Fraud confrontation dialogue generation and detection method and system based on large model and multi-agent

Abstract

The application provides a fraud confrontation dialogue generation and detection method and system based on a large model and multiple agents, wherein the method comprises the following steps: scene construction; multi-modal voice data generation; text annotation based on slow thinking and multi-round deep reasoning analysis. The application proposes a fraud confrontation dialogue generation and deep thinking analysis scheme that fuses a large language model and multiple agents, constructs a new generation of anti-fraud intelligent system with dynamic attack and defense game capabilities, constructs a specialized agent cluster, fills the key research gap in the field of multi-modal fraud detection, solves key problems such as data privacy and scene diversity, and provides technical support for promoting the development of an intelligent anti-fraud system. The application deeply fuses voice data generation technology, greatly improves the complexity and practicality of confrontation training, and approximates a real fraud scene.

Description

Technical Field

[0001] This invention belongs to the field of speech recognition technology, and relates to large model technology, specifically to a method and system for generating and detecting fraudulent adversarial dialogues based on large models and multi-agent systems. Background Technology

[0002] As telecommunications fraud techniques become increasingly sophisticated and advanced, their threat to social security and economic stability is escalating. According to a 2023 report by the non-profit Global Anti-Fraud Coalition (GASA) and data service provider ScamAdviser, global economic losses due to fraud have reached $1.02 trillion, accounting for 1.05% of global GDP. This figure represents a significant increase compared to 2020 and 2021, with over a quarter of respondents reporting experiencing fraud. Therefore, developing effective fraud detection methods has become a top priority.

[0003] Traditional fraud detection methods rely on manual verification and rule-based pattern matching, which are often inaccurate and struggle to cope with the rapidly changing strategies employed in fraudulent communications. Recently, the development of Large Language Models (LLMs), particularly their slow-thinking reasoning capabilities, has provided new solutions for combating telecom fraud. However, the primary data source for telecom fraud is voice calls, which differ significantly from text data in modality, limiting the direct application of LLMs in this field. Currently, the industry typically uses Automatic Speech Recognition (ASR) technology to convert audio data into text, and then invokes LLMs for fraud detection through carefully crafted prompts.

[0004] Existing technologies rely on real-time user authentication checks to ensure caller identity. While achieving 97.98% accuracy in 100 synthetic call tests, this technology primarily depends on text analysis and lacks the ability to directly process audio features. Previous researchers have investigated the application of Large Language Models (LLMs) in real-time telecom fraud detection, developing a framework to assess fraudulent intent in conversations and provide immediate warnings to users. Their research analyzed key factors influencing detection performance, particularly the balance between recall and timeliness. Nevertheless, this research remains text-centric and fails to fully utilize the rich information contained in audio. Another study reviewed the potential and challenges of Large Language Models (LLMs) in telecom fraud detection. They noted that while LLMs perform well in identifying fraud patterns, issues such as data bias, low recall, and hallucinations persist. This highlights the importance of innovative, high-quality, and diverse training techniques. Furthermore, some studies have revealed the vulnerability of LLMs to adversarial fraud information. By creating a comprehensive technical innovation incorporating both raw and adversarial fraud information, this study analyzed the misclassification rate of LLMs and proposed strategies to enhance their robustness. This study highlights the need to consider adversarial scenarios in the development of anti-fraud systems, although its analysis remains limited to the text domain.

[0005] While existing anti-fraud systems have performed well under specific conditions and achieved some success, they heavily rely on meticulously designed cue engineering, predefined rules, or plain text analysis, lacking the ability to integrate audio and text information for joint reasoning. Furthermore, they are typically trained and evaluated on limited-scale technological innovations, and performance differences between different models and deployment environments can significantly impact the accuracy of judgments, making them ill-equipped to combat rapidly evolving fraud tactics. Additionally, information loss during automatic speech recognition (ASR) conversion can lead to the omission of key fraud indicators, as voice features such as intonation and pauses often contain important clues about fraudulent intent. These limitations highlight the urgent need to develop large-scale, multimodal telecommunications fraud technology innovations.

