A system and method for automatically analyzing smart contract vulnerabilities by AI
By using an AI-automated analysis system that combines machine learning and deep learning technologies, the problems of time-consuming, labor-intensive, and incomplete smart contract auditing have been solved. This has enabled efficient and accurate vulnerability identification and real-time response, thereby enhancing the security of smart contracts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAMEN SLOWMIST TECHNOLOGY CO LTD
- Filing Date
- 2026-02-10
- Publication Date
- 2026-07-03
AI Technical Summary
Existing smart contract auditing methods are time-consuming and labor-intensive, making it difficult to fully cover potential risks. Automated tools lack flexibility and are unable to discover new or complex vulnerabilities, leading to security risks being overlooked.
By employing an AI-automated analysis system that combines machine learning and deep learning technologies, the system identifies and predicts known and unknown vulnerabilities in smart contracts and generates detailed vulnerability reports through data collection, preprocessing, vulnerability analysis, continuous learning, and integration service modules.
It enables comprehensive and in-depth security assessments, improves audit efficiency and accuracy, enhances developers' security awareness and protection capabilities when writing and deploying smart contracts, and can respond to emerging security risks in real time.
Smart Images
Figure FT_1 
Figure FT_2
Abstract
Description
Technical Field
[0001] This application relates to the field of blockchain technology, and more specifically to a system and method for automating the analysis of smart contract vulnerabilities using AI. Background Technology
[0002] With the widespread application of blockchain technology, smart contracts have become a core component of decentralized applications. However, the security issues of smart contracts are becoming increasingly prominent, with frequent vulnerabilities leading to significant asset losses. Existing manual auditing methods are time-consuming, inefficient, and unable to cover all potential risks, including the following:
[0003] Manually auditing smart contracts is both time-consuming and expensive, and requires a high level of expertise from the auditors.
[0004] Automated auditing tools are typically based on fixed rules, lack flexibility, and are difficult to discover new or complex vulnerabilities;
[0005] No single tool can provide a comprehensive and in-depth analysis of smart contract security, causing many potential security risks to be overlooked.
[0006] In view of this, the inventors have designed a system that uses AI to automatically analyze smart contract vulnerabilities in order to solve the above problems. Summary of the Invention
[0007] To address the shortcomings of existing technologies, the present invention aims to provide a system and method for automatically analyzing smart contract vulnerabilities using AI. Its advantages lie in providing a comprehensive and in-depth security assessment. Through machine learning and deep learning technologies, this system can identify and predict unknown vulnerabilities, thereby greatly improving audit efficiency and accuracy.
[0008] To achieve the above objectives, the present invention provides the following technical solution: a system and method for automatically analyzing smart contract vulnerabilities using AI, comprising:
[0009] The data acquisition and preprocessing module is used to automatically crawl smart contracts from the blockchain network, perform syntactic analysis on the crawled smart contracts to verify the legality of the code, and extract the structural features, logical features and behavioral features of the contracts to generate standardized feature data.
[0010] The AI vulnerability analysis module has a built-in deep learning model trained on a large number of known vulnerability smart contract samples. The deep learning model receives standardized feature data output by the data acquisition and preprocessing module, and identifies known vulnerabilities in smart contracts and predicts potential unknown vulnerabilities through feature matching and risk prediction algorithms, and outputs analysis results containing preliminary vulnerability information.
[0011] The model continuous learning module is used to collect new smart contract data and newly discovered vulnerability information in real time. It uses transfer learning technology to reuse the core parameters of the trained model and combines online learning technology to iteratively optimize the parameters and update the structure of the deep learning model to ensure the model's ability to identify new vulnerabilities.
[0012] The integration service module receives the analysis results from the AI vulnerability analysis module, classifies and organizes the vulnerability information, generates vulnerability reports, and provides standardized API interfaces for the front end to call and display the vulnerability reports.
[0013] The database module includes a smart contract information table, a known vulnerability information table, and a vulnerability report information table, which are used to store basic information and source code of captured smart contracts, detailed information and characteristics of newly discovered and historically known vulnerabilities, and vulnerability report data generated by the integrated service module, respectively.
[0014] A further preferred embodiment of the present invention: the data acquisition and preprocessing module includes a contract capture unit and a data preprocessing unit;
[0015] The contract capture unit captures smart contracts deployed in the blockchain network in batches at a preset frequency through the blockchain node API or blockchain browser data interface, supporting mainstream blockchain networks.
