Fraud-related text adaptive generation and closed-loop evolution method based on semantic blind area exploration

Abstract

The application relates to the technical field of telecommunication security, and discloses a fraud-related text adaptive generation and closed-loop evolution method based on semantic blind area exploration, which comprises the following steps: constructing a fraud-related semantic memory library, a normal semantic memory library and a semantic retrieval index; after new call text is semantically coded, double-space near neighbor retrieval and local semantic density comparison are carried out, and candidate samples possibly located in a semantic blind area are screened out; the candidate samples are subjected to semantic blind area identification and layering based on local semantic density and global semantic offset; different levels of semantic blind areas are dynamically selected to generate a fraud-related text generation strategy; a value evaluation function is constructed to quantitatively evaluate the information gain of the generated fraud-related text, and high-value samples are screened out; the screened high-value samples are subjected to artificial confirmation and incremental updating, and a dynamic adaptive closed-loop feedback mechanism is constructed. By using the method, the rapid adaptive capacity, recall capacity and long-term robustness of unknown fraud tactics are improved.

Description

Technical Field

[0001] This invention relates to the field of telecommunications security technology, specifically to a method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration. Background Technology

[0002] With the widespread adoption of communication technologies and internet applications, telecommunications fraud has become a major threat seriously endangering social security and citizens' property rights. Although existing fraud detection and sample expansion technologies have made some progress, the following shortcomings still exist: Difficulty in proactively detecting unknown fraud patterns: Existing rules, machine learning, and deep learning methods all rely on existing labeled samples for learning. When fraudulent rhetoric evolves dynamically, they lack the ability to proactively identify blind spots at the semantic space level, resulting in a lag in the detection of new fraud patterns and a high rate of missed detections.

[0003] Sample generation lacks specificity: Traditional data augmentation methods only perform lexical or syntactic perturbations, making it difficult to generate new samples with actual information gain; random generation based on large language models lacks analysis of semantic coverage gaps, easily generating a large number of semantically redundant or invalid noise samples.

[0004] Lack of proactive evolution mechanism based on semantic space structure: Existing proactive learning and incremental learning mainly focus on local optimization around uncertain samples of the model, making it difficult to identify uncovered areas from the global semantic space, and also lacking the ability to dynamically adjust exploration strategies according to the evolution of fraud patterns.

[0005] Semantic space analysis and model optimization are disconnected: existing clustering, anomaly detection and vector retrieval techniques are mostly used for static analysis or similar sample recall. The semantic structure information they generate fails to effectively guide sample generation, value assessment and model updates, and lacks a closed-loop feedback mechanism.

[0006] Lack of an effective information gain evaluation mechanism: Existing methods mostly rely on binary judgments of "whether the generation was successful" or "whether it was misclassified by the classifier" when screening generated samples. It is difficult to comprehensively evaluate the semantic novelty, blind zone coverage, and classification boundary challenge value of generated samples, resulting in a large number of low-value samples entering the training process and affecting the efficiency of model optimization.

[0007] To address this, we have invented an adaptive generation and closed-loop evolution method for fraudulent text based on semantic blind spot exploration, which solves the above-mentioned technical problems. Summary of the Invention

[0008] This invention provides a method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration, which improves the system's ability to quickly adapt to unknown fraudulent rhetoric, enhances its recall ability for unknown fraud patterns, and improves its long-term robustness.

[0009] Therefore, the present invention provides the following technical solution: A method for adaptive generation and closed-loop evolution of fraud-related text based on semantic blind spot exploration, the method comprising: Step 1: Construct a semantic memory for fraud-related issues, a semantic memory for normal information, and a semantic retrieval index; Step 2: After semantically encoding the new call text, perform dual spatial nearest neighbor retrieval and local semantic density comparison to screen out candidate samples that may be located in semantic blind spots. Step 3: Semantic blind spot identification and stratification of candidate samples based on local semantic density and global semantic offset; Step 4: Dynamically select a fraud-related text generation strategy to address different levels of semantic blind spots and generate fraud-related text. Step 5: Construct a value assessment function to quantitatively evaluate the information gain of the generated fraudulent texts and select high-value samples. Step 6: Manually confirm and incrementally update the selected high-value samples to build a dynamic and adaptive closed-loop feedback mechanism.

[0010] Optionally, step 1 includes: Step 11: Encode the confirmed historical fraudulent call texts into semantic vectors, and after deduplication, establish the fraudulent semantic memory V. db ; Step 12: Encode the confirmed historical normal call text into semantic vectors, and after deduplication, establish the normal semantic memory N. db ; Step 13: Construct a semantic retrieval index using an approximate nearest neighbor index structure.

[0011] Optionally, in step 2, after semantically encoding the new call text, a dual-space nearest neighbor search and local semantic density comparison are performed based on the fraud-related semantic memory and the normal semantic memory to filter out candidate samples that are biased towards the fraud-related semantic space and may be located in the semantic blind zone, specifically including: Step 21: Semantically encode the new call text x from the speech-to-text transcription to obtain a semantically encoded vector v. x ; Step 22: Perform dual-space nearest neighbor retrieval, searching for nearest neighbors in both the fraud-related semantic memory and the normal semantic memory to obtain the TopK historical fraud-related semantic neighborhoods that are closest to the call text x. fraud (x) and TopK semantic neighborhood of historical risk-free calls normal (x); Step 23: Perform dual-space local semantic density calculation, and calculate the local semantic density ρ of the call text x in the historical fraud space. fraud(x) and the local semantic density ρ of the call text x in the historical risk-free call semantic space. normal (x); Step 24: Perform preliminary screening by comparing the local semantic density of the two spaces; When ρ normal (x)>ρ fraud When (x), it means that the local semantic density of the call text x in the normal semantic space is higher than the local semantic density in the fraud-related semantic space. That is, the call text x is more biased towards the normal semantic space and will not enter the subsequent process. When ρ normal (x)<=ρ fraud When (x) indicates that the local semantic density of the call text x in the fraud-related semantic space is higher than that in the normal semantic space, the call text x is more biased towards the fraud-related semantic space. The call text x is then selected as a candidate sample, and the process proceeds to step 3.

