Social security anti-fraud early warning method based on big data analysis

Abstract

The invention discloses a social security anti-fraud early warning method based on big data analysis, and the method comprises the steps: automatically collecting fraud-related APP data through a big data technology, converting a text into an image form through a text graph feature extraction technology, carrying out the processing, finely extracting text features, and effectively improving the accuracy of fraud information recognition. Meanwhile, in combination with a logistic regression prediction model, high-precision distinguishing of the fraud-related APP and the secure APP is realized. In addition, the OpenCV image processing library is used for fine processing, the image quality is optimized, and the character recognition accuracy is improved. In the aspect of text processing, the problem of semantic fracture caused by cross-page is solved through an accurate calculation and alignment technology, and the readability and continuity of the text are improved. And finally, by introducing a *-shaped identifier and a *-shaped graph library, efficient and accurate screening of key words and key statements is realized, and the adaptability and expandability of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of social security anti-fraud, and in particular to a social security anti-fraud early warning method based on big data analysis. Background Art

[0002] With the continuous advancement of information technology and the in-depth application of the Internet, social media and various applications have become an important part of people's lives. But at the same time, this has also provided a new stage for fraudulent activities, making fraud methods increasingly diverse and elusive. Traditional static rule-based methods are powerless in judging fraudulent information because they often rely on known fraud patterns and are difficult to adapt to the rapid changes in fraud methods. When new fraud methods emerge, these static rules may not be updated in time, resulting in misjudgment and increasing the risk of users being defrauded.

[0003] In addition, many existing technologies require a lot of manual intervention when processing fraudulent information. This not only reduces processing efficiency, but also makes it difficult for the system to identify the latest fraudulent information in a timely manner, making it impossible to effectively respond to the rapid changes in network security threats. The limitation of manual judgment is that its response speed and accuracy are limited by human capabilities, and it is difficult to compete with the ever-changing fraud methods.

[0004] At the same time, with the explosive growth of data volume, existing technologies are unable to cope with the processing and analysis of this data. The collection, organization and analysis of massive data requires powerful data processing capabilities, but existing technologies are often insufficient in this regard. This may cause key information to be missed or misjudged, further reducing the accuracy of identifying fraudulent information.

[0005] Therefore, there is an urgent need for a new social security anti-fraud early warning method based on big data analysis. This method should be dynamically adaptable and be able to automatically learn and adapt to new fraud methods and patterns without the need for frequent manual updates of rules. At the same time, it should also be intelligent and automated to reduce manual intervention and improve processing efficiency. Most importantly, it should have efficient data processing capabilities, be able to process massive amounts of data, quickly extract key features, and ensure that the system can respond in real time and generate accurate early warning information. Such a method will help improve the ability to identify fraudulent information and the efficiency of response, and better protect users' social security and property safety. Summary of the invention

[0006] The present application provides a social security anti-fraud early warning method based on big data analysis. By introducing a social security anti-fraud early warning method based on big data analysis, the application aims to solve the problems that traditional static rule judgment is difficult to adapt to rapidly changing fraud methods, high demand for manual intervention leads to low processing efficiency, and insufficient data processing capabilities lead to omissions or misjudgments of key information. The application can achieve accurate identification and timely warning of fraudulent information, improve users' social security and property safety, and effectively solve the problems existing in the background technology.

[0007] This application provides a social security anti-fraud early warning method based on big data analysis, including:

[0008] S101, using big data technology to automatically collect and organize data related to fraudulent apps;

[0009] S102, based on the extracted features, extracting text and image features from the attachment text and email in the APP;

[0010] S103, comparing the features extracted from the text image with the text images in the keyword library to select keywords;

[0011] S104, after filtering out the keywords, further filtering out the key words and key sentences;

[0012] S105, marking the APPs with key sentences in the attachment text or email as fraudulent APPs with a mark of 1, and marking the safe APPs with a mark of 0, and creating a prediction model to distinguish between fraudulent APPs and safe APPs based on the marks;

[0013] S106, when the user uses the APP, the system sends the user's basic information and real-time behavior data to the trained model for real-time judgment;

[0014] S107, based on the output of the model, mark the APP and generate corresponding warning information.

[0015] Preferably, the S102 includes:

[0016] The text image converts each word in the text into an independent image, and processes it directly in the form of an image without performing text recognition;

[0017] Unify the format of attachment text and email text content;

[0018] Set the image size, image format and center point of each text image;

[0019] Traverse all text images and conduct preliminary screening on each text image, requiring it to meet the set pixel size and image format;

[0020] Initialize a virtual coordinate axis for each qualified text image, and use the point (10,10) in the coordinates as the expected center point of the text image;

[0021] Place the text image in the virtual coordinate axis and determine whether the origin and center of the text image are located at the origin and center of the coordinate axis (10,10);

[0022] For the corrected standard text diagram, record the coordinates of the intersection points in the text diagram.