[0016] The data preprocessing unit uses a syntax parser to perform syntax verification on smart contracts, eliminates invalid contracts with syntax errors, extracts structural features, logical features, and behavioral features from valid contracts, and maps the feature data to a preset numerical range through a data standardization algorithm to generate standardized feature data.
[0017] A further preferred embodiment of the present invention is that the deep learning model in the AI vulnerability analysis module adopts a fusion architecture of Convolutional Neural Network (CNN) + Long Short-Term Memory Network (LSTM); wherein, the CNN sub-network is used to extract local key features in the standardized feature data, and the LSTM sub-network is used to capture temporal correlation features in the feature data;
[0018] The risk prediction algorithm determines the vulnerability type and risk level by calculating the similarity between the current contract characteristics and the characteristics of known vulnerabilities, and combining the model's probability prediction of unknown risks. The output analysis results include a preliminary judgment of the vulnerability type, the risk level, and the identification of suspected vulnerability locations.
[0019] A further preferred embodiment of the present invention: the model continuous learning module includes a data collection unit and a model optimization unit;
[0020] The data collection unit builds a new sample library that is updated in real time by subscribing to vulnerability announcements from the blockchain security platform and capturing newly added contracts and unidentified vulnerability data reported by users in the blockchain network.
[0021] The model optimization unit uses transfer learning technology to freeze the parameters of the CNN and LSTM backbone networks of the deep learning model, and only trains the parameters of the feature fusion layer and the output layer. At the same time, it uses an online mini-batch gradient descent algorithm to iterate the parameters of the model once with new samples as the trigger condition. During the iteration process, an early stopping mechanism is used to avoid overfitting.
[0022] A further preferred embodiment of the present invention: the integrated service module includes a report generation unit and an interface service unit;
[0023] The report generation unit categorizes the analysis results according to the vulnerability type, clearly marks the specific location of the vulnerability in the contract, assesses the potential impact of the vulnerability, and generates remediation suggestions by matching the preset remediation strategy library.
[0024] The interface service unit provides a RESTful API interface, which allows the front-end to query the list of vulnerability reports, obtain details of individual vulnerability reports, and export report files through the interface.
[0025] A further preferred embodiment of the present invention includes a model training sample table, which stores sample data used during the training and optimization of the AI vulnerability analysis module, and supports the model continuous learning module in calling the sample data to update the model.
[0026] A method for automating smart contract vulnerability analysis based on AI includes the following steps:
[0027] S1: Data Acquisition and Preprocessing. The data acquisition and preprocessing module captures smart contracts from the blockchain network and generates standardized feature data through syntax analysis and feature extraction.
[0028] S2: AI vulnerability analysis, which uses the deep learning model of the AI vulnerability analysis module to analyze standardized feature data, identify known vulnerabilities and predict unknown vulnerabilities, and output the analysis results;
[0029] S3: Continuous model optimization. New data and vulnerability information are collected through the continuous model learning module, and transfer learning and online learning are used to optimize the deep learning model.
[0030] S4: Vulnerability report generation and service integration. The integrated service module organizes and analyzes the results to generate vulnerability reports and provides API interfaces for front-end calls.
[0031] S5: Data storage, which stores contract data, vulnerability information and vulnerability reports into the corresponding data tables of the database module.
[0032] A further preferred embodiment of the present invention: the training process of the deep learning model in step S2 includes:
[0033] S21: Sample preparation: Collect smart contract samples containing known vulnerabilities, and label each sample with the vulnerability type, vulnerability characteristics and risk level;
[0034] S22: Model initialization, constructing a CNN-LSTM fusion model;
[0035] S23: Model training uses the Adam optimizer and calculates the training error using the cross-entropy loss function. The training set is iterated for a set number of rounds. The model performance is evaluated using the validation set for a set number of rounds. When the vulnerability identification accuracy on the validation set does not improve for a set number of rounds, an early stopping mechanism is triggered. Finally, the trained deep learning model is obtained. The vulnerability identification accuracy of the model on the test set is ≥95%.
[0036] A further preferred embodiment of the present invention includes the following specific process for model optimization in step S3:
[0037] S31: New sample collection. The model continuous learning module collects new smart contract data and new vulnerability information every month. After labeling the features of the new vulnerability information, it is added to the new sample library.
