Multi-modal false news detection method and system based on multi-agent confrontation debate

Abstract

The invention discloses a multi-modal false news detection method and system based on multi-agent confrontation debate, and the method comprises the following steps: obtaining news data to be detected, including text data and image data; defining a multi-agent interaction process of which the output is a classification result and an explanatory text report; decomposing a detection task of news to be detected into a plurality of sub-tasks by using the constructed evidence mining agent; integrating observation results after the subtasks are executed into a structured evidence packet; by defining a square dialect and an inverse dialect, according to the structured evidence packet, a multi-stage field self-adaptive confrontation debate is expanded, and a complete debate history is generated; constructing an arbitration module comprising a plurality of review intelligent agents; and the arbitration module generates an explanation report of final judgment according to the complete debate history and a preset multi-dimensional scoring mechanism. According to the method, the multi-modal evidence can be deeply mined, and the method has an antagonistic logical reasoning ability and high interpretability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and network information security technology, and in particular to a hybrid source multimodal fake news detection method, system, and computer-readable storage medium based on Large Vision-Language Models (LVLMs) and Multi-Agent Systems. Background Technology

[0002] With the rapid development of internet technology and social media, multimodal news (including text and images) has become a major form of information dissemination. However, this has also led to the proliferation of fake news, seriously disrupting social order and public perception.

[0003] Traditional fake news detection techniques primarily rely on supervised learning paradigms, which involve extracting lexical features from text or visual features from images through feature engineering for binary classification. However, with the widespread adoption of generative artificial intelligence (AIGC) technology, existing detection techniques face significant challenges.

[0004] First, detecting fake news from mixed sources is difficult: Modern fake news often employs a "half-true, half-false" approach, such as using authentic text but AI-generated images, or using authentic text and images but semantically mismatched text and images. Traditional methods struggle to capture these subtle cross-modal semantic conflicts. Second, large model reasoning is prone to illusions: When using large language models directly for zero-shot detection, the model is susceptible to "illusions," such as fabricating non-existent evidence or exhibiting "compliance bias," tending to conform to the surface logic of news content without in-depth questioning. Finally, there is a lack of interpretability: Existing deep learning models are typically "black box" models, only outputting true / false labels, failing to provide convincing criteria and logical reasoning processes, and thus failing to meet the interpretability requirements of judicial evidence collection or content review. Summary of the Invention

[0005] This invention provides a multimodal fake news detection method and system based on multi-agent adversarial debate, which solves the problems of shallow reasoning, easy generation of illusions, and lack of cross-modal consistency verification capability in the existing technology when dealing with multimodal misleading information from mixed sources. This invention can deeply mine multimodal evidence, has adversarial logical reasoning capability, and has high interpretability.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, a multimodal fake news detection method based on multi-agent adversarial debate includes the following steps:

[0008] S1. Acquire the news data to be tested, including text data and image data; define the multi-agent interaction process with the output as classification results and explanatory text reports;

[0009] S2. Use the constructed evidence mining agent to decompose the detection task of the news to be tested into multiple sub-tasks; integrate the observation results after the execution of the sub-tasks into a structured evidence package;

[0010] S3. By defining affirmative and negative debaters, a multi-stage domain-adaptive adversarial debate is conducted based on a structured evidence package, generating a complete debate history.

[0011] S4. Construct an arbitration module containing multiple review agents; the arbitration module generates an explanation report of the final judgment based on the complete debate history and a preset multi-dimensional scoring mechanism.

[0012] Preferably, S2 includes:

[0013] S2.1 The evidence mining agent decomposes the news detection task into three sub-tasks: text authenticity verification, visual authenticity verification, and cross-modal consistency reasoning.

[0014] S2.2 The evidence mining agent executes each subtask by calling external toolsets through preset query statements and returns each observation result to generate a structured evidence package.

[0015] Preferably, S3 includes:

[0016] S3.1 Infer the domain of the news to be tested; use a large language model to generate character profiles of the affirmative and negative debaters based on the domain.

[0017] S3.2 The affirmative and negative debaters conduct multi-stage debates, storing the debate process of each stage in a shared memory to generate a complete debate history.

