A method and system for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration

Abstract

This invention discloses a method and system for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, belonging to the field of artificial intelligence security technology. The method includes constructing a three-dimensional dynamic trust scoring model for each agent in the multi-agent system; obtaining a comprehensive trust score for each agent based on the three-dimensional dynamic trust scoring model; abstracting the multi-agent system into a dynamic collaborative graph with trust weights; establishing an anomaly propagation quantification model based on the comprehensive trust score; and evaluating in real time the infection risk of a single anomalous agent to other agents in the system; continuously monitoring the comprehensive trust score and behavioral entropy of each agent; determining the reconstruction level when an anomaly is detected; and automatically executing a dynamic reconstruction of the collaborative topology at the corresponding level. This invention solves the technical problems in related technologies where multi-agent systems cannot effectively identify hidden anomalies, resist collusion attacks, block anomaly propagation, and lack system self-healing capabilities.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence security technology, specifically relating to a method and system for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration. Background Technology

[0002] With the rapid development of large language model capabilities, multi-agent systems (MAS), composed of multiple AI agents, are becoming a key paradigm for solving complex, long-chain, and cross-domain tasks. However, the autonomy, openness, and high interconnectivity of MAS have also introduced unprecedented security challenges.

[0003] Traditional network security technologies, such as zero-trust architecture, access control lists, and firewalls, primarily focus on static authentication and authorization between humans and machines or services and services. They are unable to effectively address security threats unique to MAS, such as: Malicious internal agent: A compromised or inherently flawed agent may send carefully crafted malicious instructions or false information to other agents through normal cooperation interfaces, inducing them to perform unauthorized operations.

[0004] The "avalanche effect" caused by the concealment and contagiousness of abnormal behavior: Due to the high semantic complexity of the behavior of intelligent agents, their abnormal behavior (such as logical errors and target deviations) is often difficult to detect through simple rule matching, and will quickly spread to other intelligent agents that depend on it through the cooperation chain, forming an "avalanche effect".

[0005] The failure of static trust models: Traditional role- or identity-based trust models are static and cannot reflect the dynamic changes in the trustworthiness of agents under different task contexts and different interaction objects.

[0006] Some existing multi-agent trust management solutions propose a two-layer trust assurance architecture, using message quality assessment and dynamic reputation updates to filter trustworthy collaborators. However, these methods have significant shortcomings: First, the trust assessment dimension is singular, relying solely on message utility / decision uncertainty, failing to identify hidden anomalies such as prompt word injection and model illusion; second, they lack dedicated anti-collusion mechanisms, failing to proactively identify and eliminate malicious agent groups engaging in score manipulation; third, the defense mechanism is passive, lacking proactive immunity capabilities at the anomaly propagation quantification model and network structure levels; and finally, the system lacks dynamic topology reconstruction and self-healing capabilities, resulting in insufficient overall resilience. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, which can solve the technical problems in related technologies such as the inability of multi-agent systems to effectively identify hidden anomalies, resist collusion attacks, block the propagation of anomalies, and lack of system self-healing capabilities.

[0008] In a first aspect, the present invention provides a method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, comprising the following steps: For each agent in a multi-agent system, a three-dimensional dynamic trust scoring model is constructed. The model includes at least the direct interaction trust based on the quality of historical interactions, the indirect recommendation trust based on third-party evaluation and after anti-collusion filtering, and the behavior entropy trust based on the predictability and consistency of behavior sequences. Based on the three-dimensional dynamic trust scoring model, the comprehensive trust score of each agent is obtained. The multi-agent system is abstracted into a dynamic collaborative graph with trust weights. An anomaly propagation quantification model is established based on the comprehensive trust score to assess the risk of infection of other agents in the system by a single anomalous agent in real time. The system continuously monitors the comprehensive trust score and behavioral entropy of each agent. When an anomaly is detected, the system determines the reconstruction level and automatically performs the corresponding level of collaborative topology dynamic reconstruction.

[0009] As an alternative implementation method, the construction and calculation methods of behavioral entropy trust include: The behavior sequence of the agent is standardized and represented in a specific way: the agent's original behavior log is parsed into a standardized sequence of atomic behavior units. Based on standardized behavior sequences, similar historical behavior sequences are retrieved based on the context of the current behavior sequence, and an n-gram probability model is constructed. The cross-entropy of the current behavior sequence is calculated using this probability model to obtain the behavior entropy value that represents the uncertainty of the behavior; The inverse mapping value of behavioral entropy is used as the behavioral entropy confidence level.

[0010] As an alternative implementation, an anti-collusion recommendation filtering algorithm is executed when calculating the agent's indirect recommendation trust level, specifically including: Obtain recommendation ratings from other agents for the current agent, and calculate the temporary indirect recommendation trust level of the current agent using the comprehensive trust ratings of other agents as weights. Calculate the variance of all recommendation ratings, and trigger collusion detection when the variance is lower than a preset abnormal consistency threshold; If a set of agents with highly consistent scores is analyzed and meets preset conditions, it is identified as a conspiracy group. In the calculation of indirect recommendation trust, all recommendation scores of the colluding group are removed, and the indirect recommendation trust is recalculated.

[0011] As an alternative implementation method, when an anomaly is detected, a reconstruction level is determined, specifically as follows: The system collects four parameters in real time: comprehensive trust score, behavioral entropy deviation, abnormal infection risk, and task criticality level of the abnormal agent. Different thresholds are set for each of these parameters, and the corresponding reconstruction level is selected according to the threshold. The system also allows low-level strategies to switch to high-level strategies based on risk escalation, and the system achieves step-by-step downgrade through recovery verification.

[0012] As an optional implementation, it also includes a system performance recovery verification step after topology dynamic reconstruction: After the reconstruction is completed, start the observation window to collect comprehensive trust score, behavior entropy, task success rate and link status data; The system sequentially determines whether four indicators—trust convergence, behavioral consistency, collaboration efficiency, and anomaly immunity—meet the preset qualification thresholds. The system will be deemed to have recovered and will exit the emergency defense mode only after all indicators have met the standards and a stable observation period has been completed.

