An artificial intelligence-based anti-fraud prevention and control intelligent assistant platform for the elderly

Abstract

The application discloses an old-age anti-fraud prevention and control intelligent assistant platform based on artificial intelligence and belongs to the technical field of network security.The platform comprises a website security evaluation module, a user behavior analysis module, a matching judgment module and a behavior risk protection module; the website security evaluation module collects website text data, determines whether a website is related to fraud after extracting fraud risk features by means of a fraud risk category model; the user behavior analysis module obtains user mobile phone daily behavior data, divides group categories and establishes individual behavior baselines; the matching judgment module performs matching analysis on the operation behavior data of the old-age group on the related fraud websites and the individual behavior baselines, generates a behavior risk score, and triggers the behavior risk protection module when the behavior risk score exceeds 30 points; and the behavior risk protection module immediately forcibly terminates the related operation and sends an alarm.The application can effectively prevent and control network frauds suffered by the old-age people and improve the anti-fraud protection capability through the collaborative work of the multiple modules.

Description

Technical Field

[0001] This invention relates to the field of network security technology, specifically to an AI-based smart assistant platform for preventing fraud among the elderly. Background Technology

[0002] With the popularization of internet technology, the elderly population's dependence on the internet has been increasing year by year. However, due to their weak ability to discern information and insufficient awareness of new fraud methods, they have become a high-risk group for online fraud.

[0003] Existing anti-fraud technologies largely rely on manual reporting, keyword filtering, or blacklisting, which suffers from problems such as strong passivity, insufficient targeting, incomplete information extraction, and poor coordination. They are mostly reactive, lacking proactive identification and pre-emptive prevention of fraudulent activities, making it difficult to block risks in real time during elderly users' operations. They fail to design personalized protection mechanisms based on the behavioral characteristics of the elderly, using uniform standards for risk assessment, which easily leads to misjudgments or omissions. Identification of website fraud risks relies heavily on text content, ignoring fraudulent information hidden in images and videos, resulting in incomplete risk feature extraction and affecting the accuracy of judgments. Furthermore, they lack dynamic matching analysis of user behavior and fraud scenarios, failing to assess real-time risks based on the individual operating habits of the elderly, resulting in low protection accuracy. Therefore, there is an urgent need for a smart anti-fraud platform that can proactively identify fraud risks, combine the behavioral characteristics of the elderly to achieve personalized protection, and achieve real-time blocking through multi-module collaboration, in order to improve the ability to prevent online fraud against the elderly. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based smart assistant platform for preventing fraud among the elderly, in order to solve the problems mentioned in the background.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: An AI-based smart assistant platform for preventing fraud among the elderly includes a website security assessment module, a user behavior analysis module, a matching judgment module, and a behavior risk protection module. The website security assessment module acquires the website's text data, performs data feature extraction processing on the text data to obtain the website's fraud risk features, and processes the website's fraud risk features through a fraud risk category model to obtain the website's corresponding fraud risk category, where the fraud risk category is either fraud-related or non-fraud-related. The user behavior analysis module acquires daily behavior data from the user's mobile phone, performs behavior analysis on the daily behavior data, and obtains the user's corresponding group category and individual behavior baseline, wherein the group category is elderly group and non-elderly group; The matching and judgment module monitors the operational behavior data of elderly users on fraudulent websites, matches and analyzes the operational behavior data of elderly users with individual behavioral baselines to obtain a behavioral risk score, and determines whether to trigger the behavioral risk protection module. The behavioral risk protection module forcibly terminates and alerts elderly users' actions on fraudulent websites.