[0006] The lack of high-quality multimodal training data poses a significant challenge to telecom fraud detection. Multimodal training data combines audio signals with reasoning-driven text analysis. As online fraud methods rapidly evolve towards intelligence, concealment, and cross-modal collaboration, traditional defense systems relying on rule matching or single-point analysis face severe challenges. Their passive response mechanisms are ill-equipped to handle dynamically evolving rhetoric and deep attacks. Existing technologies exhibit significant deficiencies in proactive prediction, deep logical deconstruction, and adaptive evolution. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes an innovative audio-text slow-thinking technology solution for detecting telecommunications fraud, specifically for small sample sizes.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A method for generating and detecting adversarial dialogues for fraud based on large models and multi-agent systems includes the following steps:

[0010] Step 1, scenario building: define common call scenarios and fraud types;

[0011] Step 2, multimodal speech data generation, including:

[0012] Step 2.1: Use Automatic Speech Recognition (ASR) to transcribe the call recording containing anonymized raw audio and generate a privacy-protected text sample;

[0013] Step 2.2: Extract typical samples from normal call text and fraudulent call text as examples to guide the large language model to imitate and expand the automatic speech recognition (ASR) results of real calls, and obtain expanded text samples.

[0014] Step 2.3: A multi-agent adversarial framework is used to generate dialogue text; the multi-agent adversarial framework includes a fraudster, a potential victim, and an administrator; the victim role is given specific identity characteristics and reaction patterns; the dialogue is conducted in a turn-based manner, with the fraudster and the victim taking turns speaking, and the administrator monitoring the entire dialogue process.

[0015] Step 2.4: Using text-to-speech (TTS) technology, the dialogue text is converted into dual-channel audio data, corresponding to the caller and the called party respectively.

[0016] Step 3, based on slow-thinking text annotation and multi-round deep reasoning analysis, includes:

[0017] Step 3.1: Input the generated audio data into the audio understanding model to extract speech features and key information;

[0018] Step 3.2 combines the audio analysis results with the Automatic Speech Recognition (ASR) text as prompts to guide the DeepSeekR1 model in reasoning and analysis, outputting call scenario classification, fraud determination, and fraud type identification. The model plays the role of a "speech analysis expert with anti-fraud expertise".

[0019] Furthermore, step 2.2 includes the following process:

[0020] Key patterns and linguistic features were extracted from real phone calls, including the structure of fraud strategies, common expressions, and interaction patterns.

[0021] By combining key patterns, linguistic features, and contextual descriptions as prompts, large-scale language models are guided to generate dialogue content that fits specific contexts.

[0022] Furthermore, for fraud-related data, the typical characteristics of fraud strategies are retained, while diversity is increased by changing the details.

[0023] Furthermore, the following joint constraint function is introduced during the generation of the dialogue content:

[0024] ,

[0025] in The original generation loss of the language model; The dialogue structure constraint loss is used to constrain the sequence of stages in fraudulent rhetoric. The semantic function distribution constraint loss is used to limit the proportion of different fraud semantic function units that appear. , Indicates the weighting coefficient;

[0026] The semantic function distribution constraint loss is defined as:

[0027]

[0028] in This represents the proportion of the k-th semantic function in the generated text. This represents the target distribution obtained based on statistics from real fraud samples.

[0029] Furthermore, step 2.3 specifically includes:

[0030] In the multi-agent adversarial framework, each agent is modeled as a finite-state decision-making process, and its set of states is defined as follows:

[0031]

[0032] in, This represents the current state of the intended message; The explicit or implicit emotional state in a conversation; This is the stage in the fraud process.

[0033] Map semantic states in the state set to continuous risk and schedule metrics:

[0034]

[0035] in, As an indicator of the progress of the fraud, it is divided into stages of the fraud process. Mapped to obtain; As a risk indicator for fraud exposure, it is determined by the state of intent in the persuasive statements. With emotional state The joint estimation yielded; and It is a mapping function;

[0036] The dialogue between the scammer and the victim is modeled as an adversarial game, with its payoff function defined as:

[0037]

[0038] in, To advance the return weighting coefficient, This is the risk penalty weighting coefficient;

[0039] The manager agent monitors the dialogue status in real time and terminates the dialogue generation when one of the following termination conditions is met:

[0040]

[0041] in, Number the current dialogue round. To the maximum number of allowed dialogue rounds, This is the dynamic threshold for risk.