[0018] As a preferred option, the external toolset includes search engine tools, visual question-answering tools, and consistency checking tools for checking for inconsistencies between text and images.

[0019] As a preferred approach, multi-stage debates include:

[0020] Phase 1, Opening Statement: Using a generative large language model based on the Transformer architecture, initial viewpoints for both sides are generated based on the news data to be tested, the structured evidence package, and the character profiles.

[0021] Phase Two: Refutation Phase. Both sides present targeted rebuttals to the opponent's initial viewpoints.

[0022] In Phase Three, the free debate phase, if one side raises new questions, the evidence mining agent will again call the search engine tool to conduct a secondary search and obtain new evidence; the new evidence, the statements of the affirmative and negative debaters will be updated in the shared memory.

[0023] Phase Four: Concluding Remarks. Both sides make concluding remarks based on the updated shared memories, forming a complete history of the debate.

[0024] As a preferred approach, the first-stage formula for multi-stage debate is expressed as follows:

[0025]

[0026] in, It represents a sequence of natural language texts that conforms to the logic of a role under given context constraints. It works in an autoregressive manner, concatenating the news data to be tested, the structured evidence package, and the role profile into vector embeddings, which are then input into a multi-layer Transformer decoder. The text response is generated by sampling by calculating the conditional probability distribution of the token under the next context constraint.

[0027] Preferably, S4 includes:

[0028] A multi-dimensional scoring mechanism is constructed; the review agent scores the opposing debater's degree of victory based on the multi-dimensional scoring mechanism; the degree of victory score is compared with a preset judgment threshold to generate classification labels.

[0029] Preferably, S4 also includes: the main review agent of the arbitration module generates a debate summary report using a large language model configured in reviewer mode based on the complete debate history; the output includes an explanation report of the final judgment, which includes classification labels and the debate summary report.

[0030] As a preferred approach, the multi-dimensional scoring mechanism includes five scoring dimensions: factual accuracy, source reliability, reasoning quality, clarity, and cross-modal consistency. For each scoring dimension, the corresponding review agent scores the opposing debater's degree of victory based on the complete debate history. Generate category tags:

[0031]

[0032] in, It is a dimension The weighting coefficients, In dimension The opposing debater's score reflects their degree of victory. It is a preset decision threshold. It is an indicator function; if the weighted total score exceeds the preset decision threshold... If the function value is 1, the news is considered false; otherwise, the function value is 0, and the news is considered true.

[0033] Secondly, a multimodal fake news detection system based on multi-agent adversarial debate includes:

[0034] The multimodal data acquisition module is used to acquire the news text and image data to be detected;

[0035] The evidence mining and task decomposition module is equipped with external tools such as search engine interfaces and visual analysis interfaces to perform text authenticity verification, visual authenticity verification and cross-modal consistency reasoning, and output structured evidence packages.

[0036] The adversarial debate module includes a role generation unit and a dialogue management unit. The role generation unit is used to generate positive and negative role profiles based on the news domain. The dialogue management unit is used to control the multi-stage domain-adaptive adversarial debate process and maintain shared memory and generate a complete debate history.

[0037] The multi-dimensional adjudication module is used to score the complete debate history from multiple dimensions, calculate the final truth or false label based on the weighted score, and generate an explanatory report of the final judgment.

[0038] The multimodal fake news detection system based on multi-agent adversarial debate is used to implement the multimodal fake news detection method and steps based on multi-agent adversarial debate as described in the first aspect.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] Deep reasoning and anti-illusion capabilities: This invention introduces an adversarial debate mechanism (DAAD) to force the model to deeply explore logical loopholes in the game between the two sides, and to conduct real-time verification by introducing an external toolset (evidence mining module MEMD), which significantly reduces the risk of illusion generated by direct reasoning in generative models and avoids "compliance bias".

[0041] Enhanced hybrid source forgery detection capability: This invention, through explicit task decomposition, specifically sets up visual authenticity verification and cross-modal consistency reasoning subtasks, which can effectively identify complex hybrid source forgery methods such as text forgery, image forgery, and inconsistencies between text and images, thus solving the blind spots of single-modal detection.