[0012] Optionally, in step 3, by jointly calculating the local semantic density of candidate samples in the fraud semantic space and the global semantic offset distance relative to the fraud semantic center, semantic blind spots that are not adequately covered by the existing system are identified, and the semantic blind spots are divided into three levels: neighborhood variant blind spots, cross-domain migration blind spots, and novel pattern blind spots.

[0013] Optionally, step 3 includes: Step 31, based on the local semantic density ρ of the candidate samples and the historical fraud space. fraud (x) performs semantic blind zone identification, when ρ fraud When (x) < θ, the local semantic density of the candidate sample is considered to be low, indicating that there are almost no historical fraud samples nearby. It belongs to a low-coverage semantic region, i.e., a semantic blind zone. Here, θ is the blind zone determination threshold, and the process proceeds to step 32. When ρ fraud When (x)>=θ, it is considered that the local semantic density of the candidate sample is high, the historical fraud semantic space has been covered by a certain number of samples, and it does not belong to the semantic blind zone, and the process ends. Step 32: Calculate the semantic center of fraud in existing historical fraud samples. :

[0014] Among them, v i is the semantic vector corresponding to the fraudulent call text in the fraudulent semantic memory; N is the total number of call texts in the existing fraudulent semantic memory. Calculate the global semantic offset distance D(x) of the candidate sample currently identified as a semantic blind spot relative to the fraud-related semantic center: D(x) = 1 - cos(v) x ,μfraud ) wherein, cos() is a cosine similarity function, and a larger D(x) indicates that the candidate sample deviates farther from the existing fraud semantic space; step 33, dividing the semantic blind area into three levels of the neighborhood variant blind area, the cross-domain migration blind area and the brand-new mode blind area according to the global semantic offset distance based on a first blind area division threshold β1 and a second blind area division threshold β2: when 0 < D(x) < β1, the global semantic offset distance is considered to be low, which indicates that the candidate sample is located in an edge region of the known fraud semantics, and the semantic blind area is divided into the neighborhood variant blind area; when β1 ≤ D(x) < β2, the global semantic offset distance is considered to be medium, which indicates that the candidate sample has multi-type fraud speech text features at the same time, and the semantic blind area is divided into the cross-domain migration blind area; when D(x) ≥ β2, the global semantic offset distance is considered to be high, which indicates that the candidate sample is far away from the known fraud semantic space, and the semantic blind area is divided into the brand-new mode blind area; wherein, the first blind area division threshold β1 is smaller than the second blind area division threshold β2.

[0015] optionally, in step 4, for the neighborhood variant blind area, a local semantic perturbation generation strategy is adopted around the existing fraud semantic neighborhood to drive a large language model to generate fraud-related texts, and specific generation directions include: performing synonym replacement on key induction words in neighboring speech texts, transforming and recombining the word order of speech text sentences, performing variant design on fraud induction modes and rewriting the overall expression style; for the cross-domain migration blind area, a multi-mode recombination generation strategy fused with semantic features of multiple fraud modes is adopted to drive a large language model to generate fraud-related texts, and specific generation directions include: mixing and splicing different types of fraud speech texts, and migrating fraud modes to new business scenarios for adaptation transformation; for the brand-new mode blind area, a long-distance semantic expansion generation strategy is adopted to drive a large language model to generate fraud-related texts, and specific generation directions include: constructing a brand-new induction logic chain, designing a new fraud scene script and arranging an unknown fraud interaction process.

[0016] optionally, in step 5, the value evaluation function constructed by fusing boundary gain, coverage gain and novelty gain is:

[0017] wherein, λ1 is the boundary gain a weighting coefficient of, λ2 is the coverage gain a weighting coefficient of, λ3 is the novelty gain The weighting coefficients, λ1+λ2+λ3=1; boundary gain Measuring the degree to which the generated fraudulent text challenges the classification boundaries; coverage gain. Indicates whether the generated fraudulent text truly fills the gap; novelty gain. Indicates the novelty gain of the generated fraudulent text; Fraudulent texts that meet the criteria of high classification boundary challenges, high blind zone coverage, and low semantic redundancy will be selected as high-value samples.

[0018] Optionally, in step 6, the selected high-value samples, after manual confirmation, enter the incremental update process, triggering a dynamic adaptive closed-loop feedback mechanism. This dynamic adaptive closed-loop feedback mechanism includes: The system includes a dynamic blind zone determination threshold adjustment mechanism, a dynamic blind zone hierarchy division threshold update mechanism, a dynamic generation strategy weight adjustment mechanism, and a dynamic sample priority mechanism based on generation value. The dynamic blind zone determination threshold adjustment mechanism is used to adjust the sensitivity of exploring unknown fraud patterns. The dynamic blind zone hierarchy division threshold update mechanism adaptively adjusts the first blind zone division threshold β1 and the second blind zone division threshold β2 based on the offset of the fraud classifier's confidence decay curve, achieving continuous optimization of the division boundaries for the three blind zone types. The dynamic generation strategy weight adjustment mechanism is used to adjust the allocation of generation resources for different types of fraud patterns. The dynamic sample priority mechanism is used to prioritize candidate samples with high information gain.

[0019] A system for adaptive generation and closed-loop evolution of fraud-related text based on semantic blind spot exploration, the system comprising: The semantic memory building unit constructs a fraud-related semantic memory, a normal semantic memory, and a semantic retrieval index. The candidate sample screening unit performs semantic encoding on the new call text, performs dual spatial nearest neighbor retrieval and local semantic density comparison, and screens out candidate samples that may be located in the semantic blind zone. The semantic blind spot identification and hierarchical unit performs semantic blind spot identification and hierarchical analysis on candidate samples based on local semantic density and global semantic offset. The fraud-related text generation unit dynamically selects fraud-related text generation strategies to generate fraud-related texts, targeting different levels of semantic blind spots. Generate value assessment units, construct value assessment functions, quantify the information gain of the generated fraudulent texts, and select high-value samples. The closed-loop feedback unit manually confirms and incrementally updates the selected high-value samples, thus constructing a dynamic and adaptive closed-loop feedback mechanism.