[0023] Preferably, the S104 includes:

[0024] After the intersection coordinates of the corrected standard text image are identified, the text image is traversed one by one according to the regular text structure of the attachment text and the email. When the intersection coordinate position of the traversed text image completely corresponds to a keyword in the keyword library, the text image is confirmed to be a keyword.

[0025] Preferably, the S104 further includes:

[0026] After the keyword is confirmed, the next text image of the keyword in the traversal order is compared with the keyword library to confirm whether the next text image also matches the keyword in the keyword library. If the coordinate position of the intersection of the next text image and the keyword in the keyword library completely corresponds, then the above two keywords are determined to be keywords;

[0027] After confirming the key words, further confirm whether the keywords traversed subsequently also match the keywords in the keyword library to filter out the key sentences.

[0028] Preferably, the S104 further includes:

[0029] S201, performing single image recognition on multiple image contents of a text page and performing stroke on the single image to remove the background color of the image;

[0030] S202, for each single image, determining a relative angle between a row center line of the multi-image text page and the page;

[0031] S203, adjusting the position of the single image in the multi-image text, and combining cross-page keywords and key sentences within the line center line and a fixed angle range.

[0032] Preferably, S203 includes: the combination of cross-page keywords and key sentences includes a cross-page combination of page breaks in which the end of the previous page and the beginning of the second page are a paragraph, and a cross-page sentence segmentation combination in which the end of a line on the first page and the beginning of the first line on the second page are a paragraph.

[0033] Preferably, the S203 further includes:

[0034] S301, performing feature processing on the key words and key sentences after traversal, screening and combination, retaining the intersection points in the text graph;

[0035] S302, generating a cross symbol at the center point of each character image;

[0036] S303, recording the coordinates of each intersection point where the cross symbol passes through the text pattern;

[0037] S304, comparing the intersection coordinates of the cross-shaped intersection with the cross-shaped intersection word library;

[0038] S305, retaining the cross-shaped text graph of the key words and key sentences according to the matching results.

[0039] Preferably, the S302, the cross mark, includes:

[0040] The cross logo can be a graphic consisting of four line segments intersecting at one point;

[0041] For each selected key word and key sentence, determine the center point position of the key word and key sentence in the virtual coordinate axis (10,10) of the text map;

[0042] At the center point, connect (0,0) and (20,20) through the coordinate axis (10,10), and connect (0,20) and (20,0) through the coordinate axis (10,10) to generate a cross symbol.

[0043] Preferably, the S305 includes:

[0044] S401, comparing the sentences in the rice character graph of the key sentences that are not completely corresponding to each other separately;

[0045] S402, superimposing the rice character graphs of the key sentence, comparing the intersections of the superimposed rice character graphs, and recording the intersections of the superimposed rice character graphs of the key sentence;

[0046] S403, determine keywords and key sentences.

[0047] Preferably, the S403 includes:

[0048] If the accuracy of the intersection of the superimposed rice character image and the intersection in the rice character image library is above the threshold of 85%, it can be determined that the screened words and sentences are keywords and key sentences. If it is lower than the threshold, it is determined whether it is a newly added keyword or key sentence, and the rice character image library is updated.

[0049] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0050] 1. It has demonstrated significant advantages and effects in the feature extraction of text graphs by setting keywords, key words, key sentences, etc. First, in the feature extraction stage, each character in the text is converted into an independent image through an innovative way of text graphics. Without performing text recognition, it is directly processed in the form of an image. This method not only avoids the errors that may be caused by text recognition, but also realizes the fine extraction of text features by recording the coordinates of the intersections in the text graph. This refined feature extraction method provides accurate basic data for the subsequent screening of keywords, key words, and key sentences, greatly improving the accuracy of fraud information identification.

[0051] 2. By integrating big data technology and web crawler technology, this method can capture and summarize fraud-related APP information from multiple channels and dimensions in real time and efficiently, greatly improving the comprehensiveness and timeliness of data collection. At the same time, using advanced text image extraction technology, the text content is converted into image form for feature extraction. This innovative measure not only effectively avoids the format difference problem in traditional text processing, but also significantly enhances the accuracy and robustness of feature recognition. In addition, by constructing a prediction model based on logistic regression and combining the comprehensive analysis of multi-dimensional feature vectors, this method achieves high-precision distinction between fraud-related APPs and safe APPs, providing a strong guarantee for the accuracy of the early warning system.

[0052] 3. By combining the powerful OpenCV image processing library, not only the background color of the image is successfully removed, but also the clear and sharp text strokes are retained. The refined processing of this step significantly improves the accuracy of text recognition. At the same time, the image quality is further optimized using the binarization algorithm and morphological operations, effectively removing noise and enhancing the contrast of the text. This refined image processing technology lays a solid foundation for subsequent text processing, ensuring accurate extraction and clear display of text information even in complex backgrounds or low contrast conditions.

