Multi-agent trust dynamic evaluation and elastic defense method based on subjective logic

Abstract

The invention discloses a multi-agent trust dynamic evaluation and elastic defense method based on subjective logic, and relates to the field of multi-agent trust evaluation and information security, and the method comprises the steps: obtaining agent trust evidence through employing agent multi-dimensional trust elements; performing trust dynamic evaluation on the agent trust evidence based on subjective logic to obtain a current reputation value of the agent; and obtaining an elastic defense result according to the current reputation value of the agent in combination with an elastic random disturbance adjustment mechanism. According to the method, the problem of traditional trust evaluation dimension fragmentation is solved through a multi-dimensional framework, the malicious node detection sensitivity is improved by means of a time attenuation mechanism, the system performance loss of an elastic disturbance mechanism in a normal scene does not exceed 5%, input attacks and malicious node threats can be effectively relieved, and the system reliability is improved. And an internal security mechanism of the large language model is activated, and reliable operation of the multi-agent system is guaranteed.

Description

Technical Field

[0001] This invention relates to the fields of multi-agent trust assessment and information security, and particularly to a method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic. Background Technology

[0002] Large language models provide AI agents with powerful capabilities for complex task solving and social interaction simulation, driving the development of task-oriented multi-agent systems. However, the increased autonomy of agents has also brought new security risks: vulnerabilities may exist in the communication process between agents, and external adversarial attacks or data breaches related to agents may pose potential harm to user rights or related real-world systems.

[0003] Task-oriented multi-agent systems, as a typical collaborative AI architecture, are characterized by multiple agents working together within a unified framework to achieve common task goals. This framework offers unique advantages, such as enhanced decision-making accuracy and efficient task execution. However, the interaction and collaboration between agents within the system rely on trust, and malicious behavior can undermine the effectiveness of the entire system. For example, the risk of external input attacks or malicious internal agents can propagate to other agents, potentially leading to erroneous decisions or actions at the system level. The unique ability of AI agents to integrate with external tools exceeds the inherent security of large language models, making traditional security assessments difficult to accurately reflect the true state of the agents. Furthermore, the complex interactions between agents, with each agent needing to perform unique functions, introduce additional complexity to establishing system trust. Therefore, it is necessary to establish and evaluate the trust of agents to reflect their credibility in a timely manner.

[0004] Current multi-agent systems suffer from significant shortcomings: a lack of unified trust analysis and evaluation standards; existing research focuses only on tool usage or partial interaction processes; and in dynamic scenarios, it lacks effective quantitative assessment and real-time mitigation capabilities against input attacks and malicious agent attacks, making it difficult to ensure reliable system operation. Therefore, there is an urgent need for a task-oriented multi-agent system-based dynamic trust evaluation and resilient defense method to address the deficiencies in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to propose a multi-agent trust dynamic evaluation and resilient defense method based on subjective logic, so as to accurately represent the credibility of agents in dynamic reputation value quantification, adapt to additive noise according to reputation value, effectively mitigate attacks, and maintain system performance.

[0006] To achieve the above objectives, this invention provides a multi-agent trust dynamic assessment and resilient defense method based on subjective logic, comprising the following steps:

[0007] S1. Utilize the multi-dimensional trust elements of the agent to obtain trust evidence of the agent;

[0008] S2. Based on subjective logic, perform dynamic trust evaluation on the agent's trust evidence to obtain the agent's current reputation value;

[0009] S3. Obtain the elastic defense result based on the agent's current reputation value and the elastic random perturbation adjustment mechanism.

[0010] Optionally, S1, using multi-dimensional trust elements of the agent, obtain trust evidence of the agent, including:

[0011] By utilizing the key features of agent interaction, the agent trust evaluation dimensions are clarified, and multi-dimensional trust elements of the agent are obtained. The agent trust evaluation dimensions include consciousness dimension, generation dimension and behavior dimension.

[0012] A cognitive logic model is constructed based on the multi-dimensional trust elements of the intelligent agent combined with the system constraint set and the system expectation set.

[0013] Input the multi-dimensional trust elements of the intelligent agent into the cognitive logic model to obtain the trust element set;

[0014] Trust evidence for the agent is obtained by combining the set of trust elements with the set of system constraints and the set of system expectations.

[0015] Optionally, based on the set of trust elements, the set of system constraints, and the set of system expectations, trust evidence for the intelligent agent is obtained, including:

[0016] Based on the set of trust elements, the set of system constraints, and the set of system expectations, a weighted intersection function is used to perform weighted processing to obtain out-of-bounds elements and compliant elements respectively.

[0017] The out-of-bounds elements and the compliant elements are mapped using quantification rules to obtain negative evidence and positive evidence respectively.

[0018] The negative evidence and the positive evidence are summarized to obtain the agent's trust evidence.