[0013] As an alternative implementation, it also includes a trust adaptive recovery step: the isolated abnormal agent is sent to a sandbox environment for diagnosis. If it is determined to be a temporary fault, its trust score is gradually restored and it is reconnected to the cooperative network after its behavior stabilizes. At the same time, the trust fusion model parameters are fine-tuned online based on global security events to achieve iterative optimization of the trust mechanism.

[0014] Secondly, the present invention provides a system for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, comprising: The model building module is configured to: build a three-dimensional dynamic trust scoring model for each agent in the multi-agent system. The model includes at least the direct interaction trust based on the quality of historical interactions, the indirect recommendation trust based on third-party evaluation and after anti-collusion filtering, and the behavior entropy trust based on the predictability and consistency of behavior sequences. The comprehensive trust scoring module is configured to obtain a comprehensive trust score for each agent based on a three-dimensional dynamic trust scoring model. The infection risk assessment module is configured to: abstract the multi-agent system into a dynamic collaborative graph with trust weights, establish an abnormal propagation quantification model based on comprehensive trust scores, and assess the infection risk of a single abnormal agent to other agents in the system in real time. The topology reconstruction module is configured to continuously monitor the comprehensive trust score and behavioral entropy of each agent, determine the reconstruction level when an anomaly is detected, and automatically perform the corresponding level of collaborative topology dynamic reconstruction.

[0015] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0016] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0017] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a method and system for constructing dynamic trust chains and blocking anomaly propagation in a multi-agent collaborative framework. By introducing behavioral entropy as the core trust dimension, it can accurately identify hidden anomalies such as prompt word injection and model illusions that are missed by traditional methods. Through a specially designed anti-collusion algorithm, it actively identifies and eliminates false recommendations from colluding groups, enhancing the system's robustness. By establishing an anomaly propagation model based on infection probability and a four-level dynamic topology reconstruction mechanism, it achieves a leap from passive defense to active immunity, effectively blocking the "avalanche" propagation of anomalies. Simultaneously, through a comprehensive recovery verification and trust adaptive recovery mechanism, it endows the system with strong self-healing capabilities and business continuity assurance. This invention forms a complete closed loop from four levels: assessment, detection, blocking, and self-healing, significantly improving the overall security and resilience of multi-agent systems. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method for constructing a dynamic trust chain and blocking anomaly propagation in multi-agent collaboration disclosed in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0021] The technical solutions disclosed in the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, including the following steps: For each agent in a multi-agent system, a three-dimensional dynamic trust scoring model is constructed. The model includes at least the direct interaction trust based on the quality of historical interactions, the indirect recommendation trust based on third-party evaluation and after anti-collusion filtering, and the behavior entropy trust based on the predictability and consistency of behavior sequences. Based on the three-dimensional dynamic trust scoring model, the comprehensive trust score of each agent is obtained. The multi-agent system is abstracted into a dynamic collaborative graph with trust weights. An anomaly propagation quantification model is established based on the comprehensive trust score to assess the risk of infection of other agents in the system by a single anomalous agent in real time. The system continuously monitors the comprehensive trust score and behavioral entropy of each agent. When an anomaly is detected, the system determines the reconstruction level and automatically performs the corresponding level of collaborative topology dynamic reconstruction.

[0023] The core idea of ​​this invention is to view a multi-agent cooperative network (i.e., a multi-agent system) as a dynamically evolving trust graph, and to introduce "behavioral entropy" as a key indicator to measure the uncertainty and potential risks of agent behavior. By constructing a three-dimensional dynamic trust scoring model of "Direct Interaction Trust (DIT) + Anti-Collusion Indirect Recommendation Trust (IRT) + Behavioral Entropy Trust (BET)," the system calculates the agent comprehensive dynamic trust score (CDTS) in real time. Based on this, an anomaly propagation model based on infection probability is established, and a four-level topology reconstruction strategy is preset. Upon detecting an anomaly, the system automatically executes the corresponding topology reconstruction and ensures business stability through recovery verification, forming a complete closed loop of "trust assessment - anomaly detection - propagation blocking - topology self-healing - recovery verification." The detailed technical solution of this invention is as follows: Step 1: Construct a three-dimensional dynamic trust scoring model for each agent in the multi-agent system.

[0024] The steps are designed for each agent in a multi-agent system (e.g., A). i Maintain a dynamic trust profile. This profile contains three core dimensions: 1. Direct Interaction Trust (DIT): Based on trust with the target agent A i The calculation of direct interaction history is performed. This embodiment uses a weighted sliding window mechanism, assigning higher weight to recent interactions. Simultaneously, weights are applied based on the task sensitivity of each interaction, with successful interactions on highly sensitive tasks resulting in a greater increase in trust level.

[0025] 2. Indirect Recommendation Trust Level (IRT): This involves collecting information from other agents regarding the target agent A. iThe evaluation (i.e., recommendation score) is used to prevent malicious collusion. This invention designs an anti-collusion recommendation filtering algorithm. This algorithm analyzes the credibility of the recommender (i.e., other agents) and the consistency of their recommendation behavior, downgrading or eliminating suspicious or highly consistent positive or negative recommendations. The core steps of the anti-collusion recommendation filtering algorithm include: a. Credibility-weighted: The recommendation scores are weighted and averaged, with the weight determined by the recommender agent A. k Self-Comprehensive Dynamic Trust Score (CDTS). (IRT) temp =Σ(CDTS(A k ) R k ) / ΣCDTS(A k This makes the recommendations of highly trustworthy agents more influential. Among them, R... k For other intelligent agents A k For A i Recommended rating.

[0026] b. Anomaly Consistency Detection: Calculate all recommendation scores {R} k The variance Var(R) of the group is a key characteristic of conspiracy attacks. Members of the conspiracy group tend to give highly consistent (minimal variance) extreme ratings. Therefore, if Var(R) is below a preset threshold T... var If this occurs, a conspiracy alert will be triggered.