[0042] High interpretability: This invention not only outputs true / false labels, but also a shared memory containing the complete debate process and a multi-dimensional adjudication report. Users can clearly see which evidence the judgment is based on and what logical deductions were made, greatly improving the credibility of the detection system and its value in judicial evidence collection.

[0043] Domain Adaptability: This invention dynamically infers the news domain and generates specific role profiles (such as medical experts and political commentators), enabling the detection system to adapt to the context and professional knowledge requirements of different domains, thereby improving the model's generalization ability in specific vertical news detection. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;

[0045] Figure 2 This is a detailed flowchart of the multimodal evidence mining and task decomposition module of Embodiment 1 of the present invention;

[0046] Figure 3 This is a five-stage debate sequence diagram of the domain-adaptive adversarial debate module of Embodiment 1 of the present invention;

[0047] Figure 4 This is a schematic diagram of the scoring logic of the multi-dimensional adjudication and arbitration module in Embodiment 1 of the present invention;

[0048] Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to Embodiment 3 of the present invention. Detailed Implementation

[0049] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments. 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.

[0050] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0051] Example 1:

[0052] like Figure 1 The multimodal fake news detection method based on multi-agent adversarial debate, as shown, includes the following steps:

[0053] S1. Acquire the news data to be tested, including text data and image data; define the multi-agent interaction process with the output as classification results and explanatory text reports;

[0054] The system receives the news data to be tested. Let the multimodal news sample of the input news to be tested be... . The text portion representing the news, such as "Breaking News: Leader of a Country Collapses in Public Due to Heart Attack," includes the headline and body text. The image portion representing the news, such as a blurry photo showing the leader lying on the ground.

[0055] The goal of this invention is to establish a mapping function. Output the authenticity tag for this news. and explanatory reports To enable deep reasoning, the mapping function is modeled as a multi-agent interaction process. Its mathematical expression is:

[0056]

[0057] In this formula: This is the final binary classification result. In this embodiment, 0 represents "Real News" and 1 represents "Fake News". It is a generated explanatory text report containing the reasons for the judgment. This refers to an external knowledge base, which is a dynamic collection of knowledge obtained in real time by external tools in subsequent steps from the Internet or specific databases, rather than static knowledge fixed during model pre-training. It represents the interaction process of the entire multi-agent system, covering the entire process from evidence collection to debate to adjudication.

[0058] S2. Use the constructed evidence mining agent to decompose the detection task of the news to be tested into multiple sub-tasks; integrate the observation results after the execution of the sub-tasks into a structured evidence package;

[0059] To address forgery from mixed sources, including cases where the image is genuine but the text is fake, the image is fake but the text is genuine, and the image and text do not match, this step introduces a specialized evidence mining agent (…). This intelligent agent does not directly judge whether something is true or false, but is responsible for collecting evidence.

[0060] S2.1 The evidence mining agent decomposes the news detection task into three sub-tasks: text authenticity verification, visual authenticity verification, and cross-modal consistency reasoning.

[0061] like Figure 2 As shown, the evidence mining agent first decomposes the complex detection task into three orthogonal (independent) subtasks:

[0062] Subtask 1 (Branch A): Text authenticity verification Focus on verifying text Whether the facts stated in the report (such as time, place, people, and events) are consistent with authoritative media reports or official statements.

[0063] Subtask 2 (Branch B): Visual Authenticity Verification (Focusing on image detection) Does the device itself exhibit signs of forgery? For example, it can use large visual models to detect whether there are AI-generated artifacts (such as distorted fingers or inconsistent lighting) or signs of Photoshop manipulation.

[0064] Subtask 3 (Branch C): Cross-modal Consistent Reasoning (This section focuses on analyzing the text.) With images Semantic relationships. For example, detecting whether the image content truly supports the text title, or whether there is a case of "copying" (such as using a real old photo to accompany a new fake news article).