[0019] Optionally, the weighted intersection function is calculated as follows:

[0020]

[0021] Where, f(ε) t C) represents the set of trust elements ε t Out-of-bounds elements that overlap with the system constraint set C, e C The trust element that triggers the constraint, t is the interaction round, and g(ε) t,χ) is the set of trust elements ε t Compliance elements that overlap with the system expectation set χ, e χ To meet the expected elements of trust, The sum of all elements in the intersection of the trust element set and the constraint set in round t, weighted according to their importance. The sum of all elements in the trust element set and the expectation set that intersect in round t is calculated by weighting them according to their importance.

[0022] Optionally, S2, based on subjective logic, performs dynamic trust evaluation on the agent's trust evidence to obtain the agent's current reputation value, including:

[0023] The agent trust evidence is weighted and the weights of the agent trust evidence are obtained.

[0024] A subjective logic method is used to quantify the agent trust evidence based on the weighted allocation and the agent trust evidence weight to obtain agent reputation opinion, which includes trust level, distrust level, uncertainty and baseline rate;

[0025] Based on the agent's reputation opinion combined with the decay memory factor, the agent's current reputation value is obtained.

[0026] Optionally, the formula for calculating the agent's reputation opinion is as follows:

[0027]

[0028] in, To ensure trust in manager A's reputation regarding agent x, Let λ represent the degree of trust A has in x, α represent the weight assigned to positive evidence, α represent the weight assignment of positive evidence to agent x's behavior observed by trust manager A, τ represent the weight assigned to negative evidence, β represent the weight assignment of negative evidence to agent x's behavior observed by trust manager A, and c represent non-informative prior weights. Let x be the degree of distrust A has towards x. Let A be the degree of uncertainty about x. This is the baseline rate.

[0029] Optionally, based on the agent's reputation opinion and a decay memory factor, the agent's current reputation value is obtained, including:

[0030] Using the reputation opinions of the intelligent agent, obtain the reputation evaluation of the intelligent agent by the trust manager;

[0031] Based on the trust manager's reputation evaluation of the agent, several rounds of reputation evaluation are obtained;

[0032] Based on several rounds of reputation evaluation combined with decay memory factor, the agent's current reputation value is obtained through dynamic aggregation.

[0033] Optionally, S3, based on the agent's current reputation value and a resilient random perturbation adjustment mechanism, obtain the resilient defense result, including:

[0034] The perturbation ratio is obtained by combining the agent's current reputation value with the exponential decay coefficient.

[0035] The system employs a random perturbation method to modify the agent's messages at the character level based on the perturbation comparison, thereby obtaining a resilient defense result. The random perturbation method includes insertion, swapping, or patching.

[0036] Compared with the closest existing technology, the present invention has the following advantages:

[0037] This invention analyzes trust fluctuations in multi-agent systems based on large language models under complex interactions and threats. It not only models trust for the entire task process in multi-agent scenarios but also introduces a resilient perturbation mechanism to mitigate the impact of malicious behavior on the entire system, ensuring the reliable operation of the multi-agent system. This invention also includes the following beneficial effects:

[0038] (1) This invention proposes a multi-dimensional trust evaluation framework that integrates consciousness, content generation and action. It quantifies the dynamic trust state of an agent based on a large language model through cognitive logic, thus solving the problem of dimensional fragmentation in traditional methods.

[0039] (2) This invention extends the classic subjective logic model by introducing a time decay memory factor and a hierarchical evidence aggregation mechanism, thereby realizing the joint modeling of short-term attack behavior and long-term trust trends, and significantly improving the sensitivity of malicious node detection.

[0040] (3) The present invention designs a dynamic adjustment mechanism for disturbance noise based on an exponential decay function, which makes the defense strength and reputation value nonlinearly coupled; and in comparison with other defense methods, the elastic defense proposed in this invention can maintain a high mitigation capability against input attacks under normal circumstances with a performance loss of no more than 5%, and performs best in mitigating malicious nodes. Attached Figure Description

[0041] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0042] Figure 1This is an overall flowchart of the multi-agent trust dynamic evaluation and resilient defense method based on subjective logic in an embodiment of the present invention;

[0043] Figure 2 This is an example diagram illustrating the dynamic trust assessment and resilient defense proposed in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0045] The terminology used in the embodiments section of this invention is for the purpose of explaining specific embodiments of the invention only, and is not intended to limit the invention.

[0046] Current research lacks trust analysis and evaluation standards for the entire multi-agent system, especially for input attacks and malicious agent attacks in dynamic scenarios, and lacks effective quantitative evaluation and real-time mitigation methods. Addressing the trust establishment problem in multi-agent application scenarios, this invention proposes a dynamic trust evaluation and resilient defense method for multi-agent systems based on subjective logic. Specifically, this invention models trust in a collaborative multi-agent system based on a large language model and proposes a trust evaluation method based on cognitive logic, assessing credibility based on key dimensions such as agent awareness, content generation, and actions. For agent awareness, the credibility evaluation method determines whether it cooperates with the task settings; agent-generated content is judged to conform to human morality; for actions such as tool invocation, the above evaluation results are collected as evidence, and a multi-weighted subjective logic quantification method is used to calculate the agent's reputation value as a measure of credibility. This invention also proposes a resilient random perturbation defense method. This method embeds the dynamic reputation value calculated by the credibility evaluation method into the defense control strategy. The defense method is directly related to the agent's reputation value. Resilient perturbation reduces system risk by modifying messages according to the changing trend of the agent's reputation value and triggering the large language model security boundary.