[0020] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to perform the steps of any of the methods for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration.

[0021] This invention provides an adaptive generation and closed-loop evolution method for fraudulent text based on semantic blind spot exploration. It constructs a blind spot identification mechanism based on semantic space analysis, proactively discovering potential fraud patterns not covered by existing fraud detection systems and implementing layered management of blind spots. Differentiated targeted generation strategies are designed for different levels of semantic blind spots to accurately supplement high-value samples. A multi-dimensional value assessment mechanism integrating semantic novelty, blind spot coverage, and classification challenge is established to screen samples with high information gain. A closed-loop adaptive evolution system of "semantic retrieval—blind spot identification—targeted generation—value assessment—incremental update" is constructed to achieve synergistic optimization of the fraud knowledge base, generation model, and classification model. Under low manual annotation costs, it efficiently discovers and confirms new fraud samples, improving the system's rapid adaptability to unknown fraudulent tactics. It continuously expands the semantic coverage and decision boundary of the fraud model, improving recall capability and long-term robustness against unknown fraud patterns. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a fraud-related text adaptive generation and closed-loop evolution system based on semantic blind spot exploration in a specific embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] like Figure 1 As shown, Figure 1 This is a flowchart of a method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration in a specific embodiment of the present invention. The method includes: Step 1: Construct a fraud-related semantic memory, a normal semantic memory, and a semantic retrieval index. This specifically includes: Step 11, Construct a fraud-related semantic memory: Encode confirmed historical fraudulent call texts into semantic vectors, and after deduplication, construct a fraud-related semantic memory V. db : V db ={v1,v2,…,v N}, where v i This is the semantic vector corresponding to the text of the fraudulent call.

[0027] This fraud-related semantic memory database only requires dozens to hundreds of samples for cold start.

[0028] Step 12, Constructing a Normal Semantic Memory: Encode confirmed historical normal call texts into semantic vectors, and after deduplication, construct a normal semantic memory N. db : N db ={n1,n2,…,n s}, where n i This is the semantic vector corresponding to the risk-free call text.

[0029] Step 13, Construct a semantic retrieval index: Adopt an Approximate Nearest Neighbor (ANN) index structure to support millisecond-level real-time retrieval.

[0030] Construct a semantic memory database for fraud and a normal semantic memory database, and combine them with near nearest neighbor retrieval to achieve continuous expansion and dynamic evolution of the fraud semantic space.

[0031] Step 2 involves semantically encoding the new call text and then performing a dual-space nearest neighbor search and density comparison based on both a fraud-related semantic memory and a normal semantic memory. This process filters candidate samples that are truly biased towards the fraud-related semantic space and may be located in semantic blind spots, improving the efficiency of subsequent generation and evolution. Specifically, this includes: Step 21: Semantically encode the new call text x from the speech-to-text transcription using the Encoder() function to obtain the semantically encoded vector v. x : v x =Encoder(x).

[0032] Step 22: Perform dual-space nearest neighbor retrieval. In both the fraud-related semantic memory and the normal semantic memory, use the Retrieve() function to perform nearest neighbor retrieval, obtaining the TopK historical fraud-related semantic neighborhoods most similar to the current call text x. fraud (x) and TopK semantic neighborhood of historical risk-free calls normal (x), which facilitates the subsequent calculation of the local semantic density of the bispace: TopK fraud (x)=Retrieve(v x V db ) TopK normal (x)=Retrieve(v x N db ).

[0033] Step 23: Calculate the local semantic density in the dual-space environment. Calculate the local semantic density ρ between the call text x and the historical fraud-related space using the cosine similarity function cos(). fraud (x), the local semantic density ρ of the call text x and the semantic space of historical risk-free calls. normal (x): ) ) Wherein: the default value of parameter K is 20. When the semantic encoder uses a 768-dimensional vector output, K=20 can provide enough statistical samples to stably estimate the local density, while ensuring that the selected neighbors still have high semantic similarity. When the sample size of the fraud-related semantic memory is less than 20, K is the total number of samples in the current memory capacity.

[0034] Local semantic density reflects whether there are a large number of semantic neighbors in the current call text that correspond to a known semantic space.

[0035] Step 24: Perform a preliminary screening by comparing the local semantic density of the two spaces.

[0036] When ρ normal (x)>ρ fraud When (x), it means that the local semantic density of the call text x in the normal semantic space is higher than the local semantic density in the fraud-related semantic space. That is, the call text x is more biased towards the normal semantic space and will not enter the subsequent process. When ρnormal (x)<=ρ fraud When (x), it means that the local semantic density of the call text x in the fraud-related semantic space is higher than that in the normal semantic space. That is, the call text x is more biased towards the fraud-related semantic space. The call text x is selected as a candidate sample, i.e., a new sample, and enters the subsequent semantic blind zone identification process.

[0037] Step 3 involves semantic blind spot identification and hierarchical classification based on local semantic density and global semantic offset. By jointly calculating the local semantic density of new samples in the fraud-related semantic space and the global semantic offset distance relative to the fraud-related semantic center, semantic blind spots that are not adequately covered by the existing system are identified and classified into three levels: neighborhood variant, cross-domain migration, and novel pattern, thereby enabling proactive discovery and hierarchical management of unknown fraud patterns.

[0038] This invention further combines local semantic density and global semantic offset to identify "semantic blind spots" and classify blind spot levels for samples that are more biased towards fraud-related semantic spaces.

[0039] The term "semantic blind spot" refers to new semantic regions of fraud that are not yet fully covered by current fraud detection systems. Compared to traditional methods based on fixed rules or single thresholds, this invention can proactively discover unknown fraud patterns at the semantic space structure level.

[0040] Semantic blind spot identification addresses the challenge of continuously evolving fraudulent tactics and the difficulty of existing models in promptly detecting unknown fraud patterns. Blind spot hierarchy division allows for the dynamic selection of corresponding targeted generation strategies, driving the large language model to generate novel fraudulent samples with high information gain. This enables continuous evolutionary learning of new fraudulent tactics, significantly improving the system's detection capability and long-term adaptability to unknown fraud patterns. Specifically, this includes: Step 31, semantic blind spot identification.

