Multi-agent distributed trusted decision making

By employing a distributed arbitration mechanism and multi-window detection, combined with hardware, behavioral, and iterative anchor point verification, the system solves the problem of secure and reliable decision-making in multi-agent systems with a decentralized architecture. It achieves efficient and reliable group decision-making and self-evolution capabilities, defends against various attack modes, and ensures system security and scalability.

CN123069374APending Publication Date: 2026-11-10李宗鹏
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611113398.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-26
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to achieve secure and reliable distributed decision-making in decentralized multi-agent systems, and are unable to effectively defend against attack modes such as Byzantine attacks, slow drift attacks, evolutionary poisoning, and hijacking of legitimate identities. Furthermore, the system's scalability and self-evolution capabilities are limited.

Method used

By employing a distributed arbitration mechanism, multi-window detection, and iterative anchoring mechanism, and through cross-verification of hardware anchors, behavioral anchors, and iterative anchors, combined with a dynamic weighted voting consensus algorithm and security boundary constraints, an intrinsic security system is formed to ensure that the node trust state is continuously reverified based on the current behavior, preventing malicious behavior and progressive attacks.

Benefits of technology

It enables efficient and reliable group decision-making in a decentralized architecture, continuously detects and defends against various attack modes, maintains the system's high scalability and self-evolution capabilities, and ensures that the system becomes more secure with use during the self-evolution process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069374A_ABST
    Figure CN123069374A_ABST
Patent Text Reader

Abstract

This invention discloses a self-evolving multi-agent cluster distributed trusted decision-making system and method, belonging to the field of artificial intelligence security technology. The system employs a decentralized peer-to-peer architecture: nodes continuously self-modify and iterate their strategies during operation; identities are cross-verified through three anchor points: hardware, behavior, and iteration; trust is continuously re-verified based on current behavior; a distributed arbitration module achieves group decision-making through dynamic weighted voting; a global behavior pattern monitoring module distinguishes between normal self-evolution and malicious slow drift using multiple time windows and a sliding baseline; and a strategy iteration optimization module adopts a three-layer iterative architecture, using dynamic elastic boundaries to ensure that iterations do not deviate from the security baseline. Security is inherent in the architecture itself, forming an intrinsic security system, suitable for scenarios such as robot clusters and self-evolving agent clusters.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security and multi-agent system technology, specifically to a self-evolving multi-agent cluster distributed trusted decision-making system and method, applicable to multi-agent collaboration platforms, robot clusters, autonomous vehicle fleets, distributed computing networks, multi-model fusion decision-making systems, intelligent recommendation systems, as well as scenarios requiring the assurance of decision security and trustworthiness in a decentralized architecture, such as self-evolving agent clusters, world model-driven robot systems, and multi-agent system security supervision and governance platforms. Background Technology

[0002] With the development of AI agent technology driven by large models, multi-agent systems are evolving from "centralized control" to "distributed collaboration." In a centralized architecture, all agents are uniformly scheduled by a central controller, and security policies are uniformly formulated and executed by a central node. The system's trustworthiness depends on the security of the central node. However, this architecture has inherent flaws: the central node is a single point of failure; once it is compromised or fails, the entire system will be paralyzed. Furthermore, as the number of agents increases, the computational and communication burden on the central node grows exponentially, limiting the system's scalability.

[0003] Meanwhile, agent systems are moving from "static deployment" to "continuous self-evolution": agent nodes in next-generation multi-agent systems generally possess online learning and self-modification capabilities, continuously modifying themselves and iterating strategies during operation, forming a self-evolving agent cluster. This continuous self-evolution of nodes introduces a triple failure scenario into the traditional static security assumption: First, the verification window fails. Traditional security systems assume that "nodes that pass security verification at the time of deployment remain trustworthy during operation"; however, when nodes continuously modify themselves during operation, the static trust assumption at the time of deployment becomes invalid for nodes in operation—the behavior of a node that passed verification at the time of deployment may no longer be consistent with that of a currently running node, and trust must be continuously re-verified based on the node's current behavior, rather than being statically granted once.