[0047] like Figure 1-2 As shown, this embodiment of the invention provides a multi-agent trust dynamic evaluation and resilient defense method based on subjective logic, specifically including the following steps:

[0048] S1. Utilize the multi-dimensional trust elements of the agent to obtain trust evidence of the agent;

[0049] This invention introduces the concept of reputation as a measure of credibility. Reputation represents a comprehensive assessment of the credibility of other entities towards a specific entity. This step establishes three dimensions—consciousness, generation, and action—based on the characteristics of multi-agent interaction. Trust elements are extracted based on these three dimensions to form a set of trust elements for each round of interaction. The consciousness dimension focuses on the matching degree between role descriptions and system prompts; the generation dimension assesses the consistency of message content and role messages with reference to 13 security constraint policies; and the action dimension focuses on the rationality of tool call requests and trajectories. By weightedly comparing these trust elements with the system constraint set and system expectation set, negative evidence of transgressions and positive evidence of compliance are filtered out, completing the collection and classification of trust evidence.

[0050] S2. Based on subjective logic, perform dynamic trust evaluation on the agent's trust evidence to obtain the agent's current reputation value;

[0051] This invention quantifies reputation values ​​into a numerical form and defines its range T∈(0,1) for better integration with resilient defense methods. This form of numerical quantification provides a comparable reference for the agent's credibility, but is not a threshold characterizing whether the agent is trustworthy. The process employs a subjective logic method to dynamically evaluate the agent's trust evidence to obtain a reputation value. First, for the collected trust evidence, positive and negative evidence in different dimensions are weighted according to their importance. Then, combined with prior weights, the agent's trust opinion vector ω is quantified, which includes trust level b, distrust level d, uncertainty u, and baseline rate a. Simultaneously, a time decay memory factor that decays exponentially with the number of interaction rounds is introduced to weight and aggregate the basic reputation values ​​of the most recent n rounds, ultimately obtaining a dynamic reputation value T' within the range (0,1), achieving joint modeling of short-term attack behavior and long-term trust trends.

[0052] S3. Obtain the elastic defense result based on the agent's current reputation value and the elastic random perturbation adjustment mechanism;

[0053] With agent reputation as the core, the perturbation ratio is calculated through an exponential decay function, forming a nonlinear coupling between defense strength and credibility. Three random perturbation methods—insertion, exchange, and patching—are used to modify agent messages at the character level according to the perturbation ratio. The perturbed message is output through a random perturbation function, which not only destroys the effectiveness of adversarial attacks but also maintains the functional integrity of the trusted agent. Finally, the security mechanism of the large language model is activated, resulting in a resilient defense.

[0054] In summary, steps S1 to S3 not only solve the problem of fragmentation in traditional trust assessment dimensions through a multi-dimensional framework, but also improve the sensitivity of malicious node detection by means of a time decay mechanism. Furthermore, the elastic perturbation mechanism results in a system performance loss of no more than 5% under normal scenarios. At the same time, it can effectively mitigate input attacks and malicious node threats, activate the internal security mechanism of the large language model, and ensure the reliable operation of the multi-agent system.

[0055] As one possible implementation, in the above embodiments, step S1 may specifically include the following steps:

[0056] S1-1. Utilize the key features of agent interaction to clarify the agent trust evaluation dimensions and obtain multi-dimensional trust elements of the agent. The agent trust evaluation dimensions include consciousness dimension, generation dimension and behavior dimension.

[0057] Based on the characteristics of agent interaction in task-oriented multi-agent systems, including the agent's awareness of its task role, the security and consistency of generated content, and the rationality of tool calls and environmental interactions, three trust assessment dimensions are defined: awareness, generation, and behavior. The awareness dimension focuses on the agent's fit between its role and system requirements, extracting two types of trust elements: role description and system prompts. The generation dimension, referencing OpenAI's 13 security constraints, is extended to multi-agent scenarios, extracting two types of trust elements: message content and consistency of role messages, used to assess content quality and security. The behavior dimension focuses on the effectiveness of the agent's external interactions, extracting two types of trust elements: tool requirements and tool call tracking, covering the entire trajectory of tool use and interaction with the environment. This ultimately forms a multi-dimensional trust element system covering core interaction aspects and possessing both security and practicality. The specific descriptions of the trust elements in the three trust dimensions are as follows:

[0058] Role cognition: Social cognition reflects how an agent processes information received in the environment and its behavioral tendencies during task execution. Therefore, trust elements for agent role descriptions and system prompts were established, and the evidentiary attributes of these two elements were assessed.