[0041] When ρ fraud When (x) < θ, the local semantic density of the sample is considered to be low, indicating that there are almost no historical fraud samples nearby, which is a low-coverage semantic region, also known as the semantic blind spot mentioned above. When ρ fraud When (x)>=θ, it is considered that the local semantic density of the sample is high, the historical fraud semantic space has been covered by a certain number of samples, it does not belong to the semantic blind zone, and there is no need to enter the subsequent blind zone level division process.

[0042] Step 32: Calculate the semantic center of fraud in existing historical fraud samples. :

[0043] Then calculate the global semantic offset distance D(x) of the sample currently determined as a semantic blind spot relative to the fraud-related semantic center: D(x)=1-cos(v x ,μ fraud ) Where: a larger D(x) indicates that the current sample deviates farther from the existing fraud semantic space.

[0044] Step 33: According to the global semantic offset, based on the first blind region division threshold β1 and the second blind region division threshold β2, the blind region is further divided into the following three categories: Neighborhood variant blind spot: when 0<D(x)<β1, it is considered that the global offset is low, indicating that the sample is located in the edge region of the known fraud semantics. Such samples usually belong to: expression variants of known fraud types; slightly rewritten speech; long-tail expressions of similar fraud.

[0045] Cross-domain migration blind spot: when β1<=D(x)<β2, it is considered that the global offset is medium, indicating that the sample has the characteristics of multiple types of fraud speech at the same time. Such samples usually belong to: fusion of multiple fraud modes; cross-scenario migration; new combined fraud.

[0046] New mode blind spot: when D(x)>=β2, it is considered that the global offset is high, indicating that the current sample is far away from the known fraud semantic space. Such samples usually correspond to: unknown fraud modes; new fraud gangs; new fraud strategies that have not been learned by the system.

[0047] The first blind region division threshold β1 and the second blind region division threshold β2 are not fixed constants set manually, but are automatically calculated based on the actual recognition performance of the fraud-related classifier at different semantic distances, and the core basis is the D(x)-confidence attenuation curve.

[0048] The construction method of the D(x)-confidence attenuation curve is as follows: the value range [0, 2] of the global offset distance D(x) is divided into M intervals with equal width (for example, M=40, and the width of each interval is 0.05). For the k-th interval [d k , d k+1 , count all historical fraud-related samples falling into this interval in the fraud-related semantic memory bank, and calculate the average predicted fraud probability of these samples by the fraud-related classifier: p local (d k )=Average{p fraud (v)|v in V db ,D(v)in[d k ,d k+1}.

[0049] In the formula, v represents the semantic vector corresponding to the fraud-related call text, and the abscissa of the curve is the median d of D(x) in each interval k (value range 0~2), the ordinate is the average fraud-related prediction probability p of the classifier for samples in this interval local (d k ) (value range 0~1). The overall curve shows a monotonically decreasing trend: the farther away from the fraud-related semantic center, the less confident the classifier is. Two key attenuation points are extracted from the curve as the blind area hierarchical division thresholds: β1 = the minimum d k such that p local (d k )drops below 0.85 for the first time; β2 = the minimum d k such that p local (d k )drops below 0.55 for the first time. The semantic meaning of β1 is: the maximum semantic distance boundary where the fraud-related classifier can still maintain high confidence (not less than 85%) for stable identification. For samples within this distance, their semantic deviation has not exceeded the natural variation range of known fraud patterns, and they belong to neighborhood variant blind areas. The semantic meaning of β2 is: the starting point of semantic distance where the recognition ability of the fraud-related classifier for fraud samples has attenuated to close to random level (not higher than 55%). For samples beyond this distance, the classifier can hardly distinguish them from normal texts, and they belong to new pattern blind areas. Samples between β1 and β2 belong to cross-domain migration blind areas. The reason why 0.85 and 0.55 are used instead of 0.90 and 0.50 is to reserve a moderate margin to avoid frequent threshold jumping caused by statistical fluctuation in a single interval, while maintaining the semantic distinguishability of the three types of blind areas. β1 is always less than β2, and the boundaries of the three blind areas are clear without overlap: neighborhood variant (0<D(x)<β1), cross-domain migration (β1<=D(x)<β2), new pattern (D(x)>=β2).

[0050] Cold start degradation strategy: In the initial cold start phase of the system, the sample size of the fraud-related semantic memory bank is insufficient (e.g., less than 200 samples), the number of samples in each D(x) interval is scarce, and the confidence attenuation curve is not smooth and reliable enough. At this time, the system adopts conservative default parameters: β1=0.10, β2=0.35. When the sample size of the semantic memory bank exceeds the preset threshold, the system automatically switches to the aforementioned confidence curve driving mode.

[0051] Step 4: Dynamic routing of directed generation strategies based on blind area hierarchy. For different levels of semantic blind areas, differentiated strategies such as local perturbation generation, multi-mode recombination generation and long-distance expansion generation are dynamically matched to achieve accurate sample supplement around semantic blind areas, and avoid semantic redundancy and inefficient amplification caused by random generation.

[0052] When a sample is identified as a semantic blind spot, the system triggers a targeted generation mechanism. This invention does not randomly generate fraudulent text, but rather dynamically selects the corresponding generation strategy based on different blind spot types, thereby avoiding the redundancy and inefficiency problems caused by traditional uniform generation methods. Specifically, it includes: Step 41, Neighborhood variant generation strategy.

[0053] To address the blind spots of neighborhood variant transformations: a local semantic perturbation generation strategy is adopted around the existing fraud semantic neighborhood to drive the large language model to generate fraud-related texts. Specific generation directions include: synonym replacement of key inducement words in neighboring phrases, transformation and reorganization of the word order of phrase sentences, variant design of fraud inducement methods, and rewriting of the overall expression style. This strategy, without changing the core logic of fraud, performs controllable mutations around known patterns, which can effectively expand the edge coverage of known fraud patterns and fill the blind spots of neighborhood variant transformations.

