Large model-based overseas fraud short message identification method

Abstract

The invention aims to provide an overseas fraud short message identification method based on a large model. The method comprises the following steps: constructing an anti-fraud large model; adding a cue word recognition function in the anti-fraud model; and optimizing the anti-fraud model according to different fraud scenes. The large model technology represented by ChatGPT has strong semantic recognition and emotion and intention analysis capabilities, fraud short message feature learning and the natural semantic understanding capability of the large model are combined by means of the large model fine tuning technology, an anti-fraud recognition dichotomous large model is provided, and the anti-fraud recognition dichotomous large model is obtained by means of the cue word engineering technology. According to the method, short message emotion and intention multi-dimensional analysis is realized, fraud short message feature deep recognition is realized, high-variation Mars fraud short messages are recognized, and a new means is provided for fraud short message treatment.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for identifying overseas fraudulent text messages based on a large model. Background Art

[0002] Since the beginning of 2023, the pre-trained large-scale model ChatGPT has demonstrated its superior capabilities in natural language processing. This article leverages big data, large-scale models, and other technologies to establish an intelligent spam SMS defense, management, and blocking system for various scenarios, capable of quickly, accurately, and intelligently intercepting fraudulent SMS messages.

[0003] Traditionally, there are two main approaches to identifying fraudulent text messages: rule-based and machine learning. Rule-based methods perform substitutions based on a predefined table of variant word correspondences. Machine learning methods utilize models such as neural networks to identify and restore variant words.

[0004] Rule-based fraudulent text message identification relies on replacing characters according to a predefined table of variant character correspondences. This approach is simple and fast, but requires manual maintenance of the variant character table, which doesn't cover all possible variant characters and is prone to incorrect substitutions. Machine learning-based fraud text message identification utilizes models like neural networks to identify and restore variant characters. This approach has the advantage of automatically learning the patterns of variant characters, but requires a large amount of labeled data for training. Furthermore, the model's limited representation and generation capabilities make it incapable of handling complex and diverse variant characters. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for identifying overseas fraudulent text messages based on a large model. The method has powerful semantic recognition, emotion and intent analysis capabilities with the help of large model technologies represented by ChatGPT. With the help of large model fine-tuning technology, the feature learning of fraudulent text messages is combined with the natural semantic understanding ability of the large model, and a large anti-fraud identification two-classification model is proposed. The prompt word engineering technology is used to realize multi-dimensional analysis of text message emotions and intentions, realize deep recognition of fraudulent text message features, and realize the recognition of highly variable Martian fraud text messages, providing a new means for the management of fraudulent text messages.

[0006] A method for identifying overseas fraudulent text messages based on a large model, comprising: Build a large anti-fraud model; Adding a prompt word recognition function to the anti-fraud model; Optimize the anti-fraud model according to different fraud scenarios.

[0007] Preferably, the construction of the anti-fraud model includes: The anti-fraud model is obtained by fusing the basic pre-trained large model and the fraud SMS monitoring adaptation layer.

[0008] Preferably, after constructing the anti-fraud big model, the anti-fraud big model is further trained, specifically: Use a specified fine-tuning method to selectively update data-scarce and newly developed modules in the anti-fraud model; Use reparameterization fine-tuning method to change the parameters of the anti-fraud model; Set the pre-trained model to: ; The generated model parameters are: ; Use incremental fine-tuning to add new layers or modules to the anti-fraud model: ; Decomposing the weight change w into a lower-rank representation, assuming w represents the weight matrix in a given neural network layer, and then using regular backpropagation, we can obtain the weight update w, which is usually calculated as the negative gradient of the loss times the learning rate: ; The updated weights are: .

[0009] Preferably, the adding of prompt word recognition function in the anti-fraud model includes: Build a prompt word project: Classify fraudulent text messages into four categories: sentiment classification, deep intent recognition, content domain classification, and judgment and quantitative classification; Set the prompt parameters; Enter different types of fraudulent text messages into the set prompt, and the prompt will output the judgment result.

[0010] Preferably, the parameters for setting the prompt include: Use top-k sampling, top-p sampling, and Temperature sampling to dynamically adjust prompt generation results; The probability of the state in Temperature sampling is expressed as: ; is the probability of the state, is the energy of the state, k is the Boltzmann constant, T is the temperature of the system, and M is the number of all quantum states that the system can reach.

