LLM-RAG-based fraud-related URL classification and identification method

Through LLM-RAG technology, the fraud-related identification model and knowledge vector library are constructed, and the identification results are verified, which solves the insufficient performance and verifiability of fraud-related URL recognition in the existing technology, and realizes more reliable and detailed fraud-related URL classification.

CN119025675AActive Publication Date: 2024-11-26SHENZHEN UNIV

Patent Information

Application Number
CN202410997005.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-26
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

Existing fraud-related URL recognition methods have insufficient performance, cannot provide detailed classification and verifiable results, and are vulnerable to noise and attacks, unable to handle complex data patterns and real-time update requirements.

Method used

Using an LLM-RAG-based method, by constructing a fraud-related identification model and a knowledge vector library, combining RAG technology to verify the identification results, we provide interpretable and verifiable fraud-related URL classification.

Benefits of technology

It realizes more reliable fraud-related URL recognition, provides detailed classification results and interpretable judgment basis, enhances the credibility of the identification results, and updates the knowledge base in real time to deal with data drift.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025675A_ABST
    Figure CN119025675A_ABST
Patent Text Reader

Abstract

The invention provides a fraud-related URL classification and identification method based on LLM-RAG. The fraud-related URL classification and identification method comprises the steps of obtaining URL data to be detected; the URL data are input into a fraud-related identification model for identification, an identification result and a judgment basis are obtained, and the fraud-related identification model is constructed by LLM; and verifying the identification result by using an RAG method in combination with the judgment basis to obtain a verification result. Compared with other methods, the verification of the identification result is realized through the RAG technology, and a more reliable identification result can be provided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of fraud-related URL identification, and in particular to a fraud-related URL classification identification method based on LLM-RAG. Background Art

[0002] A fraudulent website refers to a fake or deceptive website that is built through some existing technical methods and disguised as a normal website to conduct fraudulent activities. Uniform Resource Locator (URL) is the entrance to web application services. Through the URL, relevant resources on the Internet can be found. Fraudulent URL classification and identification is a technology that identifies the specified URL based on blacklists, rule matching, machine learning, and deep learning to confirm whether it is suspected of fraud.

[0003] At present, there are three main methods for identifying fraudulent URLs: blacklist-based identification method, rule-matching-based identification method, and artificial intelligence classification-based identification method. However, these methods have some shortcomings. The performance of the blacklist-based identification method depends on the size of the original list, and it takes a certain amount of time to update the blacklist. The rule-matching-based identification method requires manual rule formulation, cannot handle complex data patterns, and is vulnerable to attacks. The artificial intelligence-based identification method relies on data sets and feature extraction, requires the extraction of effective features and training on a large amount of data, and is easily affected by noise or interference. At the same time, because attackers continue to optimize their attack techniques and design targeted URLs to avoid detection, traditional identification methods have certain limitations.

[0004] In addition, the above identification method has two other disadvantages:

[0005] 1. The output only shows whether the URL is fraudulent, but there is no reason to explain why the URL is fraudulent. There is no verifiable strategy, and the result verification cannot be provided. The judgment is based only on the results output by the corresponding model.

[0006] 2. No more detailed classification is implemented, and URLs are simply divided into malicious / non-malicious or fraudulent / non-fraudulent.

[0007] Therefore, a technology that can solve the above problems is urgently needed in the fraud-related URL classification and identification method to achieve more detailed and reliable fraud-related URL identification. Summary of the invention

[0008] The purpose of the present invention is to provide a method for classifying and identifying fraudulent URLs based on LLM-RAG. Compared with other methods, the present invention realizes the verification of the identification results through RAG technology and can provide more reliable identification results.

[0009] To achieve the above object, the present invention provides a method for classifying and identifying fraudulent URLs based on LLM-RAG, comprising:

[0010] Get the URL data to be tested;

[0011] Input the URL data into a fraud identification model for identification, and obtain identification results and judgment basis, wherein the fraud identification model is constructed by LLM;

[0012] The RAG method is used in combination with the judgment basis to verify the recognition result and obtain a verification result.