[0006] Preferably, the method for obtaining text data from a website is as follows: Employing compliant web crawling technology based on the robots protocol, it directly captures native web text contained in the source code of website pages, while accurately extracting the storage paths of images and videos on the website. For the extracted image and video storage paths, it downloads the corresponding images and videos via HTTP or HTTPS download protocols. For the downloaded images, adaptive histogram equalization is used for preprocessing. By adaptively adjusting the pixel grayscale values ​​of different regions of the image, the local contrast of the image is enhanced, making the blurred text areas in the image clearly distinguishable from the background. At the same time, for the downloaded video, keyframe extraction is performed based on the video content change rate. By calculating the pixel difference value between consecutive video frames, video frames with a content change rate lower than a preset content change rate threshold are selected. Video frames with a content change rate lower than the preset content change rate threshold usually contain relatively stable text information. Based on this, video frames containing text are extracted, and Gaussian filtering is applied to the video frames containing text. A smooth template generated by the Gaussian function is used to perform convolution operation on the video frames containing text to remove Gaussian noise and random noise in the video frames containing text, reducing the interference of noise on text region recognition. The image with enhanced contrast and the video frame containing text after noise removal are input into the object detection model YOLOv8. The object detection model YOLOv8 accurately locates the text regions in the image and video frame through multi-scale feature fusion and bounding box regression algorithms, and outputs the coordinate information of the text regions. Based on the coordinate information of the text regions, optical character recognition (OCR) technology is used to perform character recognition and semantic parsing of the text regions, and the text information in the image and video frame is converted into editable text data, that is, multimedia derived text. By fusing native web text with multimedia-derived text, a complete website text data is formed.

[0007] The preferred process for obtaining website fraud risk characteristics: The characteristics of website fraud risk include fraud keyword density, semantic deception index, extreme sentiment value, and information integrity deficiency. To determine the density of fraud keywords, frequently occurring fraud-related words were manually labeled from the National Anti-Fraud Center's case database and texts involving elderly people who were victims of fraud. Simultaneously, the LDA topic model was used to deeply mine these texts, supplementing them with potentially deceptive latent words that could be flagged as fraudulent. These fraud-related words and deceptive latent words were then merged to obtain the fraud keywords. Based on this, a fraud keyword database was constructed. The AC automaton algorithm was then used to quickly match the text data, counting the total number of times fraud keywords appeared in the database. The total number of fraud keywords was then calculated using a word density formula to obtain the fraud keyword density, which quantifies the distribution of fraud-related words in the text data. The word density formula is:

[0008] For the semantic deception index, a large amount of texts involving elderly people being defrauded are input into the BERT-base model. Through the pre-trained parameters and contextual semantic understanding capabilities of the BERT-base model, these texts are transformed into several semantic vectors containing deep semantic information. Based on this, a clustering algorithm is used to perform cluster analysis on all semantic vectors. According to semantic similarity, all semantic vectors are classified into several categories, each category corresponding to a fraud semantic template, thus obtaining several fraud semantic templates. At the same time, the target semantic vector is generated from the text data through the BERT-base model, and the cosine similarity value between the target semantic vector and the semantic vector of each fraud semantic template is calculated using the vector space cosine similarity formula. The average of the top five cosine similarity values ​​is selected, which is the semantic deception index, used to measure the degree of matching between the text and known fraud semantic patterns. The formula for the cosine similarity of the vector space is: ; Where A is the target semantic vector, and B is the semantic vector of a certain fraud semantic template. It is the component of A in the i-th dimension. It is the component of B in the i-th dimension; For sentiment extremes, the text data is divided into several independent sentences by periods. Each sentence is fed into a pre-trained TextCNN-based sentiment analysis model. The pre-trained TextCNN-based sentiment analysis model extracts sentiment features from the sentences through convolutional layers and outputs the sentiment score of each sentence after processing by pooling and fully connected layers. The maximum and minimum values ​​are extracted from the sentiment scores of all sentences to form sentiment pairs, i.e., sentiment extremes, which are used to reflect the extreme degree of text data in terms of excessive exaggeration of positive commitment and extremely frightening negative expression. To assess the completeness of information, based on the characteristics of scams targeting the elderly, eight core information dimensions for elderly fraud were constructed. These include organization name, qualification number, risk warning, fund flow description, contact person's real name, office address, regulatory unit, unsubscription mechanism, and exit mechanism. The BERT-NER model was used to perform named entity recognition on the text data, and the number of the eight core information dimensions for elderly fraud that were not identified in the text data was counted, i.e., the number of missing dimensions. Based on this, the number of missing dimensions was calculated using the completeness missing degree formula to obtain the information completeness missing degree, which is used to assess the completeness of the text data in terms of key information disclosure. The more missing dimensions, the lower the credibility of the text and the higher the risk of fraud. The formula for the degree of completeness missing is: .