[0027] c. Social graph isolation: Once a collusion alert is triggered, the system will check and extract a candidate set with highly consistent scores. The internal average mutual trust score and external average trust score of G are calculated respectively. When the internal score is greater than or equal to the cohesion threshold and the external score is less than or equal to the externality threshold, G is determined to be a collusive group. The specific method is as follows: ① Cohesion calculation (high cohesion quantification): Define the internal average mutual trust score of set G: ; in, For the conspiratorial candidate set Number of internal agents; for right A comprehensive dynamic trust score.

[0028] Judgment threshold: Preset cohesion threshold (Recommended value) ~ ).

[0029] Judgment conditions: If these conditions are met, it indicates high cohesion.

[0030] ②Outer degree calculation (low-connection quantization): Define a set Average external trust score for external agents: ; in, For those not belonging to the system The total number of intelligent agents.

[0031] Judgment threshold: Preset external connectivity threshold (Recommended value) ~ ).

[0032] Judgment conditions: If the condition is met, it indicates a low level of connectivity.

[0033] ③ Rules for final determination of conspiracy groups: If both of the following quantification conditions are met, then it is determined that... As a conspiratorial group: Internal cohesion meets the standard: , : Default 0.9; External connectivity meets the standard: , : Default 0.4.

[0034] d. IRT Correction: After removing the recommended score of the collusion group G, the IRT is recalculated.

[0035] 3. Behavioral Entropy Trust (BET): Through continuous monitoring of target agent A i The behavioral entropy is calculated from the output logs, tool call sequences, and decision paths of the agent. This behavioral entropy, based on the Shannon entropy concept in information theory, is used to quantify the uncertainty of an agent's behavioral sequence. Behavioral entropy reflects the predictability and consistency of an agent's behavior. A normal, reliable agent should exhibit a low-entropy (high-consistency) behavioral pattern in similar contexts; while an agent that is attacked, malfunctions, or attempts to conceal malicious intent tends to behave chaotically and unpredictably, exhibiting high entropy. BET is inversely proportional to behavioral entropy; that is, the higher the behavioral entropy, the lower the BET.

[0036] The behavioral entropy described in this invention aims to quantify the predictability and consistency of AI agent behavior. The core steps of this algorithm include: a. Standardize the agent's behavior logs. All agent tool calls, function executions, decision steps, outputs, parameter modifications, and state transitions are uniformly parsed into atomic behavior units (AAUs). Each atomic behavior unit has a fixed format: [behavior type ID, target object ID, parameter hash, time slice number, context task label], and is mapped to an integer ID sequence S using a global behavior dictionary.

[0037] b. Use a pre-trained model (such as Sentence-BERT) to process the current task context C. t Encoded as vector E c And retrieve the set H of the K most similar task behavior sequences from the history database.

[0038] c. Construct an n-gram probability model P with Laplace smoothing based on set H: , Where C(...) represents the number of co-occurrences, N represents the total number of samples, and V represents the size of the behavioral dictionary. The order of the n-gram can be adaptively selected based on the sequence length.

[0039] d. Use this model to calculate the cross-entropy H(S) = -(1 / m) of the current sequence S. Σlog2P(s i |s {i-1} ), and normalized to the [0,1] interval using the relative entropy gain formula.

[0040] e. Maintain a behavioral entropy baseline for each agent based on the exponential moving average (EMA). t =β·Baseline t-1 +(1-β)·Entropy t β∈(0,1) is the attenuation coefficient, with a default β=0.9. A fixed initial baseline is set as the absolute normal reference. Absolutely normal behavior, sub-healthy behavior, and abnormal behavior are defined through a two-level judgment logic. Only absolutely normal behavior is used to update the dynamic baseline; sub-healthy and abnormal behaviors do not participate in baseline updates. The deviation between the dynamically updated baseline and the initial baseline is periodically checked. When the deviation exceeds a preset drift threshold, the dynamic baseline is calibrated using the initial baseline to prevent baseline drift. When the real-time behavior entropy exceeds three times the standard deviation of the dynamically updated baseline, it is judged as abnormal behavior, achieving stable, low-false-alarm hidden anomaly detection.

[0041] Behavioral entropy baseline dynamic update rules: The "dual baseline + hierarchical judgment + anomaly backtracking calibration" mechanism is adopted, and the specific rules are as follows: ① Dual baseline system: Each agent maintains both an "initial baseline" and a "dynamically updated baseline," which work synergistically. Initial baseline ( During the system initialization phase, based on the normal behavior sequence of the agent in the standard test scenario (provided by manual annotation or a historical reliable behavior library), the average behavior entropy of 30 consecutive time steps is calculated as the initial baseline. This baseline is fixed and serves as a reference standard for "absolute normality".

[0042] Dynamically update baseline ( ): This refers to the original EMA baseline, which is used to adapt to minor fluctuations in the normal behavior of the agent in real time and avoid false alarms. Its updates are constrained by the initial baseline to prevent drift.

[0043] ② Hierarchical judgment logic (judgment first, update later): The system's real-time behavioral entropy at each time step ( A two-level assessment is conducted to clarify the quantitative boundaries of "normal behavior," "sub-healthy behavior," and "abnormal behavior," before deciding whether to update the dynamic baseline. Level 1 Judgment (compared with the initial baseline, defining absolute normality): like ( If the standard deviation corresponding to the initial baseline is calculated from the normal behavior sequence during the initialization phase and remains fixed, then the behavior is determined to be "absolutely normal behavior" and is allowed to be used to update the dynamic baseline (by executing the original EMA update formula).

[0044] Level 2 assessment (distinguishing between "sub-healthy behavior" and "abnormal behavior" to avoid baseline drift): like ( If the behavior is not used to update the dynamic baseline (the real-time standard deviation corresponding to the baseline), it is determined to be a "sub-healthy behavior" and will not be used to update the dynamic baseline. Only the behavior will be recorded as a basis for baseline drift monitoring.

[0045] Exception detection (maintain original logic to ensure exception detection): like If the behavior is abnormal, it is considered "abnormal behavior", the baseline is not updated, and the anomaly detection process is triggered.