[0011] Preferably, the optimization of the anti-fraud model according to different fraud scenarios includes: Based on the deep data dimension characteristics of fraudulent SMS, we extract generalized features such as character UNICODE interval distribution, character symbol type, and word segmentation statistics; Maintaining the samples and updating the classifier parameters according to the generalized features; For unidentified fraudulent SMS messages, cluster analysis of SMS content and URL feature strings is performed; Determine whether it is a fraudulent text message through manual review; If so, the unidentified fraudulent text messages will be added to the fraudulent text message fingerprint sample library.

[0012] Preferably, the method of extracting generalized features of character UNICODE interval distribution, character symbol type, and word segmentation statistics based on the deep data dimension features of the fraudulent SMS includes: Extract generalized features of the UNICODE interval distribution of characters in mixed languages, rare characters, invisible characters, combined symbols, and emoji expressions; Extracting character symbol type distribution features from the underlying symbol type probability distribution of abnormal SMS messages; The statistical features of word segmentation are extracted from Chinese characters with similar shape variations, word spacing interference, phonetic similar words and English words.

[0013] A large-scale model-based overseas fraud SMS identification system, including: Construction module, used to build anti-fraud models; A function adding module, used to add a prompt word recognition function to the anti-fraud model; The optimization module is used to optimize the anti-fraud model according to different fraud scenarios.

[0014] An electronic device includes: a processor and a memory, the memory is used to store computer program code, the computer program code includes computer instructions, when the processor executes the computer instructions, the electronic device executes a method for identifying overseas fraud text messages based on a large model.

[0015] A computer-readable storage medium stores a computer program, wherein the computer program includes program instructions. When the program instructions are executed by a processor of an electronic device, the processor executes a method for identifying overseas fraud text messages based on a large model.

[0016] The beneficial effects of the present invention are: 1. Based on the pre-trained large model, the present invention adopts the large model training method of superimposing "anti-fraud special prompt words (prompt) + anti-fraud sample fine-tuning (Loray)" to form an anti-fraud large model based on spam text messages, aiming to solve the fraud problem in the text message field. By introducing the anti-fraud large model, a comprehensive, accurate and intelligent spam text message management platform has been successfully created, which can quickly identify, prevent and intercept various types of fraudulent text messages in a short time, providing users with a safer and more reliable communication environment. 2. Compared with the traditional spam text message system based on rule-based methods and machine learning methods, the anti-fraud large model has achieved significant improvements in the prediction accuracy and interception accuracy of fraudulent text messages. After a certain operation The company conducted a pilot verification on overseas text messages, and the number of overseas fraud-related text message cases decreased by 85% year-on-year, and the number of fraud-related warnings decreased by 99.98% year-on-year, demonstrating the system's remarkable achievements in the management of fraudulent text messages. This shows that the anti-fraud large model has not only demonstrated high efficiency and reliability in practice, but also provided an effective solution for addressing overseas text message security issues, and has broad application prospects and promotion value. 3. This solution can not only restore highly variable Martian text fraud messages, but also determine whether the text message is a fraud message. 4. This solution can restore characters with similar pronunciations and similar glyphs, and can also conduct evaluations from multiple dimensions such as text message semantic sentiment feature analysis, intent feature analysis, and fraud suspicion, to comprehensively and accurately predict the content of fraudulent text messages. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0019] Figure 1 This is a flow chart of a method for identifying overseas fraudulent text messages based on a large model of the present invention; Figure 2 This is a schematic diagram of the anti-fraud model structure of the present invention; Figure 3 This is a schematic diagram of fine-tuning the anti-fraud large model of the present invention; Figure 4 This is a schematic diagram of the prompt construction and test results of the present invention; Figure 5 This is a schematic diagram of the multi-dimensional evaluation of the anti-fraud model of the present invention; Figure 6This is a schematic diagram of the generalized features and applicable scenarios of the present invention; Figure 7 This is a schematic diagram of generalized feature extraction of the present invention; Figure 8 This is a schematic diagram comparing the prediction accuracy of the anti-fraud model of the present invention; Figure 9 This is a schematic diagram comparing the predicted interception rates of the anti-fraud large model system of the present invention; Figure 10 This is a schematic diagram of the comparison values ​​of fraud cases according to the present invention; Figure 11 This is a schematic diagram of the comparison value of the average daily overseas SMS warning volume of the present invention; Figure 12 The figure shows the hardware structure of an electronic device of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0021] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0022] In addition, the descriptions of "first", "second", etc. in the present invention are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0023] Rule-based fraudulent text message identification relies on replacing characters according to a predefined table of variant character correspondences. This approach is simple and fast, but requires manual maintenance of the variant character table, which doesn't cover all possible variant characters and is prone to incorrect substitutions. Machine learning-based fraud text message identification utilizes models like neural networks to identify and restore variant characters. This approach has the advantage of automatically learning the patterns of variant characters, but requires a large amount of labeled data for training. Furthermore, the model's limited representation and generation capabilities make it incapable of handling complex and diverse variant characters.