[0009] Preferably, the fraud risk category model adopts a gradient boosting tree ensemble architecture. The gradient boosting tree ensemble architecture achieves accurate determination of the fraud risk category corresponding to a website through hierarchical collaboration between the underlying model and the meta-model. The underlying model includes an XGBoost model and a LightGBM model. The XGBoost model and the LightGBM model process the input website fraud risk features in parallel. The XGBoost model captures the high-order nonlinear correlation in the website fraud risk features through regularization term design and column sampling technology, and outputs the first fraud prediction probability. The LightGBM model discretizes the website fraud risk features through the histogram algorithm and combines a one-sided gradient sampling strategy to reduce the amount of computation, mine the fine-grained patterns in the website fraud risk features, and outputs the second fraud prediction probability. The meta-model (logistic regression model) performs a weighted fusion of the first and second fraud prediction probabilities. It dynamically adjusts the weight ratio of the output results of the XGBoost model and the LightGBM model through the maximum likelihood estimation method, and maps the fusion result of the weighted fusion of the first and second fraud prediction probabilities to the interval [0,1] through the Sigmoid function. When the fusion result is not less than 0.5, the fraud risk category corresponding to the website is determined to be fraud; otherwise, it is not fraud.

[0010] The preferred method for acquiring daily behavioral data from a user's mobile phone is as follows: The SDK is embedded into the user's mobile application, thereby recording the user's daily behavioral data, including user group category characteristic data and individual user behavior characteristic data. The user group category characteristic data includes the average interval time between single screen clicks, the average screen swipe speed, the mobile phone system font size, and the proportion of time spent using health-related apps. The individual user behavior characteristic data includes the average duration of visits to fraudulent websites, the average interval between sensitive operations, the average time spent inputting sensitive information, and the distribution of operation time periods.

[0011] By using software development kit (SDK) embedding technology, the developed SDK is integrated into the user's mobile application. The SDK records daily behavioral data, including user group category characteristic data and individual user behavioral characteristic data, by calling the standardized interface provided by the mobile operating system. User group category characteristic data is used to distinguish between elderly and non-elderly groups, including the average time interval between single screen clicks, the average speed of screen swiping, the font size of the mobile phone system, and the percentage of time spent using health-related apps. User individual behavior characteristic data is used to construct individual behavior baselines, including the average duration of visits to fraudulent websites, the average interval of sensitive operations, the average time spent inputting sensitive information, and the distribution of operation time periods. Sensitive operations are those involving economic transactions, and sensitive information is information involving economic transactions.

[0012] The preferred process for performing behavioral analysis on daily behavioral data: User group category characteristic data and individual user behavior characteristic data are normalized using the min-max normalization method, mapping them to the [0,1] interval. The normalized user group category characteristic data is then input into a pre-trained random forest model to obtain the user's corresponding group category. Simultaneously, the normalized user individual behavior characteristic data is input into a pre-trained long short-term memory network model to obtain the user's corresponding average time baseline, including input sensitive information. Baseline duration of visits to fraudulent websites Sensitive operation mean interval baseline Deviation from operating time period benchmark The baseline of individual behavior.

[0013] Preferably, the process of matching and analyzing the operational behavior data of elderly users with individual behavioral baselines: The SDK obtains operation behavior data, specifically including sensitive operation intervals. Duration of visits to fraudulent websites Input time for sensitive information and actual operation period ; For actual operation period Deviation from operating time period benchmark The similarity of the operation time period is calculated using the cosine similarity formula of vector space. ,in It represents the percentage of operations performed in the i-th time unit. When the value is 1, the real-time operation period is in the i-th time unit; otherwise... =0; Similarity of operation time period Excluding actual operating time Operational behavior data and benchmarks excluding operational time periods The individual's behavioral baseline is input into the risk deviation scoring formula to calculate the behavioral risk score; The risk deviation scoring formula is as follows: ; in, It is the deviation of the frequency of sensitive operations. It is the deviation in the time spent on fraudulent websites. It is the deviation of the input speed of sensitive information. It is the deviation during the operation period. It is the weight of the deviation in the frequency of sensitive operations. It is the deviation in the time spent on fraudulent websites. It is the weight of the deviation of the input speed of sensitive information. It is the weight of the deviation during the operation period, and ; The deviation of the sensitive operation frequency is: ; The deviation in the duration of stay on the fraudulent websites is as follows: ; The deviation of the input speed of the sensitive information is: ; The deviation of the operation period is: .