[0065] S2.2 The evidence mining agent executes each subtask by calling external toolsets through preset query statements and returns each observation result to generate a structured evidence package;

[0066] For each of the above sub-tasks ( Values One of them), the evidence mining agent generates specific query statements. and call external toolsets The corresponding tool in [the document / platform]. The execution process is represented as follows:

[0067]

[0068] In this formula: This represents the execution operation of a function call by a tool. It is a collection of available tools, including . (Search engine tools): such as Google Search API or Bing Search API. (Visual Question Answering Tool): A visual analysis interface based on models such as CLIP, BLIP, or GPT-4V. (Consistency check tool): A tool used to calculate the similarity between images and text or to generate contradictory descriptions of images and text. These are natural language queries or API parameters generated for subtasks. It is the observation returned by the tool. Specific implementation examples:

[0070] For text authenticity verification Evidence mining agent generates queries =Official response from a country's leader to a heart attack, quoting The returned observation results This may include: "No mainstream media reports on the incident were found; the official account recently posted tweets about the individual attending events normally." (Regarding visual authenticity verification...) Intelligent agent generation instructions ="Detect AI generation artifacts in this image", call The returned observation results Possible explanation: "Logical errors detected in hand details; image exhibits high smoothness, suspected to be generated by Midjourney." This is relevant for cross-modal consistency inference. Intelligent agent generation instructions =="Does the image content support the text claim?". The returned observation results. The image content shows a scene of someone falling to the ground, which is consistent with the semantics of the text, but the authenticity of the image is questionable.

[0071] Finally, this step outputs a structured evidence package. As material for subsequent confrontational debates

[0072] S3. By defining the affirmative and negative debaters, a multi-stage domain adaptive adversarial debate (DAAD) is carried out based on the structured evidence package to generate a complete debate history;

[0073] Uncover deep-seated logical vulnerabilities through adversarial interactions.

[0074] S3.1 Define the affirmative and negative debaters; infer the field to which the news to be tested belongs; use a large language model to generate the role profiles of the affirmative and negative debaters based on the field;

[0075] like Figure 3 As shown, firstly, the system uses a lightweight LLM to analyze the news to be tested. Infer its field (Such as politics, healthcare, finance, technology, etc.). Next, the system is based on the domain... Two debaters are generated (affirmative side) and the opposing side Character portrait The formula for generation is:

[0076]

[0077] in, It is a generative large language model used to generate text. These are preset prompt templates used to instruct the model to generate character biographies. It is an inferred field, such as "healthcare". This refers to the characters' stances. The affirmative side is set to "support the news being true" and attempts to explain away the doubts. The negative side is set to "believe the news is false" and attempts to attack the doubts. This is the generated character description text. For example, in the medical field, the opposing character profile might be: "You are a meticulous medical doctor and fact checker, skilled at identifying pseudoscience rumors, and accustomed to refuting viewpoints with data from authoritative medical journals."

[0078] S3.2 The affirmative and negative debaters conduct multi-stage debates, storing the debate process of each stage in a shared memory to generate a complete debate history;

[0079] The debate process is divided into multiple time steps. The system maintains a shared memory. To record the history of debate.

[0080] Mt's technical implementation involves maintaining a structured list of messages. Specifically, within the multi-agent system's GPU memory or system memory, the system maintains a globally visible list of string sequences or JSON objects. The operation technically corresponds to context concatenation. At the end of each round of debate, the system serializes the generated text R and new evidence NewEvidence and appends them to this list. In the next round of reasoning, this complete list Mt serves as the new Prompt and context input for all agents, enabling both sides to 'share' and 'remember' historical debate information without requiring additional physical storage; instead, it utilizes an LLM-based context memory mechanism.

[0081] Phase One: Opening Statement Both sides based their arguments on evidence packages. Express your initial opinion.

[0082]

[0083] in: This represents the conditional probability generation function of generative large language models based on the Transformer architecture (such as the Llama series, GPT series, etc.). In terms of specific technical implementation, The function itself is not an abstract rule, but rather an existing, mature technology. This function works through an autoregressive mechanism, taking the input news content as its basis. Evidence package and character portraits The concatenation is transformed into vector embeddings, which are then input into a multi-layer Transformer decoder to calculate the conditional probability distribution of the token under the next given context constraints. Use it to sample and generate text responses. This represents a sequence of natural language text generated by the model, given the aforementioned context constraints, that conforms to the logic of the roles. For example, the affirmative side might cite... The opposing side claims the images and text match; the opposing side will cite... and It was pointed out that there was no official report and the image appeared to be fake.