[0054] Step 42, generate cross-domain migration strategy.

[0055] To address the blind spots of cross-domain migration: a multi-modal recombination generation strategy is adopted to drive a large language model to generate fraud-related text by integrating the semantic features of multiple fraud patterns. The specific generation directions include: mixing and splicing different types of fraudulent statements, and adapting fraud patterns to new business scenarios. This strategy aims to explore the potential combination relationships between different fraud patterns and proactively discover composite fraudulent statements that may appear in the blind spots of cross-domain migration. Step 43, a brand new extended generation strategy.

[0056] To address novel pattern blind spots: a long-distance semantic extension generation strategy is adopted to drive the generation of fraud-related texts using a large language model. Specific generation directions include: constructing new inducement logic chains, designing new fraud scenario scripts, and arranging unknown fraud interaction processes. This strategy breaks away from the dependence on known fraud patterns and actively explores unknown areas in the fraud semantic space in an open generation manner to discover new fraud patterns that have not yet been identified by any existing system.

[0057] Step 5: Multidimensional value assessment and screening of generated samples. A value assessment function integrating boundary gain, coverage gain, and novelty gain is constructed to quantitatively evaluate the information gain of the generated samples, and only high-value samples are retained for subsequent update processes.

[0058] Unlike traditional methods that solely rely on "successful generation," this invention further introduces a generation value assessment mechanism. The system calculates the value gain of the generated samples:

[0059] Where: λ1 is the boundary gain The weighting coefficients, where λ2 is the coverage gain. The weighting coefficients, where λ3 is the novelty gain. The weighting coefficients are λ1+λ2+λ3=1.

[0060] Boundary gain This measures how challenging the generated samples are to the classification boundary. The generated samples are evaluated offline using the basic fraud classifier described above, with the output fraud probability denoted as p. fraud (g), then:

[0061] When p fraud When =0.5, =1 indicates that the classifier is the most uncertain, meaning that the sample is closest to the decision boundary; When p fraud →0 or when p fraud →1, →0 indicates that the sample poses a low challenge to the classifier.

[0062] Coverage gain Indicates whether the generated samples truly fill the blind spot:

[0063] in, Representative of the generated samples, This represents the vector obtained by semantically encoding the generated sample. If the generated sample is highly similar to an existing fraud-related semantic memory, then: V cover →0 indicates that it is a duplicate sample; If the generated sample differs significantly from the existing fraud-related semantic memory, then V cover →1 indicates that the coverage area has been truly expanded.

[0064] Novelty gain The novelty gain of the generated sample is calculated based on the distribution difference between the generated sample and historical generated samples:

[0065] Where M represents the total number of historically generated samples, Generate samples for history. The purpose is to "generate redundancy" and avoid LLM from generating the same pattern all the time.

[0066] The system performs high-value sample screening, classifying generated samples as high-value samples only if they meet the following criteria: high classification boundary challenge, high blind zone coverage, and low semantic redundancy. This mechanism avoids duplicate generation, invalid amplification, and accumulation of low-quality noise, thus ensuring that generated samples truly enhance the system's capabilities.

[0067] Through steps 2 to 5, for a new call text x from speech-to-text transcription, the system sequentially performs the following processing steps to form a complete single-sample processing chain from semantic mapping, blind spot recognition, targeted generation to value assessment.

[0068] Step 6: Closed-loop incremental evolution and dynamic adaptive closed-loop evolution mechanism.

[0069] A closed-loop feedback mechanism is constructed, which includes dynamic blind zone determination threshold adjustment, dynamic blind zone hierarchical division threshold update mechanism, generation strategy weight update, and sample priority ranking based on generation value, so as to achieve collaborative and continuous optimization of semantic memory, generation model and fraud classification model.

[0070] High-value candidate fraud samples selected through value assessment are manually confirmed and then enter the incremental update process, triggering a dynamic adaptive adjustment mechanism to achieve the collaborative evolution of the fraud semantic memory, the generation model, and the fraud classification model.

[0071] The dynamic adaptive mechanism of this invention mainly includes a dynamic blind zone determination threshold adjustment mechanism, a dynamic blind zone hierarchical division threshold update mechanism, a dynamic generation strategy weight adjustment mechanism, and a dynamic sample priority mechanism based on generation value. Specifically, the dynamic blind zone determination threshold is used to adjust the system's sensitivity to exploring unknown fraud patterns; the dynamic blind zone hierarchical division threshold update mechanism adaptively adjusts the blind zone hierarchical division thresholds β1 and β2 based on the offset of the fraud classifier's confidence decay curve, achieving continuous optimization of the division boundaries for the three blind zone types; the dynamic generation strategy weight is used to adjust the allocation of generation resources for different types of fraud patterns; and the dynamic sample priority mechanism is used to prioritize candidate samples with high information gain. These together constitute the system's closed-loop adaptive evolution mechanism. Specifically, it includes: Step 61: Manual confirmation and incremental update mechanism.

[0072] The manual verification module in this invention is not used for traditional large-scale manual annotation, but rather as a high-value semantic verification mechanism within a closed-loop system. The system automatically filters candidate samples with high information gain through fraud-related semantic blind spot detection, blind spot stratification, targeted generation, and generation value assessment, triggering manual verification only for a small number of high-value boundary samples. Manual verification employs a low-cost binary verification method, only determining whether a sample possesses genuine fraud-related semantics, thereby preventing false fraud semantics generated by the generative model from polluting the fraud-related semantic memory.

[0073] (1a) Confirmation of fraud-related samples.

[0074] Valid fraud samples, verified by manual review, will be encoded and stored in the database, used to update the fraud semantic space, participate in subsequent vector retrieval and blind spot detection, and be used for incremental fine-tuning of the basic fraud classifier, allowing its decision boundary to evolve as new fraud semantics expand. The system uses incremental learning to gradually adapt the basic fraud classifier to new fraud patterns.

[0075] (2a) Confirmation of normal samples.

[0076] Samples that are manually confirmed to be normal text are updated to the normal semantic database to further refine the normal semantic boundaries and reduce the system's false alarm rate.