[0053] 4. In terms of text page layout and alignment, this technical process achieves the precise positioning of a single image in a multi-image text page by accurately calculating the center line of the row and determining the relative angle between the image and the page. The completion of this step benefits from the ingenious application of image registration and feature point matching technologies, as well as the precise execution of image rotation correction. The comprehensive application of these technologies ensures that all single images are presented at a unified angle, greatly enhancing the neatness of the page content and the reading experience. In addition, in terms of text integration and presentation, this technical process also realizes the flexible application of page break cross-page combination and cross-page sentence segmentation combination by intelligently identifying and processing cross-page key words and key sentences. This intelligent text integration method not only effectively solves the semantic break problem caused by page breaks in traditional text processing, but also ensures the complete presentation of key words and key sentences through the precise matching of intersection coordinates, thereby improving the readability, coherence, and the accuracy and efficiency of information transmission of the text.

[0054] 5. A cross symbol is added at the center position of the text image of each key word and key sentence. This symbol is formed by four line segments intersecting at a point, resembling the Chinese character 'cross', which neither interferes with the original intersection information nor is not eye-catching enough for subsequent identification. At the same time, the scheme also details the position coordinates where the cross symbol coincides with the intersection of the text image, that is, the 'cross intersection', and generates a text image containing the cross symbol accordingly. To further improve the screening efficiency, the scheme constructs a cross symbol library, which stores the cross intersection coordinates of known key words and key sentences. By comparing the intersection coordinates of the cross text image with the data in the library, the scheme can quickly screen out the key words and key sentences that are exactly the same as the coordinates in the library, thus achieving the goal of efficient and accurate screening.

[0055] 6. By superimposing the cross text image and comprehensively comparing the intersection situation with the cross symbol library, the system can more accurately lock those words and sentences that are truly of key significance, effectively avoiding misjudgment or omission caused by insufficient single-image information. Secondly, this scheme also demonstrates good adaptability and scalability. As the language environment changes continuously and new key words and key sentences emerge continuously, the cross symbol library in the scheme can be automatically updated according to the actual situation, ensuring that the recognition system always maintains the best state and providing a strong guarantee for long-term stable recognition performance. In summary, the technical scheme of the embodiment of this application not only improves the recognition accuracy of key words and key sentences, but also enhances the adaptive ability and scalability of the system, providing more reliable and efficient technical support for applications in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic flowchart of a social security fraud prevention and warning method based on big data analysis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] To facilitate the understanding of the present invention, the present application will be described more comprehensively below with reference to the relevant drawings; the drawings show preferred embodiments of the present invention, but the present invention can be implemented in many different forms and is not limited to the embodiments described herein; on the contrary, the purpose of providing these embodiments is to enable a more thorough and comprehensive understanding of the disclosed content of the present invention.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs; the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention; the term "and / or" used herein includes any and all combinations of one or more related listed items.

[0059] Embodiment 1:

[0060] like Figure 1 As shown, a social security anti-fraud early warning method based on big data analysis includes the following steps:

[0061] S101 uses big data technology to automatically collect and organize relevant data on fraudulent apps.

[0062] Specifically, use big data technology and web crawler technology to regularly capture information related to fraudulent apps from major app stores, social media, forums and other platforms, obtain fraudulent app data provided by relevant agencies (cybersecurity agencies, police), use user feedback and reporting mechanisms to collect fraudulent app information submitted by users, and use the collected information as historical data; the history includes basic information about the app, information about the sender of the information, user behavior data, communication attachment information, and content information;

[0063] The collected historical data is cleaned and preprocessed, and the characteristics of the information are extracted from the cleaned and preprocessed historical data to establish a historical database of information; the characteristics of the information include the sender's IP address, URL link characteristics and keywords.

[0064] S102, based on the extracted features, extract text and image features from the attachment text and emails in the APP.

[0065] Among them, the text image converts each character in the text into an independent image without performing text recognition and directly processes it in the form of an image.

[0066] Specifically, the text content of the attachments and emails is formatted in a unified way, including using a unified font, Songti, and size 4. The entire page of text content is converted into a picture format, and each word in the picture is directly formatted as a text picture (text graphics, i.e., each word is treated as a text picture) without undergoing text recognition, to ensure information consistency and reduce the risk of leaks caused by format differences.

[0067] Set the image size, image format and center point of each text image to ensure the accuracy and consistency of subsequent processing;

[0068] The image size is 20x20 pixels, the image format is PNG, the lower left corner of the text image is the origin (0,0), and the center point of the text image is located at the pixel point in the 10th row and 10th column from the lower left corner.