[0042] Furthermore, in step 2.4, the ChatTTS model is used for conversion, and different voice parameters are randomly generated to configure unique voice features for different roles, preserve natural voice features, and precisely control the time relationship between the caller and the called party's voice to ensure natural turn-taking.

[0043] Furthermore, the voice parameters include timbre, speech rate, and pitch, and the natural speech features include pauses, stress, and emotional changes, making the synthesized voice closer to a real conversation.

[0044] Furthermore, in step 3.2, the model uses specific symbols to mark its professional thinking process, demonstrating the complete reasoning chain from clue recognition to professional judgment, and marks the output results with specific symbols.

[0045] Furthermore, in step 3.2, the model outputs data in JSON format. The output for call scenario classification includes: scenario type, judgment reason, and confidence level. The output for fraud judgment includes: judgment reason, confidence level, and fraud status. The output for fraud type identification includes: fraud type, judgment reason, and confidence level.

[0046] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a method for generating and detecting fraudulent adversarial dialogues based on large models and multiple agents.

[0047] The beneficial effects of this invention are as follows:

[0048] 1. This invention creatively proposes a scheme integrating Large Language Model (LLM) with multi-agent adversarial dialogue generation and deep thinking analysis to construct a new generation of anti-fraud intelligent system with dynamic offensive and defensive game capabilities. This invention constructs a specialized intelligent agent cluster. The fraudster agent generates highly persuasive dialogues by meticulously simulating the psychological characteristics of victims; simultaneously, the victim agent deconstructs the verbal logic chain and psychological manipulation nodes in the interaction in real time; the manager agent dynamically generates countermeasures and optimizes the defense model based on multi-round adversarial feedback, filling a key research gap in the field of multimodal fraud detection, solving key issues such as data privacy and scenario diversity, and providing technical support for the development of intelligent anti-fraud systems. The system deeply integrates speech data generation technologies (including ASR transcription, LLM self-guided sampling, and multi-agent adversarial synthesis), significantly improving the complexity and practicality of adversarial training, closely approximating real fraud scenarios.

[0049] 2. By introducing a constrained text generation mechanism, multi-agent adversarial distributed expansion, and multi-round progressive reasoning structure, this invention can solve the problem that it is difficult to simultaneously take into account the authenticity, diversity, and forward-looking nature of the generated text in the prior art. It is especially suitable for application scenarios where new fraud patterns are scarce, behaviors evolve rapidly, and single features are difficult to determine.

[0050] 3. This invention designs a new data generation pipeline that, through realistic call ASR processing, LLM-based simulation, and multi-agent adversarial generation, maximizes the coverage of various fraud scenarios, forms diverse audio characteristics when generating audio data, preserves natural speech features, and, combined with precise time control, further improves the realism and diversity of audio data.

[0051] 4. This technology relies on the powerful semantic understanding and logical reasoning capabilities of large models, combined with the slow thinking two-stage analysis framework (audio feature extraction + logical reasoning), to achieve multi-dimensional analysis of fraud behavior. Through semantic tracing and intent recognition in the dialogue flow, it deeply mines the features or patterns corresponding to the victim's decision-making weaknesses, and realizes communication scenario classification, fraud determination, and fraud type analysis.

[0052] 5. This invention breaks through the limitations of static knowledge response—by actively inducing the explicitness of fraud logic through a multi-agent adversarial framework, and combining a three-level slow-thinking judgment process (scenario classification → fraud determination → type identification), it achieves a paradigm shift from "passive response" to "fraud adversarial dialogue generation." It is the first to integrate a three-in-one capability system of dynamic adversarial simulation, deep logical deconstruction, and autonomous strategy evolution, providing a technical path for the anti-fraud field. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the fraud adversarial dialogue generation and detection method based on a large model and multiple agents provided by the present invention.

[0054] Figure 2 This describes the process of processing speech and text.