[0013] Optionally, the URL data includes: web page content, page screenshots, web page icons, and ICP registration information.

[0014] Optionally, obtaining the recognition result and the judgment basis includes:

[0015] Set recognition prompt words;

[0016] Injecting the identification prompt word into the URL data;

[0017] The URL data after the recognition prompt word is injected is input into the adjusted fraud-related recognition model for recognition to obtain the recognition result and judgment basis.

[0018] Optionally, adjusting the fraud identification model includes:

[0019] Collecting preset URL data and setting the identification prompt words; wherein the preset URL data includes: fraud-related URL data and normal URL data;

[0020] Filling the recognition prompt words in the collected preset URL data to construct a fine-tuning corpus;

[0021] The fraud identification model is adjusted based on the fine-tuning corpus.

[0022] Optionally, verifying the recognition result includes:

[0023] Construct a fraud-related knowledge vector library;

[0024] According to the judgment basis, searching the knowledge vector library to obtain the corresponding knowledge vector;

[0025] After the knowledge vector is filled with verification classification prompt words, the fraud identification model is input for analysis and judgment, and the verification result is output.

[0026] Optionally, constructing a fraud-related knowledge vector library includes:

[0027] Collect text data related to fraud;

[0028] Performing semantic vectorization on the fraud-related text data;

[0029] Build an index for the vectorized text data and write it into the database to generate a fraud-related knowledge vector library.

[0030] Optionally, performing semantic vectorization on the fraud-related text data includes:

[0031] The collected fraud-related text data is segmented at the sentence granularity, the complete semantics of a sentence is retained for text segmentation, and then semantic vectorization is performed.

[0032] Optionally, according to the judgment basis, searching the knowledge vector library includes:

[0033] Input the judgment basis as a keyword into the embedding model for vectorization processing;

[0034] In the knowledge vector library, knowledge texts having a preset similarity with the semantics of the keyword vector are searched to obtain the corresponding knowledge vector.

[0035] Optionally, searching for knowledge texts having a preset semantic similarity with the keyword vector includes:

[0036] The cosine similarity is used to calculate the similarity between the semantics of the keyword vector and the knowledge vectors in the knowledge vector library, and the knowledge vectors with a preset cosine similarity are selected.

[0037] The present invention has the following beneficial effects:

[0038] The present invention provides an interpretable, verifiable, and classifiable fraud-related URL identification method. Compared with other fraud-related URL identification methods, the present invention uses the emergence ability of LLM to provide powerful understanding and reasoning capabilities. The fine-tuned LLM can identify and provide interpretable judgment basis based on the provided URL-related information; in addition, compared with other methods, the present invention realizes the verification of the identification results through RAG technology, which can provide more reliable identification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0040] Figure 1 A schematic diagram of the URL data collection process according to an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the construction process of a fraud-related knowledge vector library according to an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the fraud-related URL verification classification process based on RAG according to an embodiment of the present invention;

[0043] Figure 4 The figure is a flow chart of a fraud classification and identification method based on LLM-RAG according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0046] This embodiment provides a method for identifying fraudulent URLs with interpretable, verifiable, and classifiable results. The method collects relevant information of URLs, constructs a corpus of fraudulent URLs, and fine-tunes a large language model (LLM) to enable it to have the ability to identify fraudulent URLs, thereby achieving interpretable results. A fraudulent knowledge vector library is constructed, and retrieval-augmented generation (RAG) is used to verify and classify the recognition results, thereby achieving verifiable classification.

[0047] This embodiment proposes a method for classifying and identifying fraudulent URLs based on LLM-RAG, including:

[0048] Get the URL data to be tested;

[0049] Input the URL data into the fraud identification model for identification, and obtain the identification result and judgment basis, wherein the fraud identification model is constructed by LLM;

[0050] The RAG method is used in combination with the judgment basis to verify the recognition result and obtain the verification result.

[0051] Furthermore, the URL data includes: web page content, page screenshots, web page icon, and ICP registration information.

