A method and system for GEO anti-fraud based on multi-source contact method conflict modeling

Abstract

This invention discloses a GEO anti-fraud method and system based on multi-source contact conflict modeling, primarily relating to the field of generative artificial intelligence technology. It includes: extracting structured triples from multimodal content associated with results returned by a generative engine; constructing a cross-modal heterogeneous conflict graph using each triple as a node; assigning an initial confidence level to each node in the conflict graph based on its source authority identifier, and performing conflict-aware confidence propagation on the conflict graph, iteratively updating the confidence level of each node; and determining fraud based on the confidence levels of contact methods under the same brand obtained after propagation; and dynamically intervening in the retrieval enhancement generation process of the generative large model based on the fraud determination results to reduce the appearance of contact methods judged as fraudulent in the final generated answers. The beneficial effect of this invention is that it can achieve high-precision fraud identification by dynamically suppressing the propagation paths of mutually exclusive false information.

Description

Technical Field

[0001] This invention relates to the field of generative artificial intelligence technology, specifically a GEO anti-fraud method and system based on multi-source connection conflict modeling. Background Technology

[0002] With the widespread application of generative artificial intelligence (AIGC) technology, generative engine optimization (GEO) has emerged as a new method for information acquisition and optimization. Unlike traditional search engine optimization (SEO), GEO aims to directly influence the output of generative engines such as large language models by optimizing input content, enabling them to generate specific and standardized answers. Against this backdrop, there has emerged the practice of exploiting the characteristics of GEO technology to simultaneously distribute content containing false contact information (such as fake customer service phone numbers and fake official links) across multiple platforms. This content often mimics official language, possesses high semantic standardization, and easily induces generative engines to adopt it as a direct answer without requiring redirection for verification, posing information security and financial risks to users.

[0003] Currently, existing technologies primarily defend against misinformation in generative scenarios in two ways: first, by adopting and improving traditional static defense mechanisms based on blacklists or rule-based filtering; and second, by attempting to introduce multimodal content analysis technology, which assesses the credibility of content by integrating features from multiple modalities such as text and images. However, when faced with fraudulent activities unique to GEO scenarios and characterized by "multi-source conflicts," these existing solutions all reveal significant limitations, mainly in the following aspects:

[0004] 1. A biased understanding of the technology of "conflict" leads to the lack of core discrimination signals: Existing technologies, especially multimodal fusion analysis methods, usually regard the inconsistency of contact information from different sources as noise or interference factors affecting the generalization performance of the model, and strive to smooth or eliminate it at the feature layer. However, in the GEO fraud scenario, malicious actors deliberately create the illusion of "multi-source corroboration" by systematically feeding multiple highly similar but false contact information on multiple platforms. This "multi-source contact information conflict" is the core behavioral feature that distinguishes it from normal information promotion. Existing technologies ignore or even eliminate this feature, which is equivalent to actively giving up the most critical fraud discrimination signal. This is the root cause of its poor detection performance.

[0005] 2. The detection methods are one-sided and cannot adapt to the coordination and temporal nature of GEO attacks: Existing solutions either only analyze single-modal content, ignoring the global nature of cross-platform collaborative feeding, or although they perform multimodal fusion, they lack specific modeling for the temporal clustering of fraudulent behavior (such as dense release in a short period of time) and the contradictions between modalities (such as the same number being used with inconsistent authoritative packaging on different platforms). Their models are often static or generalized and cannot accurately capture the dynamic and organized attack patterns of GEO fraud, resulting in a serious lack of recognition accuracy in complex conflict scenarios.

[0006] 3. Rigid strategies make it difficult to achieve an effective balance between precision and recall: To avoid the risk of misjudging normal promotions as fraud, existing technologies often adopt overly conservative strategies, such as setting extremely high judgment thresholds. While this can maintain a certain level of precision, it leads to an extremely high false negative rate, making it impossible to detect a large number of actual fraudulent behaviors. This significantly reduces the overall protection effectiveness of the system. This reflects that the existing technology framework lacks flexible and adaptive risk measurement and intervention mechanisms when facing new and cunning GEO fraud methods.

[0007] Therefore, there is an urgent need for a GEO anti-fraud method and system based on multi-source connection conflict modeling to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide a GEO anti-fraud method and system based on multi-source connection conflict modeling, which can achieve high-precision fraud identification by dynamically suppressing the propagation paths of mutually exclusive false information.