[0014] Preferably, the behavior risk protection module is triggered when the behavior risk score is greater than 30; otherwise, monitoring continues.

[0015] Preferably, the process of forcibly terminating and alerting elderly users regarding their actions on fraudulent websites: By pre-installing an anti-fraud plugin in the user's mobile browser, the plugin monitors the user's operation requests in real time. When it detects that the user has initiated a critical operation request on a fraudulent website, the plugin immediately calls the browser's network request interception interface to block the critical operation request and returns a 403 Forbidden response to the user's mobile browser, preventing the operation from continuing at the network request level. At the same time, a masking layer is dynamically generated using JavaScript. The masking layer covers the entire fraudulent website page and is at a higher level than all elements on the fraudulent website page, thus blocking elderly users from clicking on fraudulent buttons on the page. Upon completing the forced termination of the operation, an alarm message is automatically pushed to the WeChat or Alipay mini-programs of the pre-associated elderly user's children by calling the official mini-program interface of WeChat or Alipay. This is to ensure that the children can be aware of the fraud risks their parents are facing in a timely manner and take appropriate intervention measures.

[0016] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: 1. This invention breaks through the limitations of single information and features, constructing a multimodal and multi-dimensional fraud identification system. Existing technologies for identifying fraudulent websites largely rely on text content and filter solely through single keywords, resulting in incomplete information extraction and insufficient identification accuracy. This invention, however, obtains the original text of a website through compliant web crawlers, further processes the multimedia content, and then extracts the textual information using YOLOv8 object detection and OCR technology, forming a complete data source of original web text and multimedia-derived text. Simultaneously, it innovatively constructs four types of website fraud risk features, combined with a gradient boosting tree integrated architecture, to achieve accurate identification of risks across multiple levels: vocabulary, semantics, sentiment, and information completeness. This effectively solves the problem of misjudgment caused by information gaps and single features in existing technologies.

[0017] 2. This invention achieves differentiated group and personalized individual analysis, improving the targeting of protection. Existing technologies use a uniform standard to assess risk, failing to consider the behavioral characteristics and individual operating habits of the elderly, resulting in insufficient targeting. This invention, however, collects user behavior data through an SDK, inputs group characteristic data into a random forest model to accurately segment the elderly and non-elderly groups, and simultaneously models individual behavioral data based on an LSTM model to generate personalized behavioral baselines. This makes risk assessment more aligned with the operating habits of the elderly, solving the problem of low accuracy in protection caused by a one-size-fits-all assessment.

[0018] 3. This invention dynamically matches real-time behavior with individual baselines, enabling proactive prevention and control. Existing technologies mostly rely on post-event interception, lacking dynamic linkage analysis between real-time operational behavior and individual habits, making timely intervention difficult when risks occur. This invention, however, compares the real-time operational data of elderly users on fraudulent websites with their individual behavioral baselines through a matching judgment module. It uses vector space cosine similarity to calculate the time-period matching degree and combines this with a risk deviation formula to generate a dynamic behavioral risk score. When the score exceeds 30 points, the behavioral risk protection module is immediately triggered, shifting from reactive post-event handling to proactive in-event blocking, significantly improving the timeliness of prevention and control. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 This is a schematic diagram of the system functional modules of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Examples, such as Figure 1 The aforementioned AI-based smart assistant platform for preventing fraud among the elderly includes a website security assessment module, a user behavior analysis module, a matching judgment module, and a behavior risk protection module, which work together to complete the task of preventing fraud among the elderly.

