False information detection method and device based on multi-agent view angle aggregation and storage medium

Abstract

The invention relates to the technical field of information content security and intelligent detection, and discloses a false information detection method and device based on multi-agent view aggregation and a storage medium, and the method comprises the steps: obtaining more than one network information instance, and constructing a multi-view feature set for each network information instance; constructing a multi-agent comprising a plurality of auditing agents, a plurality of coordination agents and a decision-making agent; the plurality of auditing agents are used for analyzing the multi-view feature set according to a preset view to obtain auditing judgments of a plurality of preset views, and the plurality of coordination agents are used for performing view perception aggregation on the auditing judgments according to the preset views to generate a plurality of intermediate coordination judgments; the decision-making agent is used for generating a judgment result according to the plurality of intermediate coordination judgments; the method has the advantages of being accurate in false information detection and high in application efficiency.

Description

Technical Field

[0001] This invention relates to the field of information content security and intelligent detection technology, specifically to a method, device, and storage medium for detecting false information based on the aggregation of multiple agent perspectives. Background Technology

[0002] With the rapid development of social media and online content platforms, user-generated content is characterized by its massive scale, rapid dissemination, and diverse forms. At the same time, misinformation is also generated and spread on these platforms in various forms, such as the manipulation of abnormal or fake accounts, misleading product reviews, and fabricated or altered news content. This type of misinformation not only misleads public perception but also damages the platform's content ecosystem and trust foundation, causing significant negative impacts in e-commerce transactions, public events, and the dissemination of public opinion.

[0003] To curb the spread of misinformation, platforms initially relied primarily on manual review for identification and handling. However, with the exponential growth of content volume, manual review alone could no longer meet the requirements of both real-time performance and comprehensive coverage. Therefore, academia and industry have proposed various automated detection methods based on statistical learning and deep learning to identify misinformation and improve review efficiency. However, existing deep learning models often suffer from opaque decision-making processes and insufficient interpretability, requiring manual verification of large numbers of samples in high-risk scenarios. This increases the review burden and introduces risks of harmful content remaining undetected and missed detections.

[0004] In recent years, large language models, with their enhanced contextual understanding and reasoning capabilities, have been gradually introduced into content moderation and misinformation detection processes, aiming to provide more interpretable judgment criteria. Building on this, large language model-driven multi-agent systems are considered capable of improving performance on complex tasks through collaborative reasoning and collective intelligence. Multi-agent solutions, in form, can aggregate multiple perspectives, providing a new technical path for misinformation detection.

[0005] Current methods for detecting misinformation mainly include: Content-based detection models: These utilize text semantics, sentiment, writing style, and other content features for binary classification, often relying on large amounts of labeled data for training; their ability to retain, amplify, and interpret fine-grained anomalous clues is limited. Social context / interaction structure-based detection models: These rely on structured signals such as comments, reposts, user relationships, and propagation networks; while they can improve performance in some scenarios, they introduce additional data acquisition and modeling costs; their performance is unstable under conditions of missing interaction information or cold start. Single-agent or multi-agent detection models based on large language models: These possess advantages in reasoning and interpretation, but in misinformation detection, they are easily limited by information overload, homogenized viewpoints, and high-cost collaboration; a systematic solution that can simultaneously amplify anomalous clues, retain diverse viewpoints, aggregate structured data, and adaptive efficiency optimization is still lacking.

[0006] More specifically, the technical problems existing in using large language model multi-agent applications for fake information detection include: (1) Information Overload Problem: False information cues are usually sparse, fragmented, and weak, while true information patterns are richer and more dominant. When existing multi-agent systems adopt a design where "each agent has full access to the entire input context" to pursue comprehensive analysis, agents are more easily swayed by dominant normal signals, resulting in "overly benign" judgments and ignoring subtle but crucial anomalous evidence. Furthermore, information exchange between agents may amplify their related judgment biases, suppressing a few but high-risk anomalous pieces of evidence during interactions, ultimately leading to a deeper concealment of false information.

[0007] (2) Lack of a multi-agent structure with "diverse perspectives - structured aggregation": False information is multifaceted and highly context-dependent, and it is difficult to capture its performance stably from a single perspective. Existing multi-agent topologies mostly focus on communication structures (such as fully connected, hierarchical, or temporal mutual evaluation), but they often lack mechanisms at the agent level to explicitly constrain "different agents to observe only different feature subsets / different perspectives", and also lack aggregation strategies at the system level to hierarchically, faithfully, and traceably aggregate multi-perspective outputs, which can easily lead to homogenization of viewpoints and cause abnormal signals to gradually disappear during the aggregation process. Summary of the Invention

[0008] The purpose of this invention is to provide a method, apparatus and storage medium for detecting false information based on multi-agent perspective aggregation, so as to solve the technical problems of information overload and lack of structured aggregation of perspective diversity when using existing large language model multi-agents for false information detection.

[0009] The technical solution adopted by this invention to solve its technical problem is: a method for detecting false information based on multi-agent perspective aggregation, comprising the following steps: S1. Obtain one or more network information instances and construct a multi-view feature set for each network information instance; S2. Construct a multi-agent system comprising several audit agents, several coordination agents, and one decision agent; the audit agents are respectively used to analyze the feature set of the multi-viewpoint according to a preset perspective to obtain several preset perspective audit judgments; the coordination agents are respectively used to perform perspective perception aggregation on the audit judgments according to the preset perspective to generate several intermediate coordination judgments; and the decision agent is used to generate a judgment result based on several intermediate coordination judgments. S3. Train multiple agents based on multi-view feature sets of one or more network information instances; S4. The trained multi-agent system judges the test information instances to obtain the judgment result.