[0024] The present invention is based on a pre-trained large model and adopts a large model training method of superimposing "anti-fraud special prompt words (prompt) + anti-fraud sample fine-tuning (Loray)" to form an anti-fraud large model based on spam text messages, aiming to solve the fraud problem in the text message field. By introducing the anti-fraud large model, a comprehensive, accurate and intelligent spam text message management platform has been successfully created, which can quickly identify, prevent and intercept various types of fraudulent text messages in a short period of time, providing users with a safer and more reliable communication environment. Compared with the traditional spam text message system based on rule-based methods and machine learning methods, the anti-fraud large model has achieved significant improvements in the prediction accuracy and interception accuracy of fraudulent text messages. After an operator targeted overseas short messages, the anti-fraud large model was used to identify, prevent and intercept all kinds of fraudulent text messages. The pilot verification of the system showed that the number of overseas fraudulent text message cases decreased by 85% year-on-year, and the number of fraud warnings decreased by 99.98% year-on-year, demonstrating the system's significant achievements in the management of fraudulent text messages. This shows that the anti-fraud model has not only demonstrated high efficiency and reliability in practice, but also provided an effective solution for addressing overseas text message security issues, with broad application prospects and promotion value. This solution can not only restore highly variable Martian text fraud messages, but also determine whether the text message is a fraudulent text message. This solution can restore characters with similar pronunciations and similar glyphs, and can also evaluate from multiple dimensions such as text message semantic sentiment feature analysis, intent feature analysis, and fraud suspicion, to comprehensively and accurately predict the content of fraudulent text messages.

[0025] Example 1 A method for identifying overseas fraud SMS based on a large model, reference Figure 1 ,include: S100, building a large anti-fraud model; This solution integrates the "Anti-Fraud Big Model" with traditional online spam message monitoring systems to achieve integrated identification, prevention, management, and combating of fraudulent text messages. Based on this "big model," spam message control uses reasoning to accurately identify various variants of fraudulent text messages, thereby improving overall recognition accuracy. The system conducts secondary learning on a pool of caller numbers for fraudulent text messages, intelligently identifying other suspicious text messages, deeply analyzing the multifaceted nature of the issue, and promptly updating the sample database with intelligent analysis results.

[0026] S200, adds prompt word recognition function to the anti-fraud model; Prompt word engineering plays a key role in improving the performance of large models. By carefully designing and optimizing prompts, large models can more accurately understand and respond to human language input, significantly improving their performance and performance. The goal of prompt word engineering is to create efficient and controllable AI systems. Carefully designed prompts can enable large models to perform specific tasks with greater accuracy and reliability. For example, in question-answering systems, optimizing prompts enables models to more accurately understand questions and provide appropriate responses. In natural language generation tasks, prompt word engineering can help models generate more fluent and logical text. When designing prompts, multiple factors must be considered, including a clear definition of the task and objectives, concise and clear expression, guidance methods for the model, and consideration of the model's capabilities. When designing prompt words for large anti-fraud models, these factors must be considered and adjusted accordingly for the task of identifying fraudulent text messages.

[0027] S300 optimizes the anti-fraud model based on different fraud scenarios.

[0028] Optimizing the anti-fraud model includes building a generalized feature model. A generalized feature model refers to a model with good generalization ability and robustness, both of which are key indicators for evaluating machine learning performance.

[0029] Generalization measures a model's ability to predict new, unseen data. A model with good generalization is able to learn general patterns from limited training data and perform well on new data. Robustness refers to a model's ability to handle perturbations or anomalies in the input data. A robust model maintains stable performance in the face of noise, missing data, or other anomalies. This allows the model to tolerate uncertainty in the input data and avoid overreacting to subtle changes or noise.