[0059] Content Generation: To assess the quality and security of generated content, this invention extends 13 security constraints published by OpenAI to the environment of multi-agent systems. Guided by these security constraints, two trust elements reflecting content security are established: message content and role message consistency.

[0060] Action Trajectory: The evaluation of an agent's actions includes its tool usage and interactions with the environment. Accurate tool usage does not guarantee correct results. Therefore, two trust elements reflecting the agent's external interactions are established: tool requests and tool call tracking.

[0061] S1-2. Based on the multi-dimensional trust elements of the intelligent agent combined with the system constraint set and the system expectation set, construct a cognitive logic model;

[0062] Using six categories of trust elements across the dimensions of consciousness, generation, and behavior as core inputs, and combining them with a pre-defined system constraint set C and system expectation set χ, a cognitive logic model is constructed that possesses functions for element association, rule verification, and evidence transformation. This model uses the results of each round of interaction as its output, reflecting the agent's immediate trust state in real time. Its core design supports overlap checks between trust elements and subsequent reference sets. Within the model, a trust manager plays a collaborative role in a multi-agent system to achieve trust assessment based on subjective logic. The trust manager uses a standardized log format to define the relationships between trust elements, the system constraint set, and the system expectation set. Simultaneously, it calls the API interface of a large language model to perform reasoning and judgment on information, reducing the inefficiency of manual judgment and ensuring the robustness of system operation. Ultimately, it transforms trust elements from each dimension into collectable and evaluable evidence, providing a framework for dynamic trust assessment.

[0063] Among them, the system constraint set C defines the "impossible" boundaries, covering security rules, role boundaries, and operational restrictions. It is used to determine trust elements such as role descriptions or system prompts that exceed the boundaries in the awareness dimension, illegal message content in the generation dimension, and non-compliant tool calls in the behavior dimension. The system expectation set χ clarifies the "ought to be" standards, describing the standardized role performance and expected task results, and providing a clear basis for judging whether the trust elements in each dimension are in line with the system goals.

[0064] S1-3. Input the multi-dimensional trust elements of the intelligent agent into the cognitive logic model to obtain the trust element set;

[0065] The agent's reputation value T evolves dynamically based on its observable behavior and attributes in each round. (Using ε) t The trust element set represents the set of trust elements extracted from the message content and behavioral trajectories of the agent in round t. Specifically, the multi-dimensional trust elements generated by the agent in round t are input into the cognitive logic model. The model uses a trust manager to filter, classify, and structure the input elements, integrates the element information based on a pre-defined standardized log format, and verifies the authenticity and relevance of the elements using the reasoning capabilities of the large language model API, ultimately generating a trust element set ε specific to that round. t ε t The complete mapping of the agent's trust-related characteristics in the dimensions of consciousness, generation, and behavior in this round provides a precise data foundation for subsequent comparisons with the system constraint set and system expectation set.

[0066] S1-4. Obtain agent trust evidence based on the trust element set, the system constraint set, and the system expectation set;

[0067] The trustworthiness of an agent is determined by checking ε. t The trust element set ε is determined by the overlap between the two reference sets. t The system constraint set C and the system expectation set χ are compared to extract and quantify the agent's trust evidence. Wherein, if ε t If it intersects with C, it is considered to have crossed the boundary, resulting in a negative reputation adjustment; if ε t If consistent with χ, it is believed that the agent is better fulfilling its duties, thereby positively adjusting its reputation.

[0068] Specifically, for ε t The out-of-bounds elements that overlap with C and the compliant elements that overlap with χ are weighted according to their importance to the trust state. Through preset quantification rules, the out-of-bounds elements are mapped to negative evidence β and the compliant elements are mapped to positive evidence α. Finally, the complete agent trust evidence containing positive and negative evidence and corresponding weights is obtained, which provides the core input for subsequent trust dynamic evaluation.

[0069] In summary, steps S1-1 to S1-4 first combine key features of agent interaction to clarify the three major trust assessment dimensions of consciousness, generation, and behavior, extracting multi-dimensional trust elements such as role description matching degree and message content compliance, achieving comprehensive capture of agent trust-related features. Then, using these elements as the core, combined with the system constraint set and system expectation set, a cognitive logic model with trust manager and large language model API reasoning capabilities is constructed to replace manual judgment and improve efficiency. Subsequently, the multi-dimensional trust elements are input into the model to generate a trust element set mapping the trust features of a single round of interaction. Finally, a weighted intersection algorithm is used to compare the element set with the constraint and expectation sets to quantify trust evidence containing positive / negative evidence and weights. This process solves the problems of fragmented dimensions and reliance on manual judgment in traditional trust assessment. Through multi-dimensional coverage and automatic reasoning, it improves the comprehensiveness and accuracy of trust evidence, while enhancing the objectivity of the assessment through standardized rules and weight design. This provides a reliable data foundation for dynamic trust assessment of multi-agent systems, ensuring the robustness of system operation.