[0009] To achieve the above objectives, the present invention employs the following technical solution:

[0010] On the one hand, this invention provides a GEO anti-fraud method based on multi-source connection conflict modeling, comprising the following steps:

[0011] Step S1: Extract structured triples from the multimodal content associated with the results returned by the generative engine. The triples include the target brand, contact information, and source authority identifier.

[0012] Step S2: Construct a cross-modal heterogeneous conflict graph using each triplet as a node; the construction of the conflict graph includes:

[0013] Establish intra-brand conflict reinforcement edges between nodes belonging to the same brand but with different contact methods, and dynamically calculate the weight of the edge based on the conflict decay function;

[0014] The conflict attenuation function calculates the weights based on at least the similarity between the two connection methods and a dynamically adjustable conflict sensitivity coefficient.

[0015] Step S3: Assign an initial confidence level to each node in the conflict graph based on its source authority identifier, and perform conflict-aware confidence propagation on the conflict graph to iteratively update the confidence level of each node;

[0016] Fraud is determined based on the confidence level of each contact method under the same brand obtained after dissemination; when the difference between the contact method with the highest confidence level and the contact method with the second highest confidence level under the brand is less than a preset threshold, GEO fraud is determined to exist.

[0017] Step S4: Based on the fraud determination results in Step S3, dynamically intervene in the retrieval enhancement generation process of the generative large model to reduce the appearance of contact information judged as fraudulent in the final generated answer.

[0018] Preferably, the process of extracting structured triples in step S1 is completed using a small sample prompt template customized for the GEO anti-fraud task;

[0019] The sample of the prompt template contains at least one keyword from "only official", "24-hour customer service", and "solemn declaration", and includes known fraud negative samples to guide the extraction model to accurately identify entities with fraud risk.

[0020] Preferably, the establishment of the intra-brand conflict reinforcement edge in step S2 includes:

[0021] Detect whether several similar contact information groups appear within a preset time window under the same brand;

[0022] If the detection is "yes", it is determined that there is a "concentrated occurrence" of GEO fraud characteristics, and the conflict sensitivity coefficient is dynamically increased.

[0023] Preferably, the preset time window is 24 hours; trigger conflict sensitivity coefficient The conditions for dynamic upgrade are: within 24 hours, three or more unique contact methods are detected under the same brand;

[0024] The conflict sensitivity coefficient Based on real-time calculation of similar contact density In a default value With a maximum value A linear adjustment is made between them, and the adjustment formula is:

[0025] when hour: ;

[0026] when hour: ;

[0027] when hour ;

[0028] in, Divide the number of contact methods by 24 hours. This is the normal density threshold. This is the fraud density threshold.

[0029] Preferably, the conflict attenuation function is specifically:

[0030] ;

[0031] in, The weight of the edge. Edit the Levenshtein distance between the two phone numbers. This is the dynamically adjusted conflict sensitivity coefficient.

[0032] Preferably, the preset threshold in step S3 is set to 0.3; when the difference between the highest confidence level and the second highest confidence level is less than 0.3, it is determined that there are multiple high-confidence but mutually exclusive fake numbers, i.e., GEO fraud.

[0033] Preferably, in step S4, the dynamic intervention in the retrieval enhancement generation process of the generative large model specifically includes:

[0034] Step S41: Store all contact information identified as fraudulent in Step S3 into a confidence mask dictionary, and associate each contact information in the dictionary with a decay factor. ,in ;

[0035] Step S42: After the retrieval output stage of the retrieval enhancement generation process, insert a confidence mask intervention layer;

[0036] Step S43: The confidence masking intervention layer scans each text segment returned by the retrieval machine. If the segment contains contact information from the masking dictionary, the original similarity score of the segment is multiplied by the attenuation factor. The new score after intervention;

[0037] Step S44: Input the list of text fragments with adjusted scores into the generator of the large model to generate the final answer.

[0038] On the other hand, the present invention also provides a GEO anti-fraud system based on multi-source connection conflict modeling, for implementing the GEO anti-fraud method based on multi-source connection conflict modeling as described above, including:

[0039] The multimodal entity extraction module is used to extract structured triples containing target brand, contact information, and source authority identifiers from multimodal content;

[0040] The conflict graph construction module is used to construct a cross-modal heterogeneous conflict graph with the triples as nodes, and to establish intra-brand conflict strengthening edges for nodes that belong to the same brand but have different contact methods, and to dynamically calculate edge weights based on the conflict decay function.