[0030] Preferably, S100, building an anti-fraud model includes: The anti-fraud model is obtained by fusing the basic pre-trained large model and the fraud SMS monitoring adaptation layer.

[0031] The training of the anti-fraud big model is implemented in the B+Z method, where B represents the basic pre-trained big model and z represents the fraud SMS monitoring adaptation layer. The pre-trained big model B has tens of billions of parameters and has powerful natural language semantic analysis, sentiment recognition, logical reasoning and induction capabilities; Z represents the fraud SMS recognition superposition model package, which collects sample data through multiple dimensions and plans special anti-fraud field training, aiming to initially build an anti-fraud big model by inputting millions of fraud SMS contents as training samples. On this basis, by superimposing Z's "anti-fraud special prompt words (prompt)... + anti-fraud sample fine-tuning (Lora)" big model training method, the goal of achieving fraud SMS recognition accuracy and recall rate of more than 95% with hundreds of thousands of training samples is achieved. At the same time, this method can also guarantee the general natural language processing capabilities of the pre-trained big model. The structure of the anti-fraud big model is as follows Figure 2 shown.

[0032] Preferably, after S100, constructing the anti-fraud big model, the method further includes S110, training the anti-fraud big model, specifically: S111 uses the specified fine-tuning method to selectively update data-scarce and newly released modules in the anti-fraud model; specified fine-tuning allows selective updating of specific parts of the model, which is suitable for sudden new tasks or data-scarce situations.

[0033] S112, uses the reparameterization fine-tuning method to change the parameters of the anti-fraud model; reparameterization fine-tuning is fine-tuning by changing the parameterization form of the model, which is more applicable when processing large-scale data or complex tasks.

[0034] While the basic pre-trained large model has natural language semantic analysis, sentiment recognition, logical reasoning, and induction capabilities, it cannot identify spam messages. Therefore, the anti-fraud large model uses an incremental approach to model training, superimposing anti-fraud samples to fine-tune the large model training method.

[0035] S113, set the pre-training model to: ; The generated model parameters are: ; S114, using incremental fine-tuning to add new layers or modules to the anti-fraud model: ; Incremental fine-tuning is done by adding new layers or modules and training them to expand model capabilities or adapt to new tasks.

[0036] Decomposing the weight change w into a lower-rank representation, assuming w represents the weight matrix in a given neural network layer, and then using regular backpropagation, we can obtain the weight update w, which is usually calculated as the negative gradient of the loss times the learning rate: ; The updated weights are: .

[0037] The core technologies in the field of pre-trained language models have made tremendous progress in recent years. This growth is driven by a combination of factors, including algorithmic improvements, increased computing power, and increased data resources. As the model size increases, PLM not only significantly improves performance on single tasks but, more importantly, demonstrates great potential in solving more complex and unknown tasks. However, as the model size increases, the challenges users face become more severe, with the consumption of computing and storage resources being a prominent issue. Traditional full-parameter fine-tuning methods require extensive GPU computing and storage resources, resulting in significant cost investment, which limits their use by researchers or businesses. Consequently, parameter-efficient methods have attracted considerable attention. These methods reduce computing and storage costs by fine-tuning only a small portion of the model's parameters while maintaining good performance.

[0038] Common incremental methods include adapter tuning, prefix tuning, prompt tuning, and LoRA. Consider the advantages of LoRA fine-tuning. LoRA (Low-Rank Adaptation) fine-tuning is an implicit low-rank transformation applied to large model weight matrices. The general idea is to find a small number of dimensional combinations in the original feature space that can capture most of the information in the dataset. LoRA proposes decomposing the weight change w into a lower-rank representation. Assuming w represents the weight matrix in a given neural network layer, conventional backpropagation can be used to obtain the weight update w, which is typically calculated as the negative gradient of the loss multiplied by the learning rate. When training fully connected layers in a neural network, the weight matrix typically has full rank, meaning that the matrix does not have any linearly dependent rows or columns. In contrast, low rank means that the matrix has redundant rows or columns compared to full rank. Therefore, although the weights of the pre-trained model have full rank on the pre-training task, the pre-trained large language model has a low "intrinsic dimensionality" when adapted to a new task. Low intrinsic dimensionality means that the data can be effectively represented or approximated by a low-dimensional space while retaining most of its essential information or structure.