[0038] S32: Transfer learning, freeze the parameters of the CNN and LSTM backbone networks of the trained model, and only unfreeze the parameters of the feature fusion layer and the output layer;
[0039] S33: Online iteration, using the Mini-BatchSGD optimizer, uses data from the new sample library to train the unfrozen layer in a set number of iterations. After the iteration is completed, the model's accuracy in identifying new vulnerabilities is tested. When the accuracy is ≥90%, the model is updated and the original model is replaced.
[0040] In summary, the present invention has the following advantages:
[0041] It provides comprehensive coverage, enabling the identification and prediction of most known and unknown vulnerabilities, and timely response to emerging security risks.
[0042] Ease of use: With this invention, testers do not need to master complex vulnerability analysis skills; they only need to provide the smart contract code to be tested to effectively perform vulnerability detection.
[0043] Security is enhanced by providing detailed vulnerability reports and remediation recommendations, which improve developers' security awareness and protection capabilities when writing and deploying smart contracts.
[0044] With real-time response, the system can detect and handle new vulnerabilities in real time, maintaining sensitivity and responsiveness to the latest risks. Attached Figure Description
[0045] Figure 1 This is a structural diagram of the present invention;
[0046] Figure 2This is a flowchart of the present invention. Detailed Implementation
[0047] The present invention will be further described in detail below with reference to the accompanying drawings.
[0048] As shown in the figure, a system that uses AI to automatically analyze smart contract vulnerabilities includes:
[0049] The data acquisition and preprocessing module is used to automatically crawl smart contracts from the blockchain network, perform syntactic analysis on the crawled smart contracts to verify the legality of the code, and extract the structural features, logical features and behavioral features of the contracts to generate standardized feature data.
[0050] The AI vulnerability analysis module has a built-in deep learning model trained on a massive number of known vulnerability smart contract samples. The deep learning model receives standardized feature data output by the data acquisition and preprocessing module, and through feature matching and risk prediction algorithms, identifies known vulnerabilities in smart contracts and predicts potential unknown vulnerabilities, outputting analysis results containing preliminary vulnerability information.
[0051] The model continuous learning module is used to collect new smart contract data and newly discovered vulnerability information in real time. It uses transfer learning technology to reuse the core parameters of the trained model and combines online learning technology to iteratively optimize the parameters and update the structure of the deep learning model to ensure the model's ability to identify new vulnerabilities.
[0052] The integration service module receives the analysis results from the AI vulnerability analysis module, classifies and organizes the vulnerability information, generates a vulnerability report that includes the vulnerability type, the location of the vulnerability in the contract, the potential impact of the vulnerability, and targeted remediation suggestions, and provides a standardized API interface for the front end to call to display the vulnerability report.
[0053] The database module includes a smart contract information table, a known vulnerability information table, and a vulnerability report information table. These are used to store basic information and source code of captured smart contracts, detailed information and characteristics of newly discovered and historically known vulnerabilities, and vulnerability report data generated by the integration service module, respectively.
[0054] Preferably, the data acquisition and preprocessing module includes a contract capture unit and a data preprocessing unit. The contract capture unit captures smart contracts deployed in the blockchain network in batches once per hour at a preset frequency through the blockchain node API or blockchain explorer data interface, supporting mainstream blockchain networks including Binance Smart Chain and Polygon. The data preprocessing unit uses a syntax parser to perform syntax verification on the smart contracts, eliminating invalid contracts with syntax errors, and extracting structural features, logical features, and behavioral features from valid contracts. Structural features include the number of functions, the type of state variables, and contract inheritance relationships. Logical features include the number of conditional branches, the number of nested loops, and function call relationships. Behavioral features include fund transfer logic, access control logic, and the frequency of external contract interactions. The feature data is then mapped to a preset numerical range using a data standardization algorithm to generate standardized feature data.