[0046] ③ Baseline drift monitoring and calibration mechanism (addressing baseline drift issues caused by sub-health behaviors): Every 10 time steps, the system performs a drift check on the dynamically updated baseline. The specific process is as follows: Calculate the deviation between the dynamically updated baseline and the initial reference baseline: ; Preset drift threshold (A value of 0.1 is recommended, but can be configured according to business scenarios). The dynamic baseline was determined to be drifting. Drift calibration: Resets the dynamically updated baseline to By pulling the initial baseline back to a reasonable range; If drift occurs in three consecutive verifications, the system will automatically trigger a retrospective analysis of sub-health behaviors to investigate whether there are any hidden slow abnormalities. At the same time, the attenuation coefficient will be adjusted to 0.95 to reduce the sensitivity of dynamic baseline updates and further suppress drift.

[0047] When real-time behavior entropy > Baseline t A value of +3σ (three standard deviations) is considered "abnormal behavior." Here, σ represents the real-time standard deviation corresponding to the dynamically updated baseline, calculated from the normal behavior entropy (containing only "absolutely normal behavior") over the most recent 50 time steps. This mechanism, combining dual baselines and hierarchical judgment, can reliably identify hidden anomalies such as hallucinations, cue injection, and internal tampering, while avoiding missed detections due to baseline drift. BET is inversely proportional to behavioral entropy.

[0048] Step 2: Based on the three-dimensional dynamic trust scoring model, obtain the comprehensive trust score for each agent.

[0049] By combining the three dimensions from step 1 (i.e., DIT, IRT, and BET) through a learnable fusion function, agent A is obtained. i The Comprehensive Dynamic Trust Score (CDTS) is derived. The learnable fusion function is designed as follows: This invention employs a lightweight multilayer perceptron (MLP) as the fusion function F. fuse This MLP receives three normalized scalar values—DIT, IRT, and BET—as input and outputs the final CDTS. Its specific structure is as follows: Input layer: 3 neurons, corresponding to [DIT] norm IRT norm BET norm ].

[0050] Hidden layer: 1 hidden layer containing 8 neurons, with ReLU activation function.

[0051] Output layer: 1 neuron, using the Sigmoid activation function to ensure that the output value is within the range [0, 1].

[0052] The parameters (weights and biases) of this MLP are generated using an initialization optimization strategy and are not completely random. Furthermore, an online learning mechanism is designed to automatically acquire supervision signals and ensure convergence. The specific implementation is as follows: 1. MLP parameter initialization optimization.

[0053] To avoid the learning inefficiency caused by random initial parameters, this invention adopts a strategy combining "domain priors + Xavier initialization" to ensure that the initial parameters have a reasonable foundation, achieving effective fusion without relying on subsequent sparse events. Specific steps are as follows: a. Domain Priority Weighting: Based on the common characteristics of multi-agent collaboration scenarios, an initial weight range is preset. The initial weight of Behavioral Entropy Trust (BET) is 40%, Direct Interaction Trust (DIT) is 35%, and Indirect Recommendation Trust (IRT) is 25% (this ratio is set based on domain experience in secure multi-agent collaboration and can be flexibly adjusted to adapt to different scenarios).

[0054] b. Xavier initialization fine-tuning: Based on the prior weights mentioned above, the Xavier normal initialization algorithm is used to fine-tune the weights and biases, so that the variance of the input and output of each layer is consistent, avoiding gradient vanishing or exploding, and ensuring that the initial model has basic nonlinear fusion capabilities, which can be directly used for the initial operation of the system.

[0055] This initialization strategy combines domain experience with mature deep learning initialization methods, solving the problem of "poor initial fusion results caused by completely random initialization" and laying a stable foundation for subsequent online learning.

[0056] 2. Mechanism for automatic acquisition of online learning supervision signals.

[0057] To address the characteristics of unsupervised / semi-supervised environments in multi-agent collaborative settings, this invention designs a dual-source supervision signal acquisition mechanism that combines "objective anomaly evidence-driven + manual verification." This mechanism automatically generates effective supervision signals without relying on manually labeled "standard answers." Specific implementation details are as follows: a. Automatic monitoring signal generation (core, accounting for over 90%): The system automatically determines "missed" and "false" events and generates monitoring signals by capturing the following quantifiable objective abnormal evidence, without requiring manual intervention: ① False alarm judgment (error isolation of normal agents): If an agent marked as abnormal has stable behavioral entropy within the dynamic baseline ±σ range and exhibits no abnormal behavior for three consecutive observation periods (each period consists of 10 time steps) in the sandbox environment, and its success rate in simulated collaborative tasks is ≥95%, then it is judged as a "false alarm" and a supervision signal of "the agent is normal" is generated.

[0058] ② Missed Detection (No Malicious Agent Detected): If an agent not marked as anomalous appears, and any of the following quantifiable evidences are present, it is determined as a "missed detection," and a monitoring signal indicating "the agent is malicious" is generated: (a) Its behavioral entropy is higher than the dynamic baseline 3σ for 5 time steps, and causes the downstream agent's behavioral entropy to increase synchronously (the increase is ≥50%). (b) The conflict rate between its output decision results and the system's preset rule base (such as compliance rules and security policies) is ≥80%; (c) Being included in a conspiracy group (identified by the anti-conspiracy algorithm of this invention), or having high-frequency interactions with confirmed malicious agents (interaction frequency ≥ 3 times the average normal interaction frequency).

[0059] b. Manual Assistance Verification (Supplement, for extreme scenarios): For the very few ambiguous cases that cannot be automatically determined through objective evidence (≤10%), the system will trigger a manual verification process. Administrators will quickly label these cases as "normal / malicious" based on quantitative data such as system-provided behavior logs, trust data, and abnormal propagation trajectories, generating a monitoring signal. The manual labeling results will be simultaneously fed back to the system to optimize the automatic judgment rules and gradually reduce the proportion of manual intervention.