[0077] Step 62, Dynamic blind zone determination threshold adjustment mechanism.

[0078] This invention is based on the local semantic density ρ related to fraud. fraud (x) identifies low-coverage areas in the fraud semantic space.

[0079] in: ) ρ fraud (x) represents the local semantic density related to fraud in the current sample. When ρ fraud When (x) < θ, where θ is the blind zone determination threshold, the system considers the current sample to be located in a low-coverage area of ​​fraud semantics and enters the subsequent blind zone stratification and targeted generation process.

[0080] (1b) Initial threshold setting.

[0081] During system initialization, a basic blind zone threshold θ0=0.2 is set based on business experience and data analysis. This means that when the local semantic density of a sample involving fraud is lower than 0.2, it is considered that there is a lack of historical fraud samples covering its vicinity.

[0082] (2b) Statistics on the growth rate of blind spots.

[0083] The system continuously counts the number of newly added blind zone samples within the most recent time window. Let N be the number of blind zone samples in the current time window. blind (t), where N is the number of samples in the blind zone of the previous time window. blind (t-1), then define the blind zone growth rate:

[0084] in To prevent smooth terms with a denominator of zero.

[0085] (3b) Dynamic threshold update.

[0086] The system dynamically adjusts the blind zone threshold based on the blind zone growth rate.

[0087] Where α is the adaptive adjustment coefficient.

[0088] (4b) Dynamic adjustment logic.

[0089] When the system detects a continuous increase in the blind spot growth rate, it indicates the recent emergence of a large number of new rhetoric techniques not covered by the historical fraud semantic space. At this point, R... growth As R(t) increases, θ(t) increases accordingly. The system automatically improves the identification range of low-coverage areas, allowing more edge samples to enter the blind zone exploration process, thereby enhancing the system's sensitivity to new fraud patterns. Conversely, when fraud patterns tend to stabilize, R... growth As θ(t) decreases, the system gradually restores the default exploration intensity, thereby reducing the computational and manual confirmation overhead caused by invalid exploration.

[0090] Step 63, dynamic blind zone hierarchy division threshold update mechanism.

[0091] In addition to the dynamic adjustment of the blind zone determination threshold θ, the blind zone hierarchy classification thresholds β1 and β2 are also adaptively updated during the closed-loop evolution of the system. β1 and β2, together with θ, constitute the system's dual threshold system: θ controls the sensitivity of "whether to enter the blind zone for identification," while β1 and β2 control the classification boundaries of "how to classify after entering the blind zone." This includes: Update trigger condition: When the semantic memory undergoes a significant increase—that is, when the number of new samples reaches 10% of the current memory capacity—the system triggers the update process for β1 and β2. This trigger condition is aligned with the time window for the blind zone sample growth rate statistics, ensuring that the system adjusts the hierarchical division boundaries in a timely manner when perceptible changes occur in the semantic space.

[0092] Update execution process: The system reconstructs the D(x)-confidence decay curve, and calculates p for each interval based on the predictive ability of the updated fraud classifier. local (d k ); β1 and β2 (i.e., p) are re-extracted from the curve using the method described above. local (d k (The first time the value falls below 0.85 is designated as β1, and the first time it falls below 0.55 is designated as β2). If the current database capacity is still below the cold start threshold (e.g., 200 entries), the system maintains the default parameters β1=0.10 and β2=0.35.

[0093] Dynamic adjustment logic: As the semantic memory bank continues to expand and the fraud classifier is incrementally trained, the classifier's ability to identify known fraud patterns gradually increases, its confidence decay curve shifts to the right overall, and β1 automatically increases—this means that more samples are classified into "neighborhood variants," and the system concentrates resources to explore more distant cross-domain migration and entirely new pattern blind spots. Simultaneously, β2 is also dynamically updated accordingly, ensuring that the judgment boundary for "completely new patterns" always aligns with the actual capability boundary of the classifier. This mechanism allows β1 and β2 to naturally relax their tolerance for "known variants" as the system's cognitive ability improves—the more "experienced" the system, the more accurate its blind spot hierarchy classification.

[0094] Step 64: Dynamically generate a strategy weight adjustment mechanism.

[0095] This invention sets different targeted generation strategies for different blind zone types. Let S1 be the neighborhood variant generation strategy, S2 be the cross-domain migration generation strategy, and S3 be the novel extension generation strategy, with corresponding weights: W(t) = {w1(t), w2(t), w3(t)} in: w1 + w2 + w3 = 1 The system dynamically adjusts the corresponding weights based on the effective value of recently generated samples from different generation strategies.

[0096] (1c) Initial policy weights.

[0097] System initialization settings:

[0098] This indicates that the three generation directions have the same initial priority.

[0099] (2c) Strategy effectiveness statistics.

[0100] The system continuously calculates the average generation value of each generation strategy within the most recent time window.

[0101] Let G be the set of samples generated by the i-th generation strategy. i Then its average generated value is:

[0102] Where V(g) is the generated sample value function:

[0103] (3c) Dynamic weight update.

[0104] The system uses a normalized exponential update method to dynamically adjust the generation strategy weights:

[0105] Where η is the weight update rate coefficient.

[0106] (4c) Dynamic adjustment logic.

[0107] If a certain generation strategy consistently produces high boundary gain, high blind zone coverage gain, and high semantic novelty gain, then the corresponding Growth, and thus w i As the number of fraud patterns increases, the system automatically increases the trigger probability and resource allocation ratio of the generation strategy in that direction. This allows the system to dynamically adjust the semantic exploration direction based on the evolution trend of fraud patterns.

[0108] Step 65, a dynamic sample priority mechanism based on generated value.

[0109] The system sorts the candidate generated samples according to the comprehensive value function V(g) of the generated samples, and prioritizes samples with high boundary gain, high blind zone coverage gain, and high semantic novelty gain to enter the manual confirmation and incremental update process.

[0110] Among them, high boundary gain samples help optimize the decision boundary of the classifier; high blind zone coverage gain samples help expand the semantic coverage of fraud; and high semantic novelty samples can effectively avoid the repeated generation and invalid amplification of samples.