[0052] Furthermore, obtaining the recognition result and the judgment basis includes:

[0053] Set prompt words;

[0054] Add prompt words to URL data;

[0055] The URL data after adding the prompt words is input into the adjusted fraud identification model for identification to obtain the identification result and judgment basis; the fraud identification model adopts an open source basic visual large model.

[0056] Furthermore, the adjustment of the fraud identification model includes:

[0057] Collect URL data and set recognition prompt words;

[0058] Fill the collected URL data with recognition prompt words to build a fine-tuning corpus;

[0059] Fine-tune the fraud identification model based on the fine-tuning corpus.

[0060] Specifically, in this embodiment, a URL data collection module is used to collect URL data to be tested. Given a URL, various information is collected, including web page content, page screenshots, web page icons, and ICP registration information.

[0061] The data collection process is as follows Figure 1 As shown in the figure, we first implemented a web crawler to automatically access websites and collect information using the Chrome DevTools Protocol. Given an input URL, the web crawler retrieves the URL of the web page reached (the final destination after any redirects), loads the HTML after JavaScript execution, obtains the page icon, and captures a screenshot image. The reason for obtaining HTML after JavaScript execution is that some fraudulent websites use obfuscated JavaScript to generate DOM elements as a means of evading analysis. Before JavaScript is executed, it becomes challenging to analyze any traces of fraud based solely on HTML. When crawling, the configuration simulates two different environments: Google Chrome on Windows and Safari on iPhone. These configurations include specifying the User-Agent and browser size for each environment. Then call the ICP registration information query API to crawl the registration information of the domain name corresponding to the URL. Finally, the crawled data is stored in a csv file.

[0062] Specifically, in this embodiment, the construction of the fraud-related URL identification module based on LLM includes:

[0063] The dataset generated by the URL data collection template is processed. First, recognition prompt words are designed, URL data is filled into the prompt words, and a fine-tuning corpus is constructed. Then, an open source basic large model is selected for fine-tuning to realize fraudulent URL recognition based on LLM.

[0064] As the direct input of LLM, the prompt word is one of the key factors affecting the accuracy of model output. The prompt word design guides LLM through instructions according to task requirements, then takes the relevant information of the URL as input and the URL label and judgment basis as output.

[0065] In this embodiment, the step of setting the recognition prompt word is the work of the prompt word engineering. How to set the prompt word can be as follows:

[0066] The recognition prompt words are designed through prompt word engineering. Based on the Chain of Thought (CoT), the fraudulent URL identification task is divided into four subtasks to facilitate the execution of specific reasoning processes. The purpose is to provide LLM with relevant information of the URL to be detected and control the LLM reasoning process to determine whether it is a fraudulent URL.

[0067] 1. Analyze whether the site is suspected of being a counterfeit website. Counterfeit websites usually imitate the appearance and content of official websites, but there may be slight differences or errors. Check whether the URL has spelling errors, special characters and other abnormal information, whether the site icon is wrong, and whether there are slight differences in the page content. If these factors exist, it indicates that the URL is likely to be a fraudulent URL.

[0068] 2. Analyze whether there are fraudulent elements on the web page. Fraudulent websites are disguised through various methods and it is difficult to identify them from a single piece of information. Therefore, this study comprehensively checks whether there are fraudulent elements on the web page from various relevant aspects such as page content and title.

[0069] 3. Analyze whether the domain name corresponding to the URL is suspected of being a malicious domain name, and determine whether the domain name is compliant through ICP information. If there is any illegal information, it means that the domain name is at risk.

[0070] 4. Based on the above analysis, determine whether the URL to be tested is fraudulent and clearly state the reasons for the judgment. LLM is required to provide a detailed explanation to improve the accuracy of the subsequent verification and classification stage.

[0071] Through the above 4 subtasks, detailed analysis thinking is given to LLM, so that LLM can analyze the provided URL related information more comprehensively and carefully, thereby achieving better recognition. In summary, the template for identifying prompt words is shown in the following table:

[0072]

[0073]

[0074] This embodiment fills the information corresponding to the URL data set into the recognition prompt words on the basis of the recognition prompt words to construct a multimodal fine-tuning corpus.