[0055] Figure 3 This is a diagram of a multi-agent adversarial framework architecture. Detailed Implementation

[0056] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0057] The fraud adversarial dialogue generation and detection method based on large models and multi-agent systems provided by this invention has the following process: Figure 1 As shown, it includes the following steps:

[0058] Step 1, Scene Construction

[0059] This invention defines seven common calling scenarios: "food service", "customer consultation", "appointment service", "transportation inquiry", "daily shopping", "ride-hailing service" and "food delivery service", as well as seven major types of fraud: "investment fraud", "phishing fraud", "identity theft", "lottery fraud", "bank fraud", "extortion fraud" and "customer service fraud".

[0060] Step 2, Multimodal speech data generation

[0061] To construct high-quality, diverse, and small-sample technological innovations, it is first necessary to create rich voice data. This invention employs three complementary innovative methods to generate dialogue content, which is then converted into real call recordings using text-to-speech technology, thereby generating high-quality and diverse telephone call voice data covering various scenarios of telecom fraud and normal conversations. Specifically, this includes:

[0062] Step 2.1 uses Automatic Speech Recognition (ASR) to transcribe the call recording containing anonymized original audio, generating a privacy-protected text sample. The text obtained in this step will be regenerated into a speech sample using a Text-to-Speech (TTS) model in subsequent steps to ensure its authenticity and consistency with real-world language expression, thereby improving its applicability in practical applications.

[0063] Step 2.2 employs a self-guided paradigm to acquire dialogue text, utilizing a large-scale language model (LLM) to mimic and extend the automatic speech recognition (ASR) results of real-life calls obtained in Step 2.1, resulting in extended text samples. This invention designs specific prompt templates, extracting typical samples from normal and fraudulent call texts as a small set of examples to guide the LLM in generating a large number of dialogue scripts. These scripts, while following similar patterns, differ in content. Specifically, key patterns and linguistic features are first extracted from real-life calls, including the structure of fraudulent strategies (the phased organization of fraudulent calls in multi-turn dialogues), common expressions (high-frequency semantic function types that repeatedly appear in different fraud scenarios), and interaction patterns (the stable response relationship and rhythm control methods formed between the fraudster and the called party in the dialogue). Subsequently, the linguistic structural features, common semantic patterns, and interaction rules extracted from real-life calls are combined with the predefined call scenario description from Step 1 to construct structured prompt information, guiding the large-scale language model to generate dialogue text that conforms to specific contexts. The prompt information includes at least scenario constraints, role settings, and generation focus control information. For generating normal call texts, the focus is on ensuring the integrity of the service process and the naturalness of the language. For generating fraudulent call texts, while retaining the typical structural features and interaction patterns of fraudulent strategies, non-critical details are adjusted to enhance sample diversity. The resulting normal and fraudulent call texts maintain authenticity while possessing high coverage and usability.

[0064] In its implementation, the dialogue script generation process is not solely based on natural language prompts, but rather controls the model output by constructing a computable generation constraint mechanism. To achieve precise control over the generated dialogue structure and semantic distribution, this invention introduces a formalized generation constraint function into the self-guided generation process. Let the dialogue text sequence be:

[0065]

[0066] in Let represent the content of the t-th round of speech, and this patent introduces a joint constraint objective function in the process of generating a large-scale language model:

[0067] ,

[0068] in The original generation loss of the language model; The dialogue structure constraint loss is used to constrain the sequence of stages in fraudulent rhetoric. The semantic function distribution constraint loss is used to limit the proportion of different fraud semantic function units that appear. , This represents the weighting coefficient.

[0069] The semantic function distribution constraint is defined as follows:

[0070]

[0071] in This represents the proportion of the k-th semantic function in the generated text. This represents the target distribution obtained based on statistics from real fraud samples.

[0072] The constraint generation mechanism should include at least the following:

[0073] (1) Dialogue structure constraints, used to limit the order of appearance and dependencies of different functional stages in the dialogue;

[0074] (2) Semantic function distribution constraints, used to control the proportion and combination of different semantic function units in the dialogue;

[0075] (3) Interaction response constraints are used to limit the subsequent dialogue behaviors that can be triggered under different user feedback conditions.

[0076] The above constraints ensure that the generated text maintains structural consistency for specific behavioral patterns while preserving diversity.