[0039] like Figure 3 As shown in the figure, after applying fine-tuning techniques, the anti-fraud model can understand and identify some fraudulent text messages. Given the potential for inference errors, it is necessary to introduce prompt word engineering.

[0040] Preferably, S200, adding a prompt word recognition function to the anti-fraud model includes: S210, build prompt word engineering: classify fraudulent text messages into four categories: sentiment classification, deep intent recognition, content domain classification, and judgment quantification classification; S220, set prompt parameters; S230, input different types of fraudulent text messages into the set prompt, and the prompt outputs the judgment result.

[0041] In AI scenarios, a prompt is an initial input or hint given to a model, guiding it to generate a specific output. In chat interactions, users can provide a question or topic as a prompt, prompting the model to generate a corresponding response. In text generation tasks, a prompt can specify a preceding text, allowing the model to generate a coherent text based on it.

[0042] The existing prompt methodology framework is CRISPE, where each letter means the following: CR: Capabilities and roles, that is, what role you want the big model to play.

[0043] I: Insight, clearly articulate background and contextual information.

[0044] S: State what you want the large model to do.

[0045] P: Personality, how you hope the big model will answer the question.

[0046] E: Experiments, testing the answers provided by the large model.

[0047] According to the above methodology, the prompt is constructed. The specific results are as follows Figure 4 As shown in the figure, prompts play a crucial role in the field of artificial intelligence, especially in large-scale model applications. They are not only key to the model's understanding of input text and generating appropriate responses, but also a key factor in evaluating model performance. As input to the model, prompts provide task guidance and specific execution requirements. Well-designed prompts can accurately and clearly convey the user's intent and needs, guiding the model to generate the desired response.

[0048] A well-designed prompt helps the model better understand the task, improving its accuracy and efficiency. Therefore, when designing a prompt, consider the model's parameter settings to ensure that the prompt fully utilizes the model's capabilities and adapts to its workings. The selection and setting of parameters has a significant impact on the model's performance, optimizing its accuracy, generalization, and computational efficiency.

[0049] Preferably, in S220, setting prompt parameters includes: Use top-k sampling, top-p sampling, and Temperature sampling to dynamically adjust prompt generation results; The probability of the state in Temperature sampling is expressed as:

[0050] is the probability of the state, is the energy of the state, k is the Boltzmann constant, T is the temperature of the system, and M is the number of all quantum states that the system can reach.

[0051] Designing prompt parameters is a complex and sophisticated process, with the choice of decoding strategy being particularly crucial. Greedy decoding and random sampling are two commonly used decoding strategies, each with its own unique characteristics and applicable scenarios. Greedy decoding uses the word with the highest current probability as its output. This is efficient and deterministic, enabling rapid text generation, but can result in monotonous and repetitive text that lacks diversity and innovation. In contrast, random sampling selects words randomly based on a probability distribution, increasing the diversity and creativity of the generated text. However, this can result in incoherent or meaningless text, so careful control of sampling probability is crucial.

[0052] Top-k, top-p, and temperature are commonly used parameters in generative models, controlling different aspects of the model's text generation. Top-k and top-p sampling, situated between greedy decoding and random sampling, are important approaches for generating text. Top-k sampling optimizes greedy decoding by sampling from the top k ranked tokens, increasing the diversity and creativity of generated text. The length of the sampling list can be controlled by adjusting the value of k. In essence, greedy decoding is top-k sampling with k=l.

[0053] In top-k sampling, only the k words with the highest probability are randomly sampled at each step, avoiding the appearance of low-probability words while retaining a certain degree of randomness and creativity.

[0054] Although top-k sampling can dynamically adjust the number of candidate words, controlling the diversity and quality of generated text, it can also lead to text that lacks common sense or logic. To address this limitation, a top-p sampling strategy has emerged, which dynamically adjusts the size of the candidate list. Top-p sampling randomly samples only from the smallest set of words whose cumulative probability exceeds a threshold p, avoiding irrelevant words while retaining interesting or creative words.