[0041] The conflict-aware reasoning module is used to assign an initial confidence level to a node, perform confidence propagation on the conflict graph and iteratively update the node confidence level, and perform GEO fraud determination based on the confidence level difference after propagation of each contact method under the same brand.

[0042] The generation intervention module is used to dynamically intervene in the retrieval enhancement generation process of the generative large model based on the fraud determination results, thereby reducing the output of fraudulent contact information.

[0043] Preferably, the conflict graph construction module includes:

[0044] The time-sensitive analysis unit is used to monitor the density of contact information appearing within a preset time window for the same brand.

[0045] The dynamic parameter adjustment unit is used to dynamically adjust the conflict sensitivity coefficient in the conflict attenuation function according to the contact density.

[0046] Preferably, the intervention generation module includes:

[0047] The mask dictionary management unit is used to store contact information that is identified as fraudulent and its associated decay factor;

[0048] The intervention layer execution unit is embedded in the retrieval enhancement generation process to scan the retrieved text fragments and recalculate the similarity scores;

[0049] The intervention layer execution unit uses a regular expression engine that integrates multiple phone number patterns to quickly match and scan text fragments.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. In response to the fundamental flaw in existing technologies that mistakenly treat "multi-source contact conflict" as noise, leading to performance degradation, this invention proposes a completely opposite technical approach: establishing this conflict as the core discrimination signal and designing a special conflict graph model for forward modeling and mining. This shift in thinking enables the system to directly capture the essential characteristics of GEO fraud behavior, achieving a disruptive innovation in detection principles, thereby achieving a recognition performance improvement that traditional methods cannot match.

[0052] 2. By constructing a heterogeneous conflict graph that integrates intra-brand conflict, time window sensitivity, and cross-modal consistency verification, this invention can accurately characterize the "cooperative feeding" and "temporal aggregation" characteristics of GEO fraud. The system's detection mechanism is no longer one-sided or rigid, but can actively discover and amplify the conflict patterns left by malicious behavior, thereby significantly improving the ability to detect various variant frauds, effectively reducing the risk of missed detections, and achieving broader monitoring coverage of fraudulent activities.

[0053] 3. This invention achieves a refined and dynamic assessment of information credibility through a conflict-aware confidence propagation algorithm and an adaptive threshold judgment logic. The system can effectively distinguish between maliciously created "high-confidence conflicts" and reasonable differences in normal commercial promotion, avoiding misjudgments caused by oversensitivity. Therefore, this invention can maintain a high level of judgment accuracy while greatly expanding the detection range, overcoming the problem of traditional solutions that sacrifice precision and recall.

[0054] 4. This invention differs from traditional post-event filtering or interception by dynamically intervening in the Retrieval Enhancement Generation (RAG) process through a confidence mask of the fraud determination result. This allows for suppression and guidance at the source of generative answer formation. This integrated "detection-intervention" design achieves closed-loop governance of GEO fraud, which is more in line with the security protection needs of the generative AI era and provides direct and effective technical support for ensuring user experience and information security. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0057] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0058] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0059] Example:

[0060] like Figure 1 As shown, this embodiment provides a GEO anti-fraud method based on multi-source contact conflict modeling, including the following steps:

[0061] Step S1: Extract structured triples from the multimodal content associated with the results returned by the generative engine. The triples include the target brand, contact information, and source authority identifier.

[0062] Step S2: Construct a cross-modal heterogeneous conflict graph using each triplet as a node; the construction of the conflict graph includes:

[0063] Establish intra-brand conflict reinforcement edges between nodes belonging to the same brand but with different contact methods, and dynamically calculate the weight of the edge based on the conflict decay function;

[0064] The conflict attenuation function calculates the weights based on at least the similarity between the two connection methods and a dynamically adjustable conflict sensitivity coefficient.

[0065] Step S3: Assign an initial confidence level to each node in the conflict graph based on its source authority identifier, and perform conflict-aware confidence propagation on the conflict graph to iteratively update the confidence level of each node;

[0066] Fraud is determined based on the confidence level of each contact method under the same brand obtained after dissemination; when the difference between the contact method with the highest confidence level and the contact method with the second highest confidence level under the brand is less than a preset threshold, GEO fraud is determined to exist.

[0067] Step S4: Based on the fraud determination results in Step S3, dynamically intervene in the retrieval enhancement generation process of the generative large model to reduce the appearance of contact information judged as fraudulent in the final generated answer.