[0010] The significant advantages of this invention are as follows: In this solution, the auditing agent analyzes and detects the input information through a preset perspective. This allows the auditing agent to analyze only specific content of the input information, avoiding the situation where the auditing agent performs a comprehensive analysis of the entire context of the input information, which could lead to the analysis results being influenced by dominant normal signals and resulting in overly benign judgments. This ensures that the auditing agent can accurately capture key abnormal evidence in the input information during the analysis process. At the same time, by coordinating the agent and the decision-making agent to perform multi-perspective perception aggregation in sequence, the final output judgment result is a multi-perspective analysis and detection of the input information, improving the accuracy of false information detection.

[0011] Furthermore, step S1 also includes: decomposing the multi-perspective feature set of each network information instance according to a preset perspective to obtain several feature subsets, wherein the feature subsets cover complementary clue dimensions, and the audit agent generates an audit judgment based on the preset perspective according to the feature subsets.

[0012] Furthermore, step S2 also includes initializing the multi-agent system: Each audit agent is configured with an audit role profile, audit experience memory, and audit action memory; wherein, the audit role profile is used to constrain the preset perspective of the audit agent's input and audit judgment output, the audit experience memory is used to store the audit agent's identification rules, and the audit action memory is used to record the audit agent's historical audit judgments; Each coordinating agent is configured with a coordinating role profile, a coordinating confidence memory, and a coordinating action memory. The coordinating role profile stores the agent's role attributes and behavioral characteristics; the coordinating confidence memory stores the agent's experiential confidence information; and the coordinating action memory records the agent's historical intermediate coordination decisions. The coordinating agent supervises a subset of auditing agents comprised of multiple auditing agents. The decision-making agent is configured with a decision role profile, decision experience memory, decision confidence memory, and decision action memory. The decision role profile is used to store the decision-making agent's role attributes and decision characteristics. The decision experience memory is used to store the decision-making agent's experience memory information. The decision confidence memory is used to store the decision-making agent's confidence memory information. The decision action memory is used to store the decision-making agent's historical judgment results.

[0013] By setting role profiles for each agent, when analyzing input information, each agent can be constrained in its perspective and characteristics when processing information based on its role profile. This ensures that each agent can only analyze and process information from a preset perspective or rule, suppressing information overload and homogenization of viewpoints caused by the full context. Furthermore, by setting experience memory, confidence memory, or action memory for each agent to record the content of each reasoning, traceable explanation can be achieved. Furthermore, in step S3, adaptive optimization of the multi-agent system is also performed: A1. Calculate the overall score of each audit agent and tailor the audit agents according to the overall score; A2. Update the audit agent, coordination agent, or decision agent that made the incorrect judgment.

[0014] The beneficial effects of the above technical solution are as follows: By introducing an adaptive optimization mechanism during the training phase, the contribution of the agents is evaluated and pruned, and the audit agents, coordination agents or decision agents that make incorrect judgments are updated in a targeted manner, so that the multi-agent structure can be dynamically adjusted according to the difficulty of the samples and the risk mode, reducing redundant reasoning and information stacking, suppressing the homogenization of opinions, and improving the accuracy, robustness and reasoning efficiency of false information detection.

[0015] Further, in step A1, a comprehensive score is calculated for each audit agent. If the comprehensive score of an audit agent is less than a score threshold, the audit agent is removed. The formula for calculating the comprehensive score of an audit agent is:

[0016] In the formula, For the first The overall score of each auditing agent. To improve the accuracy of current multi-agent architectures, To coordinate the set of audit agents under the intelligent agent, The penalty coefficient is... For the first Each agent's prediction vector and the Each agent's prediction vector cosine similarity, To coordinate the removal of subordinates of the intelligent agent A collection of intelligent agents for auditing. To coordinate the number of audit agents under the intelligent agent.

[0017] The beneficial effects of the above technical solution are as follows: by using a comprehensive score to quantitatively evaluate the agent, it is possible to simultaneously characterize the marginal contribution of a single agent to the overall accuracy and the degree of redundancy of its output with other audit agents, thereby automatically removing low-contribution or highly homogeneous audit agents, reducing the risk of information overload and waste of computing resources, and improving the efficiency and scalability of multi-agent collaborative reasoning while ensuring detection performance.

[0018] Furthermore, in step A2, the formula for updating audit experience memory or decision-making experience memory is:

[0019] In the formula, For the purpose of memorizing updated audit experience, For the current multi-view features to be detected, To create a profile of the audit role, To memorize auditing experience prior to the update, For the purpose of recording the audit operation, For the audit decision made by the auditing agent in the current inference round, This is a summary of the reasoning process and decision-making basis related to the audit decision. For final supervision labeling, The operator is updated based on self-reflection based on experience; The formula for updating decision experience memory is:

[0020] In the formula, For the updated memory of decision-making experience, To create a profile of decision-making roles, To remember the decision-making experience before the update, For decision-making and action memory, The decision result formed by the decision-making agent in the current inference round. A summary of the reasoning process and decision-making basis related to the judgment result; The formula for coordinating confidence memory update is:

[0021] In the formula, For the updated reconciliation confidence memory, To trust the self-reflective update operator, For the reconciled confidence memory before the update, To coordinate action memory, To coordinate character portraits, To coordinate the intermediate agreement decisions formed by the agents in the current inference round, A summary of the reasoning process and decision-making basis related to the intermediate agreement determination; The formula for calculating decision confidence memory update is:

[0022] In the formula, For the updated decision confidence memory, To trust the self-reflective update operator, Confidence memory of decisions made before the update To coordinate action memory.