[0075] In this embodiment, the open source basic visual large model Qwen-VL with 7 billion parameters is selected, and then it is fine-tuned using LoRA in parameter efficient fine-tuning, so as to adjust the ability of Qwen-VL to the ability of fraud-related URL recognition. LoRA will freeze all pre-trained weight parameters of the original model Qwen-VL, and introduce a low-rank matrix to approximate the changes in the model weights after fine-tuning on a specific task while keeping the original pre-trained weights unchanged, thereby reducing the number of parameters and computing resources required for training.

[0076] Determine fine-tuning parameters based on hardware resources. The process of fine-tuning using LoRA involves hyperparameters such as LoRA rank and learning rate. The main parameters are shown in Table 1 below:

[0077] Table 1 Main parameters

[0078]

[0079]

[0080] The process of identifying whether a URL is a fraudulent URL is as follows: input the URL to be tested, first collect the relevant information of the URL to be tested through the URL data collection module, then inject it into the identification prompt word, and then pass it to the fraud identification LLM, and finally output the result. The result not only includes the identification result of the URL to be tested, but also includes the judgment basis.

[0081] Furthermore, verifying the recognition result includes:

[0082] Construct a fraud-related knowledge vector library;

[0083] According to the judgment basis, the knowledge vector library is searched to obtain the corresponding knowledge vector;

[0084] After filling the knowledge vector with verification classification prompt words, the knowledge vector is input into the fraud identification model for analysis and judgment, and the verification result is output.

[0085] Furthermore, constructing a fraud-related knowledge vector library includes:

[0086] Collect text data related to fraud;

[0087] Perform semantic vectorization on fraud-related text data;

[0088] Build an index for the vectorized text data and write it into the database to generate a fraud-related knowledge vector library.

[0089] Furthermore, semantic vectorization of fraud-related text data includes:

[0090] The collected fraud-related text data is segmented at the sentence granularity, and the complete semantics of a sentence is retained for text segmentation, and then semantic vectorization is performed.

[0091] Further, according to the judgment basis, the search knowledge vector library includes:

[0092] The keywords used as the basis for judgment are input into the embedding model for vectorization processing;

[0093] In the knowledge vector library, search for knowledge texts with preset semantic similarity to the keyword vector to obtain the corresponding knowledge vector.

[0094] Furthermore, searching for knowledge texts with a preset semantic similarity to the keyword vector includes:

[0095] The cosine similarity is used to calculate the similarity between the semantics of the keyword vector and the knowledge vector in the knowledge vector library, and the knowledge vector with the preset cosine similarity is selected.

[0096] Specifically, in this embodiment, the fraud-related URL verification classification module based on RAG is constructed including:

[0097] A large fraud identification model is used to identify whether a URL is fraudulent. There is no suitable method to verify the correctness of the identification result, and manual verification is unrealistic. Therefore, the present invention uses RAG technology for verification and realizes more specific classification of fraud types. In addition, the use of RAG technology can update the latest knowledge base in real time, which can effectively avoid the problem of data drift to a certain extent.

[0098] RAG is one of the most popular LLM application solutions, which can effectively solve the problems of LLM knowledge limitation and illusion. RAG retrieves relevant knowledge and integrates it into prompt words, so that LLM can refer to the corresponding knowledge and give reasonable answers. Therefore, the core of RAG can be understood as "retrieval + generation". The former mainly uses the efficient storage and retrieval capabilities of the vector database to recall the target knowledge; the latter uses LLM and prompt word engineering to reasonably use the recalled knowledge and generate the target answer.

[0099] To implement fraud-related URL identification, verification and classification based on RAG, we must first build a fraud-related knowledge vector library, then search the fraud-related knowledge vector library based on the judgment basis in the output results of the fraud-related identification model, fill in the prompt words, and pass it to the fraud-related identification model for analysis and judgment again, so as to output the verification result. If the result is correct, the fraud-related type is returned.

[0100] Construction of Fraud-Related Knowledge Vector Database

[0101] The process of constructing a fraud-related knowledge vector library is as follows: Figure 2 shown.