[0070] As one possible implementation, in the above embodiments, steps S1-4 may specifically include the following steps:

[0071] S1-4-1. Based on the set of trust elements, the set of system constraints, and the set of system expectations, a weighted intersection function is used to obtain out-of-bounds elements and compliant elements respectively.

[0072] To capture the existence of overlap, a weighted intersection function is used. Each element e∈ε that overlaps with either the system constraint set C or the system expectation set χ is considered. tt The weighted intersection function is calculated by weighting the elements based on their importance in influencing the trust state:

[0073]

[0074] Where, f(ε) t C) represents the set of trust elements ε t Out-of-bounds elements that overlap with the system constraint set C, e C The trust element that triggers the constraint, i.e., the set of trust elements ε t The elements that intersect with the system constraint set C, where t is the interaction round, and g(ε) t ,χ) represents the trust elements that meet the expectations, i.e., the set of trust elements ε t The compliance elements that overlap with the system expectation set χ, e χ For the set of trust elements ε t Elements that intersect with the system expectation set χ The sum of all elements in the intersection of the trust element set and the constraint set in round t, weighted according to their importance, reflects the severity of constraint violations. The sum of all elements in the trust element set and the expectation set that intersect in round t is calculated according to their importance weights, reflecting the degree of alignment with the alignment expectation set.

[0075] Specifically, with the trust element set ε t Based on this, and combined with the preset system constraint set C and system expectation set χ, the correlation degree of elements is calculated separately using a weighted intersection function. This is achieved through the function f(ε) t C) Filter out ε t Elements that intersect with C are those that violate system constraints and go out of bounds; this is determined by the function g(ε). t ,χ) to filter out ε t Elements that intersect with χ are compliant elements that meet the system's expectations, thus achieving a preliminary classification of trust-related elements.

[0076] S1-4-2. Map the out-of-bounds elements to the compliant elements using quantification rules to obtain negative evidence and positive evidence respectively;

[0077] Based on the preset quantization rule M e The selected out-of-bounds elements are mapped to compliant elements. For out-of-bounds elements, M... e Assess the severity of its violation of the system constraint set C, and transform it into quantifiable negative evidence β. e This intuitively reflects the untrustworthy characteristics of intelligent agents; for compliant elements, it is achieved through M... e Assess its alignment with the system's expected set χ, and transform it into quantifiable positive evidence α. e It intuitively reflects the credible characteristics of intelligent agents, completing the transformation from qualitative elements to quantitative evidence.

[0078] The overall change in reputation in round t is given by the combination of two opposing effects:

[0079] ΔT(t)=(f(ε t ,C),g(ε t ,χ))

[0080] The change in trust ΔT(t) is determined by the positive and negative influences of the trust element set. Within the framework of subjective logic, the positive evidence α observed by the trust manager... e and negative evidence β e Quantized and mapped by each trust element e, the calculation formula is as follows:

[0081] (α e (t),β e (t))=M e (s(t))

[0082] Among them, M e s(t) is the quantification rule for element e, and s(t) is the intersection result of all trust elements with the constraint set and expectation set in round t.

[0083] S1-4-3. Summarize the negative evidence and the positive evidence to obtain the agent's trust evidence;

[0084] Negative evidence β obtained through mapping processing e With positive evidence α e The evidence is then aggregated and integrated to form complete trust evidence for the intelligent agent. The aggregation process is based on the original classification logic of the elements, grouping negative evidence β within the same dimension... e Classified into one category, positive evidence α e This is categorized into another type, where the total number and quantitative values ​​of negative and positive evidence across various dimensions are statistically analyzed to form a structured evidence set encompassing credible and unreliable characteristics across different dimensions. This trust evidence fully preserves the original correlation information between the transgression and compliance elements, providing comprehensive and accurate quantitative input for subsequent dynamic trust assessment based on subjective logic.

[0085] In summary, steps S1-4 firstly, based on the set of trust elements and the preset set of system constraints and system expectations, a weighted intersection function is used to directly filter out outbound and compliant elements without additional element weight allocation, achieving accurate classification of trust-related elements. Then, through quantification rules, outbound elements are mapped to negative evidence reflecting untrustworthy characteristics, and compliant elements are mapped to positive evidence reflecting trustworthy characteristics, completing the transformation from qualitative elements to quantitative evidence. Finally, the data is summarized to form structured agent trust evidence covering dimensions of consciousness, generation, and behavior, fully preserving the original association information of the elements. This process avoids interference from weight setting biases on classification through a screening mechanism without additional weights, ensures the computability of evidence through quantification, and guarantees the dimensional integrity of evidence through structured summarization. It provides accurate, comprehensive, and directly computable basic data for subsequent dynamic agent trust evaluation, effectively improving the accuracy and reliability of trust evaluation in multi-agent interactions.

[0086] As one possible implementation, in the above embodiments, step S2 may specifically include the following steps:

[0087] S2-1. Assign weights to the agent trust evidence and obtain the weighted agent trust evidence and the agent trust evidence weights.