[0055] Temperature sampling is inspired by statistical thermodynamics; higher temperatures mean a higher probability of encountering lower-energy states. In probabilistic models, logits play the role of energy. Temperature sampling can be achieved by dividing the logits by the temperature, which is then input into Sottrnax to obtain the sampling probability. Lower temperatures make the model more confident in its preferred choice, while temperatures above 1 decrease confidence. A temperature of 0 is equivalent to the argmax likelihood, while an infinite temperature is equivalent to uniform sampling. The temperature adjusts the probability distribution of the word. The ultimate macro effect is that the model becomes more deterministic at lower temperatures and less so at higher temperatures.

[0056] The anti-fraud model uses top-k, top-p, and temperature together in the order of top-k > top-p > temperature. Random sampling can be performed from the above distribution to select a word as the final generated result. After determining the parameters, the generated result is as follows: Figure 5 shown.

[0057] Preferably, in S300, optimizing the anti-fraud model according to different fraud scenarios includes: S310, extracting generalized features such as character UNICODE interval distribution, character symbol type, and word segmentation statistics based on the deep data dimension features of the fraudulent text message; S320, maintaining the samples and updating the classifier parameters according to the generalized features; S330, for unidentified fraudulent SMS messages, cluster analysis of SMS content and URL feature strings is performed on the messages; S340, determining whether the text message is fraudulent through manual review; S350: If yes, the unidentified fraudulent SMS is added to the fraudulent SMS fingerprint sample library.

[0058] Fraudulent text messages are characterized by regional and temporal variations, and rapid mutation. They employ counter-monitoring strategies such as character variants, multilingual mixed encoding, similar shapes and sounds, and symbolic interference. Traditional monitoring rules and publicly available neural network models utilize specific Chinese characters, symbols, and word segmentation features from text messages to configure rules or train models. However, these models are short-lived, requiring frequent rule updates and model retraining.

[0059] Therefore, the anti-fraud model uses generalized features to train the model robustly. Generalized features refer to general features that are applicable to multiple scenarios and can help the model better adapt to different types of fraud SMS scenarios, such as Figure 6 Take the text message "Baijiajie video sports electronics, etc., share the excitement with you, withdraw (money) quickly without limit!" as an example. Figure 7 shown.

[0060] Preferably, in step S310, based on the deep data dimension features of the fraudulent SMS, the generalized features of character UNICODE interval distribution, character symbol type, and word segmentation statistics are extracted, including: S311, extracts generalized features of the UNICODE interval distribution of characters in mixed languages, rare characters, invisible characters, combined symbols, and emoji expressions; S312, extracting character symbol type distribution features from the underlying symbol type probability distribution of the abnormal SMS; S313, extracting word segmentation statistical features from Chinese character variants, word spacing interference, phonetic similar words and English words.

[0061] Based on neural network sample training and intelligent learning recognition technology, combined with semantic association information, it can automatically identify generalized features in fraudulent text messages, establish a neural network monitoring model, and achieve real-time interception of fraudulent text messages. This is specifically divided into two aspects: online classification and prediction. The system mines deep statistical data dimensional features of fraudulent text messages, selecting generalized features such as text character Unicode interval distribution, symbol type, and word segmentation statistics. This improves the accuracy and adaptability of fraudulent text message identification and reduces the required model training cycle. The model also supports online classification and prediction, as well as continuous classifier learning and updating, including sample maintenance and classifier parameter updates. Cluster analysis and real-time monitoring: For undetected fraudulent text messages, the system performs cluster analysis on message content and URL feature strings, and uses the clustering results for real-time monitoring. The top N messages in the cluster are manually reviewed to determine whether they are fraudulent. If so, they are added to the fraudulent text message fingerprint sample library. The system supports automatic fraud detection based on set thresholds, enabling the establishment of a more robust feature model.

[0062] Use the anti-fraud big model to enable intelligent analysis of overseas SMS messages and make comparison predictions with the spam SMS system that does not use the big model. Figure 8 、 Figure 9 As shown. Comparing the above results, we can see that: The prediction accuracy rate for overseas fraud text messages remains stable at over 99%, which is significantly higher than the 75% level of traditional spam text message systems.

[0063] The accuracy rate of intercepting overseas fraudulent text messages is between 80.63% and 99.60%, which is significantly higher than the 51% level of traditional spam text message systems. Before the anti-fraud large model was activated, the average daily overseas text message volume of a certain operator was 500,000 to 600,000 (including normal text messages and fraudulent text messages that were not identified by the existing spam text message system); one week after the activation, it dropped sharply to an average of 20,000 to 30,000 normal text messages per day. Since the launch of the fraud SMS intelligent prevention and management system, the number of overseas fraud SMS cases has decreased by 64% month-on-month and 85% year-on-year; the number of fraud warnings has decreased by 91.23% month-on-month and 99.98% year-on-year, and the effect is very significant. The number of fraud cases and the average daily overseas SMS warning volume are as follows: Figure 10 、 Figure 11 shown.