[0023] The beneficial effects of the above technical solution are as follows: based on the supervision label, the self-reflective update is triggered for the erroneous sample, so as to realize the continuous correction of the erroneous rules and unreliable judgment mode and the accumulation of reliable experience. This enables the intelligent agent to dynamically calibrate the credibility and decision basis organization under different perspectives, reduce the recurrence of misjudgment, improve the generalization ability and stability of the system under different scenarios and different attack types, and enhance the traceability and auditability of the judgment basis.

[0024] Furthermore, in step S4, when determining the instance of the information to be tested, confidence-guided routing is performed, and the decision agent selects a coordinating agent for evaluation according to the confidence weight: When one or more audit agents and one or more coordinating agents activated by the confidence-guided routing have reached a consensus, and the margin of the consensus reaches the margin threshold, output the judgment result and terminate the inference. Otherwise, the remaining audit agents and coordinating agents are activated in descending order of confidence level, and the current judgment conclusion and margin are recalculated until the margin reaches the margin threshold or the audit agents and coordinating agents reach the preset maximum number of activations, at which point the judgment result is output.

[0025] The beneficial effects of the above technical solution are as follows: By guiding the routing with confidence, a progressive activation and early termination mechanism is achieved. When the conclusions of the activated agents are consistent and the margin reaches the threshold, the result can be directly output and the inference can be terminated, thereby reducing the number of inference rounds and computational overhead while ensuring the reliability of the judgment. When the consistency is insufficient, the remaining agents are activated step by step according to the confidence level from high to low for supplementary evaluation. This can prioritize the allocation of computing resources to high-confidence perspectives and suppress the introduction of low-value information, reduce information overload and inference redundancy, and improve the real-time performance of online detection and the stability of the final judgment.

[0026] Furthermore, the formula for calculating confidence-guided routing is:

[0027] In the formula, To determine the result, The reasons for the judgment are related to the judgment result. For intelligent agents The test results For intelligent agents The reason for the test, The set of currently active agents. For inference operators that take confidence weights into account, To coordinate the confidence weights in confidence memory, This is the margin threshold.

[0028] A fake information detection device based on multi-agent perspective aggregation includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the fake information detection method based on multi-agent perspective aggregation described above.

[0029] A computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, and when the computer program is executed by one or more processors, implements the above-described method for detecting false information based on multi-agent perspective aggregation. Attached Figure Description

[0030] Figure 1 This is a flowchart of the multi-agent false information detection method in an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the multi-agent structure in an embodiment of the present invention.

[0032] Figure 3 This is a graph showing the performance comparison between the multi-agent system and the baseline in an embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of the ablation experiment results in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing token consumption in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0035] See appendix Figure 1 The fake information detection method based on multi-agent perspective aggregation shown includes the following steps: S1. Obtain multiple network information instances and construct a multi-perspective feature set for each network information instance. These network information instances include user behavior sequences, product review text, news articles, and their metadata awaiting detection. Specifically, in this embodiment, network information instances are obtained from publicly available benchmark data from Amazon, DeRev2018, and PolitiFact as the network information to be detected. Amazon is geared towards e-commerce scenarios and is used to identify abnormal or fake user accounts involved in manipulative behaviors such as fake reviews, spam posting, or collaborative injection. DeRev2018 is a product review dataset with authenticity annotations, used to detect deceptive or misleading reviews. PolitiFact is used for detecting fake news or political misinformation. Furthermore, the network information to be detected obtained from each benchmark dataset includes both real and fake data. The quantity and distribution of each dataset are shown in Table 1. Table 1 Dataset Distribution Information

[0036] When constructing a multi-perspective feature set, network information instances are transformed into structured information that can be understood by a large language model. Specifically, this involves acquiring statistical features, textual semantic features, time-series features, and propagation source features from the network information instances. Statistical information includes the rating distribution, frequency, and length of the network information instances. Textual semantic features include the emotional polarity and intensity, subjectivity, inconsistencies, and template reuse of the network information instances. Time-series features include the distribution characteristics of the network information instances over time, such as whether they are sudden, periodic, or drifting. Propagation source features include the forwarding links and source credibility of the network information instances. The statistical features, textual semantic features, time-series features, and propagation source features of the network information instances are combined to obtain the complete multi-perspective feature set. The specific methods for acquiring each feature of the network information instances are existing technologies and will not be elaborated upon here.

[0037] In step S1, each network information instance in the multi-view feature set is further divided into several feature subsets. Each feature subset corresponds to the input constraints of an audit agent. The feature subsets cover complementary cue dimensions. Specific feature subsets can be decomposed by focusing only on score statistical anomalies, focusing only on text sentiment extremism, focusing only on cross-comment consistency or repeated templates, focusing only on time mutations, etc., so as to avoid all agents being exposed to the full amount of information at the same time, resulting in attention concentration and information overload.