[0069] Traverse all text images, perform preliminary screening on each text image, and ensure that it meets the set pixel size and image format requirements;

[0070] Initialize a virtual coordinate axis for each qualified text image, and use the point (10,10) in the coordinates as the expected center point of the text image;

[0071] Place the text image in the virtual coordinate axis and determine whether the origin and center of the text image are located at the origin and center of the coordinate axis (10,10); if not, adjust the text image through image transformation (such as translation, rotation) to align its center point with point (10,10) to ensure that all text images are in the same direction;

[0072] For the corrected standard text image, the intersection coordinates in the text image are recorded (ie, the position coordinates of the intersection points of the text strokes in the text image, which are used for feature extraction and recognition).

[0073] S103, comparing the features extracted from the text image with the text images in the keyword library to filter keywords.

[0074] Among them, the keyword library contains a database of commonly used words in fraud and their text-graph features, which is used for screening keywords, key words and key sentences.

[0075] S104, after filtering out the keywords, further filtering out the key words and key sentences.

[0076] Specifically, after the intersection coordinates of the corrected standard text image are identified, the text image is traversed one by one according to the conventional text structure of the attachment text and the email (from left to right, from top to bottom). When the intersection coordinate position of the traversed text image completely corresponds to a keyword in the keyword library, the text image is confirmed to be a keyword;

[0077] After the keyword is confirmed, the next text image of the keyword in the traversal order is compared with the keyword library to confirm whether the next text image also matches the keyword in the keyword library. If the coordinate position of the intersection of the next text image and the keyword in the keyword library completely corresponds, then the above two keywords are determined to be keywords;

[0078] After confirming the key words, further confirm whether the keywords traversed subsequently also match the keywords in the keyword library to filter out the key sentences.

[0079] Among them, the keywords are words commonly used in fraud (such as: congratulations, congratulations, winning, prize, involved, suspected, credit, use, loan, verification, certificate, code). By combining the words, we can get key words (such as: congratulations, winning, suspected, credit loan, verification code) and key sentences (Congratulations on winning the prize). The specific keywords can be further confirmed according to the specific application scenarios and fraud types. The keyword library should be regularly supplemented and increased according to the updates of fraud methods.

[0080] S105, marking the APPs with key sentences in the attachment text and email as fraud-related APPs with a mark of 1, and marking the safe APPs with a mark of 0, and creating a prediction model to distinguish between fraud-related APPs and safe APPs based on the marks.

[0081] Specifically, key information features are extracted from historical fraud-related apps, including the sender's attachment text, email IP address, URL link, and the frequency of occurrence of keywords;

[0082] Convert the sender's IP address into a numerical feature, implement feature encoding on the URL link, and count the frequency of occurrence of keywords;

[0083] The above features are integrated into a feature vector A, expressed as A = {a1, a2, a3}; where a1 represents the numerical feature of the sender's IP address, a2 is the feature code of the URL link, and a3 is the frequency feature of the keyword;

[0084] Perform standardization on the feature vector A and apply the K-means clustering algorithm to classify the historical fraud-related apps. Apps marked as 1 are considered fraud information, and those marked as 0 are considered safe information.

[0085] According to the classification results of historical data, the average value of all fraud information feature vectors marked as 1 is calculated as the control sample A_1; the similarity between the fraud information marked as 1 and the control sample A_1, and the similarity between the safety information marked as 0 and the control sample A_1 are calculated respectively;

[0086] The logistic regression model is constructed with the value of feature vector A as input and similarity Z as output; the model expression is Z=w1*a1+w2*a2+w3*a3, where w1, w2, and w3 represent the weight coefficients of the sender's IP address numerical feature, URL link feature encoding information, and keyword frequency feature, respectively;

[0087] The similarity Z is used to convert the logistic regression results into probability values ​​through an activation function to form a fraud information prediction model; the calculation formula is P=1 / (1+e^(-Z)). According to the classification results of historical data, the average probability P1 that the information is safe information and the average probability P2 that the information is fraud information are obtained, and P1<P2.

[0088] S106, when the user uses the APP, the system sends his basic information and real-time behavior data to the trained model for real-time judgment.

[0089] S107, based on the output of the model, mark the APP and generate corresponding warning information.

[0090] Among them, the marked APPs include: fraud-related APPs, suspected fraud-related APPs, and safe APPs.

[0091] Specifically, the system receives the output from the prediction model, which represents the probability of fraud (P value) of information when the user uses the APP;

[0092] According to the classification results of historical data and actual application requirements, two thresholds are set: one is the threshold for judging as "suspected fraud APP" (for example, P = 0.5), and the other is the threshold for judging as "fraud APP" (for example, P = 0.8);

[0093] When the P value is greater than or equal to the "fraud-related APP" threshold, the APP is marked as "fraud-related APP"; when the P value is less than the "fraud-related APP" threshold but greater than or equal to the "suspected fraud-related APP" threshold, the APP is marked as "suspected fraud-related APP"; when the P value is less than the "suspected fraud-related APP" threshold, the APP is marked as "safe APP";

[0094] For "fraud-related APPs", emergency warning information is generated, such as "Warning: The APP you are using has a high risk of fraud. Please stop using it immediately and report it to the relevant departments." For "suspected fraud-related APPs", general warning information is generated, such as "Reminder: The APP you are using may have a risk of fraud. Please operate with caution and pay attention to subsequent prompts." For "safe APPs", no warning information is generated or a confirmation message is generated, such as "Your APP is safe to use, please feel free to use it."