[0060] 3. Convergence and stability guarantees for online MLP learning To ensure the convergence, stability, and efficiency of MLP parameter fine-tuning, this invention designs the following targeted mechanisms to avoid learning failure caused by sparse supervision signals: a. Small step gradient update strategy: After each acquisition of a supervision signal, gradient backpropagation is performed with a small step size (learning rate η = 0.001~0.005), and only the MLP parameters are fine-tuned to avoid model instability caused by parameter mutations; at the same time, a learning rate decay mechanism is set, and the learning rate is gradually reduced as the system running time increases (updated once every 100 supervision signals, with a decay coefficient of 0.9) to ensure that the model gradually converges.

[0061] b. Parameter constraints and regularization: During MLP training, L2 regularization constraints are applied to the weight parameters to prevent overfitting; at the same time, the weight ratio of each dimension's confidence is limited to between 10% and 60% to avoid imbalance in the fusion result due to excessive weight of a single dimension, thus ensuring model stability.

[0062] c. Batch update and smoothing: When the number of supervision signals accumulates to a preset threshold (e.g., 20), batch parameter updates are performed instead of updating individual signals immediately, reducing the impact of sparse signals on the model; at the same time, the updated parameters are smoothed by exponential moving average (EMA, decay coefficient β=0.9) to further improve model stability.

[0063] d. Convergence determination and backtracking mechanism: The system monitors the matching degree (accuracy) between the CDTS output by the MLP and the actual anomaly determination results in real time. When the accuracy is stable above 90% for three consecutive update cycles, the model is determined to have converged, and the learning rate is automatically reduced to 0.0001 with only minor adjustments. If the accuracy drops below 70%, parameter backtracking is triggered to restore the previous stable state and the learning strategy is re-optimized.

[0064] Through the above mechanism, MLP can gradually converge under sparse supervision signals and stably learn the nonlinear fusion relationship between DIT, IRT and BET, ensuring the accuracy and robustness of the Comprehensive Dynamic Trust Score (CDTS).

[0065] In summary, this step obtains the Comprehensive Dynamic Trust Score (CDTS) through a learnable fusion function. In a preferred embodiment, the fusion function is a lightweight multilayer perceptron (MLP) whose input layer receives normalized [DIT, IRT, BET], the hidden layer is an 8-neuron ReLU layer, and the output layer is a 1-neuron Sigmoid layer, with the output CDTS value in the [0,1] interval. This MLP can be fine-tuned online using false positive / false negative events.

[0066] Step 3: Abstract the multi-agent system into a dynamic collaborative graph with trust weights. Establish an anomaly propagation quantification model based on comprehensive trust scores to assess the risk of infection to other agents in the system from a single anomalous agent in real time. Specifically: The entire multi-agent system is abstracted as a directed graph G=(V, E), where node V represents an agent and edge E represents the cooperative relationship between agents. Each edge (A... i A j A trust weight is assigned to the receiver A. j For sender A i The CDTS decision.

[0067] Based on this graph, this invention defines an anomaly propagation model. This model assumes that abnormal states (such as error messages and malicious commands) can propagate through cooperative edges, and the propagation strength is negatively correlated with the trust weight of the edge. Specifically, if A i It was identified as an anomaly source, and it affected neighbor A. j The probability of infection P infect (A j The model is as follows: P infect (A j )=f(CDTS(A i ),S ctx ), where f is a monotonically increasing function, S ctx The similarity between the current task context and the abnormal historical context is used; the higher the similarity, the greater the risk of infection. Its specific design is as follows: a. Define risk factors: Risk(A) i )=1-CDTS(A i ).

[0068] b. Weighted linear combination: Z = α Risk(A i )+(1-α) S ctx , where α is a configurable weight.

[0069] c. Sigmoid nonlinear mapping: Where k is the kurtosis coefficient, which controls the sensitivity of the function, and Z0 is the offset, which is usually set to 0.5.

[0070] This design ensures the monotonically increasing nature of the function and provides good interpretability and engineering controllability.

[0071] Step 4: Continuously monitor the comprehensive trust score and behavioral entropy of each agent. When an anomaly is detected, determine the reconstruction level and automatically perform the corresponding level of collaborative topology dynamic reconstruction.

[0072] Multi-agent systems deploy a central coordinator or distributed agent that periodically performs the following operations: 1. Anomaly Detection: Check whether the CDTS of each agent is lower than the safety threshold (Master Trust Threshold T). primary , is the basic security criterion for Agent Integrated Dynamic Trust Score (CDTS), used to determine whether an agent is a suspected anomalous node, or whether its behavioral entropy has a sudden change (exceeding the baseline 3σ). If either condition is met, it is marked as a suspected anomalous node.

[0073] 2. Impact Scope Assessment: For each suspected anomalous node, calculate its potential infection probability P to other agents using the propagation model from step 3. infect .

[0074] 3. Dynamic Topology Reconstruction: The system automatically selects and executes one of four reconstruction strategies based on four parameters: CDTS, behavioral entropy offset, infection probability, and task criticality level. Level 1 Restructuring (Lightweight Trust Correction): Only reduces the output trust weight of abnormal nodes, without severing the link. Specifically: Triggering condition: CDTS at the primary trust threshold T primary The percentage fluctuates between 90% and 100%, the behavioral entropy is in the 1σ-2σ range of the dynamic baseline, the infection probability is <0.3, and the mission criticality level is non-critical.

[0075] Target: Agents exhibiting slight trust fluctuations, a small increase in behavioral entropy but not exceeding the dynamic baseline, and no risk of infecting other nodes.

[0076] Execution content: Maintain the normal connection relationship of the agent in the cooperative network, only reduce its trust weight in external output, and reduce its influence in task allocation, message voting, and decision fusion; do not cut off the communication link, and do not change the overall cooperative topology; at the same time, start short-term behavior enhancement monitoring and continuously observe the trend of trust and behavior entropy changes.

[0077] Secondary restructuring (local trust chain pruning): This involves severing the collaborative edges between this node and highly sensitive, core decision-making nodes, and marking it as a restricted node. Specifically: Triggering condition: CDTS < Master Trust Threshold T primary The behavioral entropy is in the 2σ-3σ range of the baseline, the infection probability is 0.3-0.65, and the mission criticality level is general.