[0077] The above method enables the generation of a large amount of high-quality normal call text (DNS2) and fraudulent call text (DFS2), significantly expanding technological innovation. The advantage of the self-guided method lies in its ability to substantially increase data volume and diversity while maintaining the original data characteristics. The generation process is efficient and controllable, and our experiments demonstrate that this method can effectively generate high-quality data.

[0078] Step 2.3: Generate dialogue text using a multi-agent adversarial framework.

[0079] This invention simulates various emerging fraud methods and expands the data distribution. Using the text data obtained in steps 2.1 and 2.2 as the reference and constraint basis for adversarial generation and distribution expansion, a multi-agent adversarial framework is employed to generate novel dialogue texts that deviate from the existing distribution. Although existing telecom fraud call recordings are authentic and reliable, their dialogue patterns and scene distributions are relatively concentrated, making it difficult to cover diverse fraud scenarios and new technologies. To address this issue, this invention designs a multi-agent adversarial framework. In this framework, each agent is modeled as a finite-state decision-making process, and its state set is defined as:

[0080]

[0081] in, This represents the current state of the intended message; The explicit or implicit emotional state in a conversation; This is the stage in the fraud process.

[0082] In the specific implementation, to facilitate quantitative control of the dialogue generation process, the semantic states in the state set are mapped to continuous risk and progress indicators. Its specific form is as follows:

[0083]

[0084] in, As an indicator of the progress of the fraud, it is divided into stages of the fraud process. Mapped to obtain; As a risk indicator for fraud exposure, it is determined by the state of intent in the persuasive statements. With emotional state The joint estimation yielded; and It is a mapping function.

[0085] The dialogue between the scammer and the victim is modeled as an adversarial game, with its payoff function defined as:

[0086]

[0087] To achieve a dynamic trade-off between the progress of fraud and the risk of exposure, parameters were defined. and ,in To advance the return weighting coefficient, These are risk penalty weight coefficients, with values ​​between 0 and 1. Both parameters can be initialized using historical real fraud data and dynamically adjusted during adversarial generation using gradient updates or Bayesian optimization methods, thus achieving adaptive parameter evolution.

[0088] The manager agent monitors the dialogue status in real time and terminates the dialogue generation when one of the following termination conditions is met:

[0089]

[0090] Where t is the current dialogue round number; This represents the maximum number of allowed dialogue rounds. This is a dynamic risk threshold, which can be adaptively adjusted according to the data distribution.

[0091] By simulating fraudulent behavior in different business scenarios, the data distribution space was effectively expanded. Based on the comparative analysis of the text generated in steps 2.1, 2.2, and 2.3, the analysis demonstrates the effectiveness of multi-agent adversarial generation in expanding data distribution and simulating new fraud patterns.

[0092] like Figure 3As shown, the framework consists of three collaborative roles: two roles act as the scammer and the potential victim, respectively, while the third role acts as the manager. In this framework, information about the scam type is first provided to the scammer's large language model to make the scam more purposeful and realistic. Furthermore, specific identity characteristics and reaction patterns are set for the victim role to ensure more authentic interaction. The dialogue proceeds in a turn-based manner, with the scammer and victim taking turns speaking, while the manager monitors the entire dialogue process to ensure it conforms to the preset scenario and remains natural and fluent. To ensure the diversity and broader distribution of the generated data, multiple scenario-scam type combinations are designed based on seven common call scenarios and seven scam types in real-world business environments. These combinations guide the multi-agent system to generate dialogues covering these scenarios. This approach not only effectively expands the distribution of the original data but also fills gaps in the existing data distribution. Multi-turn dialogues are then conducted based on the above framework.

[0093] To achieve a natural end to the call, this invention designs a special termination signal mechanism. Both the caller and the called party can indicate their intention to hang up by sending the special symbol ##END_SIGNAL## (this symbol can be adjusted as needed). Furthermore, the administrator can decide to terminate the conversation based on its progress to prevent invalid loops or excessive deviation from the topic.

[0094] Using a multi-agent adversarial framework, a new set of normal call data (DNS3) and fraudulent call data (DFS3) was generated. The semantic embeddings of all generated data were visualized in this step. The results show that the data generated by the multi-agent framework significantly expands the distribution space of the original data, especially in novel fraud types and complex scenarios, providing more comprehensive training data for the model.