[0068] like Figure 2 As shown, this embodiment also provides a GEO anti-fraud system based on multi-source contact conflict modeling, including:

[0069] The multimodal entity extraction module is used to extract structured triples containing target brand, contact information, and source authority identifiers from multimodal content;

[0070] The conflict graph construction module is used to construct a cross-modal heterogeneous conflict graph with the triples as nodes, and to establish intra-brand conflict strengthening edges for nodes that belong to the same brand but have different contact methods, and to dynamically calculate edge weights based on the conflict decay function.

[0071] The conflict-aware reasoning module is used to assign an initial confidence level to a node, perform confidence propagation on the conflict graph and iteratively update the node confidence level, and perform GEO fraud determination based on the confidence level difference after propagation of each contact method under the same brand.

[0072] The generation intervention module is used to dynamically intervene in the retrieval enhancement generation process of the generative large model based on the fraud determination results, thereby reducing the output of fraudulent contact information.

[0073] The implementation of the above method and system will be described in detail below with reference to a specific scenario:

[0074] A well-known electronics manufacturer, "Brand A," has an official customer service number of 400-123-4567, which has been verified through its official website and official stores on major e-commerce platforms. However, attackers exploited Generative Engine Optimization (GEO) technology to trick the engine into returning a fake number they controlled when users searched for "Brand A's official customer service number." This represents a typical GEO attack strategy.

[0075] 1. Multiple counterfeit numbers and coordinated distribution: Within 24 hours, dozens of posts containing highly similar official scripts were published almost simultaneously on multiple social media platforms, short video platforms, and Q&A communities. The fake numbers embedded in these posts were not single numbers, but a group of highly similar numbers, such as: 400-123-4568, 400-123-4569, and 400-123-4560.

[0076] 2. Conflicting cross-modal packaging: The text post claims "Brand A's only 24-hour after-sales hotline: 400-123-4568"; the accompanying image uses a counterfeit official poster style to indicate "Solemn declaration: National service hotline 400-123-4569"; the audio and subtitles of the short video indicate "The latest official customer service has changed to 400-123-4560";

[0077] 3. Avoiding traditional verification: None of this content contains links that redirect to the official website of "Brand A", so users cannot verify it instantly by clicking on them.

[0078] In response to the aforementioned GEO attack strategy employed by attackers, the specific execution of the method and system in this embodiment is as follows:

[0079] Step S1: Multimodal structured entity extraction:

[0080] After receiving a user query, the generative search engine's internal Retrieval Enhancement Generation (RAG) process first retrieves multimodal content fragments related to "Brand A" and "customer service phone number" from the associated index. These fragments may come from news websites, forum posts, social media images and text, video captions, etc. Before sending these fragments into the large model to generate the answer, they are first sent to the system of this invention for processing.

[0081] First, the multimodal entity extraction module is activated. The core of this module is a large language model (LLM) or information extraction model guided by specially customized few-shot prompting. The design of this customized prompting template is one of the key innovations. It is not a general named entity recognition template, but is highly specialized for GEO fraud scenarios.

[0082] The template will contain several carefully crafted example pairs, such as:

[0083] Example 1 (Positive Sample - Fraud Features):

[0084] Enter the text: "'Solemn Declaration' Brand X's only 24-hour official after-sales hotline is now in use: 400-555-0001. The old number has been discontinued."

[0085] Output triple: (Brand X, 4005550001, Authority equals 0.2), where the authority of 0.2 stems from the fact that the content source is an unverified personal self-media account;

[0086] Example 2 (Negative Samples - Normal Promotion):

[0087] Enter the text: "Brand Y's official flagship store on JD.com has a customer service hotline of 400-666-8888, and the store has a 'Brand Direct Sales' certification mark."

[0088] Output triple: (Brand Y, 4006668888, Authority equals 0.9), where the authority of 0.9 comes from mentioning an official store with platform credit certification;

[0089] Example 3 (Positive Sample - Keyword Guidance):

[0090] Enter the text: "Attention! Brand Z's nationwide repair center phone number is 400-777-9999. Do not trust any other numbers."