[0102] (1) Collect fraud-related information. Collect authoritative, high-quality text data related to fraud through channels such as fraud case notices, fraud news disclosed by the media, and fraud-related laws and regulations to ensure the effectiveness of the knowledge vector library.

[0103] (2) Semantic vectorization: the collected fraud-related text data is segmented at the granularity of "sentences", the complete semantics of a sentence is retained for text segmentation, and then semantic vectorization is performed. Vectorization is a process of converting text data into a vector matrix, which directly affects the effect of subsequent retrieval. The present invention selects the open source Embedding model M3E to vectorize text data.

[0104] (3) Data storage: The process of constructing an index after vectorizing text data and writing it into the database can be summarized as the data storage process. The present invention uses the vector database Chroma to store data. Chroma is a vector database for building AI applications with embedded vectors. It can represent vector data of modalities such as text and images, and is small and lightweight.

[0105] Fraudulent URL Verification Classification Based on RAG

[0106] According to the recognition results of the fraud identification model, the most relevant knowledge is recalled through efficient retrieval methods, and verification and classification prompt words are injected; the large model refers to the current question and related knowledge to generate corresponding answers, thereby verifying and classifying the fraud URL recognition results. The key links include data retrieval, injection of prompt words, and recognition result verification and classification. Verification and classification prompt words belong to the scope of RAG technology, and no thinking chain is used.

[0107] Data Retrieval

[0108] After the preparations in the above steps, you can start to process the related queries of the fraud identification results. First, the keywords used as the basis for the judgment of fraudulent URL identification will be input into the embedding model for vectorization. Then, the system will search the vector database for knowledge texts that are semantically similar to the keyword vector, that is, calculate the similarity score between the query vector and all stored vectors, and return the top-k records with high scores.

[0109] When the present embodiment uses cosine similarity to calculate similarity, cosine similarity is data used to measure the similarity between the associated corpus and the related corpus. The related corpus refers to a corpus with a higher cosine similarity (e.g., greater than a cosine threshold) with the associated corpus in the vectorized corpus. In some embodiments, the text generation module can use a corpus with a cosine similarity greater than a cosine threshold as a related corpus. The higher the cosine similarity, the greater the similarity between the two corpuses.

[0110] For vector A and vector B, Where A·B represents the dot product of vector A and vector B, and ||A|| and ||B|| are the norms of vector A and vector B respectively.

[0111] Injection prompt word

[0112] In the RAG scenario, prompt words generally include task description, background knowledge (retrieved knowledge), task instructions (fraudulent URL identification results and judgment basis), etc. According to the task scenario and LLM performance, other instructions can also be appropriately added to the prompt words to optimize the LLM output. The prompt words for the fraudulent URL identification and verification classification based on RAG are shown in the following table:

[0113]

[0114] Recognition result verification classification

[0115] In this embodiment, LLM uses its emergence ability and strong semantic understanding ability to learn the background knowledge in the prompt words, and then understands the fraudulent URL identification provided by this example, and logically thinks about the background knowledge and identification situation, so as to determine whether the identification result is consistent with its corresponding authoritative knowledge. If the identification result is correct, the specific fraudulent type of the fraudulent URL is further determined based on the background knowledge. The complete fraudulent URL verification and classification process is shown in the attached figure. Figure 3 shown.

[0116] Through the construction of the above three modules, a complete fraud-related URL classification and identification method can be constructed. The process is as follows: Figure 4 First, the URL to be tested input by the user is obtained, and then the relevant information of the URL is collected and passed to the fraud identification model for identification, the identification result and judgment basis are obtained, and then the judgment basis is passed to the RAG module for verification and classification, thereby realizing the classification and identification of fraudulent URLs.

[0117] This embodiment proposes a fraud-related URL identification and classification method based on LLM-RAG, which realizes identifiable, explainable, verifiable, and classifiable identification. There are two key points:

[0118] 1. Use LLM to make fraudulent URLs identifiable and explainable. Collect the pages, icons, webpage screenshots, ICP registration and other information of the URL to be tested as raw data, then inject prompt words, use the fine-tuned LLM to achieve fraudulent identification, and output the judgment basis at the same time.

