unit adopts a stacked generalization strategy, using the output of the base classifier as the input of the meta classifier, learning the optimal combination weights and outputting the final true probability value. The interpretability analysis module is used to generate a list of core anomaly features and their corresponding contributions, providing a quantitative basis for judgment.

[0020] The detection methods of the above detection system are as follows: S1. Data Preprocessing and Structured Extraction: The system receives the transaction files (images, PDFs, etc.) to be detected, uses high-precision OCR technology to extract the transaction records (date, amount, payee, summary, etc.) in a structured manner, and performs format unification and missing value processing on the extracted data. The summary text is also standardized and vectorized to prepare for subsequent feature engineering. S2. Advanced Intelligent Feature Engineering and Feature Space Construction: Extract macroscopic statistical and distribution features, microscopic temporal and behavioral features, local anomalies and combination features, and construct a multi-dimensional identification feature space.

[0021] Specifically, when extracting macro-statistical and distributional features, the Gini coefficient is calculated to measure the uniformity of transaction amount distribution, the coefficient of variation is calculated to detect balance fluctuations, and the transaction type information entropy is calculated to measure transaction diversity. When extracting micro-temporal and behavioral features, the transaction time interval sequence is analyzed and skewness and kurtosis are calculated. Fourier transform is used to detect periodic features, and a transaction graph is constructed and a centrality index is calculated. When extracting local anomalies and combination features, the local anomaly factor of each transaction is calculated, and highly discriminative cross features are constructed through feature cross-conversion.

[0022] S3. Multimodal Ensemble Learning and Interpretability Detection: This step employs a hybrid model architecture to achieve optimal judgment robustness, and specifically includes the following steps: Base classifier training: Temporal dependency modeling: Using Long Short-Term Memory (LSTM) networks or Transformer models to model the time-series features of transactions and capture long-term temporal context information; Non-temporal statistical modeling: using gradient boosting trees (such as LightGBM) to efficiently learn statistical, distributional, and correlational features, capturing complex nonlinear decision boundaries in the feature space; Integration and fusion judgment: A stacked generalization strategy is adopted, using the outputs of the time series model and the statistical model as inputs to a meta-classifier (such as logistic regression). This meta-classifier learns the optimal combination weights and outputs the final true probability value $P$. Below the preset threshold The system identified it as "high risk / forged"; Explainability analysis: The system uses interpretability techniques such as SHAP values ​​to automatically generate a list of core anomaly features and their corresponding contributions, providing auditors with scientific and quantitative evidence of "why it was determined to be forgery".

[0023] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A bank transaction record authenticity detection system based on intelligent analysis, characterized in that, include: The data preprocessing and structured extraction module is used to receive the transaction files to be inspected, use OCR technology to extract structured data such as date, amount, payee and payee, and summary from the transaction records, and perform format unification and missing value handling on the extracted data, while standardizing and vectorizing the summary text. The advanced intelligent feature engineering module is used to construct a multi-dimensional discriminative feature space, extracting macro-statistical and distributional features, micro-temporal and behavioral features, and local anomalies and combination features. Among them, macro-statistical and distributional features include transaction concentration indicators, balance fluctuation and retention indicators, and transaction type information entropy; micro-temporal and behavioral features include higher-order moments of timestamps, periodic patterns, transaction graph centrality, and correlation measures; local anomalies and combination features include local anomaly factors (LOF) and highly discriminative features constructed by feature cross-construction. The multimodal ensemble learning module includes a base classifier training unit and an ensemble fusion judgment unit. The base classifier training unit uses LSTM or Transformer models for temporal dependency modeling and gradient boosting trees for non-temporal statistical modeling. The ensemble fusion judgment unit adopts a stacked generalization strategy, using the output of the base classifier as the input of the meta classifier, learning the optimal combination weights and outputting the final true probability value. The interpretability analysis module is used to generate a list of core anomaly features and their corresponding contributions, providing a quantitative basis for judgment.

2. The bank transaction record authenticity detection system based on intelligent analysis according to claim 1, characterized in that, The transaction concentration index is the Gini coefficient of the account's transaction amount; the balance volatility and retention index is the ratio of the standard deviation to the mean of the account's daily balance; and the transaction type information entropy is expressed by the formula... The calculation yielded the result.

3. The bank transaction record authenticity detection system based on intelligent analysis according to claim 1, characterized in that, The higher-order moments of the timestamp include skewness and kurtosis. The periodic pattern is detected by decomposing the transaction sequence using Fourier transform. The transaction graph centrality index is used to quantify the degree of correlation between the account to be detected and the abnormal trading counterparty.

4. The bank transaction record authenticity detection system based on intelligent analysis according to claim 1, characterized in that, The highly discriminative features constructed by the feature cross-construction include the proportion of large transactions occurring outside of working hours and the ratio of transaction frequency to balance change.

5. A detection method based on the detection system according to any one of claims 1-4, characterized in that, Includes the following steps: S1. Data Preprocessing and Structured Extraction: Receive the bank transaction log files to be inspected, use OCR technology to extract the date, amount, payee and payee information and summary information from the transaction records to form structured data, perform format unification and missing value processing on the structured data, and standardize and vectorize the summary text. S2. Advanced Intelligent Feature Engineering and Feature Space Construction: Extract macroscopic statistical and distribution features, microscopic temporal and behavioral features, local anomalies and combination features, and construct a multi-dimensional discriminative feature space; S3. Multimodal Ensemble Learning and Interpretability Detection: Temporal dependency modeling is performed using LSTM / Transformer models, and non-temporal statistical modeling is performed using gradient boosting trees. The outputs of the two models are integrated using a stacked generalization strategy. The authenticity probability value is output through a meta-classifier. When the authenticity probability value is lower than a preset threshold, it is judged as high risk or fake transaction. A list of core abnormal features and their contribution are generated based on interpretability analysis technology to provide quantitative judgment basis.

6. The method for detecting the authenticity of bank transaction records based on intelligent analysis according to claim 5, characterized in that, In step S2, when extracting macroeconomic statistics and distribution characteristics, the Gini coefficient is calculated to measure the uniformity of transaction amount distribution, the coefficient of variation is calculated to detect balance fluctuations, and the transaction type information entropy is calculated to measure transaction diversity.

7. The method for detecting the authenticity of bank transaction records based on intelligent analysis according to claim 5, characterized in that, In step S2, when extracting micro-time series and behavioral features, the transaction time interval sequence is analyzed and skewness and kurtosis are calculated. Periodic features are detected by Fourier transform, and a transaction graph is constructed and a centrality index is calculated.

8. The method for detecting the authenticity of bank transaction records based on intelligent analysis according to claim 5, characterized in that, In step S2, when extracting local anomalies and combined features, the local anomaly factor for each transaction is calculated, and a high-discrimination cross feature is constructed through feature cross-conversion.

Images