[0091] Output triple: (Brand Z, 4007779999, Authority equals 0.3), the model learns fraud risk from misleading keywords such as "Caution!" and "Do not believe";

[0092] These sample samples selected keywords that are frequently used in GEO fraud, such as "only official," "24-hour customer service," and "solemn declaration," but also included negative samples of known fraud cases. This guides the model to achieve two goals: first, to accurately identify entities with seemingly attractive but inherently questionable contact information; and second, to conduct a preliminary quantitative assessment of the authority of their source based on objective evidence (such as whether ICP filing, platform certification labels, website domain authority, etc. are mentioned).

[0093] Extraction: This module applies the above-mentioned customized template to extract entities from all relevant text fragments retrieved, text extracted from images via OCR, and speech-to-text content extracted from videos.

[0094] Output: For the query "Brand A", assuming the extraction module outputs the following five structured triples:

[0095] 1. (Brand A, 4001234567, Authority equals 1.0): From the official website page of Brand A that has been registered;

[0096] 2. (Brand A, 4001234568, Authority equals 0.3): A text post from a personal social media account, claiming to be the "only hotline";

[0097] 3. (Brand A, 4001234569, Authority equals 0.4): Image OCR text from a post on another platform, with the phrase "Solemn Declaration";

[0098] 4. (Brand A, 4001234560, Authority equals 0.3): Sourced from a short video subtitle and audio transcription;

[0099] 5. (Brand A, 4001234567, Authority equals 0.8): From the official flagship store page of Brand A on a major e-commerce platform.

[0100] Step S2: Construct a cross-modal heterogeneous conflict graph:

[0101] The conflict graph construction module receives these five triples and, using each triple as a node, begins constructing a cross-modal heterogeneous conflict graph. The construction rules of the conflict graph fully reflect the modeling of GEO fraud characteristics:

[0102] 1. Node initialization: Each node stores its triplet information: (brand, contact information, authority);

[0103] 2. Edge construction and weight calculation:

[0104] The system first groups the nodes according to "brand", and all nodes of "brand A" are grouped together;

[0105] For any two distinct nodes in this group and If their contact methods are different (i.e.) If so, then an intra-brand conflict reinforcement edge is established between them;

[0106] Weight of this edge It is not a fixed value, but is dynamically calculated through a conflict decay function:

[0107] ;

[0108] in, This is the Levenshtein edit distance between two phone numbers. Edit distance refers to the minimum number of single-character edits (insertion, deletion, replacement) required to convert one string into another. In this embodiment:

[0109] (Replace '8' with '9');

[0110] (Replace '8' with '0');

[0111] (Replace '8' with '7');

[0112] This distance This method intuitively measures the similarity of phone numbers; the closer the numbers are, the more likely they are to be "mass-generated" counterfeit numbers.

[0113] The conflict sensitivity coefficient is a key parameter for dynamic adjustment.

[0114] 3. The dynamic adjustment mechanism (GEO-specific characteristics) is as follows:

[0115] The system sets up a 24-hour dynamic monitoring window to check the number of unique contact methods appearing under "Brand A";

[0116] In this embodiment, four unique numbers appeared in the window: 4001234567, 4001234568, 4001234569, and 4001234560. Since the number is greater than or equal to three, the event is triggered. Adjustment conditions;

[0117] Calculate the density of similar contact methods : It equals the number of contact methods divided by 24 hours. ;

[0118] The system has a preset normal density threshold. (Approximately 2 per 24 hours) and fraud density threshold (Approximately 5 per 24 hours);

[0119] because ,but Dynamically adjusted based on the linear interpolation formula:

[0120] ;

[0121] Substitute the preset value ( , ):

[0122] ;

[0123] This adjustment means that the system has detected an unusually high density of similar numbers for "Brand A" within a short period of time, which is highly consistent with the characteristics of a GEO attack. Therefore, it automatically increases the sensitivity of conflict detection. From 1.5 to approximately 2.508);

[0124] 4. Calculate edge weights:

[0125] For the edge between node 2 (4001234568) and node 3 (4001234569): , Weight ;

[0126] For the edge between node 2 (4001234568) and node 1 (4001234567): , The weight calculation is also approximately 0.081;

[0127] Important rule: If either of the two connected nodes has an official certification (e.g., authority equal to 1.0), the edge weight will be specially handled (e.g., set to an extremely low value or 0, indicating that the conflict is not valid or can be ignored), because official sources have the highest credibility. However, in the conflict graph construction logic of this embodiment, a more direct approach is: if any node's source is officially registered (authority equal to 1.0), the above decay function will not be triggered, and the edge weight may be set to a fixed low value or handled separately. For the sake of simplicity, this embodiment still uses a formula for calculation, but in the actual system, official sources will be protected.

