Malicious cognitive attack organization mining method and device based on text detection and weight distribution
The method improves the identification and clustering of malicious cognitive attack organizations through data collection, advanced text classification, and organizational clustering, addressing the inadequacies of existing security defenses in detecting organized cognitive threats.
Patent Information
- Application Number
- CN202411011073.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-12-06
AI Technical Summary
Existing security defenses are inadequate to identify and block complex cognitive attack scenarios, particularly those driven by organized groups, which pose hidden and wide-ranging threats to social stability and public psychology, necessitating improved methods to detect and understand the operating mode and organizational structure of malicious organizations.
A method involving data collection, formation of a comprehensive black word library, use of a HAN model for text classification, and Louvain algorithm for organizational clustering, with weight allocation rules to identify and cluster malicious cognitive attack organizations based on user interactions and topic relationships.
Enhances the accuracy and efficiency of identifying malicious cognitive attack organizations by leveraging advanced data collection, timely keyword libraries, model training, and effective clustering, ensuring high reliability and applicability.
Smart Images

Figure CN119088957A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of natural language processing and social network cognitive attack, and in particular to a method and device for mining malicious cognitive attack organizations based on text detection and weight distribution. Background Art
[0002] In today's society where global informatization and networking are developing rapidly, cyberspace has become an indispensable part of information exchange, business transactions, social activities, etc. With the advancement of network technology, various network security threats have also emerged and evolved. Among them, cognitive attacks have become an emerging and extremely challenging security threat. Cognitive attackers influence the cognitive judgment of individuals or groups by manipulating network information, spreading false information, disguising identities, etc., thereby achieving the purpose of attack. Its harm is not limited to the information level, but is more likely to have a long-term impact on social stability and public psychology. Compared with traditional network security threats, cognitive attacks are more hidden and have a wider impact.
[0003] With the popularity of social media and various online platforms, information is spreading faster and faster, and people have more and more channels to obtain information, which provides attackers with a wider space to play.
[0004] In reality, malicious activities are often not carried out independently by a single user. Deeper cyber threats are usually driven by highly organized groups with clear goals. Therefore, we need to pay more attention to those malicious organizations. These malicious organizations use multiple means to control and manipulate online public opinion, which may cause greater social impact and chaos. These organizations are often hidden in the massive amount of information on the Internet, and have their own unique operating mode and complex organizational structure, which brings greater challenges to the field of network security.
[0005] Therefore, with the diversification of cognitive attack methods and the specificity of targets, how to accurately identify and timely block cognitive attack activities from massive network information has become an urgent problem to be solved in the current network security field. In addition, digging out the malicious organizations behind them and understanding their operating mode and organizational structure are of great significance to completely eliminate the root causes of cognitive attacks.
[0006] In this context, traditional security defense means are no longer sufficient to cope with increasingly complex cognitive attack scenarios. New methods and technologies are needed to identify, analyze and defend against this new type of network threat. This invention is proposed based on this background, aiming to identify and mine the malicious cognitive attack organizations behind through effective data collection, model training, weight allocation and organizational clustering, and provide a new defense strategy for network security. Summary of the invention
[0007] In response to the above technical problems, the present invention provides a method and device for mining malicious cognitive attack organizations based on text detection and weight allocation, which can effectively mine malicious cognitive attack organizations.
[0008] The present invention provides the following technical solutions:
[0009] The first aspect of the present invention relates to a method for mining malicious cognitive attack organizations based on text detection and relationship weight allocation, comprising the following steps:
[0010] (1) collecting information data related to cognitive attacks from the Internet, wherein the information data includes a collection of text information and user information;
[0011] (2) Since cognitive attack speech often contains specific keywords, we collected existing black words from the Internet and combined them with the latest black words to form a relatively complete black word library.
[0012] (3) Through keyword matching algorithms combined with a small amount of manual correction, most of the real cognitive attack texts and corresponding users are found.
[0013] (4) The HAN model is trained and tested multiple times using most of the known positive samples and unknown samples, and finally all cognitive attack sample data is identified and the corresponding cognitive attack users are found.