[0078] Execution target: CDTS is below the primary trust threshold T primary An agent whose behavioral entropy is consistently higher than the dynamic baseline and poses a slight risk of infection to neighboring agents, but does not cause interruption of the task chain.

[0079] Execution content: Retain the basic registration status of the agent in the system, actively sever its collaborative edges with highly sensitive tasks and core decision-making agents; retain limited connections with non-critical nodes; mark the agent as a restricted node, prohibiting it from initiating new collaborative requests and participating in critical path calculations; synchronously reduce the trust weight of adjacent nodes affected by the agent to prevent the spread of local risks.

[0080] Level 3 Restructuring (Secure Rerouting): The original path is discarded, and a new, high-trust, low-latency alternative path is planned based on the CDTS ranking of the remaining healthy agents. Specifically: Triggering condition: CDTS is much lower than the primary trust threshold T primary Behavioral entropy > baseline 3σ, infection probability ≥ 0.65, original collaborative path interrupted, mission criticality level is important.

[0081] Targets: Agents on the critical path malfunction, the original collaborative link is interrupted, or the infection probability exceeds the secondary infection threshold T. secondary (Secondary infection threshold T) secondary It serves as a quantitative benchmark for assessing the risk of abnormal propagation (default value 0.65), used to evaluate the infection risk level of abnormal nodes to other intelligent agents, and scenarios that prevent tasks from executing according to the original process.

[0082] Execution content: Discard the original collaborative path containing the current abnormal node, and based on the CDTS ranking of the remaining healthy agents, replan an alternative collaborative path with no abnormalities, high trust, and low latency; reallocate execution nodes according to task dependencies, completely excluding abnormal nodes from the new path; perform trust consistency verification on the new path to ensure that the CDTS of all nodes on the path is higher than the primary trust threshold T. primary The behavior entropy is within the normal range; after the new path takes effect, all related connections of the original path are closed and resources are released.

[0083] Level 4 Restructuring (Global Restructuring): Isolate the entire abnormal subgraph, clear the original dynamic trust graph, and rebuild the global collaboration relationship based on all remaining trusted nodes. Specifically: Triggering conditions: A conspiracy group is identified, the proportion of abnormal nodes is ≥15%, the average global behavioral entropy increases, the risk of infection spreads over a large area, and the task is classified as critical.

[0084] Target scenarios: Scenarios where a collusion group of intelligent agents is detected, anomalies spread over a large area, trust mutations occur at multiple nodes, the overall entropy of system behavior increases, and there is a risk of avalanche propagation.

[0085] Execution content: Isolate the identified abnormal nodes and collusion subgraphs as a whole, disconnecting all their incoming and outgoing edges; clear the original dynamic trust graph and reconstruct the global cooperation relationship based on the remaining trusted nodes; recalculate the trust edge weights between trusted nodes and optimize task allocation and message passing structure; perform security verification on the reconstructed global topology to ensure that there are no abnormal node remnants, no high-risk connections, and no collusion subgraphs; switch to the new topology after reconstruction is completed.

[0086] The system employs a mutual exclusion mechanism, allowing only one reconstruction strategy to take effect at any given time. Lower-level strategies can be switched to higher-level strategies based on risk escalation, while lower-level strategies must be gradually downgraded through recovery verification.

[0087] For core tasks, a level 3 or 4 strategy is mandatory to ensure security; for non-critical tasks, a level 1 or 2 strategy is preferred to ensure execution efficiency.

[0088] Step 5: Restore verification and trust adaptive recovery.

[0089] To ensure the system can operate stably, efficiently, and securely after topology reconstruction, this invention sets up a standardized performance recovery verification mechanism, which is automatically activated after each topology reconstruction. It collects data such as CDTS, behavior entropy, and task success rate within the observation window. Through multi-dimensional indicator monitoring, threshold determination, and continuous observation, the system can only exit the emergency defense mode after confirming that it has returned to a reliable, stable, and efficient operating state.

[0090] A. Design of recovery verification indicators: Trust Convergence Metric: The mean CDTS of all online agents is higher than the master trust threshold T. primary The CDTS of affected nodes has recovered and stabilized within the normal range; there are no nodes with continuous trust mutations; and the fluctuation of trust updates is less than the set threshold.

[0091] Behavioral consistency indicators: The average global behavioral entropy of the system has fallen back to the normal range; the behavioral entropy of all agents except abnormal nodes and isolated nodes has returned to the range of the initial baseline (Baselineinit) ± σ; there are no new nodes with sudden behavioral entropy changes.

[0092] Collaboration efficiency metrics: Task execution success rate recovers to over 90% of pre-abnormality levels; average decision latency and message transmission latency return to normal levels; task completion rate and link utilization reach system baseline values.

[0093] Abnormal immune indicators: No new abnormal alarms, no new infection risks, and no signs of collusion appeared within the continuous observation period.

[0094] B. Resume the verification execution process: Activation conditions: The new collaborative topology or new collaborative path officially takes effect after any level of the four-level topology reconstruction is completed.

[0095] Data Acquisition: Within a fixed observation window, continuously collect CDTS, behavioral entropy, task execution results, link status, infection probability, and message interaction records for each agent.

[0096] Item-by-item judgment: The four indicators of trust convergence, behavioral consistency, collaboration efficiency, and anomaly immunity are judged in turn. Only when all of them meet the threshold requirements can we proceed to the next stage.

[0097] Continuous and stable observation: After the indicators initially meet the standards, the observation period is extended to confirm that there is no rebound, no fluctuation, and no trend of re-deterioration in the indicators.

[0098] Recovery Confirmation: After all the judgment conditions are met and stable observation is passed, the system automatically determines that it has "recovered", exits the emergency defense state, closes the temporary restriction policy, and returns to the normal dynamic trust update and operation mode.

[0099] C. Criteria for restoring compliance: Behavioral entropy: The behavioral entropy of all trusted agents in the system has fallen back to the initial baseline, with no nodes remaining persistently high.