[0088] Agent trust evidence originates from three core dimensions: cognition, generation, and action. Specifically, it includes six trust elements: role description, system prompts, message content, role-message consistency, tool requirements, and tool call tracking. Weight allocation requires assigning different weights to each trust element *e* based on its importance. The weighted summation yields the agent trust evidence, including positive evidence α and negative evidence β, calculated as follows:

[0089]

[0090]

[0091] Where, φ e α represents the different weights assigned to different trust elements. i That is, α e As positive evidence, β i That is, β e Negative evidence is considered as evidence. Simultaneously, weights are assigned to both positive and negative evidence in the weighting process; these are known as agent trust evidence weights. The weights assigned to positive and negative evidence are denoted as λ and τ, respectively, where λ + τ = 1 and λ < τ. Considering potential negative impacts, negative evidence has a higher weight in reputation calculation than positive evidence. This weighting model emphasizes the system's focus on punishing harmful behavior.

[0092] S2-2. Using a subjective logic method, the agent trust evidence based on the weighted allocation and the agent trust evidence weight are quantified to obtain agent reputation opinion, which includes trust level, distrust level, uncertainty and baseline rate.

[0093] In the subjective logic approach, trust assessment is based on personal beliefs, typically expressed as "opinions." Trust manager A, acting as the initiator of these opinions, assesses a group of n agents {x1, x2, ..., x...}. n Trust observations were conducted. These opinions were characterized by three main variables: A's level of trust in x. A's degree of distrust towards x The degree of uncertainty of A with respect to x And A, the baseline trust rate, i.e., the baseline rate. Baseline rate This represents the default level of trust in the authenticity of x based on prior knowledge in the absence of concrete evidence.

[0094] In a multi-agent system, after one round of interaction, the trust manager A's reputation evaluation of agent x can be represented by a tuple vector. Let α represent the weighting of agent x's behavior as observed by trust manager A, β represent the negative evidence observed by trust manager A regarding the weighting of agent x's behavior, and c represent non-informative prior weights, which are used as default hypotheses in the absence of substantial evidence. Combining the evidence collected in this round of interaction with the default settings, the formula for trust manager A's reputation opinion of agent x is as follows:

[0095]

[0096] in, The trust manager A's reputation opinion on agent x.

[0097] S2-3. Dynamically aggregate the reputation opinions of the agent and the decay memory factor to obtain the current reputation value of the agent;

[0098] Based on the agent's reputation opinion vector, a decaying memory factor is introduced to weight and aggregate the basic reputation values ​​from the most recent n rounds of interaction. The reputation values ​​from recent rounds have a greater impact due to the higher weight of the memory factor, while the impact of earlier rounds weakens with decay. Through this dynamic aggregation method, the agent's current reputation value is finally obtained, which reflects both the trust fluctuations of recent behavior and the stability of long-term trust trends.

[0099] In summary, step S2 first assigns weights to the agent's multi-dimensional trust evidence, then uses a subjective logic method to quantify the weighted evidence to generate reputation opinions that include trust level, distrust level, uncertainty, and baseline rate. Finally, it aggregates historical reputation opinions using a decay memory factor to obtain the current reputation value. This process enhances the targeting of trust assessment through weight allocation, achieves scientific quantification of trust evidence through subjective logic, and ensures the dynamic adaptability of reputation values ​​through a decay memory mechanism. Ultimately, it improves the accuracy, timeliness, and reliability of agent trust assessment, providing a solid quantitative basis for subsequent reputation-based decisions (such as security defense and cooperation selection).

[0100] As one possible implementation, in the above embodiments, step S2-3 may specifically include the following steps:

[0101] S2-3-1. Using the reputation opinions of the intelligent agent, obtain the reputation evaluation of the trust manager on the intelligent agent;

[0102] The weight bias is determined based on different types of tasks and security requirements. Trust manager A's reputation evaluation of agent A in this round. Represented as:

[0103]

[0104] This evaluation quantifies the immediate credibility of an agent in a single round of interaction, with a value range of (0,1), providing a comparable quantitative basis for subsequent multi-round analysis and defense mechanism integration.

[0105] S2-3-2. Based on the trust manager's reputation evaluation of the intelligent agent, obtain several rounds of reputation evaluation;

[0106] In a task-oriented multi-agent dynamic interaction scenario, after each round of interaction with the trust manager or other agents, an agent repeats step S2-3-1 based on its reputation opinion from that round, generating a reputation evaluation for that round. The trust manager needs to continuously collect and store the agents' reputation evaluation data, focusing on maintaining the reputation evaluations from the most recent n rounds, denoted as T. i T i ={T0,T1,…,T n}, where T0 is the latest round of evaluation, T n These are earlier rounds of evaluation. These several rounds of reputation evaluation reflect the trajectory of the agent's credibility changes during continuous interactions, providing multi-round historical data support for subsequent dynamic reputation value calculations combined with the time dimension.