[0095] It should be noted that, Figure 2 ( Figure 2 To simplify the process (slightly different from the process described in the text of this invention), the agent module shown is used to generate and expand the dialogue text. This module corresponds to the large language model expansion generation process in step 2.2 and the multi-agent adversarial generation process in step 2.3. The dialogue text generated and expanded above, together with the automatic speech recognition (ASR) text obtained in step 2.1, constitutes the input data for the subsequent speech synthesis stage. Specifically, as follows:

[0096] Step 2.4: Based on the ASR text obtained in Step 2.1, the extended text generated in Step 2.2, and the dialogue text generated by the multi-agent adversarial process in Step 2.3, high-quality text-to-speech (TTS) technology is used to convert these texts into dual-channel audio data, corresponding to the caller and the called party respectively. To achieve a highly realistic speech synthesis effect, the advanced open-source TTS model ChatTTS is adopted, which is optimized for dialogue scenarios. This model can produce natural and expressive synthesized speech, supports multi-speaker scenarios, and allows fine-grained control over prosodic features, surpassing most open-source TTS models in terms of naturalness.

[0097] Furthermore, the present invention implements the following strategies during the speech synthesis process to further improve the realism and diversity of audio data:

[0098] 1. Different voice parameters, including timbre, speech rate, and pitch, were randomly generated to ensure that the synthesized voice has diverse auditory features. Unique voice features were configured for different characters, making it easier to distinguish between the two parties in a conversation.

[0099] 2. During the TTS process, natural speech features such as pauses, stress, and emotional changes are preserved, making the synthesized voice more like a real conversation. Especially in fraud scenarios, specific emotional expressions commonly used in fraud tactics, such as a sense of urgency, authority, or false affinity, are enhanced.

[0100] 3. Precise control over the timing of the caller's and callee's voices ensures natural turn-taking, including appropriate response delays, overlapping speech, and interruptions—all common phenomena in real conversations. This further enhances the realism of the audio data.

[0101] Step 3: Text annotation based on slow thinking and multi-round deep reasoning analysis

[0102] This step simulates the analysis process and professional judgment of anti-fraud experts on voice calls. To generate high-quality text annotations, this invention designs a two-stage processing workflow:

[0103] Step 3.1, Audio Analysis Stage: The generated speech data is input into a professional audio understanding model to extract speech features and key information, such as emotional changes, intonation features, pause patterns, and other audio-level anti-fraud clues. The professional audio understanding model can employ any one or a combination of existing publicly available speech representation learning models, emotion recognition models, or prosodic analysis models. For the input speech signal x(t), the audio understanding model extracts multi-level feature vectors:

[0104]

[0105] in, These are prosodic features extracted based on fundamental frequency and energy variations; This represents the probability distribution of emotions output by the emotion classification model. The pause duration, frequency, and location characteristics are used. The final audio features are unified to a fixed-dimensional space through linear mapping:

[0106]

[0107] in, The mapped unified audio feature vector is used to concatenate with text features, can also be used as an inference prompt vector, and participate in the calculation of the risk scoring function; It is a trainable linear mapping matrix used to achieve audio-text semantic alignment.

[0108] The above formulas and methods are used to encode high-dimensional features of speech signals.

[0109] Step 3.2, Expert Reasoning Stage: The audio analysis results are combined with the Automatic Speech Recognition (ASR) text obtained in Step 2.1 to obtain speech-text pairs. These pairs serve as prompts to guide the DeepSeekR1 model, enabling it to act as a "speech analysis expert with anti-fraud expertise" and perform systematic deep reasoning and analysis. The model needs to use specific symbols. <think>< / think> To mark their professional thinking process, showcasing the complete reasoning chain from clue identification to professional judgment, and using " <answer>< / answer> Use "" to mark the output results.