[0128] 5. At this point, a conflict graph containing 5 nodes and multiple weighted edges is completed. In the conflict graph, the nodes between fake numbers and between fake numbers and real official numbers are all connected by different contact methods and do not have official certification at the same time. The specific edge weights are calculated by the conflict decay function. The smaller the weight value (such as 0.081), the stronger the connection (which has a greater impact in subsequent propagation). This reflects the modeling of the "conflict" relationship.

[0129] Step S3: Run conflict-aware belief propagation:

[0130] The conflict-aware reasoning module takes over the constructed conflict graph and performs the following operations:

[0131] 1. Initialize node confidence: The initial confidence level of each node. Set it directly as its source authority score, for example:

[0132] Node 1 (Official Website): ,

[0133] Node 2 (Personal Account): ,

[0134] Node 3 (Image and Text Post): ,

[0135] Node 4 (Short Video): ,

[0136] Node 5 (Official Flagship Store): ;

[0137] 2. Iteratively execute belief propagation:

[0138] The system uses an iterative formula to update the confidence level of each node. In the round of iteration, nodes confidence level The update is as follows:

[0139] ;

[0140] in, It is a node In iteration Confidence level at that time; These are the edge weights calculated in step S2; It is a decay factor (e.g., 0.8) used to control the strength of confidence propagation and prevent numerical overflow;

[0141] Propagation logic: A node's confidence level is "pulled" by the confidence levels of all its neighboring nodes; if its edge weights with its neighbors are... If the edge weight is large (i.e., the conflict relationship is weak), then the confidence of the neighbors will have a significant impact on it; if the edge weight is small (i.e., the conflict relationship is strong), then the confidence of the neighbors will have a significant impact on it. If the value is close to 0, then the influence of neighbors will be passed on. The damage is severe, primarily due to the differences between fake number nodes and between fake numbers and official numbers. The values ​​are all very small (approximately 0.081), and their "negative impact" on each other will be strongly transmitted. Although the official node (confidence 1.0) will also affect the fake node through the edges, the fake nodes will lower each other's confidence because they form a tight "conflict cluster".

[0142] 3. Iterative convergence and fraud detection:

[0143] After several iterations (e.g., 10 iterations), the confidence of all nodes in the graph tends to stabilize.

[0144] GEO's unique judgment logic: The system focuses on all contact information nodes under "Brand A", extracts the highest confidence level corresponding to each contact method after the propagation has stabilized (a number may appear in multiple nodes, so the highest value is taken), assuming the result is:

[0145] 4001234567 (Official): Confidence level 0.95 (slightly lower than the initial 1.0 due to the impact of fake node conflicts).

[0146] 4001234568: Confidence level 0.28

[0147] 4001234569: Confidence level 0.35

[0148] 4001234560: Confidence level 0.27;

[0149] Calculate the difference between the highest confidence level (0.95) and the second highest confidence level (0.35): ,

[0150] Compare this difference with a preset threshold (default 0.3), which is based on statistical prediction of a typical GEO scenario of "multiple high-confidence but mutually exclusive fake numbers". It aims to identify the "evenly matched" conflict situation that black market operators carefully create to confuse large models.

[0151] Due to the difference The current assessment indicates that no typical "multiple confidence conflicts" type GEO fraud has been detected, reflecting the results of the system's first assessment: the official number still leads by a significant margin.

[0152] In a real-world system, in addition to the confidence difference, the topological features of the graph may also be considered for a comprehensive judgment. For example, the system may simultaneously detect that in the "Brand A" subgraph, there is a tightly connected "cluster" with very low weights consisting of nodes 2, 3, and 4, and the initial authority of these nodes is low. This "low authority-high intensity conflict cluster" is itself a strong GEO fraud signal. Therefore, the system may trigger a secondary alarm or record that the brand has a highly suspicious conflict pattern, even if the confidence difference does not exceed the threshold.

[0153] This example assumes an extreme scenario where there are no genuine official numbers (nodes 1 and 5) (the fraud gang feeds the system extensively before official information is retrieved). After propagation, the possible distribution would be: 4001234568: 0.38, 4001234569: 0.40, 4001234560: 0.37. The difference between the highest and second-highest values ​​in this case is... The system will accurately identify it as GEO fraud and mark {4001234568, 4001234569, 4001234560} as a set of fraudulent numbers.