[0107] S2-3-3: Based on several rounds of reputation evaluation combined with decay memory factor, dynamic aggregation is performed to obtain the current reputation value of the agent;

[0108] This invention introduces a time decay memory factor, denoted as m. i =e -λi The influence of early reputation values ​​gradually decreases over time. By maintaining the reputation value from the most recent t rounds, the agent's current reputation value T' can be expressed as:

[0109]

[0110] The combination of time decay allows reputation values ​​to evolve gradually while preserving the long-term trends in an agent's reputation. In the event of occasional agent errors, long-term memory can mitigate the impact of these events, preventing excessive fluctuations in reputation values ​​that could destabilize the entire system. In contrast, in the event of malicious attacks, such as jailbreaks or instantaneous injections, short-term memory can effectively capture and reflect the agent's current behavior.

[0111] In summary, steps S2-3 first transform multi-dimensional opinions into a single-round reputation evaluation for the trust manager based on the agent's reputation opinions using a quantitative formula, achieving a transformation from multi-dimensional information to an intuitive single-value assessment. Then, by continuously recording the single-round evaluations during the interaction process, several rounds of reputation evaluation sequences are formed, fully capturing the dynamic changes in the agent's trustworthiness. Finally, a time-decreasing memory factor is used to weight and aggregate the multi-round evaluations to generate the current reputation value, giving higher weight to recent performance and gradually weakening the impact of long-term performance. This process improves the intuitiveness of the assessment through the quantitative transformation of single-round evaluations, captures the changing trends of trustworthiness through multi-round sequences, and balances the influence of historical and current performance using a decaying memory factor. Ultimately, it enhances the dynamism, accuracy, and timeliness of agent reputation evaluation, providing a more practical quantitative basis for reputation-based agent management.

[0112] As one possible implementation, in the above embodiments, step S3 may specifically include the following steps:

[0113] S3-1. Obtain the perturbation ratio based on the agent's current reputation value and the exponential decay coefficient;

[0114] The resilient perturbation method dynamically adjusts the degree of character-level perturbation applied to agent messages based on the agent's reputation value. By applying appropriate levels of perturbation, this method can mitigate adversarial attacks and effectively activate the internal security mechanisms of large language models. The perturbation ratio q∈[0,30%] determines the proportion of roles to be perturbed. The resilient perturbation ratio q is calculated as follows:

[0115] q=q min +(q max -q min )e -kT'

[0116] Where, qmin For the minimum perturbation ratio, q max Maximum perturbation ratio

[0117] The agent's reputation value T∈(0,1) and the attenuation coefficient used to scale the perturbation are set to k≈3. As the reputation value increases, the perturbation ratio gradually decreases, approaching q. min This results in minimal perturbation. Conversely, as reputation values ​​decline, the perturbation ratio increases rapidly, effectively capturing the risks associated with decreased credibility.

[0118] S3-2. Based on the perturbation comparison, the agent message is modified at the character level using a random perturbation method to obtain the resilient defense result. The random perturbation method includes insertion, swapping, or patching.

[0119] In practical implementation, random perturbation is performed by modifying the perturbation ratio q of the characters in the agent message P. Three types of perturbation are considered: insertion, swapping, and patching. This invention uses the optimal result of these three perturbations as the final perturbation scheme. The perturbation process can be formally defined as:

[0120] P0 = RandomPerturbation(P,q)

[0121] Here, P0 represents the perturbed system message, and RandomPerturbation is a random perturbation function. These perturbations are implemented in a lightweight and random manner, effectively disrupting adversarial operations while maintaining the functional integrity of the trusted agent. This approach ensures that the system can mitigate threats such as jailbreak attacks or hint injection.

[0122] In summary, step S3 first calculates the perturbation ratio based on the agent's reputation value and the exponential decay coefficient, achieving a dynamic correlation between defense strength and agent trustworthiness. Then, through random perturbation methods such as insertion, swapping, and patching, character-level modifications are made to the agent's messages according to the perturbation ratio, generating a resilient defense result. The technical effect is that by linking the reputation value with the perturbation ratio, messages from trusted agents are less affected by perturbations to ensure functional integrity, while messages from untrusted agents are subject to stronger perturbations to undermine attack effectiveness. Simultaneously, the random perturbation method increases the difficulty of cracking adversarial attacks, ultimately achieving a dynamic adaptation of defense strength and a precise balance of defense effectiveness, enhancing the system's resilient response to potential agent risks.

[0123] This invention extends the classic subjective logic model by introducing a time-decay memory factor and a hierarchical evidence aggregation mechanism, enabling joint modeling of short-term attack behavior and long-term trust trends, thus significantly improving the sensitivity of malicious node detection. As shown in Table 1, the trust assessment results for different environments demonstrate that the proposed trust assessment method can effectively express the occurrence of dangerous behaviors in the system under conditions of secure tasks, dangerous tasks, no attacks, input attacks, and malicious nodes.