[0055] Preferably, the deep learning model in the AI vulnerability analysis module adopts a fusion architecture of "Convolutional Neural Network (CNN) + Long Short-Term Memory Network (LSTM)". Among them, the CNN sub-network is used to extract local key features from the standardized feature data (including abnormal function call patterns and missing segments of permission judgment), and the LSTM sub-network is used to capture temporal correlation features in the feature data, including logical vulnerability paths caused by multi-step operations. The risk prediction algorithm determines the vulnerability type and risk level (high, medium, low) by calculating the similarity between the current contract features and known vulnerability features, combined with the model's probability prediction of unknown risks. The output analysis results include a preliminary judgment of the vulnerability type, risk level, and identification of suspected vulnerability locations.
[0056] Preferably, the model continuous learning module includes a data collection unit and a model optimization unit. The data collection unit subscribes to vulnerability announcements from the blockchain security platform and captures newly added contracts and unidentified vulnerability data reported by users in the blockchain network to build a new sample library that is updated in real time. The model optimization unit uses transfer learning technology to freeze the CNN and LSTM backbone network parameters of the deep learning model, and only trains the parameters of the feature fusion layer and the output layer. At the same time, it uses the online learning mini-batch gradient descent algorithm (Mini-Batch SGD) to iterate the model parameters once every 1,000 new samples. During the iteration process, an early stopping mechanism is adopted. If the accuracy of the validation set does not improve after 5 consecutive iterations, the iteration stops to avoid overfitting.
[0057] Preferably, the integrated service module includes a report generation unit and an interface service unit. The report generation unit categorizes the analysis results according to vulnerability type, including reentrancy vulnerabilities, integer overflows, privilege escalation, and logic vulnerabilities. It clearly marks the specific location of the vulnerability in the contract, including the function name and logical branch in the corresponding contract structure, assesses the potential impact of the vulnerability, including whether it only affects a single user's assets, causes overall contract malfunction, or leads to large-scale on-chain asset losses, and generates remediation suggestions by matching a preset remediation strategy library, including adding permission verification logic and optimizing the fund transfer process. The interface service unit provides a RESTful API interface, allowing the front-end to query the vulnerability report list, obtain details of individual vulnerability reports, and export report files, including PDF and Excel formats.
[0058] Preferably, the database module also includes a model training sample table, which stores sample data used in the training and optimization of the AI vulnerability analysis module, including labeled vulnerability features and contract features. The sample data is organized in the format of "contract ID - vulnerability type - feature tag - labeling time", and supports the model continuous learning module to call the sample data to update the model.
[0059] As shown in the figure, a method for automating smart contract vulnerability analysis based on AI includes the following steps:
[0060] S1: Data Acquisition and Preprocessing. The data acquisition and preprocessing module captures smart contracts from the blockchain network and generates standardized feature data through syntax analysis and feature extraction.
[0061] S2: AI Vulnerability Analysis. This module uses a deep learning model to analyze standardized feature data, identify known vulnerabilities, predict unknown vulnerabilities, and output the analysis results.
[0062] S3: Continuous model optimization. New data and vulnerability information are collected through the continuous model learning module, and transfer learning and online learning are used to optimize the deep learning model.
[0063] S4: Vulnerability report generation and service integration. The integration service module organizes and analyzes the results to generate vulnerability reports and provides API interfaces for front-end calls.
[0064] S5: Data storage, which stores contract data, vulnerability information and vulnerability reports into the corresponding data tables of the database module.
[0065] Preferably, the training process of the deep learning model in step S2 includes:
[0066] S21: Sample preparation: Collect 100,000+ smart contract samples containing known vulnerabilities. Each sample is labeled with the vulnerability type, vulnerability characteristics and risk level, and divided into training set, verification set and test set in a ratio of 7:2:1.
[0067] S22: Model initialization, constructing a CNN-LSTM fusion model, setting the CNN sub-network to 5 layers, including 3 convolutional layers and 2 pooling layers, the LSTM sub-network to 3 layers, the hidden layer dimension to 256, the output layer to be a fully connected layer, and the output dimension to correspond to 20+ types of vulnerabilities;
[0068] S23: Model training uses the Adam optimizer with an initial learning rate of 0.001 and weight decay of 0.01. The training error is calculated using the cross-entropy loss function. The training set is trained for 100 rounds. The model performance is evaluated every 10 rounds using the validation set. An early stopping mechanism is triggered when the vulnerability identification accuracy on the validation set does not improve for 15 consecutive rounds. Finally, the trained deep learning model is obtained, and the vulnerability identification accuracy of the model on the test set is ≥95%.