[0119] 2. Use RAG technology to make the identification of fraud-related URLs verifiable and classifiable. By building an authoritative knowledge vector library related to fraud, the judgment basis of fraud-related URLs is used as keywords. Use RAG to retrieve records with high similarity and inject them into the background knowledge of the verification classification prompt words. Use LLM for verification. If the verification is successful, classification is performed and the classification results are output.

[0120] This embodiment provides a method for identifying fraudulent URLs that is identifiable, explainable, verifiable, and classifiable. Compared with other methods for identifying fraudulent URLs, this embodiment uses the emergence capability of LLM to provide powerful understanding and reasoning capabilities. The fine-tuned LLM can identify and provide an interpretable basis for judgment based on the provided URL-related information; in addition, compared with other methods, this embodiment verifies the recognition results through RAG technology, which can provide more reliable recognition results.

[0121] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for classifying and identifying fraudulent URLs based on LLM-RAG, characterized in that: include: Get the URL data to be tested; Input the URL data into a fraud identification model for identification, and obtain identification results and judgment basis, wherein the fraud identification model is constructed by LLM; The RAG method is used in combination with the judgment basis to verify the recognition result and obtain a verification result.

2. The fraud-related URL classification and identification method based on LLM-RAG according to claim 1 is characterized in that: The URL data includes: web page content, page screenshots, web page icon, and ICP registration information.

3. The fraud-related URL classification and identification method based on LLM-RAG according to claim 1 is characterized in that: Obtaining identification results and judgment basis include: Set recognition prompt words; Injecting the identification prompt word into the URL data; The URL data after the recognition prompt word is injected is input into the adjusted fraud-related recognition model for recognition to obtain the recognition result and judgment basis.

4. The fraud-related URL classification and identification method based on LLM-RAG according to claim 3 is characterized in that: Adjusting the fraud identification model includes: Collecting preset URL data and setting the identification prompt words; wherein the preset URL data includes: fraud-related URL data and normal URL data; Filling the recognition prompt words in the collected preset URL data to construct a fine-tuning corpus; The fraud identification model is adjusted based on the fine-tuning corpus.

5. The fraud-related URL classification and identification method based on LLM-RAG according to claim 1 is characterized in that: Verifying the recognition result includes: Construct a fraud-related knowledge vector library; According to the judgment basis, searching the knowledge vector library to obtain the corresponding knowledge vector; After the knowledge vector is filled with verification classification prompt words, the fraud identification model is input for analysis and judgment, and the verification result is output.

6. The fraud-related URL classification and identification method based on LLM-RAG according to claim 5 is characterized in that: Building a fraud-related knowledge vector library includes: Collect text data related to fraud; Performing semantic vectorization on the fraud-related text data; Build an index for the vectorized text data and write it into the database to generate a fraud-related knowledge vector library.

7. The fraud-related URL classification and identification method based on LLM-RAG according to claim 6 is characterized in that: The semantic vectorization of the fraud-related text data includes: The collected fraud-related text data is segmented at the sentence granularity, the complete semantics of a sentence is retained for text segmentation, and then semantic vectorization is performed.

8. The fraud-related URL classification and identification method based on LLM-RAG according to claim 5 is characterized in that: According to the judgment basis, searching the knowledge vector library includes: Input the judgment basis as a keyword into the embedding model for vectorization processing; In the knowledge vector library, knowledge texts having a preset similarity with the semantics of the keyword vector are searched to obtain the corresponding knowledge vector.

9. The fraud-related URL classification and identification method based on LLM-RAG according to claim 8 is characterized in that: Searching for knowledge texts with preset semantic similarity to keyword vectors includes: The cosine similarity is used to calculate the similarity between the semantics of the keyword vector and the knowledge vectors in the knowledge vector library, and the knowledge vectors with a preset cosine similarity are selected.

Citation Information

Patent Citations

  • Anti-phishing method based on multi-factor comprehensive assessment method

    CN104202291A

  • Fraud website identification method and device based on multi-modal fusion

    CN116722992A

Cited By

  • Method and device for training large language model to identify high-risk webpage service, electronic equipment, medium and program product

    CN119807687A