[0023] The website security assessment module acquires the website's text data, performs data feature extraction processing on the text data to obtain the website's fraud risk features, and processes the website's fraud risk features through a fraud risk category model to obtain the website's corresponding fraud risk category, where the fraud risk category is either fraud-related or non-fraud-related. The user behavior analysis module acquires daily behavior data from the user's mobile phone, performs behavior analysis on the daily behavior data, and obtains the user's corresponding group category and individual behavior baseline, wherein the group category is elderly group and non-elderly group; The matching and judgment module monitors the operational behavior data of elderly users on fraudulent websites, matches and analyzes the operational behavior data of elderly users with individual behavioral baselines to obtain a behavioral risk score, and determines whether to trigger the behavioral risk protection module. The behavioral risk protection module forcibly terminates and alerts elderly users' actions on fraudulent websites.

[0024] Furthermore, the working principle of this invention will be illustrated below using an example of an elderly user, Ms. Zhang, accessing a senior care investment website via her mobile phone: The platform uses a compliant web crawler based on the robots.txt protocol to scrape the native web text of the pension investment website, such as the main page content and button text. It also extracts the storage paths of three promotional images and one introductory video from the website's pages and downloads the images and video via HTTPS. Adaptive histogram equalization is used to enhance the contrast of images, making the blurred government certification text distinguishable from the background. For videos, video frames containing text are extracted based on the content change rate. After removing noise through Gaussian filtering, YOLOv8 is used to locate text regions such as "zero risk" and "guaranteed profit" in the video frames. Then, OCR technology is used to extract the text, resulting in multimedia derived text. Based on this, the original web text and multimedia derived text are merged to form complete website text data.

[0025] By combining manual annotation with LDA topic modeling, and drawing from the National Anti-Fraud Center's case database and cases involving the elderly... The text was analyzed to extract fraudulent keywords, including but not limited to those related to retirement investment, high interest rates, and guaranteed profits. A fraud keyword database was constructed, and the AC automaton algorithm was used to count the total frequency of these fraudulent keywords in the text data, which was found to be 28 times. The total number of words in the text was 200, and the keyword density was calculated to be 14%. The text data involving elderly people and fraud was input into a BERT-base model to generate semantic vectors. K-means clustering was used to obtain semantic templates for fraudulent retirement projects. The text data was then processed through the BERT-base model to generate target semantic vectors, and cosine similarity scores were calculated between these vectors and the five most similar templates. The top five cosine similarity scores were 0.82, 0.79, 0.75, 0.73, and 0.71, with the average value representing a semantic deception index of 0.76. The text data was then segmented by period. The data was divided into 12 sentences, and sentiment scores were calculated using a TextCNN-based sentiment analysis model. The sentiment scores ranged from -10 to 10, with a maximum of 9.2 (meaning easy money) and a minimum of -8.5 (meaning missing out means loss), forming extreme sentiment values ​​(9.2, -8.5). Simultaneously, eight core information dimensions for elderly fraud were constructed. Using the BERT-NER model to identify text data, four types of information were found to be missing: "qualification number," "regulatory unit," "fund flow explanation," and "unsubscription mechanism." The number of missing dimensions was four, and the information completeness missingness was calculated to be 50%. Based on this, website fraud risk characteristics were obtained, including a fraud keyword density of 14%, a semantic deception index of 0.76, extreme sentiment values ​​(9.2, -8.5), and an information completeness missingness of 50%.

[0026] The website's fraud risk characteristics were input into the fraud risk category model. The underlying XGBoost model output a fraud prediction probability of 0.85, and the LightGBM model output a fraud prediction probability of 0.82. The meta-model adjusted the XGBoost weight to 0.52 and the LightGBM weight to 0.48 through maximum likelihood estimation. The fusion result was 0.836. Since 0.836 is greater than 0.5, the website was determined to be in the fraud risk category of fraud.