[0110] This design embodies a systematic and professional analytical framework for anti-fraud experts, including voice pattern recognition, technical fraud identification, risk assessment, and type determination. The model uses specific symbols… <think>< / think> The system uses a tag to mark the professional thought process, including extracting key clues from the dialogue, analyzing speaking patterns, and conducting risk assessments, ultimately providing a clear fraud judgment. Our designed prompt template first provides the dialogue content and audio feature analysis, then requires the model to output three key pieces of information in JSON format: reason, confidence, and Boolean fraud judgment (is_fraud). The model first needs to... <think>< / think> The "Tag" section displays the full analysis process, then... <answer>< / answer> The final judgment is given under the marker. This structure allows the model to simulate the thought process of anti-fraud experts, comprehensively demonstrating the entire process from evidence collection to professional judgment.

[0111] The slow-thinking reasoning in this invention employs a phased discriminant function, rather than a single-round generative judgment. The overall fraud risk score is defined as:

[0112]

[0113] in, Scenario consistency score; Score for the completeness of the fraud logic; Audio anomaly score; This represents the Sigmoid function. When... The call was immediately identified as a scam. In the multi-round reasoning process, the intermediate results of each round were explicitly recorded and used as input constraints for the next round, thus forming a causal and progressive reasoning chain.

[0114] Multi-round deep analysis tasks consist of three consecutive steps. Instead of simply repeating model reasoning, multi-round deep analysis constructs a causal, progressive analytical chain. Each round identifies local anomalous features in speech or text; each subsequent round uses the intermediate conclusions of the previous round as input or constraints to further abstract behavioral patterns, ultimately forming a comprehensive judgment about the overall behavioral intent. The specific implementation of deep analysis can be completed through one or more model calls. This progressive reasoning structure avoids misjudgments or information omissions caused by single-round analysis.

[0115] 1. Call Scenario Classification: The model first needs to analyze the basic scenario and topic of the call to lay the foundation for subsequent judgments. Based on the various scenarios created in step 1, the model needs to determine the most appropriate scenario category by analyzing the dialogue content and voice features.

[0116] 2. Fraud Determination: Based on the scenario classification results, the model further determines whether fraudulent behavior exists in the call and explains the basis for its determination. This step requires the model to apply professional anti-fraud knowledge to analyze professional indicators such as voice characteristics, the rationality of requests, and information disclosure patterns in order to provide a clear fraud determination and a detailed chain of evidence.

[0117] 3. Fraud Type Identification: For calls identified as fraudulent, the model needs to further determine the specific fraud type. Based on the various fraud types defined in Step 1, the model needs to determine the most matching fraud type according to professional classification standards and explain the classification basis.

[0118] In the multi-round deep analysis task, each step requires the model to provide JSON-formatted output, including key judgment results and confidence levels. For call scenario classification, the output includes scenario type, judgment reason, and confidence level; the fraud determination stage requires the judgment reason, confidence level, and a Boolean value representing the fraud status (whether it is fraud); fraud type identification includes fraud type, judgment reason, and confidence level. In each step, the model first demonstrates its professional thought process and then provides structured results. This step-by-step analysis design not only conforms to the workflow of professional anti-fraud personnel but also helps the model build the necessary contextual understanding before making complex decisions, thereby improving the accuracy and reliability of the final judgment. By recording the thought process at each step, rich analytical samples are provided for subsequent research.

[0119] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the fraud adversarial dialogue generation and detection method based on large models and multi-agents provided by the present invention.

[0120] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A method for generating and detecting adversarial dialogues for fraud based on large models and multi-agent systems, characterized in that, Includes the following steps: Step 1, scenario building: define common call scenarios and fraud types; Step 2, multimodal speech data generation, including: Step 2.1: Use Automatic Speech Recognition (ASR) to transcribe the call recording containing anonymized raw audio and generate a privacy-protected text sample; Step 2.2: Extract typical samples from normal call text and fraudulent call text as examples to guide the large language model to imitate and expand the automatic speech recognition (ASR) results of real calls, and obtain expanded text samples. Step 2.3: A multi-agent adversarial framework is used to generate dialogue text; the multi-agent adversarial framework includes a fraudster, a potential victim, and an administrator; the victim role is given specific identity characteristics and reaction patterns; the dialogue is conducted in a turn-based manner, with the fraudster and the victim taking turns speaking, and the administrator monitoring the entire dialogue process. Step 2.4: Using text-to-speech (TTS) technology, the dialogue text is converted into dual-channel audio data, corresponding to the caller and the called party respectively. Step 3, based on slow-thinking text annotation and multi-round deep reasoning analysis, includes: Step 3.1: Input the generated audio data into the audio understanding model to extract speech features and key information; Step 3.2 combines the audio analysis results with the Automatic Speech Recognition (ASR) text as prompts to guide the DeepSeekR1 model in reasoning and analysis, outputting call scenario classification, fraud determination, and fraud type identification. The model plays the role of a "speech analysis expert with anti-fraud expertise".