[0100] Trust score: The global mean of CDTS is higher than the primary trust threshold T. primary No node is below the critical safety value.

[0101] Collaboration performance: Task success rate ≥90%, latency returns to baseline range, no blocking, no packet loss, and no abnormal retries.

[0102] Safety status: Zero new anomalies, zero risk of infection, and zero signs of collusion within the continuous observation period.

[0103] Topological stability: The reconstructed trust graph shows no structural changes during the observation period, and the trust weights are updated smoothly.

[0104] D. Non-compliance handling mechanism: If the verification fails, the system will automatically re-enter the anomaly detection process. Minor non-compliance: Re-execute the level 2 or level 3 restructuring strategy; Severe non-compliance: Directly upgrade to Level 4 global topology reconstruction; If recovery verification fails twice in a row: Force isolation of all suspicious nodes and activation of the backup agent cluster to take over the task.

[0105] For isolated agents, they are placed in a sandbox environment for in-depth diagnostics. If the diagnosis indicates a temporary fault, their trust score is gradually restored and they are reconnected to the network after stabilizing in the sandbox. Simultaneously, the system's trust fusion model parameters (such as the weights of the MLP) can be fine-tuned online based on global security events, achieving adaptive evolution.

[0106] The following specific application examples will further illustrate the solution of the present invention: financial risk control scenario.

[0107] The following describes four intelligent agents (Customer, Customer, and Controller) used by banks to process cross-border M&A loan applications. Agent Risk Agent Compliance Agent Approval Agent Taking [example] as an example, the specific working process of the present invention will be demonstrated.

[0108] Custom Agent Responsible for verifying customer identity, verifying qualifications, and screening for fraud.

[0109] Risk Agent Responsible for assessing the probability of default, risk level, and loss estimation of loan projects.

[0110] Compliance Agent : Responsible for checking whether transactions comply with domestic and foreign regulatory rules and industry compliance requirements.

[0111] Approval Agent Based on the output of the preceding intelligent agent, execute the final loan approval / rejection decision.

[0112] The normal collaboration process is as follows: Customer verification agent (CustomAgent → Risk assessment agent (Risk) Agent → Compliance Review Agent Agent → Approval Decision-Making Intelligent Agent (Approval) Agent ).

[0113] First, perform system initialization configuration.

[0114] 1. Trust model initialization: For each agent, the initial values ​​are: Direct Interaction Trust (DIT) = 0.90, Indirect Recommendation Trust (IRT) = 0.85, and Behavioral Entropy Trust (BET) = 0.88. The MLP-based integrated dynamic trust score (CDTS) is approximately 0.86. Master Trust Threshold T primary =0.5, secondary infection threshold T secondary =0.65; The behavioral entropy dynamic baseline is updated using an EMA exponential moving average with a decay coefficient β=0.9.

[0115] 2. Initialization of the behavioral entropy model: Establish a dictionary of atomic behavior units for financial risk control scenarios, and symbolize behaviors such as identity verification, risk calculation, compliance lookup, and decision output as unique IDs; An adaptive n-gram model is adopted, using bigrams for short sequences and trigrams for long sequences, and Laplace smoothing is performed. Initialize the behavioral entropy baseline, and maintain the normal behavioral entropy range between 0.10 and 0.30.

[0116] 3. Topology Reconstruction and Recovery Verification Initialization: The Level 4 restructuring strategy has been activated, and the critical level of the task has been set to the core level. After reconstruction, the verification window length is 20 time steps, the task success rate threshold is ≥90%, and the behavior entropy regression threshold is ≤baseline ±1σ.

[0117] Scenario 1: Single point of failure (injection of prompt words).

[0118] 1. Attack: Attackers control Risk by injecting prompt words. Agent This leads to the output of false low-risk assessment reports.

[0119] 2. Detection: The system of this invention detects Risk. Agent The behavioral entropy surged from a normal 0.22 to 0.92, BET dropped to 0.08, and CDTS plummeted from 0.86 to 0.42, falling below the master trust threshold T. primary =0.5. Meanwhile, the propagation model predicts downstream compliance.Agent The probability of infection P infect =0.72>Secondary infection threshold T secondary =0.65.

[0120] 3. Refactoring: The system triggers a level 3 refactoring. Immediately isolate the abnormal risk. Agent Activate the backup agent Risk Agent_Backup And reconstruct the secure collaboration path: Custom Agent →Risk Agent_Backup →Compliance Agent →Approval Agent .

[0121] 4. Recovery: After reconstruction, the recovery verification showed that the task approval success rate reached 96%, and the system returned to normal operation.

[0122] Scenario 2: Collusive attack by intelligent agents.

[0123] 1. Attack: Malicious AI agent Risk Agent_Mal and Compliance Agent_Mal They have been accumulating IRT by giving each other perfect scores for a long time, and are preparing to work together to get a high-risk loan approved.

[0124] 2. Detection: The system calculates the Risk. Agent_Mal When the IRT (Information Technology Response) is triggered, the anti-collusion recommendation filtering algorithm is activated. It detects that the recommendation score variance is extremely low, and the Risk... Agent_Ma l and Compliance Agent_Mal Subgraphs that exhibit high internal trust and low external connectivity are identified as a conspiracy group.

[0125] 3. IRT Correction and Isolation: The system removes false recommendation scores, Ris kAgent_Mal The IRT dropped sharply from 0.92 to 0.45, and the CDTS was recalculated to 0.48, which is below the primary trust threshold T. primary Two malicious agents were simultaneously identified and flagged.

[0126] 4. Reconstruction: Triggers a level 4 global topology reconstruction. The entire system isolates the collusion subgraph, clears the original graph, and reconstructs based on the remaining trusted Custom subgraphs. Agent Approval Agent And the backup node rebuilds the global collaborative topology.

[0127] 5. Recovery: After global reconstruction, the recovery verification is successful, the fraudulent loan application is correctly rejected, and the system achieves self-healing.