[0069] Preferably, the specific process of model optimization in step S3 includes:
[0070] S31: New sample collection. The model continuous learning module collects 20,000+ new smart contract data and 500+ new vulnerability information every month. After labeling the features of the new vulnerability information, it is added to the new sample library.
[0071] S32: Transfer learning, freeze the parameters of the CNN and LSTM backbone networks of the trained model, and only unfreeze the parameters of the feature fusion layer and output layer;
[0072] S33: Online iteration, using the Mini-Batch SGD optimizer, batch size 32, learning rate 0.0001, the unfrozen layer is trained for 20 rounds using data from the new sample library. After the iteration is completed, the accuracy of the model in identifying new vulnerabilities is tested. When the accuracy is ≥90%, the model is updated and the original model is replaced.
[0073] The working process and beneficial effects of this invention are as follows:
[0074] This invention enables the identification and prediction of various potential vulnerabilities in smart contracts, including known and unknown security risks. It utilizes deep learning and machine learning technologies for comprehensive analysis of smart contracts, employs automated tools to crawl and preprocess on-chain smart contracts (including syntax analysis and feature extraction), maintains a real-time updated vulnerability database to ensure the system is responsive to the latest vulnerabilities, and features a user-friendly report generation interface, allowing ordinary users to easily perform vulnerability detection and risk assessment. Its key features include:
[0075] It provides comprehensive coverage, enabling the identification and prediction of most known and unknown vulnerabilities, and timely response to emerging security risks.
[0076] Ease of use: With this invention, testers do not need to master complex vulnerability analysis skills; they only need to provide the smart contract code to be tested to effectively perform vulnerability detection.
[0077] Security is enhanced by providing detailed vulnerability reports and remediation recommendations, which improve developers' security awareness and protection capabilities when writing and deploying smart contracts.
[0078] With real-time response, the system can detect and handle new vulnerabilities in real time, maintaining sensitivity and responsiveness to the latest risks.
[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the design concept of the present invention should be included within the protection scope of the present invention.
Claims
1. A system for automatically analyzing smart contract vulnerabilities using AI, characterized in that, include: The data acquisition and preprocessing module is used to automatically crawl smart contracts from the blockchain network, perform syntactic analysis on the crawled smart contracts to verify the legality of the code, and extract the structural features, logical features and behavioral features of the contracts to generate standardized feature data. The AI vulnerability analysis module has a built-in deep learning model trained on a large number of known vulnerability smart contract samples. The deep learning model receives standardized feature data output by the data acquisition and preprocessing module, and identifies known vulnerabilities in smart contracts and predicts potential unknown vulnerabilities through feature matching and risk prediction algorithms, and outputs analysis results containing preliminary vulnerability information. The model continuous learning module is used to collect new smart contract data and newly discovered vulnerability information in real time. It uses transfer learning technology to reuse the core parameters of the trained model and combines online learning technology to iteratively optimize the parameters and update the structure of the deep learning model to ensure the model's ability to identify new vulnerabilities. The integration service module receives the analysis results from the AI vulnerability analysis module, classifies and organizes the vulnerability information, generates vulnerability reports, and provides standardized API interfaces for the front end to call and display the vulnerability reports. The database module includes a smart contract information table, a known vulnerability information table, and a vulnerability report information table, which are used to store basic information and source code of captured smart contracts, detailed information and characteristics of newly discovered and historically known vulnerabilities, and vulnerability report data generated by the integrated service module, respectively.
2. The system for automatically analyzing smart contract vulnerabilities using AI according to claim 1, characterized in that: The data acquisition and preprocessing module includes a contract capture unit and a data preprocessing unit; The contract capture unit captures smart contracts deployed in the blockchain network in batches at a preset frequency through the blockchain node API or blockchain browser data interface, supporting mainstream blockchain networks. The data preprocessing unit uses a syntax parser to perform syntax verification on smart contracts, eliminates invalid contracts with syntax errors, extracts structural features, logical features, and behavioral features from valid contracts, and maps the feature data to a preset numerical range through a data standardization algorithm to generate standardized feature data.