[0154] Step S4: Dynamically intervene in the large model generation process based on conflict-aware reasoning results:

[0155] Regardless of whether the final judgment of step S3 is "fraud" or "highly suspicious", the output "list of suspicious / fraudulent contact information" and related confidence information will be passed to the generation intervention module for proactive intervention before the generative answer is formed.

[0156] 1. The system stores all the "contact information judged as fraudulent or highly suspicious" output in step S3 (in this complete scenario, assuming it includes {4001234568, 4001234569, 4001234560}) as keys in a temporary confidence mask dictionary, assigning each key a very small decay factor. ( For example, the default This means that any content containing these numbers will have its importance reduced to 5% or even less of its original value;

[0157] 2. In the RAG process of the search engine, after the Retriever retrieves the Top-K text fragments most relevant to the user query from the massive knowledge base (with a similarity score, such as cosine similarity), before sending these fragments to the Generator for answer synthesis, the confidence mask intervention layer of this invention is inserted.

[0158] 3. Execute dynamic mask:

[0159] The search engine returned, for example, 10 relevant fragments, which might include:

[0160] Excerpt 1: "Brand A's official website displays customer service number: 400-123-4567" (high similarity score, such as 0.92);

[0161] Segment 2: "According to user sharing, the latest customer service number for brand A is 400-123-4568" (similarity score, such as 0.75);

[0162] Segment 3: "An online image states 'Brand A solemnly declares: Hotline 400-123-4569'" (similar score, e.g., 0.70);

[0163] Segment 4: "A video mentions that to contact brand A, please call 400-123-4560" (similar score, e.g., 0.68);

[0164] The confidence mask intervention layer quickly scans each of the 10 segments. The scan uses a high-efficiency regular expression engine that integrates various phone number patterns (such as 400 / 800 number segments, 95XXX number segments, and mobile phone number formats). It can quickly identify and extract the contact information strings in the segments and normalize them into pure numeric format for matching.

[0165] For fragment 1, no number was matched in the mask dictionary, so its score remains at 0.92;

[0166] For fragment 2, a match was found for 4001234568, and its score was updated as follows:

[0167] ;

[0168] For segment 3, a match was found to be 4001234569, and the score was updated to [score missing]. ;

[0169] For segment 4, a match was found to be 4001234560, and the score was updated to [score missing]. ;

[0170] 4. Generate safe answers:

[0171] The intervention layer returns all segments and their adjusted scores to the RAG process;

[0172] In subsequent sorting or reordering stages, because the scores of fragments 2, 3, and 4 are greatly diminished (from 0.7 to 0.035), they will be ranked far behind fragment 1 (score 0.92).

[0173] Ultimately, the large model generator primarily uses the highest-scoring segment containing the real official number (segment 1) to organize and generate answers. Although false information exists in the context information "seen" by the generator, it has been marked as having extremely low relevance, so the probability of it being adopted as the core basis of the answer is negligible.

[0174] The answer returned to the user was: "According to information on Brand A's official website, its official customer service number is 400-123-4567," thus successfully preventing the output of GEO fraudulent information.

[0175] This embodiment forms a closed-loop, proactive GEO anti-fraud solution, from custom entity extraction based on GEO features, to constructing a conflict graph with dynamic weights to positively model "multi-source conflicts", to performing confidence propagation for conflict awareness to make accurate judgments, and finally to achieving source intervention through a dynamic masking mechanism seamlessly integrated into the RAG process.