[0095] The generated warning information will be sent to users in a timely manner through push notifications, SMS, emails, etc., to ensure that users can receive it in time and take corresponding actions;

[0096] Record the sending of warning information, including sending time, sending method, user feedback, etc., so as to facilitate subsequent analysis and optimization of the warning system. According to user feedback and actual conditions, continuously adjust and optimize the generation and sending strategies of warning information to improve the accuracy and effectiveness of the warning system.

[0097] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages:

[0098] 1. It has demonstrated significant advantages and effects in the feature extraction of text graphs by setting keywords, key words, key sentences, etc. First, in the feature extraction stage, each character in the text is converted into an independent image through an innovative way of text graphics. Without performing text recognition, it is directly processed in the form of an image. This method not only avoids the errors that may be caused by text recognition, but also realizes the fine extraction of text features by recording the coordinates of the intersections in the text graph. This refined feature extraction method provides accurate basic data for the subsequent screening of keywords, key words, and key sentences, greatly improving the accuracy of fraud information identification.

[0099] 2. By integrating big data technology and web crawler technology, this method can capture and summarize fraud-related APP information from multiple channels and dimensions in real time and efficiently, greatly improving the comprehensiveness and timeliness of data collection. At the same time, using advanced text image extraction technology, the text content is converted into image form for feature extraction. This innovative measure not only effectively avoids the format difference problem in traditional text processing, but also significantly enhances the accuracy and robustness of feature recognition. In addition, by constructing a prediction model based on logistic regression and combining the comprehensive analysis of multi-dimensional feature vectors, this method achieves high-precision distinction between fraud-related APPs and safe APPs, providing a strong guarantee for the accuracy of the early warning system.

[0100] Embodiment 2:

[0101] The above-mentioned embodiment 1, the security and anti-fraud warning method of APP based on big data analysis, not only has comprehensive and automatic data collection and organization capabilities, but also realizes effective identification and warning of fraud-related APPs through innovative text image extraction features and logistic regression model construction technology. This method improves the accuracy and timeliness of warnings and provides users with a safer and more reliable social environment. When the key words and key sentences that need to be identified in the attachment text and mailbox are not on the same page, but multiple pages are stored in the text in the form of multiple images, single-image recognition of a separate page is used, and the row center line or angle of the two pages is found and combined to improve the accuracy and speed of recognition and improve the timeliness of the warning. Now, step S104 of embodiment 1 is further improved, specifically:

[0102] S201, performing single image recognition on multiple image contents of a text page and performing stroke on the single image to remove the background color of the image.

[0103] Specifically, the text in each picture is recognized using image recognition OCR (optical character recognition) technology, and the recognized pictures are processed to remove the background color of the picture and retain the text stroke. The image processing library OpenCV is used to read the picture and convert it into a grayscale image. The binarization algorithm Otsu's threshold method is used to convert the grayscale image into a binary image, in which the text part is retained in black and the background part is changed to white. The binary image is further processed through morphological operations (such as corrosion and expansion) to remove noise and small holes around the text and retain the text stroke. The image recognition technologies involved in the above text are all existing technologies and will not be described in detail in this article.

[0104] S202: For each single image, determine the relative angle between the row center line of the multi-image text page and the page.

[0105] Specifically, use image processing technology to detect text lines in a single image, calculate the center point of each line of text, and then use least squares to fit a straight line as the line centerline. Use image registration technology or feature point matching technology to determine the relative angle between a single image and a multi-image text page, and record the line centerline and relative angle information of each single image. Use image rotation correction technology to adjust the relative angle between single images, and correct all single images to the same angle to ensure the alignment of page content.

[0106] S203, adjusting the position of the single image in the multi-image text, and combining cross-page keywords and key sentences within the line center line and a fixed angle range.

[0107] Among them, the combination of cross-page keywords and key sentences includes a cross-page combination of page breaks (the end of the previous page and the beginning of the second page is a paragraph) and a cross-page sentence segmentation combination (the end of a line on the first page and the beginning of the first line on the second page is a paragraph).

[0108] Specifically, according to the row center line and relative angle information determined in S202, the position of each single image in the multi-image text page is adjusted to ensure that the single image maintains a consistent row center line and a uniform relative angle in the page;

[0109] Find the order of pages. If the key words or sentences are split across pages, reassemble them using the intersection coordinates.