[0038] S2. Construct a multi-agent system comprising several audit agents, several coordination agents, and decision-making agents, and initialize the multi-agent system. The agents can be implemented by combining the reasoning module of a natural language processing model or a generative model with prompt templates; as shown in the appendix. Figure 2As shown, several audit agents are located at the bottom layer of the multi-agent system. The audit agents are configured to perform independent reasoning analysis on the feature set of multiple perspectives based on a preset analysis perspective, and output the corresponding audit judgment and its confidence level.

[0039] Several coordinating agents reside in the intermediate layer of the multi-agent system. These coordinating agents are configured to receive audit results and their confidence levels from several auditing agents, perform consistency assessments and weight adjustments on the audit results, and generate aggregated results for subsequent judgments. Each coordinating agent supervises a subset of auditing agents comprised of multiple auditing agents; this subset is the set of subordinate auditing agents of the coordinating agent. When the number of auditing agents is large or when perspective dimensions require grouping, the coordinating agents can be hierarchically configured. The lower-level coordinating agents aggregate the audit judgments, and the upper-level coordinating agents perform a secondary aggregation of the intermediate coordinating judgments output by the lower-level agents. Finally, the decision agent outputs the final result. The specific details of this hierarchical configuration are existing technologies and will not be elaborated upon here.

[0040] The decision-making agent sits at the top level of the multi-agent system. It is configured to perform comprehensive judgments based on the aggregated results generated by the coordinating agent and output the final detection conclusion for the information to be detected. In other words, it supervises the highest-level coordinating agent and outputs the final result.

[0041] The initialization of multi-agent systems specifically includes: For each audit agent, an audit role profile, audit experience memory, and audit action memory are set based on a feature subset. The audit role profile is used to constrain the feature subset input by the audit agent and to constrain the audit agent to only look for abnormal evidence and provide reasons from the preset perspective corresponding to the feature subset. The audit experience memory is used to store generalizable audit agent identification rules, such as the identification rule that "emotionally strong language is more likely to be manipulative behavior". The audit action memory is used to record the specific judgment results and judgment reasons of the audit agent on historical samples, so as to facilitate the traceability and subsequent correction of audit judgments. The initial state of the audit action memory is empty.

[0042] For each coordinating agent, a coordinating role profile, a coordinating confidence memory, and a coordinating action memory are set according to a feature subset, enabling it to make judgments and generate evidence solely based on its corresponding perspective. The coordinating role profile characterizes the coordinating agent's role attributes and behavioral features, including at least the corresponding analytical perspective, judgment preferences, capability constraints, or historical performance statistics. The coordinating role profile is generated according to preset rules during the initialization phase and then updated in subsequent optimization phases. The coordinating confidence memory stores experiential confidence information related to the coordinating agent, i.e., the confidence weight for each subordinate auditing agent. The initial confidence weight for the auditing agent is 1, and it is updated in subsequent optimization phases. The coordinating action memory records the specific judgment results and reasons for the coordinating agent on historical samples, facilitating the tracing and subsequent correction of intermediate judgments. The initial state of the coordinating action memory is empty.

[0043] The system sets up decision-making agent profiles, decision experience memories, decision confidence memories, and decision action memories. The decision role profile characterizes the agent's role attributes and decision-making characteristics, including at least full feature configurations, fusion strategies, and historical decision performance statistics. This profile is generated during system initialization according to preset rules and updated during subsequent optimization phases. The decision experience memory stores relevant experience information, including at least intermediate coordination judgment experiences, decision results, and feedback information formed during historical reasoning. This memory supports subsequent decision reasoning and experience reuse. It is generated during initialization and updated during optimization. The decision confidence memory stores the agent's confidence information, specifically the confidence weights of the agent's assessment of the outputs of each coordinating agent. The initial confidence weights for the agent's assessment of coordinating agents are all 1. The decision action memory stores the agent's action patterns or decision behaviors during historical reasoning, facilitating decision path tracing and subsequent correction. This memory is empty during initialization and updated during optimization.

[0044] S3. Train and test the multi-agent system based on the multi-view feature set; specifically including the following steps: S301. Divide the multi-view feature set of multiple network information instances into a test set and a training set; both the test set and the training set adopt a binary classification setting, that is, both the test set and the training set contain real information and false information, and each data is labeled as real information or false information. Both the test set and the training set contain multi-view feature sets obtained from Amazon, DeRev2018 and PolitiFact data; the training set is used to train the multi-agent, and the test set is used to test the trained multi-agent; the specific content of dividing the data in the multi-view feature set into the test set and the training set is existing technology and will not be described in detail here.

[0045] S302. The audit agent generates audit judgments based on the multi-perspective feature set in the training set through audit experience memory and audit action memory. Multiple audit agents operate in parallel, and each agent reads a feature subset corresponding to the audit role profile. Based on audit experience memory and audit action memory, it generates a false information judgment result and a textual explanation of the judgment basis. Simultaneously, the output judgment result and the textual explanation of the judgment basis are saved to the audit action memory for audit traceability. The expression for the audit judgment is:

[0046] In the formula, For information audit judgment, among which 1 indicates false information (or abnormal / fraudulent / misleading content), and 0 indicates true information (or normal content). The textual interpretation that forms the basis for audit decisions. For auditing intelligent agents, To audit the reasoning operations of intelligent agents, For the current multi-view features to be detected, To create a profile of the audit role, For the purpose of memorizing auditing experience, For assignment operations, the output result calculated by the function (operator) on the right is assigned to the variable on the left to update the memory or state represented by the left. S303. The coordinating agent generates intermediate coordination decisions based on the audit decisions of the auditing agents. Specifically, each coordinating agent collects the audit decisions output by the auditing agents in its subordinate set, analyzes the audit decisions based on the coordination confidence memory and coordination action memory to obtain intermediate coordination decisions, and saves the output intermediate coordination decision results and the textual explanations of the basis for the intermediate coordination decisions to the audit action memory. The expression for the intermediate coordination decision is as follows:

[0047] In the formula, For intermediate coordination and determination, among which 1 indicates false information (or abnormal / fraudulent / misleading content), and 0 indicates true information (or normal content). The textual interpretation that serves as the basis for intermediate coordination and judgment. To coordinate the intelligent agents, To coordinate the reasoning operations of the intelligent agent, The sample to be tested, For intelligent agents The test results For intelligent agents The reason for the detection To coordinate the set of audit agents under the intelligent agent, To coordinate character portraits, To coordinate the confidence memory, it includes the confidence level for each downstream audit agent; S304. The decision-making agent obtains the judgment result based on the intermediate coordination judgment result; the specific decision-making agent receives the intermediate coordination judgment and its confidence weight (i.e., decision confidence memory) output by all coordinating agents, and combines it with the full set of features corresponding to the data in the multi-view feature set and the decision experience memory to generate the judgment result and explanation. At the same time, the output judgment result and the textual explanation of the judgment basis are saved to the decision action memory for auditing and traceability; the expression of the judgment result is:

[0048] In the formula, For the determination result, among which 1 indicates false information (or abnormal / fraudulent / misleading content), and 0 indicates true information (or normal content). The textual interpretation on which the judgment is based For decision-making intelligent agents, For the reasoning operations of the decision-making agent, For the current multi-view features to be detected, For all decision-making agents, the set of downstream agents. To create a profile of decision-making roles, For the memory of decision-making experience, The decision confidence memory includes the confidence level for each downstream coordinating agent.

[0049] S305. Update the multi-agent system based on the judgment results of each agent, and determine whether the multi-agent training is complete. If training is complete, proceed to S306; otherwise, return to S302. Optimizing the parameters of the multi-agent system includes structural optimization and parameter updates. When updating the parameters, the comparison between the judgment results output by each round of inference and the true labels is used as feedback to calculate the classification loss. Forward computation and error backpropagation are performed in mini-batch mode in each training round to update the memories of the relevant agents, thereby ensuring training optimization. The specific content of updating the parameters of the multi-agent system based on the judgment results is existing technology and will not be elaborated here.

[0050] After the multi-agent system judges the test data and obtains the judgment result, an adaptive inference strategy is executed based on the current structure and experience memory of the multi-agent system. The confidence-guided routing prioritizes the activation of high-confidence branches, and the inference is terminated in advance when the majority margin reaches the threshold, so as to reduce unnecessary communication and inference calls, thereby reducing latency, computing power consumption and collaboration overhead while maintaining detection accuracy.

[0051] When adaptively optimizing the structure of a multi-agent system, the judgment results are compared with the target labels, and the resulting correct or incorrect information is used as supervisory feedback to trigger the multi-agent adaptive optimization. During adaptive optimization, the multi-agent cooperative topology is dynamically adjusted (e.g., redundant agents or weakly contributing branches are pruned), and the experiential and confidence memories of misjudged agents are updated in a targeted manner, thereby forming a multi-agent structure and memory state that matches the data distribution. The specific multi-agent adaptive optimization includes the following steps: A1. Redundant agents in the multi-agent cooperative topology are pruned to obtain the updated multi-agent structure; specifically, the comprehensive score of each subordinate auditing agent of the coordinating agent is calculated based on the marginal accuracy contribution:

[0052] In the formula, For the first The overall score of each auditing agent. To improve the accuracy of current multi-agent architectures, To coordinate the set of audit agents under the intelligent agent, The penalty coefficient is... For the first Each agent's prediction vector and the Each agent's prediction vector cosine similarity, To coordinate the removal of subordinates of the intelligent agent A collection of intelligent agents for auditing. To coordinate the number of audit agents under the intelligent agent.

[0053] The system determines whether the overall score of the audit agent is less than the score threshold. If the overall score is less than the score threshold, the audit agent's contribution is insufficient and its redundancy is high. The audit agent is then removed from the subordinate set of the coordinating agent to prune redundant agents and reduce the cost of repeated calls. Specifically, in this embodiment, the score threshold is 0.

[0054] A2. Update the memory of agents whose judgment results do not match the true labels to learn from experience. Specifically, adopt a result-driven and error-triggered optimization update strategy. The system first back-evaluates the output of each agent in this round based on the judgment result labels (which may come from manual review, authoritative source verification, or high-confidence consensus results) to identify the set of agents that made misjudgments. Only the memory corresponding to the set is updated. For agents that made correct judgments, no writing is performed or only necessary statistical counts are maintained, thereby ensuring that the memory content mainly carries error correction experience rather than repetitive information.

[0055] When updating the agent's memory, the audit action memory, coordination action memory, and decision-making action memory do not need to be updated; only the memory used to update the other two types of memory needs to be recorded. The formula for updating audit experience memory is:

[0056] In the formula, For the purpose of memorizing updated audit experience, For the current multi-view features to be detected, To create a profile of the audit role, To memorize auditing experience prior to the update, For the purpose of recording the audit operation, For the audit decision made by the auditing agent in the current inference round, This is a summary of the reasoning process and decision-making basis related to the audit decision. The final oversight label can be derived from manual review, verification by authoritative information sources, or high-confidence consensus results. The self-reflective update operator is configured to update upon obtaining the final supervision label. Then, compare the judgment results of the decision-making agent. With the final supervision label, backtrack and analyze key information related to misjudgment in the current inference process, and based on this, analyze the input information to be detected. Extract experiential content for error correction from the agent's role profile and reasoning process information, and write or update it to new experiential memory. This is to support the reuse of experience and avoidance of misjudgments in subsequent reasoning processes.