[0111] Step 66: Closed-loop collaborative evolution mechanism.

[0112] The system continuously records: the growth rate of low-coverage areas, the sample value gain corresponding to different generation strategies, and the improvement effect of different types of samples on detection performance. It also dynamically adjusts the blind zone determination threshold, generation strategy weights, and sample confirmation priorities, thus forming a closed-loop adaptive mechanism for the coordinated evolution of the "fraud-related semantic memory—generation model—fraud-related classification model." This mechanism enables the system to continuously expand the fraud semantic space as fraud patterns change and gradually improve its ability to identify unknown fraudulent statements. Simultaneously, after a significant increase in the semantic memory (new samples reaching 10% of the current capacity), the system automatically recalculates the blind zone hierarchy thresholds β1 and β2 based on the updated D(x)-confidence decay curve, keeping them synchronized with the continuous expansion of the semantic space and the dynamic evolution of the classifier's capabilities.

[0113] In summary, this invention constructs a closed-loop adaptive mechanism of "semantic retrieval - blind spot identification - targeted generation - value assessment - incremental evolution", which is carried out in three layers: offline basic construction (preparation before system cold start), single sample online processing and blind spot targeted generation (complete processing link for each new input text), and closed-loop incremental evolution and dynamic adaptive adjustment (continuous self-optimization of the system in the time dimension).

[0114] Accordingly, embodiments of the present invention also provide an adaptive generation and closed-loop evolution system for fraudulent text based on semantic blind spot exploration, such as... Figure 2 The diagram shown is a structural schematic of the system. This system, which adaptively generates and evolves fraudulent text based on semantic blind spot exploration, includes the following modules: Semantic memory construction unit 201 constructs a fraud-related semantic memory, a normal semantic memory, and a semantic retrieval index; The candidate sample screening unit 202 performs semantic encoding on the new call text, performs dual spatial nearest neighbor retrieval and local semantic density comparison, and screens out candidate samples that may be located in the semantic blind zone. The semantic blind spot identification and layering unit 203 performs semantic blind spot identification and layering on candidate samples based on local semantic density and global semantic offset; Fraudulent text generation unit 204 dynamically selects fraudulent text generation strategies to generate fraudulent text for different levels of semantic blind spots; Generate value assessment unit 205, construct value assessment function, quantitatively assess the information gain of the generated fraudulent text, and screen out high-value samples. The closed-loop feedback unit 206 performs manual verification and incremental updates on the selected high-value samples to construct a dynamic and adaptive closed-loop feedback mechanism.

[0115] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0116] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it is run. Figure 1 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0117] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.

[0118] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and systems of the present invention, and are merely some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive generation and closed-loop evolution of fraud-related text based on semantic blind spot exploration, characterized in that, Said method comprises: Step 1, constructing a fraud-related semantic memory bank, a normal semantic memory bank, and a semantic retrieval index; Step 2, after performing semantic encoding on a new call text, performing dual-space nearest neighbor retrieval and local semantic density comparison, and screening candidate samples that may be located in semantic blind spots; Step 3, performing semantic blind spot identification and layering on the candidate samples based on local semantic density and global semantic offset; Step 4, dynamically selecting a fraud-related text generation strategy according to semantic blind spots of different levels to generate fraud-related texts; Step 5, constructing a value evaluation function to quantitatively evaluate the information gain of the generated fraud-related texts, and screening high-value samples; Step 6, performing manual confirmation and incremental update on the screened high-value samples, and constructing a dynamically adaptive closed-loop feedback mechanism.

2. The method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration according to claim 1, characterized in that, Step 1 comprises: Step 11: Encode the confirmed historical fraudulent call texts into semantic vectors, and after deduplication, establish the fraudulent semantic memory V. db ; Step 12: Encode the confirmed historical normal call text into semantic vectors, and after deduplication, establish the normal semantic memory N. db ; Step 13, adopting an approximate nearest neighbor index structure to construct a semantic retrieval index.

3. The method for adaptive generation and closed-loop evolution of fraud-related text based on semantic blind spot exploration according to claim 1, characterized in that, In step 2, after performing semantic encoding on the new call text, performing dual-space nearest neighbor retrieval and local semantic density comparison based on the fraud-related semantic memory bank and the normal semantic memory bank, and screening candidate samples that are biased towards the fraud-related semantic space and may be located in semantic blind spots, which specifically comprises: Step 21: Semantically encode the new call text x from the speech-to-text transcription to obtain a semantically encoded vector v. x ; Step 22: Perform dual-space nearest neighbor retrieval, searching for nearest neighbors in both the fraud-related semantic memory and the normal semantic memory to obtain the TopK historical fraud-related semantic neighborhoods that are closest to the call text x. fraud (x) and TopK semantic neighborhood of historical risk-free calls normal (x); Step 23: Perform dual-space local semantic density calculation, and calculate the local semantic density ρ of the call text x in the historical fraud space. fraud (x) and the local semantic density ρ of the call text x in the historical risk-free call semantic space. normal (x); Step 24, performing pre-screening for dual-space local semantic density comparison; When ρ normal (x)>ρ fraud When (x), it means that the local semantic density of the call text x in the normal semantic space is higher than the local semantic density in the fraud-related semantic space. That is, the call text x is more biased towards the normal semantic space and will not enter the subsequent process. When ρ normal (x)<=ρ fraud When (x) indicates that the local semantic density of the call text x in the fraud-related semantic space is higher than that in the normal semantic space, the call text x is more biased towards the fraud-related semantic space. The call text x is then selected as a candidate sample, and the process proceeds to step 3.

4. The method for adaptive generation and closed-loop evolution of fraud-related text based on semantic blind spot exploration according to claim 3, characterized in that, In step 3, by jointly calculating the local semantic density of the candidate sample in the fraud-related semantic space and the global semantic offset distance relative to the fraud-related semantic center, identifying semantic blind spots with insufficient coverage by the existing system, and dividing the semantic blind spots into three levels: neighborhood variant blind spots, cross-domain migration blind spots and brand-new pattern blind spots.