[0014] (5) A designed weight allocation rule is used to assign the most reasonable weights to the cognitive attack user behavior relationship and topic relationship, and the classic Louvain algorithm is used to organize clustering.
[0015] In the present invention, the design coverage of the weights described in step (5) includes:
[0016] (a) Following relationship: Following usually means a long-term connection between two users;
[0017] (b) Like: Like may indicate that the user likes and approves a piece of content;
[0018] (c) Reply: This may indicate the user’s level of participation in a topic, and more likely indicates the user’s desire to keep the topic going;
[0019] (d) The number of interactions between two users within a certain period of time indicates the closeness of their connection during this period, and the weight will increase as the number of interactions increases;
[0020] (e) the influence of the comments (the sum of the number of likes, replies, and reposts);
[0021] (f) Topic overlap between two users: the number of topics of the same type between users;
[0022] (g) If the number of common topics between two users exceeds a certain threshold within a period of time, the weight will be appropriately increased.
[0023] In step (1), information data related to cognitive attacks is collected from the Internet, including:
[0024] The users are expanded from the follow lists and fan lists of ten manually identified cognitive attack users, and the comments made by the users and the number of likes, reposts, and comments corresponding to each comment and the corresponding user name are collected.
[0025] In step (2), existing black word libraries are collected from the Internet and combined with the latest black words collected to form a relatively complete black word library. This includes: black word libraries collected by scholars on the Internet many years ago, and the latest black words summarized from the latest cognitive attack remarks on the Internet, and finally a relatively complete black word library that adapts to the latest current events is formed.
[0026] In step (3), the keyword matching algorithm combined with a small amount of manual correction is used to find most of the real cognitive attack texts and corresponding users. This includes:
[0027] The speech texts in all the collected data are matched with the sensitive words in the above-mentioned black word library through the keyword matching algorithm, and are identified as cognitive attack texts. The users who post, like, forward, and follow this text are also cognitive attack users.
[0028] In step (4), the HAN model is trained and tested multiple times, including
[0029] (4-i) The text containing sensitive words is fixed as the positive sample, and sample 1 equal to the positive sample is selected from the remaining N unknown samples as the negative sample each time to obtain the training set to train the HAN model.
[0030] (4-ii) The remaining samples N-1 in the above unknown samples are used as the test set. After each training, the test set is tested;
[0031] Repeat the above process multiple times, taking different unknown samples as negative samples. Each unknown sample will have N-1 prediction probabilities. We select the average of the N-1 probabilities as the final prediction probability of this sample.
[0032] (4-iii) Set the test threshold to 0.8, and mark the samples with a probability greater than 0.8 as positive samples, otherwise mark them as negative samples;
[0033] (5) Assign reasonable weights to different behavior relationships and topic relationships through weight allocation rules, and use the Louvain algorithm to organize clustering, including:
[0034] (5-I) Weight design rules:
[0035] Post influence: a
[0036] The number of interactions between two users in one month is b
[0037] Number of posts sent by a user in a month c
[0038] User A's topic type is topic a (An array containing all topic types in which user A participates) [1, 2, 4, 6]
[0039] User B's topic type is topic b (An array containing all topic types in which user B participates)[4,6]
[0040] The number of posts with the same topic is d
[0041] Initial weight: W0
[0042] Attention weight between users: W1
[0043] Weight of likes between users: W2
[0044] Reply weight between users: W3
[0045] Forwarding weight between users: W4
[0046] Topic weight between users: W5
[0047] The total weight between users: W = W1 + W2 + W3 + W4 + W5
[0048] 1. If there is one-way attention between two users, then W1 = W0.
[0049] 2. If there is two-way attention between two users, then W1=2W0.