3. The system for automatically analyzing smart contract vulnerabilities using AI according to claim 1, characterized in that, The deep learning model in the AI vulnerability analysis module adopts a fusion architecture of Convolutional Neural Network (CNN) + Long Short-Term Memory Network (LSTM); wherein, the CNN sub-network is used to extract local key features from the standardized feature data, and the LSTM sub-network is used to capture temporal correlation features in the feature data; The risk prediction algorithm determines the vulnerability type and risk level by calculating the similarity between the current contract characteristics and the characteristics of known vulnerabilities, and combining the model's probability prediction of unknown risks. The output analysis results include a preliminary judgment of the vulnerability type, the risk level, and the identification of suspected vulnerability locations.
4. The system for automatically analyzing smart contract vulnerabilities using AI according to claim 1, characterized in that: The model continuous learning module includes a data collection unit and a model optimization unit; The data collection unit builds a new sample library that is updated in real time by subscribing to vulnerability announcements from the blockchain security platform and capturing newly added contracts and unidentified vulnerability data reported by users in the blockchain network. The model optimization unit uses transfer learning technology to freeze the parameters of the CNN and LSTM backbone networks of the deep learning model, and only trains the parameters of the feature fusion layer and the output layer. At the same time, it uses an online mini-batch gradient descent algorithm to iterate the parameters of the model once with new samples as the trigger condition. During the iteration process, an early stopping mechanism is used to avoid overfitting.
5. A system for automatically analyzing smart contract vulnerabilities using AI according to claim 1, characterized in that: The integrated service module includes a report generation unit and an interface service unit; The report generation unit categorizes the analysis results according to the vulnerability type, clearly marks the specific location of the vulnerability in the contract, assesses the potential impact of the vulnerability, and generates remediation suggestions by matching the preset remediation strategy library. The interface service unit provides a RESTful API interface, which allows the front-end to query the list of vulnerability reports, obtain details of individual vulnerability reports, and export report files through the interface.
6. The system for automatically analyzing smart contract vulnerabilities using AI according to claim 1, characterized in that: The database module also includes a model training sample table, which stores sample data used in the training and optimization process of the AI vulnerability analysis module, and supports the model continuous learning module to call the sample data to update the model.
7. A method for AI-automated analysis of smart contract vulnerabilities based on the system described in claim 1, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing. The data acquisition and preprocessing module captures smart contracts from the blockchain network and generates standardized feature data through syntax analysis and feature extraction. S2: AI vulnerability analysis, which uses the deep learning model of the AI vulnerability analysis module to analyze standardized feature data, identify known vulnerabilities and predict unknown vulnerabilities, and output the analysis results; S3: Continuous model optimization. New data and vulnerability information are collected through the continuous model learning module, and transfer learning and online learning are used to optimize the deep learning model. S4: Vulnerability report generation and service integration. The integrated service module organizes and analyzes the results to generate vulnerability reports and provides API interfaces for front-end calls. S5: Data storage, which stores contract data, vulnerability information and vulnerability reports into the corresponding data tables of the database module.
8. The system and method for automatically analyzing smart contract vulnerabilities using AI according to claim 7, characterized in that, The training process of the deep learning model in step S2 includes: S21: Sample preparation: Collect smart contract samples containing known vulnerabilities, and label each sample with the vulnerability type, vulnerability characteristics and risk level; S22: Model initialization, constructing a CNN-LSTM fusion model; S23: Model training uses the Adam optimizer and calculates the training error using the cross-entropy loss function. The training set is iterated for a set number of rounds. The model performance is evaluated using the validation set for a set number of rounds. When the vulnerability identification accuracy on the validation set does not improve for a set number of rounds, an early stopping mechanism is triggered. Finally, the trained deep learning model is obtained. The vulnerability identification accuracy of the model on the test set is ≥95%.
9. The system and method for automatically analyzing smart contract vulnerabilities using AI according to claim 7, characterized in that, The specific process of model optimization in step S3 includes: S31: New sample collection. The model continuous learning module collects new smart contract data and new vulnerability information every month. After labeling the features of the new vulnerability information, it is added to the new sample library. S32: Transfer learning, freeze the parameters of the CNN and LSTM backbone networks of the trained model, and only unfreeze the parameters of the feature fusion layer and the output layer; S33: Online iteration, using the Mini-BatchSGD optimizer, uses data from the new sample library to train the unfrozen layer in a set number of iterations. After the iteration is completed, the model's accuracy in identifying new vulnerabilities is tested. When the accuracy is ≥90%, the model is updated and the original model is replaced.