[0057] The formula for updating decision experience memory is:

[0058] In the formula, For the updated memory of decision-making experience, To create a profile of decision-making roles, To remember the decision-making experience before the update, For decision-making and action memory, The decision result formed by the decision-making agent in the current inference round. A summary of the reasoning process and decision-making basis related to the judgment result; The formula for coordinating confidence memory update is:

[0059] In the formula, For the updated reconciliation confidence memory, To trust the self-reflective update operator, For the reconciled confidence memory before the update, To coordinate action memory, To coordinate character portraits, To coordinate the intermediate agreement decisions formed by the agents in the current inference round, Summary of reasoning process information and decision-making basis related to intermediate agreement determination.

[0060] The formula for calculating decision confidence memory update is:

[0061] In the formula, For the updated decision confidence memory, To trust the self-reflective update operator, Confidence memory of decisions made before the update To coordinate action memory, the update logic for decision confidence memory is basically the same as that for experience memory, except that the object of confidence memory update is no longer experience content used for error correction, but confidence parameters or weight information used to characterize the reliability of the agent.

[0062] S306. Evaluate the decision-making performance of the trained multi-agent system using a test set. The evaluation of multi-agent decision-making performance involves comprehensively measuring various evaluation metrics, including accuracy, precision, recall, F1 score, and AUC, to fully reflect the model's detection performance on different datasets. The specific calculation process for each evaluation metric is based on existing technology and will not be elaborated here.

[0063] S4. The trained multi-agent system judges the information instance to be tested to obtain the judgment result. Specifically, the information instance to be tested is inferred and judged by the audit agent, coordination agent and decision agent of the multi-agent system in turn, and the final judgment result is output by the decision agent. The specific content of the inference and judgment by the multi-agent system is the same as the inference and judgment steps of each agent in step S3, and will not be repeated here.

[0064] In step S4, when determining the instance of the information to be tested, confidence-guided routing is performed, prioritizing the activation of high-confidence branches. Inference is terminated early when the majority margin reaches the margin threshold. The expression is as follows:

[0065] In the formula, To determine the result, The reasons for the judgment are related to the judgment result. For intelligent agents The test results For intelligent agents The reason for the test, The set of currently active agents. For inference operators that take confidence weights into account, To coordinate the confidence weights in confidence memory, This is the margin threshold.

[0066] The decision-making agent first selects a small number of coordinating agents for evaluation, ranked from highest to lowest confidence weight. When several branches formed by one or more auditing agents and one or more coordinating agents activated according to confidence-guided routing reach a consistent conclusion, and the majority margin corresponding to this conclusion reaches a preset threshold, the decision result is output and the inference is terminated early. Otherwise, subsequent branches are activated step by step in descending order of confidence, and the current aggregation conclusion and its majority margin are recalculated after each new branch is added, until the majority margin threshold is met or the preset maximum number of activated agents is reached, at which point the decision result is output. This avoids a full call to all agents and reduces collaboration overhead. The activated branches include the coordinating agent and several subordinate auditing agents corresponding to that coordinating agent.

[0067] Specifically, in this embodiment, the fake information detection method based on multi-agent perspective aggregation in this solution is compared and tested with existing traditional deep learning detection models and representative conventional multi-agent collaborative methods; the traditional deep learning detection models selected include: The multilayer perceptron-based text classification model (MLP) uses fixed-dimensional features as input and is a representative of traditional deep learning methods.

[0068] The NFGCN-TIA method for detecting fake information based on graph convolutional networks captures structural information by modeling the relationships between nodes, and is used for the identification of abnormal behavior and fake content.

[0069] SIPUL, a deep learning-based detection method for comments and user behavior, focuses on uncovering potential patterns between users and content.

[0070] The BREAK detection model, designed to detect misinformation in news and politics, utilizes joint modeling of structured and semantic features.

[0071] The commonly used multi-agent cooperation methods include: Vanilla-LLM: It directly uses a single large language model to judge input samples without introducing an explicit agent interaction mechanism.

[0072] Vanilla-Agent: Encapsulates large language models as independent intelligent agents, but does not introduce structured cooperation or aggregation strategies.

[0073] Chain: It uses a chain structure to organize multiple intelligent agents, and information is transmitted in sequence and the judgment results are corrected step by step.

[0074] Star: Centered around a central agent, with other agents interacting around the center, and the central node making the final decision.

[0075] Tree: Organizes intelligent agents using a tree-like hierarchical structure, aggregating multi-perspective information in a top-down or bottom-up manner.

[0076] Graph: Organizes agents into a graph structure, allowing many-to-many interactions between agents to fully model the complex relationships between perspectives.

[0077] Layer: A multi-agent approach based on a hierarchical structure, where agents at different levels are responsible for analysis at different levels of abstraction.

[0078] DyDAN: A dynamic multi-agent decision network that can adaptively adjust the interaction patterns between agents according to task requirements.