[0050] 3. If a post a<1000, and b>3 between two users, then W2=W2+0.2W0, W3=W3+0.4W0, W4=W4+0.6W0
[0051] 4. If a post a<1000, and b≤3 between two users, then W2=W2+0.1W0, W3=W3+0.2W0, W4=W4+0.3W0
[0052] 5. If a post a≥1000, and b / c>0.2, then W2=W2+0.2W0, W3=W3+0.4W0, W4=W4+0.6W0
[0053] 6. If a post a≥1000, and b / c≤0.2, then W2=W2+0.1W0, W3=W3+0.2W0, W4=W4+0.3W0
[0054] 7. If topic a ∩topic b >2, d>3, then W5=W5+0.4W0
[0055] 8. If topic a ∩topic b >2, d≤3, then W5=W5+0.3W0
[0056] 9. If topic a ∩topic b ≤2, d>3, then W5=W5+0.2W0
[0057] 10. If topic a ∩topic b ≤2, d≤3, then W5=W5+0.1W0
[0058] (5-II) applying the weight design rule to cognitive attack user data to calculate the interaction relationship weights between all users;
[0059] (5-III) Apply the obtained interaction relationship weights to the Louvain algorithm to obtain the corresponding cognitive attack group clustering, and finally measure the clustering effect by modularity, including:
[0060] (a) In the first stage, the Louvain algorithm tries to maximize modularity by moving nodes to the communities to which their neighboring nodes belong.
[0061] ΔQ=(∑in+2k i,in )-(∑tot+k i ) 2 -(∑tot) 2
[0062] where ∑tot is the total weight of the edges within community C; k i,in is the total weight of the edge between node i and the nodes in community C; ∑tot is the total degree of nodes in community C; k i is the degree of node i;
[0063] (b) The Louvain algorithm aggregates the communities identified in the first phase as new nodes, constructs a new network, and repeats the process of the first phase on this new network.
[0064] The second aspect of the present invention relates to a malicious cognitive attack organization mining device based on text detection and weight allocation, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the malicious cognitive attack organization mining based on text detection and weight allocation of the present invention.
[0065] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a malicious cognitive attack organization mining based on text detection and weight distribution of the present invention.
[0066] The working principle of the present invention is (analyzing the reasons for the advantages of the present invention): The present invention realizes the mining of malicious cognitive attack organizations through the following steps. First, information data related to cognitive attacks are collected from the Internet, including text information and user information, user data is expanded and interactive data is collected. Next, a relatively complete black word library is formed by combining the existing black word library and the latest black words to ensure the comprehensiveness and timeliness of the keywords. Then, most of the cognitive attack texts and corresponding users are identified through the keyword matching algorithm combined with a small amount of manual correction. After that, the HAN model is trained using the method of fixed positive samples and multiple selections of negative samples, and unknown samples are tested and the average prediction probability is calculated to identify all cognitive attack samples. Finally, by designing weight allocation rules, reasonable weights are given to the behavioral relationships and topic relationships of cognitive attack users, and the Louvain algorithm is used for organizational clustering, thereby effectively identifying malicious cognitive attack organizations.
[0067] The advantages of the present invention are: the present invention significantly improves the accuracy and efficiency of identifying malicious cognitive attack organizations through comprehensive data collection, the latest and complete black word library, efficient text recognition, advanced model training, reasonable weight allocation and effective organization clustering. Specifically, by collecting user interaction data, a comprehensive understanding of user relationships; combining the latest black word library to ensure the accuracy of recognition; through keyword matching and a small amount of manual correction, a large number of attack texts can be quickly identified; the HAN model is used for multiple training and testing to improve the generalization ability of the model; weight allocation rules are designed according to the relationship between behavior and topic to ensure the rationality of weights; and the Louvain algorithm is used for efficient clustering to identify malicious organizations. In summary, the present invention has high efficiency, accuracy and reliability, and has significant application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] Figure 1 This is a flow chart of a method for mining malicious cognitive attack organizations based on text detection and weight allocation according to the present invention;
[0069] Figure 2 It is a flow chart of the data collection module of the present invention;
[0070] Figure 3 A schematic diagram of the process of cognitive attack user detection of the present invention;
[0071] Figure 4 A schematic diagram of the process of cognitive attack organization mining of the present invention. DETAILED DESCRIPTION
[0072] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.
[0073] Example 1
[0074] This embodiment provides a method for mining malicious cognitive attack organizations based on text detection and weight allocation. Figure 1 As shown, including:
[0075] The data collection and black word library construction module collects information data that may be related to cognitive attacks from the Internet, including the text content of comments posted by users and behavioral relationships such as following, liking, forwarding, and replying; in addition, the existing black word library and the latest black words collected manually are combined into a relatively complete and popular black word library.