2. The fraud adversarial dialogue generation and detection method based on large models and multi-agent systems according to claim 1, characterized in that, Step 2.2 includes the following process: Key patterns and linguistic features were extracted from real phone calls, including the structure of fraud strategies, common expressions, and interaction patterns. By combining key patterns, linguistic features, and contextual descriptions as prompts, large-scale language models are guided to generate dialogue content that fits specific contexts.

3. The fraud adversarial dialogue generation and detection method based on large models and multi-agent systems according to claim 2, characterized in that, For fraud-related data, retain the typical characteristics of fraud strategies while increasing diversity by changing details.

4. The fraud adversarial dialogue generation and detection method based on large models and multi-agent systems according to claim 2, characterized in that, The following joint constraint function is introduced during the generation of the dialogue content: ， in The original generation loss of the language model; The dialogue structure constraint loss is used to constrain the sequence of stages in fraudulent rhetoric. The semantic function distribution constraint loss is used to limit the proportion of different fraud semantic function units that appear. , Indicates the weighting coefficient; The semantic function distribution constraint loss is defined as: in This represents the proportion of the k-th semantic function in the generated text. This represents the target distribution obtained based on statistics from real fraud samples.

5. The method for generating and detecting fraudulent adversarial dialogues based on large models and multi-agent systems according to claim 1, characterized in that, Step 2.3 specifically includes: In the multi-agent adversarial framework, each agent is modeled as a finite-state decision-making process, and its set of states is defined as follows: in, This represents the current intent state of the speech; The explicit or implicit emotional state in a conversation; This is the stage in the fraud process. Map semantic states in the state set to continuous risk and schedule metrics: in, As an indicator of the progress of the fraud, it is divided into stages of the fraud process. Mapped to obtain; As a risk indicator for fraud exposure, it is determined by the state of intent in the persuasive statements. With emotional state The joint estimation yielded; and It is a mapping function; The dialogue between the scammer and the victim is modeled as an adversarial game, with its payoff function defined as: in, To advance the return weighting coefficient, This is the risk penalty weighting coefficient; The manager agent monitors the dialogue status in real time and terminates the dialogue generation when one of the following termination conditions is met: in, Number the current dialogue round. To the maximum number of allowed dialogue rounds, This is the dynamic threshold for risk.

6. The method for generating and detecting fraudulent adversarial dialogues based on large models and multi-agent systems according to claim 1, characterized in that, In step 2.4, the ChatTTS model is used for conversion, and different voice parameters are randomly generated to configure unique voice features for different roles, preserve natural voice features, and accurately control the time relationship between the caller and the called party's voice to ensure natural turn-taking.

7. The method for generating and detecting fraudulent adversarial dialogues based on large models and multi-agent systems according to claim 6, characterized in that, The sound parameters include timbre, speech rate, and pitch, and the natural speech features include pauses, stress, and emotional changes, making the synthesized voice closer to a real conversation.

8. The method for generating and detecting fraudulent adversarial dialogues based on large models and multi-agent systems according to claim 1, characterized in that, In step 3.2, the model uses specific symbols to mark its professional thinking process, demonstrating the complete reasoning chain from clue recognition to professional judgment, and marks the output results with specific symbols.

9. The method for generating and detecting fraudulent adversarial dialogues based on large models and multi-agent systems according to claim 1, characterized in that, In step 3.2, the model outputs data in JSON format. The output for call scenario classification includes: scenario type, reason for judgment, and confidence level. The output for fraud judgment includes: reason for judgment, confidence level, and fraud status. The output for fraud type identification includes: fraud type, reason for judgment, and confidence level.

10. A computer system comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the fraud adversarial dialogue generation and detection method based on a large model and multiple agents as described in any one of claims 1-9.

Images