[0064] This solution, based on a pre-trained large model, employs a large model training method that superimposes "anti-fraud prompts" and "fine-tuning" anti-fraud samples (Loray) to form an anti-fraud large model based on spam text messages, aiming to address the problem of SMS fraud. By introducing the anti-fraud large model, a comprehensive, accurate, and intelligent spam SMS management platform has been successfully established, capable of rapidly identifying, preventing, and intercepting all types of fraudulent text messages, providing users with a more secure and reliable communication environment. Compared to traditional spam SMS systems, the anti-fraud large model has achieved significant improvements in the prediction and interception accuracy of fraudulent text messages. A pilot test conducted by a certain operator on overseas text messages showed that the number of overseas fraudulent text message cases decreased by 85% year-on-year, and the number of fraudulent warnings decreased by 99.98% year-on-year, demonstrating the system's significant achievements in fraudulent text message management. This demonstrates that the anti-fraud large model not only demonstrates efficiency and reliability in practice, but also provides an effective solution to addressing overseas SMS security issues, with broad application prospects and promotional value.

[0065] Example 2 A large-scale model-based overseas fraud SMS identification system, including: Construction module, used to build anti-fraud models; Function addition module, used to add prompt word recognition function to the anti-fraud model; The optimization module is used to optimize the anti-fraud model according to different fraud scenarios.

[0066] Example 3 An electronic device includes: a processor and a memory, the memory is used to store computer program code, the computer program code includes computer instructions, when the processor executes the computer instructions, the electronic device executes a method for identifying overseas fraud text messages based on a large model.

[0067] refer to Figure 12The electronic device 2 includes a processor 21, a memory 22, an input device 23, and an output device 24. The processor 21, the memory 22, the input device 23, and the output device 24 are coupled via a connector, which may include various interfaces, transmission lines, or buses, etc., but this is not limited in the present embodiment. It should be understood that in various embodiments of the present invention, coupling refers to mutual connection in a specific manner, including direct connection or indirect connection through other devices, such as connection via various interfaces, transmission lines, buses, etc.

[0068] The processor 21 may be one or more graphics processing units (GPUs). If the processor 21 is a GPU, the GPU may be a single-core GPU or a multi-core GPU. Alternatively, the processor 21 may be a processor group consisting of multiple GPUs, with the multiple processors coupled to each other via one or more buses. Alternatively, the processor may be another type of processor, and this is not limited in this embodiment of the present invention.

[0069] The memory 22 can be used to store computer program instructions and various computer program codes, including program codes for executing the embodiments of the present invention. Optionally, the memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or compact disc read-only memory (CD-ROM), and is used for related instructions and data.

[0070] The input device 23 is used to input data and / or signals, and the output device 24 is used to output data and / or signals. The output device 24 and the input device 23 can be independent devices or an integrated device.

[0071] Example 4 A computer-readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor of an electronic device, the processor executes a method for identifying overseas fraud text messages based on a large model.

[0072] The present invention is based on a pre-trained large model and adopts a large model training method of superimposing "anti-fraud special prompt words (prompt) + anti-fraud sample fine-tuning (Loray)" to form an anti-fraud large model based on spam text messages, aiming to solve the fraud problem in the text message field. By introducing the anti-fraud large model, a comprehensive, accurate and intelligent spam text message management platform has been successfully created, which can quickly identify, prevent and intercept various types of fraudulent text messages in a short period of time, providing users with a safer and more reliable communication environment. Compared with the traditional spam text message system based on rule-based methods and machine learning methods, the anti-fraud large model has achieved significant improvements in the prediction accuracy and interception accuracy of fraudulent text messages. After an operator targeted overseas short messages, the anti-fraud large model was used to identify, prevent and intercept all kinds of fraudulent text messages. The pilot verification of the system showed that the number of overseas fraudulent text message cases decreased by 85% year-on-year, and the number of fraud warnings decreased by 99.98% year-on-year, demonstrating the system's significant achievements in the management of fraudulent text messages. This shows that the anti-fraud model has not only demonstrated high efficiency and reliability in practice, but also provided an effective solution for addressing overseas text message security issues, with broad application prospects and promotion value. This solution can not only restore highly variable Martian text fraud messages, but also determine whether the text message is a fraudulent text message. This solution can restore characters with similar pronunciations and similar glyphs, and can also evaluate from multiple dimensions such as text message semantic sentiment feature analysis, intent feature analysis, and fraud suspicion, to comprehensively and accurately predict the content of fraudulent text messages.