[0076] The cognitive attack user identification module uses a rule-based sensitive word matching method and a HAN neural network model to classify the collected comment content and obtain cognitive attack users.
[0077] The cognitive attack organization clustering module assigns different weights to different interactive relationships by designing weight distribution rules to calculate the weight between every two users, and then uses the Louvain algorithm to cluster users into cognitive attack organizations according to the weights.
[0078] In the data collection and black word library construction module, the following steps are included:
[0079] (1) The method for collecting information data is as follows: based on the 10 manually identified cognitive attack users, their follow and fan lists are expanded, and their corresponding likes, reposts, and replies are extended to the corresponding users, and the users involved, the corresponding text content and related information are crawled.
[0080] (2) Before using the keyword matching method to identify cognitive attack users, it is necessary to prepare a blacklist containing the latest sensitive words, including the following steps:
[0081] (2-1) We searched for open-source sensitive word libraries from previous relevant research papers. Most of these sensitive word libraries were created several years ago, and the black words involved were mainly common black words from many years ago.
[0082] (2-2) Manually search for some latest cognitive attack text content and related reports from Twitter and search engines, summarize the latest sensitive words that conform to the current situation from them, add them to the existing sensitive word library, and combine them into a new complete sensitive word library.
[0083] Cognitive attack user identification: After obtaining the text that may be involved in cognitive attacks and the latest sensitive word library, it is necessary to classify whether these texts are cognitive attacks, including the following steps:
[0084] (1) All the collected speech texts usually include various unstructured data, such as punctuation marks, stop words, special symbols, etc. Before performing keyword matching, it is necessary to preprocess the text. The specific steps include:
[0085] (1-1) Remove all punctuation marks and special characters in the text to standardize the text content.
[0086] (1-2) Use the natural language processing tool jieba to split the text into individual words (tokens) for keyword matching.
[0087] (1-3) Remove stop words: Use the open-source stop word library on the Internet to remove common stop words (such as "de", "shi", "zai", etc.), which have no practical significance for identifying cognitive attacks.
[0088] (2) Since the content of cognitive attacks often contains specific sensitive words, the preprocessed text is matched and identified according to the above-mentioned sensitive word library through a sensitive word matching algorithm, including the following steps:
[0089] (2-1) For all the collected review text content, apply the keyword matching algorithm, traverse each preprocessed speech text, and for each word in each text, check whether the word exists in the black word library.
[0090] (2-2) If the text contains any sensitive vocabulary in the black word library, mark the text as a cognitive attack text and record the user information corresponding to the text, including users who perform behaviors such as publishing, liking, forwarding, and following.
[0091] (3) Since the pure keyword matching method may have misjudgments and missed judgments, and will ignore the semantic information in the text, it is necessary to combine the HAN neural network model for identification. The specific steps include:
[0092] (3-1) Fix the text containing sensitive vocabulary as positive samples. The selection of positive samples is based on the cognitive attack texts identified in the previous step. Randomly select samples with the same number as the positive samples from the remaining unknown samples as negative samples. This process will be repeated multiple times to cover different negative sample sets and ensure the reliability of the prediction.
[0093] (3-2) HAN (Hierarchical Attention Networks) model is a deep learning model for text classification. It encodes the hierarchical structure of text through a hierarchical attention mechanism. Its main structure includes word-level attention layer and sentence-level attention layer. The positive samples and selected negative samples form a training set. The HAN model is trained using the training set, and the model parameters are optimized through the back-propagation algorithm to minimize the classification error. Different negative samples are used in each training to ensure that the model can learn different types of non-attack text features.
[0094] (3-3) The remaining samples that are not selected as negative samples in each training are used as the test set. The model after each training is tested using the test set. The predicted probability of each test sample is recorded. Repeat the training and testing multiple times, and each sample will have multiple predicted probabilities. Calculate the average of the predicted probabilities of each test sample as the final predicted probability.