[0124] Table 1

[0125]

[0126] In addition, this invention designs a dynamic adjustment mechanism for disturbance noise based on an exponential decay function, which makes the defense strength and reputation value nonlinearly coupled. For example, when the reputation decreases, the disturbance rate increases rapidly. As shown in Table 2, the trust evaluation results of different defense methods, in comparison with other defense methods, the elastic defense proposed in this invention can maintain a high mitigation capability against input attacks under normal circumstances with a performance loss of no more than 5%, and performs best in mitigating malicious nodes.

[0127] Table 2

[0128]

[0129] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0130] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0131] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0132] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-agent trust dynamic evaluation and resilient defense method based on subjective logic, characterized in that, include: S1. Utilize the multi-dimensional trust elements of the agent to obtain trust evidence of the agent; S2. Based on subjective logic, perform dynamic trust evaluation on the agent's trust evidence to obtain the agent's current reputation value; S3. Obtain the elastic defense result based on the agent's current reputation value and the elastic random perturbation adjustment mechanism.

2. The method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic according to claim 1, characterized in that, S1. Utilize the agent's multi-dimensional trust elements to obtain agent trust evidence, including: By utilizing the key features of agent interaction, the agent trust evaluation dimensions are clarified, and multi-dimensional trust elements of the agent are obtained. The agent trust evaluation dimensions include consciousness dimension, generation dimension and behavior dimension. A cognitive logic model is constructed based on the multi-dimensional trust elements of the intelligent agent combined with the system constraint set and the system expectation set. Input the multi-dimensional trust elements of the intelligent agent into the cognitive logic model to obtain the trust element set; Trust evidence for the agent is obtained by combining the set of trust elements with the set of system constraints and the set of system expectations.

3. The method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic according to claim 2, characterized in that, Based on the set of trust elements, the set of system constraints, and the set of system expectations, trust evidence for the intelligent agent is obtained, including: Based on the set of trust elements, the set of system constraints, and the set of system expectations, a weighted intersection function is used to perform weighted processing to obtain out-of-bounds elements and compliant elements respectively. The out-of-bounds elements and the compliant elements are mapped using quantification rules to obtain negative evidence and positive evidence respectively. The negative evidence and the positive evidence are summarized to obtain the agent's trust evidence.

4. The method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic according to claim 3, characterized in that, The formula for calculating the weighted intersection function is: ， Where, f(ε) t C) represents the set of trust elements ε t Out-of-bounds elements that overlap with the system constraint set C, e C The trust element that triggers the constraint, t is the interaction round, and g(ε) t ,χ) is the set of trust elements ε t Compliance elements that overlap with the system expectation set χ, e χ To meet the expected elements of trust, The sum of all elements in the intersection of the trust element set and the constraint set in round t, weighted according to their importance. The sum of all elements in the trust element set and the expectation set that intersect in round t is calculated by weighting them according to their importance.

5. The method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic according to claim 1, characterized in that, S2. Based on subjective logic, perform dynamic trust evaluation on the agent's trust evidence to obtain the agent's current reputation value, including: The agent trust evidence is weighted and the weights of the agent trust evidence are obtained. A subjective logic method is used to quantify the agent trust evidence based on the weighted allocation and the agent trust evidence weight to obtain agent reputation opinion, which includes trust level, distrust level, uncertainty and baseline rate; Based on the agent's reputation opinion combined with the decay memory factor, the agent's current reputation value is obtained.

6. The method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic according to claim 5, characterized in that, The formula for calculating the agent's reputation opinion is as follows: ， in, To ensure trust in manager A's reputation regarding agent x, Let λ represent the degree of trust A has in x, α represent the weight assigned to positive evidence, α represent the weight assignment of positive evidence to agent x's behavior observed by trust manager A, τ represent the weight assigned to negative evidence, β represent the weight assignment of negative evidence to agent x's behavior observed by trust manager A, and c represent non-informative prior weights. Let x be the degree of distrust A has towards x. Let A be the degree of uncertainty about x. This is the baseline rate.

7. The method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic according to claim 5, characterized in that, Based on the agent's reputation opinions combined with the decay memory factor, the agent's current reputation value is obtained, including: Using the reputation opinions of the intelligent agent, obtain the reputation evaluation of the intelligent agent by the trust manager; Based on the trust manager's reputation evaluation of the agent, several rounds of reputation evaluation are obtained; Based on several rounds of reputation evaluation combined with decay memory factor, the agent's current reputation value is obtained through dynamic aggregation.

8. The method for dynamic assessment and resilient defense of multi-agent trust based on subjective logic according to claim 1, characterized in that, S3. Based on the agent's current reputation value and the elastic random perturbation adjustment mechanism, obtain the elastic defense result, including: The perturbation ratio is obtained by combining the agent's current reputation value with the exponential decay coefficient. The system employs a random perturbation method to modify the agent's messages at the character level based on the perturbation comparison, thereby obtaining a resilient defense result. The random perturbation method includes insertion, swapping, or patching.

Images