[0176] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A GEO anti-fraud method based on multi-source contact information conflict modeling, characterized in that, The method comprises the following steps: Step S1: extracting structured triples from the multi-modal content associated with the returned results of the generative engine, the triples comprising a target brand, a contact method, and a source authority identifier; Step S2: constructing a cross-modal heterogeneous conflict graph with each triple as a node; The construction of the conflict graph comprises: establishing an intra-brand conflict strengthening edge between nodes belonging to the same brand and having different contact methods, and dynamically calculating the weight of the edge based on a conflict decay function; The conflict decay function calculates the weight according to at least the similarity of the two contact methods and a dynamically adjustable conflict sensitivity coefficient; Step S3: assigning an initial confidence to each node in the conflict graph based on its source authority identifier, and performing conflict-aware confidence propagation on the conflict graph to iteratively update the confidence of each node; Based on the confidence of each contact method under the same brand obtained after propagation, fraud is determined; when the difference between the contact method with the highest confidence and the contact method with the second highest confidence under the brand is less than a preset threshold, it is determined that there is GEO fraud; Step S4: based on the fraud determination result of step S3, dynamically intervening in the retrieval enhancement generation process of the generative large model to reduce the appearance of contact methods determined to be fraudulent in the final generated answer; The process of extracting structured triples in step S1 is completed using a small sample prompt template customized for the GEO anti-fraud task; The sample of the prompt template contains at least one of the keywords "unique official", "24-hour customer service", and "serious statement", and contains known fraudulent negative samples to guide the extraction model to accurately identify entities with fraud risk; The establishment of the intra-brand conflict strengthening edge in step S2 comprises: detecting whether a number of similar contact method groups appear within a preset time window under the same brand; If the detection is "yes", it is determined that there is a "centralized appearance" GEO fraud feature, and the dynamic adjustment of the conflict sensitivity coefficient is triggered; The preset time window is 24 hours; trigger conflict sensitivity coefficient The condition for dynamic up-regulation is that in 24 hours, more than or equal to 3 unique contact methods under the same brand are detected; The conflict sensitivity coefficient According to the real-time calculated similar contact method density , a default value , and a maximum value , the adjustment formula is: When Time: ; When Time: ; When Time ; wherein, is the number of contacts divided by 24 hours, is the normal density threshold, is the fraud density threshold; The conflict decay function is specifically: ； wherein, is the weight of the edge, is the Levenshtein edit distance between two phone numbers, is the dynamically adjusted conflict sensitivity coefficient; In step S3, the preset threshold is set to 0.3; when the difference between the highest confidence and the second highest confidence is less than 0.3, it is determined that there are multiple high-confidence but mutually exclusive false numbers, i.e., GEO fraud.

2. The GEO anti-fraud method based on multi-source contact information conflict modeling according to claim 1, characterized in that, In step S4, the dynamic intervention in the retrieval enhancement generation process of the generative large model specifically comprises: Step S41: Store all contact details determined as fraudulent in step S3 in a confidence mask dictionary and associate a decay factor with each contact detail in the dictionary wherein ; Step S42: inserting a confidence mask intervention layer after the retriever output stage of the retrieval enhancement generation process; Step S43: The confidence mask intervention layer scans each text segment returned by the retriever, and if it finds a contact in the segment, it multiplies the original similarity score of the segment by the decay factor to obtain a new score after intervention . Step S44: inputting the text segment list with adjusted scores into the generator of the large model to generate the final answer.

3. A GEO anti-fraud system based on multi-source contact information conflict modeling, used to implement the GEO anti-fraud method based on multi-source contact information conflict modeling according to any one of claims 1-2, characterized in that, It comprises: A multi-modal entity extraction module for extracting structured triples containing a target brand, a contact method, and a source authority identifier from multi-modal content; A conflict graph construction module for constructing a cross-modal heterogeneous conflict graph with the triples as nodes, establishing an intra-brand conflict strengthening edge between nodes belonging to the same brand and having different contact methods, and dynamically calculating the edge weight based on a conflict decay function; A conflict perception reasoning module is configured to assign an initial confidence level to a node, perform belief propagation on the conflict graph, and iteratively update the confidence level of the node, and determine the GEO fraud based on the difference in the confidence level of each contact method under the same brand after the propagation; A generation intervention module is configured to dynamically intervene in the retrieval enhancement generation process of the generative large model based on the fraud determination result, and reduce the output of fraudulent contact methods.

4. The GEO anti-fraud system based on multi-source contact information conflict modeling according to claim 3, characterized in that, The conflict graph construction module comprises: A time-sensitive analysis unit is configured to monitor the density of contact methods under the same brand within a preset time window; A dynamic parameter adjustment unit is configured to dynamically adjust the conflict-sensitive coefficient in the conflict decay function according to the contact method density.

5. The GEO anti-fraud system based on multi-source contact information conflict modeling according to claim 3, characterized in that, The generation intervention module comprises: A mask dictionary management unit is configured to store the contact methods determined to be fraudulent and their associated decay factors; An intervention layer execution unit is implanted into the retrieval enhancement generation process, configured to scan the retrieved text segment and recalculate the similarity score; The intervention layer execution unit uses a regular expression engine integrated with multiple phone number patterns to quickly match and scan the content of the text segment.

Images