[0073] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is intended to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for identifying overseas fraudulent text messages based on a large model, characterized in that: include: Build a large anti-fraud model; Adding a prompt word recognition function to the anti-fraud model; Optimize the anti-fraud model according to different fraud scenarios.

2. According to the large model-based overseas fraud SMS identification method of claim 1, it is characterized in that: The construction of the anti-fraud model includes: The anti-fraud big model is obtained by integrating the basic pre-trained big model and the fraud SMS monitoring adaptation layer.

3. According to the large model-based overseas fraud SMS identification method of claim 1, it is characterized in that: After the anti-fraud big model is constructed, the anti-fraud big model is also trained, specifically: Use the specified fine-tuning method to selectively update the data-scarce modules and newly released modules in the anti-fraud model; The reparameterization fine-tuning method is used to change the parameters of the anti-fraud model; Set the pre-trained model to: ; The generated model parameters are: ; Use incremental fine-tuning to add new layers or modules to the anti-fraud model: ； Decomposing the weight change w into a lower-rank representation, assuming that w represents the weight matrix in a given neural network layer, and then using regular backpropagation, we can obtain the weight update w, which is usually calculated as the negative gradient of the loss times the learning rate: ; The updated weights are: .

4. According to the large model-based overseas fraud SMS identification method of claim 1, it is characterized in that: The function of adding prompt word recognition to the anti-fraud model includes: Construct prompt word engineering: classify fraudulent SMS into four categories: sentiment classification, deep intent recognition, content domain classification, and judgment quantification classification; Set the prompt parameters; Input different types of fraudulent text messages into the set prompt, and the prompt will output the judgment result.

5. According to the large model-based overseas fraud SMS identification method of claim 4, it is characterized in that: The parameters for setting the prompt include: Use top-k sampling, top-p sampling and Temperature sampling to dynamically adjust the prompt generation result; The probability of the state in Temperature sampling is expressed as: ; is the probability of the state, is the energy of the state, k is the Boltzmann constant, T is the temperature of the system, and M is the number of all quantum states that the system can reach.

6. According to the large model-based overseas fraud SMS identification method of claim 1, it is characterized in that: The optimization of the anti-fraud model according to different fraud scenarios includes: Based on the deep data dimension characteristics of fraudulent SMS, the generalized features of character UNICODE interval distribution, character symbol type, and word segmentation statistics are extracted; Maintaining the samples and updating classifier parameters according to the generalized features; For unidentified fraudulent SMS, cluster analysis of SMS content and URL feature strings is performed; Determine whether it is a fraudulent text message through manual review; If so, the unidentified fraudulent text messages will be added to the fraudulent text message fingerprint sample library.

7. According to claim 6, a method for identifying overseas fraudulent text messages based on a large model is characterized in that: The generalized features of extracting character UNICODE interval distribution, character symbol type, and word segmentation statistics based on the deep data dimension features of the fraudulent SMS include: Extract generalized features of the distribution of UNICODE intervals of characters in mixed languages, rare characters, invisible characters, combined symbols, and emoji expressions; Extracting character symbol type distribution features from the underlying symbol type probability distribution of abnormal SMS; The statistical features of word segmentation are extracted from Chinese characters with similar shape variations, word spacing interference, phonetic similar words and English words.

8. A large-scale model-based overseas fraud SMS identification system, characterized in that: include: Construction module, used to build anti-fraud big model; A function adding module, used to add a prompt word recognition function to the anti-fraud model; The optimization module is used to optimize the anti-fraud model according to different fraud scenarios.

9. An electronic device, characterized in that: include: A processor and a memory, wherein the memory is used to store computer program code, wherein the computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method for identifying overseas fraud SMS based on a large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor of an electronic device, the processor executes a large-model-based method for identifying overseas fraud text messages as described in any one of claims 1 to 7.

Images