[0110] If the text content on both sides of the page break is semantically coherent and meets the rules of cross-page combination (such as not destroying the integrity of the sentence, not causing reading difficulties, etc.), then cross-page combination is performed to merge the text content on both sides of the page break into one sentence or paragraph;

[0111] If the text content on both sides of the page break is not semantically coherent, or does not conform to the rules of cross-page combination (such as splitting key words or sentences, or causing incoherence in reading, etc.), cross-page sentence segmentation and combination will be performed. This means that we need to segment the text content on both sides of the page break as two sentences or two paragraphs, and ensure that each sentence or paragraph is complete and coherent on its own page.

[0112] Among them, when combining cross-page keywords and key sentences, page break cross-page combination is first performed. If it meets the combination rules, cross-page combination is performed. If it does not meet the page break cross-page combination, cross-page sentence segmentation combination is performed.

[0113] For example, first, combine the page breaks across pages. If the word "Congratulations" is at the end of the first page, and "You won the prize" is at the beginning of the second page, then confirm whether the number of intersection coordinates of "Congratulations" and the number of intersections of the two key words "You won the prize" on the second page are the same as the number of intersections of the key sentence "Congratulations on your winning the prize" in the keyword library. If the number of intersections is the same, then the two single images are considered to be combined, and context splicing can be performed. If the word "Congratulations" is at the end of the first line of the first page, and "You won the prize" is at the beginning of the first line of the second page, then confirm whether the number of intersection coordinates of "Congratulations" and the number of intersections of the two key words "You won the prize" on the second page are the same as the number of intersections of the key sentence "Congratulations on your winning the prize" in the keyword library. If the number of intersections is the same, then the two single images are considered to be combined, and left and right page margin splicing can be performed.

[0114] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages:

[0115] 1. By combining the powerful OpenCV image processing library, not only the background color of the image is successfully removed, but also the clear and sharp text strokes are retained. The refined processing of this step significantly improves the accuracy of text recognition. At the same time, the image quality is further optimized using the binarization algorithm and morphological operations, effectively removing noise and enhancing the contrast of the text. This refined image processing technology lays a solid foundation for subsequent text processing, ensuring accurate extraction and clear display of text information even in complex backgrounds or low-contrast conditions.

[0116] 2. In terms of text page layout and alignment, this technical process achieves accurate positioning of a single image in a multi-image text page by accurately calculating the center line of the row and determining the relative angle between the image and the page. This step is achieved thanks to the clever use of image registration and feature point matching technology, as well as the precise execution of image rotation correction. The combined application of these technologies ensures that all single images are presented at a uniform angle, greatly improving the neatness of the page content and the reading experience. In addition, in terms of text integration and presentation, this technical process also achieves flexible application of page break combinations and cross-page sentence segmentation combinations by intelligently identifying and processing cross-page keywords and key sentences. This intelligent text integration method not only effectively solves the semantic break problem caused by paging in traditional text processing, but also ensures the complete presentation of keywords and key sentences through the precise matching of intersection coordinates, thereby improving the readability, coherence of the text, and the accuracy and efficiency of information transmission.

[0117] Embodiment three:

[0118] The above-mentioned embodiment two is through OCR recognition, image processing and accurate text positioning technology, and this method not only removes the picture background color and retains clear text stroke, but also accurately determines the position and direction of each picture in the page. The most important thing is, by optimizing the combination strategy of cross-page keywords and key sentences, including page break cross-page combination and cross-page sentence segmentation combination, this method realizes the combination of cross-page keywords and key sentences, ensures the accurate extraction and combination of keywords and key sentences, improves the accuracy and speed of screening keywords, keywords, key sentences, and provides powerful guarantee for the accuracy of early warning system. In order to further improve the rapid identification of keywords and key sentences, the step S203 of the embodiment two is now further improved, specifically as follows:

[0119] S301, performing feature processing on the key words and key sentences after traversal, screening and combination to retain the intersection points in the text graph.

[0120] Specifically, for each selected keyword and key sentence, feature extraction is performed using OpenCV image processing technology. This includes identifying the intersection points in the text image, that is, the places where character strokes intersect or touch, and saving the coordinates of the identified intersection points.

[0121] S302, at the center point position of each text image, generate a cross-shaped identifier.

[0122] Among them, the cross-shaped identifier can be a small graphic, composed of four line segments intersecting at one point, similar to the shape of the Chinese character "米" (mi). This identifier is required not to block the intersection points in the text image but be obvious enough for subsequent processing.

[0123] Specifically, for each selected keyword and key sentence, determine its center point position in the text image (in the virtual coordinate axis (10, 10)), and generate a cross-shaped identifier at the center point position (connecting (0, 0) and (20, 20) passing through the coordinate axis (10, 10), and connecting (0, 20) and (20, 0) passing through the coordinate axis (10, 10)), similar to the shape of the Chinese character "米" (mi).

[0124] S303, record the coordinates of the positions where each cross-shaped identifier passes through the intersection points of the text image.