[0079] Traditional deep learning detection models and conventional multi-agent collaborative methods were comprehensively evaluated on the test set using metrics such as accuracy, precision, recall, F1 score, and AUC. The comparison results are attached. Figure 3 As shown, compared to traditional deep learning methods and existing multi-agent system baselines, the adaptive multi-agent approach of this scheme achieves superior detection performance on most datasets and evaluation metrics. This demonstrates that the hierarchical aggregation mechanism based on perspective awareness can more effectively highlight and retain anomalous cues, thereby improving the accuracy and stability of false information detection. Furthermore, the Vanilla-LLM method, which directly calls a large language model for detection, performs similarly to, or even better than, the simple multilayer perceptron (MLP) baseline in some metrics. Even when encapsulated as a single agent and incorporating domain knowledge or evolutionary memory, the overall improvement remains limited, making it difficult to adequately address the deceptive and context-dependent characteristics of false information.

[0080] Meanwhile, existing multi-agent topologies exhibit varying performance: simple structures such as chain and star topologies typically achieve performance comparable to or slightly improved by multilayer perceptrons (MLPs), and still fall short of deep learning methods designed for specific scenarios in several tasks; while more complex collaborative organization methods such as graph structures, hierarchical structures, and dynamic collaborative networks demonstrate more significant performance gains, further confirming the necessity and effectiveness of introducing structured, multi-level collaborative aggregation designs in the context of false information detection.

[0081] To further verify the contribution of various efficiency-oriented strategies to overall performance, the adaptive multi-agent system in this embodiment, which includes three strategies—adaptive optimization (topology adaptation and target correction) and confidence-guided routing—is decomposed and verified separately. Specifically, five control variants are constructed: PAMAS without Enhancements, and simultaneously disable the above three strategies; PAMAS with / o Topology Adaptation retains target correction and route reasoning but removes topology adaptation; PAMAS w / o Target Refinement disables target refinement but retains topology adaptation and route reasoning. PAMAS with no Routing removes confidence-guided routing but retains the rest of the policies; A full PAMAS that integrates three strategies.

[0082] like Figure 4As shown, experimental results demonstrate that even without any enhancement strategies, the fake information detection method based on multi-agent perspective aggregation still achieves competitive detection results and has lower token consumption compared to complex multi-agent structures such as FullyGraph, Fully Layer, and DyDAN. This indicates that the hierarchical perspective aggregation framework of this invention inherently possesses significant efficiency advantages and detection capabilities. Furthermore, the three strategies of topology adaptation, target correction, and confidence-guided routing all deliver quantifiable improvements, enhancing both detection accuracy and token utilization efficiency. Topology adaptation dynamically adjusts the number of effectively participating agents, reducing redundant collaboration and enhancing system robustness and accuracy. Target correction selectively retrains agents with historical misjudgments, achieving targeted correction of systemic weaknesses and further improving detection performance. Confidence-guided routing prioritizes the activation of high-confidence branches and focuses on reliable signals during the inference phase, reducing unnecessary communication and calls, thereby improving inference efficiency.

[0083] A comprehensive comparison reveals that, among the three strategies, topology adaptation offers the most significant improvement in accuracy across different datasets, while target correction has the most prominent effect on reducing token consumption. Confidence-guided routing, while continuously reducing inference overhead, also brings a small but stable increase in consistency performance. This makes full PAMAS, which integrates all strategies, the optimal balance between accuracy and efficiency, enabling it to make good use of source data and adapt more effectively to new event data.

[0084] The false information detection method based on multi-agent perspective aggregation in this scheme was compared and evaluated with existing multi-agent collaborative topologies in terms of token consumption. The results are as follows: Figure 5 As shown, experimental results demonstrate that the fake information detection method based on multi-agent perspective aggregation achieves the lowest token consumption while maintaining optimal detection performance. This is because the present invention, through mechanisms such as topology adaptation, target correction, and confidence-guided routing, can dynamically control the number of effective agents participating in inference based on sample difficulty and collaboration requirements, reducing unnecessary updates and redundant activations. This significantly reduces token overhead without sacrificing accuracy, demonstrating high efficiency and scalability in large-scale fake information detection scenarios.

[0085] Meanwhile, existing complex multi-agent systems typically exhibit a trade-off between higher token costs and stronger performance: for example, fully connected graphs, hierarchical organizations, or DyDAN topologies offer stronger detection capabilities compared to chain or star structures, but require significantly higher token costs. While such complex structures are reasonable for general inference tasks, excessive token overhead reduces practical application efficiency in scenarios like false information detection, which emphasize robustness and scalable deployment. This invention achieves a better overall balance between robustness and scalability by introducing a hierarchical structure and an adaptive collaboration mechanism tailored to the characteristics of this task.

[0086] The present invention also aims to provide a fake information detection device based on multi-agent perspective aggregation, including a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program stored in the memory, the steps of the above-mentioned fake information detection method based on multi-agent perspective aggregation can be implemented.

[0087] A computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, and when the computer program is executed by one or more processors, implements the above-described method for detecting false information based on multi-agent perspective aggregation.

[0088] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting false information based on multi-agent perspective aggregation, characterized in that, Includes the following steps: S1. Obtain one or more network information instances and construct a multi-view feature set for each network information instance; S2. Construct a multi-agent system comprising several audit agents, several coordination agents, and one decision agent; the audit agents are respectively used to analyze the feature set of the multi-viewpoint according to a preset perspective to obtain several preset perspective audit judgments; the coordination agents are respectively used to perform perspective perception aggregation on the audit judgments according to the preset perspective to generate several intermediate coordination judgments; and the decision agent is used to generate a judgment result based on several intermediate coordination judgments. S3. Train multiple agents based on multi-view feature sets of one or more network information instances; S4. The trained multi-agent system judges the test information instances to obtain the judgment result.