[0128] Example 2 This embodiment provides a system for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, including: The model building module is configured to: build a three-dimensional dynamic trust scoring model for each agent in the multi-agent system. The model includes at least the direct interaction trust based on the quality of historical interactions, the indirect recommendation trust based on third-party evaluation and after anti-collusion filtering, and the behavior entropy trust based on the predictability and consistency of behavior sequences. The comprehensive trust scoring module is configured to obtain a comprehensive trust score for each agent based on a three-dimensional dynamic trust scoring model. The infection risk assessment module is configured to: abstract the multi-agent system into a dynamic collaborative graph with trust weights, establish an abnormal propagation quantification model based on comprehensive trust scores, and assess the infection risk of a single abnormal agent to other agents in the system in real time. The topology reconstruction module is configured to continuously monitor the comprehensive trust score and behavioral entropy of each agent, determine the reconstruction level when an anomaly is detected, and automatically perform the corresponding level of collaborative topology dynamic reconstruction.

[0129] It should be noted that the above modules correspond to the steps in Embodiment 1, and the examples and application scenarios implemented by the above modules and their corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules can be executed in a computer system as part of the system.

[0130] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0131] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0132] A computer-readable storage medium for storing computer instructions that, when executed by a processor, perform the method of Embodiment 1.

[0133] The method in Example 1 can be directly executed by a hardware processor, or it can be executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0134] A computer program product includes a computer program that, when executed by a processor, implements the method in Embodiment 1.

[0135] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0136] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0137] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0138] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0139] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, characterized in that, Includes the following steps: A three-dimensional dynamic trust scoring model is constructed for each agent in a multi-agent system. The model includes at least direct interaction trust based on the quality of historical interactions, indirect recommendation trust based on third-party evaluation and after anti-collusion filtering, and behavioral entropy trust based on the predictability and consistency of behavioral sequences. Based on the three-dimensional dynamic trust scoring model, the comprehensive trust score of each agent is obtained. The multi-agent system is abstracted into a dynamic collaborative graph with trust weights. An anomaly propagation quantification model is established based on the comprehensive trust score to assess the risk of infection of other agents in the system by a single anomalous agent in real time. The system continuously monitors the comprehensive trust score and behavioral entropy of each agent. When an anomaly is detected, the system determines the reconstruction level and automatically performs the corresponding level of collaborative topology dynamic reconstruction.

2. The method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration as described in claim 1, characterized in that, The method for constructing and calculating the behavioral entropy trust degree includes: The behavior sequence of the agent is standardized and represented in a specific way: the agent's original behavior log is parsed into a standardized sequence of atomic behavior units. Based on standardized behavior sequences, similar historical behavior sequences are retrieved based on the context of the current behavior sequence, and an n-gram probability model is constructed. The cross-entropy of the current behavior sequence is calculated using this probability model to obtain the behavior entropy value that represents the uncertainty of the behavior; The inverse mapping value of the behavioral entropy is used as the behavioral entropy confidence level.

3. The method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration as described in claim 1, characterized in that, When calculating the indirect recommendation trust level of an agent, an anti-collusion recommendation filtering algorithm is executed, which specifically includes: Obtain recommendation ratings from other agents for the current agent, and calculate the temporary indirect recommendation trust level of the current agent using the comprehensive trust ratings of other agents as weights. Calculate the variance of all recommendation ratings, and trigger collusion detection when the variance is lower than a preset abnormal consistency threshold; If a set of agents with highly consistent scores is analyzed and meets preset conditions, it is identified as a conspiracy group. In the indirect recommendation trust score calculation, all recommendation scores of the colluding group are removed, and the indirect recommendation trust score is recalculated.

4. The method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration as described in claim 1, characterized in that, When an anomaly is detected, the reconstruction level is determined as follows: The system collects four parameters in real time: comprehensive trust score, behavioral entropy deviation, abnormal infection risk, and task criticality level of the abnormal agent. Different thresholds are set for each of these parameters, and the corresponding reconstruction level is selected according to the threshold. The system also allows low-level strategies to switch to high-level strategies based on risk escalation, and the system achieves step-by-step downgrade through recovery verification.

5. The method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration as described in claim 1, characterized in that, It also includes a system performance recovery verification step after dynamic topology reconstruction: After the reconstruction is completed, start the observation window to collect comprehensive trust score, behavior entropy, task success rate and link status data; The system sequentially determines whether four indicators—trust convergence, behavioral consistency, collaboration efficiency, and anomaly immunity—meet the preset qualification thresholds. The system will be deemed to have recovered and will exit the emergency defense mode only after all indicators have met the standards and a stable observation period has been completed.

6. The method for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration as described in claim 1, characterized in that, It also includes a trust adaptive recovery step: the isolated abnormal agent is sent to the sandbox environment for diagnosis. If it is determined to be a temporary fault, its trust score is gradually restored and it is reconnected to the collaborative network after its behavior stabilizes. At the same time, the parameters of the trust fusion model are fine-tuned online according to global security events to achieve iterative optimization of the trust mechanism.

7. A system for constructing dynamic trust chains and blocking anomaly propagation in multi-agent collaboration, characterized in that, include: The model building module is configured to: build a three-dimensional dynamic trust scoring model for each agent in the multi-agent system. The model includes at least direct interaction trust based on the quality of historical interactions, indirect recommendation trust based on third-party evaluation and after anti-collusion filtering, and behavioral entropy trust based on the predictability and consistency of behavioral sequences. The comprehensive trust scoring module is configured to obtain a comprehensive trust score for each agent based on a three-dimensional dynamic trust scoring model. The infection risk assessment module is configured to: abstract the multi-agent system into a dynamic collaborative graph with trust weights, establish an abnormal propagation quantification model based on comprehensive trust scores, and assess the infection risk of a single abnormal agent to other agents in the system in real time. The topology reconstruction module is configured to continuously monitor the comprehensive trust score and behavioral entropy of each agent, determine the reconstruction level when an anomaly is detected, and automatically perform the corresponding level of collaborative topology dynamic reconstruction.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images