[0125] Among them, the coordinates of the positions where the cross-shaped identifier passes through the intersection points of the text image are called "cross intersection points", and the text image with the cross-shaped identifier added is called a cross text image.

[0126] S304, compare based on the intersection coordinates of the cross intersection points and the cross intersection point library.

[0127] Specifically, based on the keyword library, upgrade and construct a cross image library, which contains the cross intersection coordinates of known keywords and key sentences.

[0128] S305, retain the cross text images of the keywords and key sentences according to the matching results.

[0129] Specifically, if the intersection coordinates of the cross text image completely correspond to the coordinates of the keywords and key words in the cross image library, retain the cross text image.

[0130] The technical solutions in the embodiments of the present application described above have at least the following technical effects or advantages:

[0131] 1. At the center position of the text graph of each keyword and key sentence, a cross mark is added. This mark is formed by the intersection of four line segments at a point, resembling the Chinese character'mi' (rice). It neither interferes with the original intersection information nor is it too conspicuous to facilitate subsequent identification. At the same time, the scheme also details the position coordinates where the cross mark coincides with the intersection of the text graph, that is, the 'cross intersection', and generates a text graph containing the cross mark based on this. To further improve the screening efficiency, the scheme constructs a cross mark library, which stores the cross intersection coordinates of known keywords and key sentences. By comparing the intersection coordinates of the cross text graph with the data in the library, the scheme can quickly screen out the keywords and key sentences that are exactly the same as the coordinates in the library, thus achieving the goal of efficient and accurate screening.

[0132] Example 4:

[0133] In Example 3 above, the OpenCV technology is used to retain the intersections in the text graph, and then a unique cross mark is added at the center of the text graph. The coordinates where the cross coincides with the intersection are recorded, and a cross mark library is constructed. By comparing the cross text graph with the library coordinates, efficient and accurate screening of keywords and sentences is achieved. However, due to inaccurate cross mark positions in all the above operations, there may be a situation where the intersection coordinates of individual cross text graphs in key sentences do not correspond to those in the cross mark library. To solve this problem, an improvement is made based on step S305, specifically as follows:

[0134] S401. Compare the sentences in the cross text graph of the key sentence that do not completely correspond separately.

[0135] Here, 'not completely corresponding' means that the coordinates of the cross text graph of the key sentence cannot correspond 100% to the cross text graph coordinate points of the key sentence in the cross mark library.

[0136] S402. Overlay the cross text graphs of the key sentences, compare the intersections of the overlaid cross text graphs, and record the intersections of the overlaid cross text graphs of the key sentences.

[0137] For example, if there are 3 cross intersections in the cross text graph of the key sentence 'Congratulations on your winning' that do not correspond to those in the cross mark library, then the cross text graph of 'Congratulations on your winning' is overlaid with the center point as the overlay point. The resulting effect is that all the cross intersections of the 5 cross text graphs of 'Congratulations on your winning' are shown on the cross text graph.

[0138] S403. Determine the keywords and key sentences.

[0139] Specifically, if the accuracy rate of the comparison between the intersection points of the superimposed cross-shaped text diagrams and the intersection points in the cross-shaped diagram library is above the threshold of 85%, it can be determined that the selected words and sentences are keywords and key sentences. If it is below the threshold, it is determined whether they are newly added keyword phrases or key sentences, and the cross-shaped diagram library is updated.

[0140] The technical solutions in the embodiments of the present application at least have the following technical effects or advantages:

[0141] 1. For those key sentences that do not fully match when compared with the cross-shaped diagram library, this solution does not simply exclude or ignore them, but adopts a more scientific and comprehensive processing method. Specifically, it superimposes the cross-shaped text diagrams corresponding to these key sentences. The ingenuity of this step is that it can effectively integrate the key information in multiple images, making the feature points that may have been ignored due to individual differences stand out, providing a richer and more accurate visual basis for subsequent judgments.

[0142] 2. By superimposing the cross-shaped text diagrams and comprehensively comparing the intersection points with the cross-shaped diagram library, the system can more accurately lock the words and sentences that are truly of key significance, effectively avoiding misjudgments or omissions caused by insufficient single-image information. Secondly, this solution also demonstrates good adaptability and scalability. As the language environment changes continuously and new keyword phrases and key sentences emerge continuously, the cross-shaped diagram library in the solution can be automatically updated according to the actual situation, ensuring that the recognition system always maintains the best state and providing a strong guarantee for long-term stable recognition performance. In summary, the technical solutions of the embodiments of the present application not only improve the recognition accuracy of keywords and key sentences, but also enhance the adaptive ability and scalability of the system, providing more reliable and efficient technical support for applications in related fields.