[0095] (3-4) Set the probability threshold for predicting a positive sample to 0.8. If the final predicted probability of a sample is greater than 0.8, it is marked as a positive sample (cognitive attack text), otherwise it is marked as a negative sample (non-cognitive attack text).
[0096] In the cognitive attack organization clustering module, weights are assigned to cognitive attack users through weight assignment rules, and the Louvain algorithm is applied for clustering, including the following steps:
[0097] (1) Weight design rules include:
[0098] (a) If there is one-way attention between two users, then W1 = W0.
[0099] (b) If there is a two-way attention between two users, then W1 = 2W0.
[0100] (c) If a post a < 1000, and b > 3 between two users, then W2 = W2 + 0.2W0, W3 = W3 + 0.4W0, W4 = W4 + 0.6W0
[0101] (d) If a post a<1000, and b≤3 between two users, then W2=W2+0.1W0, W3=W3+0.2W0, W4=W4+0.3W0
[0102] (e) If a post a ≥ 1000 and b / c > 0.2, then W2 = W2 + 0.2W0, W3 = W3 + 0.4W0, W4 = W4 + 0.6W0
[0103] (f) If a post a ≥ 1000 and b / c ≤ 0.2, then W2 = W2 + 0.1W0, W3 = W3 + 0.2W0, W4 = W4 + 0.3W0
[0104] (g) If topic a ∩topic b >2, d>3, then W5=W5+0.4W0
[0105] (h) If topic a ∩topic b >2, d≤3, then W5=W5+0.3W0
[0106] (i) If topic a ∩topic b ≤2, d>3, then W5=W5+0.2W0
[0107] (j) If topic a ∩topic b ≤2, d≤3, then W5=W5+0.1W0
[0108] Total weight W = W1 + W2 + W3 + W4 + W5
[0109] (2) Apply the Louvain algorithm for clustering, including the following steps:
[0110] (2-1) Apply the weight design rules to the cognitive attack user data and calculate the interaction relationship weights between all users.
[0111] (2-2) In the first stage, the Louvain algorithm attempts to maximize modularity by moving nodes to the communities to which their neighboring nodes belong.
[0112] ΔQ=(∑in+2k i,in )-(∑tot+k i ) 2 -(∑tot) 2
[0113] where ∑tot is the total weight of the edges within community C; k i,in is the total weight of the edge between node i and the nodes in community C; ∑tot is the total degree of nodes in community C; k i is the degree of node i;
[0114] (2-3) In the second stage, the communities identified in the first stage are aggregated as new nodes to build a new network, and the first stage process is repeated on the new network until the modularity no longer increases significantly. The clustering effect is measured by the final modularity to obtain the final cognitive attack group clustering results.
[0115] Example 2
[0116] The present embodiment relates to a malicious cognitive attack organization mining device based on text detection and weight allocation, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a malicious cognitive attack organization mining based on text detection and weight allocation of embodiment 1.
[0117] Example 3
[0118] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the malicious cognitive attack organization mining based on text detection and weight distribution of Embodiment 1 is implemented.
[0119] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for mining malicious cognitive attack organizations based on text detection and weight allocation, comprising the following steps: (1) collecting information data related to cognitive attacks from the Internet, wherein the information data includes a collection of text information and user information; (2) Collect existing black words and combine them with the latest black words to form a relatively complete black word library; (3) Through keyword matching algorithms and a small amount of manual correction, most of the real cognitive attack texts and corresponding users are found; (4) Use most of the known positive samples and unknown samples to train the HAN model multiple times and conduct tests, and finally identify all cognitive attack sample data and find the corresponding cognitive attack users; (5) Through the designed weight distribution rules, reasonable weights are assigned to the behavioral relationships and topic relationships of cognitive attack users, and the Louvain algorithm is used to organize and cluster them.
2. The method according to claim 1, characterized in that Step 1 specifically includes: expanding users from the follower lists and fan lists of N manually identified cognitive attack users, collecting the comments made by the users and the number of likes, reposts, comments and corresponding user names for each comment.
3. The method according to claim 1, characterized in that: Step 2 specifically includes: collecting the black word library collected by scholars on the Internet many years ago, and summarizing the latest black words from the latest cognitive attack remarks on the Internet, to form a relatively complete black word library that adapts to the latest current events.