[0084] Phase Two: Refutation Phase Both sides must provide targeted rebuttals to the opponent's initial statements and must not speak unilaterally.

[0085] Phase Three: Free Debate and Dynamic Evidence Acquisition This is the key innovation of this invention. During the debate, if one party raises a new question (e.g., "This image might be a still from a 2015 movie"), the intelligent agent has the right to invoke the tool again. Conduct a "secondary search" to obtain new evidence. .

[0086] The formula for updating shared memory is:

[0087]

[0088] in: It is the context of the debate at the previous moment. This indicates a string concatenation or vector concatenation operation, which concatenates the statements of the affirmative side in this round. Opposing side's remarks And newly unearthed evidence This is added to memory. This mechanism ensures that the debate is dynamic and in-depth, rather than based on the repetition of static information.

[0089] Phase Four: Summary and Presentation After the debate, both sides based on To summarize and present the final, complete history of the debate. .

[0090] S4. Multidimensional Adjudication and Arbitration (MDA)

[0091] To avoid bias in decisions made by a single model, this step introduces a "jury" composed of multiple judge agents, such as... Figure 4 As shown.

[0092] (1) Generate a debate summary

[0093] First, using system prompts pre-defined in the LLM (Leadership Management System), a lead judge generates a debate summary report that is as neutral as possible. :

[0094]

[0095] In this formula: It is a large language model configured for reviewer mode. It is a summary instruction that requires the model to summarize the core arguments and strength of evidence from both sides. It is the complete debate history output by step S3.

[0096] (2) Calculate authenticity labels

[0097] The judging panel employs a multi-dimensional scoring mechanism, rather than directly outputting true or false. Five pre-defined scoring dimensions are used. For each dimension The reviewers based on Score the degree of victory for the opposing side (the questioning side). The score range is, for example, 0-7 points. The higher the score, the more valid the opposing argument is, meaning the more fake the news is.

[0098] The final determination is calculated using the following formula:

[0099]

[0100] In this formula: It is a dimension The weighting coefficients. For example, "factual" and "cross-modal consistency" are usually given higher weights. In dimension The opposing team's score. It is a preset decision threshold. It is an indicator function. If the weighted total score exceeds a threshold... If the function value is 1, the news is considered fake; otherwise, the function value is 0, and the news is considered real. Using the above formula, the system ultimately outputs category labels. and including Explanation report.

[0101] Example 2: A multimodal fake news detection system based on multi-agent adversarial debate, comprising:

[0102] Multimodal data acquisition module: used to acquire news text and image data to be detected.

[0103] Evidence mining and task decomposition module: It is equipped with external tools such as search engine interface and visual analysis interface to perform text verification, visual verification and cross-modal consistency analysis, and output structured evidence packages.

[0104] The adversarial debate module includes a character generation unit and a dialogue management unit. The character generation unit generates character profiles for both sides based on the news subject; the dialogue management unit controls the flow of the opening statement, rebuttal, free debate, and concluding statement, and maintains shared memory.

[0105] Multidimensional adjudication module: used to score the debate history from multiple dimensions, calculate the final true / false label based on the weighted score, and generate an explanatory report.

[0106] Example 3: A computer-readable storage medium

[0107] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1. The storage medium may include a read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0108] Example 4: An electronic device

[0109] This embodiment provides an electronic device, such as... Figure 5 As shown, the device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the method described in Embodiment 1. This electronic device can be a server, workstation, or high-performance personal computer; in particular, it is typically equipped with a GPU accelerator card to support inference operations on large visual language models.

Claims

1. A multimodal fake news detection method based on multi-agent adversarial debate, characterized in that, Includes the following steps: S1. Acquire the news data to be tested, including text data and image data; define the multi-agent interaction process with the output as classification results and explanatory text reports; S2. Use the constructed evidence mining agent to decompose the detection task of the news to be tested into multiple sub-tasks; The observation results after the sub-tasks are executed are integrated into a structured evidence package; S3. By defining affirmative and negative debaters, a multi-stage domain-adaptive adversarial debate is conducted based on a structured evidence package, generating a complete debate history. S4. Construct an arbitration module that includes multiple review agents; The arbitration module generates an explanatory report of the final judgment based on the complete debate history and a preset multi-dimensional scoring mechanism.

2. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 1, characterized in that, S2 include: S2.1 The evidence mining agent decomposes the news detection task into three sub-tasks: text authenticity verification, visual authenticity verification, and cross-modal consistency reasoning. S2.2 The evidence mining agent executes each subtask by calling external toolsets through preset query statements and returns each observation result to generate a structured evidence package.

3. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 1, characterized in that, S3 include: S3.1 Determine the field to which the news to be tested belongs; Using a large language model, generate character profiles of the affirmative and negative debaters based on their respective fields; S3.2 The affirmative and negative debaters conduct multi-stage debates, storing the debate process of each stage in a shared memory to generate a complete debate history.

4. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 2, characterized in that, External toolsets include search engine tools, visual question-answering tools, and consistency checking tools for checking for inconsistencies between text and images.

5. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 3, characterized in that, The multi-stage debate includes: Phase 1, Opening Statement: Using a generative large language model based on the Transformer architecture, initial viewpoints for both sides are generated based on the news data to be tested, the structured evidence package, and the character profiles. Phase Two: Refutation Phase. Both sides present targeted rebuttals to the opponent's initial viewpoints. In Phase Three, the free debate phase, if one side raises new questions, the evidence mining agent will again call the search engine tool to conduct a secondary search and obtain new evidence; the new evidence, the statements of the affirmative and negative debaters will be updated in the shared memory. Phase Four: Concluding Remarks. Both sides make concluding remarks based on the updated shared memories, forming a complete history of the debate.

6. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 5, characterized in that, The first-stage formula for multi-stage debate is expressed as follows: ； in, It represents a sequence of natural language texts that conforms to the logic of a role under given context constraints. It works in an autoregressive manner, concatenating the news data to be tested, the structured evidence package, and the role profile into vector embeddings, which are then input into a multi-layer Transformer decoder. The text response is generated by sampling by calculating the conditional probability distribution of the token under the next context constraint.

7. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 1, characterized in that, S4 include: A multi-dimensional scoring mechanism is constructed; the review agent scores the opposing debater's degree of victory based on the multi-dimensional scoring mechanism; the degree of victory score is compared with a preset judgment threshold to generate classification labels.

8. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 7, characterized in that, S4 also includes: the main reviewer agent of the arbitration module generates a debate summary report using a large language model configured in reviewer mode based on the complete debate history; the output includes an explanation report of the final judgment, which includes classification labels and the debate summary report.

9. The multimodal fake news detection method based on multi-agent adversarial debate according to claim 7, characterized in that, The multi-dimensional scoring mechanism includes five scoring dimensions: factual accuracy, source reliability, reasoning quality, clarity, and cross-modal consistency. For each scoring dimension, the corresponding review agent scores the opposing debater's degree of victory based on the complete debate history. Generate category tags: ； in, It is a dimension The weighting coefficients, In dimension The opposing debater's score reflects their degree of victory. It is a preset decision threshold. It is an indicator function; if the weighted total score exceeds the preset decision threshold... If the function value is 1, the news is considered false; otherwise, the function value is 0, and the news is considered true.

10. A multimodal fake news detection system based on multi-agent adversarial debate, characterized in that, include: The multimodal data acquisition module is used to acquire the news text and image data to be detected; The evidence mining and task decomposition module is equipped with external tools such as search engine interfaces and visual analysis interfaces to perform text authenticity verification, visual authenticity verification and cross-modal consistency reasoning, and output structured evidence packages. The adversarial debate module includes a role generation unit and a dialogue management unit; the role generation unit is used to generate positive and negative role profiles based on the news field. The dialogue management unit is used to control the multi-stage domain-adaptive adversarial debate process and maintain shared memory and generate a complete debate history; The multi-dimensional adjudication module is used to score the complete debate history from multiple dimensions, calculate the final truth or false label based on the weighted score, and generate an explanatory report of the final judgment. The multimodal fake news detection system based on multi-agent adversarial debate is used to implement the multimodal fake news detection method and steps based on multi-agent adversarial debate as described in claim 1.

Images