[0143] The above is only the preferred embodiment of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A social security anti-fraud early warning method based on big data analysis, characterized in that: include: S101, using big data technology to automatically collect and organize data related to fraudulent apps; S102, based on the extracted features, extracting text and image features from the attachment text and email in the APP; S103, comparing the features extracted from the text image with the text images in the keyword library to select keywords; S104, after filtering out the keywords, further filtering out the key words and key sentences; S105, marking the APPs with key sentences in the attachment text or email as fraudulent APPs with a mark of 1, and marking the safe APPs with a mark of 0, and creating a prediction model to distinguish between fraudulent APPs and safe APPs based on the marks; S106, when the user uses the APP, the system sends the user's basic information and real-time behavior data to the trained model for real-time judgment; S107, based on the output of the model, mark the APP and generate corresponding warning information.

2. A social security anti-fraud early warning method based on big data analysis as claimed in claim 1, characterized in that: The S102 includes: The text image converts each word in the text into an independent image, and processes it directly in the form of an image without performing text recognition; Unify the format of attachment text and email text content; Set the image size, image format and center point of each text image; Traverse all text images and conduct preliminary screening on each text image, requiring it to meet the set pixel size and image format; Initialize a virtual coordinate axis for each qualified text image, and use the point (10,10) in the coordinates as the expected center point of the text image; Place the text image in the virtual coordinate axis and determine whether the origin and center of the text image are located at the origin and center of the coordinate axis (10,10); For the corrected standard text diagram, record the coordinates of the intersection points in the text diagram.

3. A social security anti-fraud early warning method based on big data analysis as claimed in claim 1, characterized in that: The S104 includes: After the intersection coordinates of the corrected standard text image are identified, the text image is traversed one by one according to the regular text structure of the attachment text and the email. When the intersection coordinate position of the traversed text image completely corresponds to a keyword in the keyword library, the text image is confirmed to be a keyword.

4. A social security anti-fraud early warning method based on big data analysis as claimed in claim 1, characterized in that: The S104 further includes: After the keyword is confirmed, the next text image of the keyword in the traversal order is compared with the keyword library to confirm whether the next text image also matches the keyword in the keyword library. If the coordinate position of the intersection of the next text image and the keyword in the keyword library completely corresponds, then the above two keywords are determined to be keywords; After confirming the key words, further confirm whether the keywords traversed subsequently also match the keywords in the keyword library to filter out the key sentences.

5. A social security anti-fraud early warning method based on big data analysis as claimed in claim 1, characterized in that: The S104 further includes: S201, performing single image recognition on multiple image contents of a text page and performing stroke on the single image to remove the background color of the image; S202, for each single image, determining a relative angle between a row center line of the multi-image text page and the page; S203, adjusting the position of the single image in the multi-image text, and combining cross-page keywords and key sentences within the line center line and a fixed angle range.

6. A social security anti-fraud early warning method based on big data analysis as claimed in claim 5, characterized in that: The S203 includes: the combination of cross-page keywords and key sentences includes a cross-page combination of page breaks where the end of the previous page and the beginning of the second page are a paragraph, and a cross-page sentence segmentation combination where the end of a line on the first page and the beginning of the first line on the second page are a paragraph.

7. A social security anti-fraud early warning method based on big data analysis as claimed in claim 5, characterized in that: The S203 further includes: S301, performing feature processing on the key words and key sentences after traversal, screening and combination, retaining the intersection points in the text graph; S302, generating a cross symbol at the center point of each character image; S303, recording the coordinates of each intersection point where the cross symbol passes through the text pattern; S304, comparing the intersection coordinates of the cross-shaped intersection with the cross-shaped intersection word library; S305, retaining the cross-shaped text graph of the key words and key sentences according to the matching results.

8. A social security anti-fraud early warning method based on big data analysis as claimed in claim 7, characterized in that: The S302, the cross mark, includes: The cross logo can be a graphic consisting of four line segments intersecting at one point; For each selected key word and key sentence, determine the center point position of the key word and key sentence in the virtual coordinate axis (10,10) of the text map; At the center point, connect (0,0) and (20,20) through the coordinate axis (10,10), and connect (0,20) and (20,0) through the coordinate axis (10,10) to generate a cross symbol.

9. A social security anti-fraud early warning method based on big data analysis as claimed in claim 7, characterized in that: The S305 includes: S401, comparing the sentences in the rice character graph of the key sentences that are not completely corresponding to each other separately; S402, superimposing the rice character graphs of the key sentence, comparing the intersections of the superimposed rice character graphs, and recording the intersections of the superimposed rice character graphs of the key sentence; S403, determine keywords and key sentences.

10. A social security anti-fraud early warning method based on big data analysis as claimed in claim 9, characterized in that: The S403 includes: If the accuracy of the intersection of the superimposed rice character image and the intersection in the rice character image library is above the threshold of 85%, it can be determined that the screened words and sentences are keywords and key sentences. If it is lower than the threshold, it is determined whether it is a newly added keyword or key sentence, and the rice character image library is updated.

Images