[0027] By embedding the SDKs of Aunt Zhang's mobile social and health apps, we collected her daily behavior data for 30 days. Specifically, this included group category characteristic data: the average interval between single screen clicks was 1.8 seconds, the average screen swipe speed was 0.3 m / s, the mobile phone system font size was "extra large", and the usage time of health apps accounted for 35%; individual behavior characteristic data: the average time spent visiting fraudulent websites was 4.2 minutes, the average interval between sensitive operations such as entering bank card numbers was 65 seconds, the average time spent entering sensitive information was 70 seconds, and the operation time was concentrated between 8:00 and 20:00.

[0028] After performing min-max normalization on the group category characteristic data, it is input into a pre-trained random forest model consisting of 100 decision trees to determine that Aunt Zhang belongs to the elderly group. After normalizing the individual behavioral characteristic data, it is input into a pre-trained LSTM model to generate individual behavioral baselines, specifically including the average time baseline for inputting sensitive information. Baseline for 70-second visits to fraudulent websites The baseline for sensitive operation mean interval is 7.2 minutes. The deviation benchmark is 65 seconds and the operation period. It is the frequency percentage vector from 0:00 to 24:00.

[0029] Obtain data on Ms. Zhang's actions on the aforementioned fraudulent websites, specifically including the intervals between sensitive operations. The duration of a visit to a fraudulent website is 15 seconds (after repeatedly clicking the "Invest Now" button). For 10 minutes, the time required to input sensitive information The time allotted is 25 seconds (for entering bank card number) and the actual operation period. The actual operation period is 22:30 (belonging to the 22:00-23:00 time unit). Converted into a single-dimensional vector, with a value of 1 for the 22:00-23:00 dimension and 0 for the rest, and the deviation from the benchmark during the operation period. The similarity of the operation time period was calculated using the cosine similarity formula. The similarity of the operation time period is 0.03. Excluding actual operating time Operational behavior data and benchmarks excluding operational time periods Individual behavioral baselines are input into the risk deviation scoring formula. The calculated behavioral risk score is approximately 78 points.

[0030] Because the behavioral risk score of 78 points is greater than 30 points, the protection mechanism, namely the behavioral risk protection module, was triggered. The mobile browser's anti-fraud plugin intercepted the critical request to submit an investment order, returned a 403 Forbidden response, and used JavaScript to generate a semi-transparent overlay layer to cover the page, blocking click events of fraudulent buttons such as "Pay Now" and "Confirm Submission". At the same time, it called the WeChat mini-program interface to push an alarm message to Ms. Zhang's daughter's WeChat: "Your mother is visiting a fraudulent website. The operation is of high risk. The operation has been blocked. Please see details." The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based intelligent assistant platform for preventing fraud among the elderly, characterized in that: include: The website security assessment module is used to acquire text data from the website, extract data features from the text data to obtain website fraud risk features, and process the website fraud risk features through a fraud risk category model to obtain the corresponding fraud risk category of the website, where the fraud risk category is fraud-related and non-fraud-related. The user behavior analysis module is used to acquire daily behavior data from users' mobile phones, perform behavior analysis on the daily behavior data, and obtain the user's corresponding group category and individual behavior baseline. The group category is elderly group and non-elderly group. The matching and judgment module is used to monitor the operational behavior data of elderly users on fraudulent websites. It matches and analyzes the operational behavior data of elderly users with individual behavioral baselines to obtain a behavioral risk score, and determines whether to trigger the behavioral risk protection module. The behavioral risk protection module is used to forcibly terminate and alert elderly users on fraudulent websites.

2. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 1, characterized in that, The method for obtaining text data from a website is as follows: A compliant web crawler based on the robots.txt protocol is used to obtain the website's native text, as well as the storage paths of images and videos, and then the images and videos are downloaded according to their storage paths. The image contrast is enhanced by adaptive histogram equalization. At the same time, based on the video content change rate, video frames containing text are extracted from the video. The video frames containing text are then filtered by Gaussian filtering to remove noise. Based on this, the image with enhanced contrast and the video frames containing text with noise removed are used to identify text regions using the YOLOv8 object detection model. The text regions are then extracted using OCR technology to obtain multimedia derived text. The multimedia-derived text and the native web text are merged to obtain text data.

3. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 2, characterized in that, The process of obtaining website fraud risk characteristics: By manually annotating and combining LDA topic modeling, fraud keywords were mined from the National Anti-Fraud Center's case library and texts involving fraud among the elderly, and a fraud keyword library was constructed. Based on this, the total number of fraud keywords appearing in the fraud keyword library in the text data was counted using the AC automaton algorithm, and the total number of fraud keywords was used to calculate the fraud keyword density using the word density formula. For texts involving fraud involving the elderly, semantic vectors are generated by processing them with the BERT-base model. The semantic vectors are then clustered to form several fraud semantic templates. Based on this, the cosine similarity value of the target semantic vector generated by processing the text data with the BERT-base model is calculated with the semantic vector of each fraud semantic template. The average of the top five cosine similarity values ​​is taken as the semantic deception index. The text data is divided into several sentences by periods. Each sentence is processed by a pre-trained sentiment analysis model based on TextCNN to obtain a sentiment score. The maximum and minimum sentiment scores are then used to form a sentiment pair, i.e., the extreme values ​​of sentiment tendency. Eight core information dimensions for scams targeting the elderly were constructed. The number of unidentified core information dimensions for scams targeting the elderly was counted using the BERT-NER model on the text data. The number of missing dimensions was then used to calculate the information integrity missingness using the integrity missingness formula. The eight core information dimensions for scams targeting the elderly include: organization name, qualification number, risk warning, explanation of fund flow, real name of contact person, office address, regulatory unit, unsubscription mechanism, and exit mechanism. The characteristics of website fraud risk include fraud keyword density, semantic deception index, extreme sentiment value, and information integrity deficiency.

4. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 3, characterized in that, The fraud risk category model adopts a gradient boosting tree integrated architecture that includes a bottom-level model and a meta-model. The bottom-level model includes the XGBoost model and the LightGBM model, and the meta-model is a logistic regression model.

5. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 4, characterized in that, The process of acquiring daily behavioral data from users' mobile phones: The SDK is embedded into the user's mobile application, thereby recording the user's daily behavioral data, including user group category characteristic data and individual user behavior characteristic data. The user group category characteristic data includes the average interval time between single screen clicks, the average screen swipe speed, the mobile phone system font size, and the proportion of time spent using health-related apps. The individual user behavior characteristic data includes the average duration of visits to fraudulent websites, the average interval between sensitive operations, the average time spent inputting sensitive information, and the distribution of operation time periods.

6. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 5, characterized in that, The process of performing behavioral analysis on daily behavioral data: User group category characteristic data and user individual behavior characteristic data are normalized. The normalized user group category characteristic data is then input into a pre-trained random forest model to obtain the user's corresponding group category. The normalized user individual behavior characteristic data is then input into a pre-trained long short-term memory network model to obtain the user's corresponding individual behavior baseline.

7. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 6, characterized in that, The process of matching and analyzing the operational behavior data of elderly users with individual behavioral baselines: The operational behavior data includes sensitive operational intervals. Duration of visits to fraudulent websites Input time for sensitive information and actual operation period Individual behavioral baselines include the average time spent inputting sensitive information. Baseline duration of visits to fraudulent websites Sensitive operation mean interval baseline Deviation from operating time period benchmark ; For actual operation period Deviation from operating time period benchmark The similarity of the operation time period is calculated using the cosine similarity formula of vector space. Based on this, the similarity of the operation time period is determined. Excluding actual operating time Operational behavior data and benchmarks excluding operational time periods The individual's behavioral baseline is input into the risk deviation scoring formula to calculate the behavioral risk score.

8. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 7, characterized in that, When the behavioral risk score is greater than 30, the behavioral risk protection module is triggered.

9. The AI-based smart assistant platform for preventing fraud among the elderly as described in claim 8, characterized in that, The process of forcibly terminating and alerting elderly users regarding their actions on fraudulent websites: The anti-fraud plugin for mobile browsers intercepts critical operation requests from fraudulent websites and returns a 403 Forbidden response. At the same time, it dynamically generates a masking layer using JavaScript to block elderly users from clicking on fraudulent buttons, and calls WeChat or Alipay mini-program interfaces to push alarm information to notify their children.

Images