5. The method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration according to claim 4, characterized in that, Step 3 comprises: Step 31, based on the local semantic density ρ of the candidate samples and the historical fraud space. fraud (x) performs semantic blind zone identification, when ρ fraud When (x) < θ, the local semantic density of the candidate sample is considered to be low, indicating that there are almost no historical fraud samples nearby. It belongs to a low-coverage semantic region, i.e., a semantic blind zone. Here, θ is the blind zone determination threshold, and the process proceeds to step 32. When ρ fraud When (x)>=θ, it is considered that the local semantic density of the candidate sample is high, the historical fraud semantic space has been covered by a certain number of samples, and it does not belong to the semantic blind zone, and the process ends. Step 32: Calculate the semantic center of fraud in existing historical fraud samples. : Among them, v i is the semantic vector corresponding to the fraudulent call text in the fraudulent semantic memory; N is the total number of call texts in the existing fraudulent semantic memory. Calculating a global semantic offset distance D(x) of a current candidate sample determined as a semantic blind spot relative to a fraud-related semantic center: D(x)=1-cos(v x ,m fraud ) wherein cos() is a cosine similarity function, and a larger D(x) indicates that the candidate sample deviates farther from the existing fraud semantic space; Step 33, dividing the semantic blind spots into three levels, namely the neighborhood variant blind spots, the cross-domain migration blind spots and the brand-new pattern blind spots according to the global semantic offset distance based on a first blind spot division threshold β1 and a second blind spot division threshold β2: When 0<D(x)<β1, the global semantic offset distance is considered to be low, which indicates that the candidate sample is located in an edge area of known fraud semantics, and the semantic blind spot is divided into the neighborhood variant blind spot; When β1<=D(x)<β2, the global semantic offset distance is considered to be medium, which indicates that the candidate sample has multi-type fraud speech characteristics at the same time, and the semantic blind spot is divided into the cross-domain migration blind spot; When D(x)>=β2, the global semantic offset distance is considered to be high, which indicates that the candidate sample is far away from the known fraud semantic space, and the semantic blind spot is divided into the brand-new pattern blind spot; wherein the first blind spot division threshold β1 is smaller than the second blind spot division threshold β2.

6. The method for adaptive generation and closed-loop evolution of fraud-related text based on semantic blind spot exploration according to claim 4, characterized in that, In step 4, for the neighborhood variant blind spot, a local semantic perturbation generation strategy is used to drive the large language model to generate fraud-related text around the existing fraud semantic neighborhood. The specific generation directions include: replacing key inducement words in the neighboring words with synonyms, changing and reorganizing the word order of the words, designing variants of fraud inducement methods, and rewriting the overall expression style. To address the aforementioned cross-domain migration blind spots, a multi-modal recombination generation strategy is adopted to drive a large language model to generate fraud-related text by integrating the semantic features of multiple fraud patterns. Specific generation directions include: mixing and splicing different types of fraudulent rhetoric, and adapting fraud patterns to new business scenarios. To address the aforementioned new pattern blind spots, a long-distance semantic extension generation strategy is adopted to drive the large language model to generate fraud-related texts. Specific generation directions include: constructing new inducement logic chains, designing new fraud scenario scripts, and arranging unknown fraud interaction processes.

7. The method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration according to claim 1, characterized in that, In step 5, the value evaluation function for the constructed fusion boundary gain, coverage gain, and novelty gain is: Where λ1 is the boundary gain The weighting coefficients, where λ2 is the coverage gain. The weighting coefficients, where λ3 is the novelty gain. The weighting coefficients, λ1+λ2+λ3=1; boundary gain Measuring the degree to which the generated fraudulent text challenges the classification boundaries; coverage gain. Indicates whether the generated fraudulent text truly fills the gap; novelty gain. Indicates the novelty gain of the generated fraudulent text; Fraudulent texts that meet the criteria of high classification boundary challenges, high blind zone coverage, and low semantic redundancy will be selected as high-value samples.

8. The method for adaptive generation and closed-loop evolution of fraud-related text based on semantic blind spot exploration according to claim 5, characterized in that, In step 6, the selected high-value samples, after manual confirmation, enter the incremental update process, triggering a dynamic adaptive closed-loop feedback mechanism. This dynamic adaptive closed-loop feedback mechanism includes: The system includes a dynamic blind zone determination threshold adjustment mechanism, a dynamic blind zone hierarchy division threshold update mechanism, a dynamic generation strategy weight adjustment mechanism, and a dynamic sample priority mechanism based on generation value. The dynamic blind zone determination threshold adjustment mechanism is used to adjust the sensitivity of exploring unknown fraud patterns. The dynamic blind zone hierarchy division threshold update mechanism adaptively adjusts the first blind zone division threshold β1 and the second blind zone division threshold β2 based on the offset of the fraud classifier's confidence decay curve, achieving continuous optimization of the division boundaries for the three blind zone types. The dynamic generation strategy weight adjustment mechanism is used to adjust the allocation of generation resources for different types of fraud patterns. The dynamic sample priority mechanism is used to prioritize candidate samples with high information gain.

9. A system for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration, characterized in that, The system includes: The semantic memory building unit constructs a fraud-related semantic memory, a normal semantic memory, and a semantic retrieval index. The candidate sample screening unit performs semantic encoding on the new call text, performs dual spatial nearest neighbor retrieval and local semantic density comparison, and screens out candidate samples that may be located in the semantic blind zone. The semantic blind spot identification and hierarchical unit performs semantic blind spot identification and hierarchical analysis on candidate samples based on local semantic density and global semantic offset. The fraud-related text generation unit dynamically selects fraud-related text generation strategies to generate fraud-related texts, targeting different levels of semantic blind spots. Generate value assessment units, construct value assessment functions, quantify the information gain of the generated fraudulent texts, and select high-value samples. The closed-loop feedback unit manually confirms and incrementally updates the selected high-value samples, thus constructing a dynamic and adaptive closed-loop feedback mechanism.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the method for adaptive generation and closed-loop evolution of fraudulent text based on semantic blind spot exploration as described in any one of claims 1 to 8.

Images