4. The method according to claim 1, characterized in that Step 3 specifically includes: matching all collected speech texts with sensitive words in the black word library through a keyword matching algorithm, and identifying them as cognitive attack texts. The users who post, like, forward, and follow the texts are also identified as cognitive attack users.
5. The method according to claim 1, characterized in that Step 4 specifically includes: (4-i) Fix the text containing sensitive words as positive samples, and select the same number of samples as positive samples from the remaining unknown samples as negative samples each time to obtain the training set to train the HAN model; (4-ii) Use the remaining samples from the above unknown samples as the test set, and test the test set after each training; (4-iii) Repeat the above process multiple times, taking different unknown samples as negative samples. Each unknown sample will have multiple prediction probabilities, and take the average value as the final prediction probability of this sample; (4-iv) Set the test threshold to 0.8, and mark the samples with a probability greater than 0.8 predicted as positive samples as positive samples, otherwise they are marked as negative samples.
6. The method according to claim 1, characterized in that Step 5 specifically includes: (5-i) Weight design rules: Post influence: a The number of interactions between two users in one month is b Number of posts sent by a user in a month c User A's topic type is topic a (An array containing all topic types in which user A participates) [1, 2, 4, 6] User B's topic type is topic b (An array containing all topic types in which user B participates)[4,6] The number of posts with the same topic is d Initial weight: W0 Attention weight between users: W1 Weight of likes between users: W2 Reply weight between users: W3 Forwarding weight between users: W4 Topic weight between users: W5 The total weight between users: W = W1 + W2 + W3 + W4 + W5 If there is a one-way attention between two users, then W1=W0 If there is a two-way attention between two users, then W1=2W0 If a post a<1000, and b>3 between two users, then W2=W2+0.2W0, W3=W3+0.4W0, W4=W4+0.6W0 If a post a<1000, and b≤3 between two users, then W2=W2+0.1W0, W3=W3+0.2W0, W4=W4+0.3W0 If a post a≥1000, and b / c>0.2, then W2=W2+0.2W0, W3=W3+0.4W0, W4=W4+0.6W0 If a post a≥1000, and b / c≤0.2, then W2=W2+0.1W0, W3=W3+0.2W0, W4=W4+0.3W0 If topic a ∩topic b >2, d>3, then W5=W5+0.4W0 If topic a ∩topic b >2, d≤3, then W5=W5+0.3W0 If topic a ∩topic b ≤2, d>3, then W5=W5+0.2W0 If topic a ∩topic b ≤2, d≤3, then W5=W5+0.1W0 (5-ii) applying the weight design rule to cognitive attack user data to calculate the interaction relationship weights between all users; (5-iii) Apply the obtained interaction relationship weights to the Louvain algorithm to obtain the corresponding cognitive attack group clustering, and finally measure the clustering effect by modularity, including: In the first phase, the Louvain algorithm tries to maximize modularity by moving nodes to the communities to which their neighbor nodes belong; In the second stage, the Louvain algorithm aggregates the communities identified in the first stage as new nodes, builds a new network, and repeats the process of the first stage on this new network.
7. The method according to claim 1, characterized in that The design coverage of the weights described in step (5) includes: (a) Following relationship: Following means a long-term connection between two users; (b) Like: Like indicates that the user likes and approves a piece of content; (c) Reply: indicates the user’s level of participation in a topic and their desire to keep the topic going; (d) The number of interactions between two users within a certain period of time indicates the closeness of their connection during this period, and the weight increases as the number of interactions increases; (e) The influence of the comment content, calculated as the sum of the number of likes, replies and reposts; (f) Topic overlap between two users: the number of topics of the same type between users; (g) If the number of common topics between two users exceeds a certain threshold within a period of time, their weight is increased.
8. A malicious cognitive attack organization mining device based on text detection and weight allocation, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement a malicious cognitive attack organization mining based on text detection and weight allocation as described in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the malicious cognitive attack organization mining based on text detection and weight distribution as described in any one of claims 1-7 is implemented.