2. The method for detecting false information according to claim 1, characterized in that, Step S1 further includes: decomposing the multi-perspective feature set of each network information instance according to a preset perspective to obtain several feature subsets, wherein the feature subsets cover complementary clue dimensions, and the audit agent generates an audit judgment based on the preset perspective according to the feature subsets.

3. The method for detecting false information according to claim 1, characterized in that, In step S2, the multi-agent system is also initialized: Each audit agent is configured with an audit role profile, audit experience memory, and audit action memory; wherein, the audit role profile is used to constrain the preset perspective of the audit agent's input and audit judgment output, the audit experience memory is used to store the audit agent's identification rules, and the audit action memory is used to record the audit agent's historical audit judgments; Each coordinating agent is configured with a coordinating role profile, a coordinating confidence memory, and a coordinating action memory. The coordinating role profile stores the agent's role attributes and behavioral characteristics; the coordinating confidence memory stores the agent's experiential confidence information; and the coordinating action memory records the agent's historical intermediate coordination decisions. The coordinating agent supervises a subset of auditing agents comprised of multiple auditing agents. The decision-making agent is configured with a decision role profile, decision experience memory, decision confidence memory, and decision action memory. The decision role profile is used to store the decision-making agent's role attributes and decision characteristics. The decision experience memory is used to store the decision-making agent's experience memory information. The decision confidence memory is used to store the decision-making agent's confidence memory information. The decision action memory is used to store the decision-making agent's historical judgment results.

4. The method for detecting false information according to claim 3, characterized in that, In step S3, adaptive optimization of the multi-agent system is also performed: A1. Calculate the overall score of each audit agent and tailor the audit agents according to the overall score; A2. Update the audit agent, coordination agent, or decision agent that made the incorrect judgment.

5. The method for detecting false information according to claim 4, characterized in that, In step A1, the overall score of each audit agent is calculated. If the overall score of an audit agent is less than a score threshold, the audit agent is removed. The formula for calculating the overall score of an audit agent is: In the formula, For the first The overall score of each auditing agent. To improve the accuracy of current multi-agent architectures, To coordinate the set of audit agents under the intelligent agent, The penalty coefficient is... For the first Each agent's prediction vector and the Each agent's prediction vector cosine similarity, To coordinate the removal of subordinates of the intelligent agent A collection of intelligent agents for auditing. To coordinate the number of audit agents under the intelligent agent.

6. The method for detecting false information according to claim 4, characterized in that, In step A2, the formula for updating the audit experience memory is: In the formula, For the purpose of memorizing updated audit experience, For the current multi-view features to be detected, For assignment operation, To create a profile of the audit role, To memorize auditing experience prior to the update, For the purpose of recording the audit operation, For the audit decision made by the auditing agent in the current inference round, This is a summary of the reasoning process and decision-making basis related to the audit decision. For final supervision labeling, The operator is updated based on self-reflection based on experience; The formula for updating decision experience memory is: In the formula, For the updated memory of decision-making experience, To create a profile of decision-making roles, To remember the decision-making experience before the update, For decision-making and action memory, The decision result formed by the decision-making agent in the current inference round. A summary of the reasoning process and decision-making basis related to the judgment result; The formula for coordinating confidence memory update is: In the formula, For the updated reconciliation confidence memory, To trust the self-reflective update operator, For the reconciled confidence memory before the update, To coordinate action memory, To coordinate character portraits, To coordinate the intermediate agreement decisions formed by the agents in the current inference round, A summary of the reasoning process and decision-making basis related to the intermediate agreement determination; The formula for calculating decision confidence memory update is: In the formula, For the updated decision confidence memory, To trust the self-reflective update operator, Confidence memory of decisions made before the update To coordinate action memory.

7. The method for detecting false information according to claim 1, characterized in that, In step S4, when judging the information instance to be tested, confidence-guided routing is performed, and the decision agent selects a coordinating agent for evaluation according to the confidence weight: When one or more audit agents and one or more coordinating agents activated by the confidence-guided routing have reached a consensus, and the margin of the consensus reaches the margin threshold, output the judgment result and terminate the inference. Otherwise, the remaining audit agents and coordinating agents are activated in descending order of confidence level, and the current judgment conclusion and margin are recalculated until the margin reaches the margin threshold or the audit agents and coordinating agents reach the preset maximum number of activations, at which point the judgment result is output.

8. The method for detecting false information according to claim 7, characterized in that, The formula for calculating confidence-guided routing is: In the formula, To determine the result, The reasons for the judgment are related to the judgment result. For intelligent agents The test results For intelligent agents The reason for the test, The set of currently active agents. For inference operators that take confidence weights into account, To create a profile of decision-making roles, For the memory of decision-making experience, For decision-making confidence memory, To coordinate the confidence weights in confidence memory, This is the margin threshold.

9. A false information detection device based on multi-agent perspective aggregation, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the false information detection method based on multi-agent perspective aggregation as described in any one of claims 1-8.

10. A computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, characterized in that, When the computer program is executed by one or more processors, it implements the false information detection method based on multi-agent perspective aggregation as described in any one of claims 1-8.

Images