[0004] Secondly, delayed feedback leads to failure. The security policy adjustments of static security systems mainly rely on manual auditing and periodic updates. The feedback cycle is much longer than the policy iteration cycle of self-evolving systems. Security feedback continuously lags behind system evolution. By the time risks are discovered manually, the system may have already iterated through multiple versions.

[0005] Third, the exploration-security paradox. Self-evolving systems need to retain the ability to explore in order to achieve continuous optimization, but exploration means trying behaviors that have not yet been verified, which may introduce security risks; if exploration is prohibited in order to ensure security, it is equivalent to crippling the system's self-evolutionary ability. How to ensure security without stifling evolution is a problem that static security frameworks have never faced.

[0006] While distributed multi-agent architectures solve single point of failure and scalability issues, they introduce new security challenges: First, there's the issue of decision-making credibility. In a decentralized architecture, each agent makes independent decisions. How can we ensure that the decision-making results are both secure and efficient? Traditional majority voting mechanisms are easily controlled by malicious nodes—if an attacker controls more than half of the nodes, they can dominate the entire system's decision-making process. Existing Byzantine fault-tolerant technologies (such as PBFT) can tolerate a certain percentage of malicious nodes, but they suffer from high communication overhead, high consensus latency, and can only handle simple attack patterns like "intentionally sending incorrect information," failing to address more covert attack methods.

[0007] Second, slow drift attacks are difficult to detect. Traditional security detection focuses on sudden, obvious abnormal behavior, such as a node suddenly outputting malicious commands. However, a more insidious attack is the "boiling frog" slow drift attack—a malicious node gradually changes the baseline behavior of the entire system through small, incremental behavioral shifts without triggering alarms, ultimately achieving the goal of manipulating the system. Because this type of attack changes slowly and the amount of each shift is small, it is easily mistaken for normal system evolution and thus overlooked.

[0008] Third, there is the risk of poisoning the evolutionary direction. Distributed systems with self-learning and self-evolution capabilities face the risk of their evolutionary direction being poisoned—attackers can create false security events and induce incorrect feedback signals, causing the system's evolutionary direction to gradually deviate from the security baseline. Traditional security solutions, because they are based on static rules, do not address the evolutionary issue and have never considered this attack pattern of "targeted contamination of the security DNA."

[0009] Fourth, the vulnerability of identity and trust. Most existing distributed systems are based on cryptographic authentication—as long as one possesses the correct key or certificate, they are considered a trusted node. However, this model cannot cope with "malicious behavior of legitimate identities"—a node may be legitimately registered, but if it is controlled by an attacker, it can be used to launch attacks using a legitimate identity. Although zero-trust architecture proposes the concept of "never trust, always verify," there is still a lack of feasible technical solutions for achieving efficient and continuous verification in distributed scenarios.

[0010] The common flaw in existing technologies lies in their design of distributed security solutions based on a "centralized control" or "static rules" mindset. They fail to establish a truly reliable, dynamically evolving decision-making mechanism that matches the distributed architecture, and they lack a continuous re-verification mechanism for the aforementioned three failure scenarios in self-evolving systems. When system scale increases, attack methods evolve, and agents acquire autonomous evolution capabilities, existing technologies either lack sufficient security strength, incur excessive performance overhead, or are unable to adapt to dynamically evolving scenarios.

[0011] From a technological paradigm perspective, existing distributed security solutions are essentially still "distributed implementations of a centralized control mindset"—either relying on a centralized root of trust or on globally unified static rules, similar to the master-slave architecture in early distributed systems, which cannot truly unleash the inherent security potential of distributed systems.

[0012] In contrast, the multi-head attention mechanism in the Transformer architecture has proven that by fusion of global information through parallel computation of multiple independent attention heads, understanding and decision-making capabilities far exceeding those of a single module can emerge. Applying this idea to the field of distributed security means that each agent node can become a "security attention head," generating more powerful security detection and trustworthy decision-making capabilities than any single node through distributed behavioral observation and consensus arbitration. However, in the field of multi-agent security, there is currently no mature solution that can combine distributed observation mechanisms with evolutionary direction anchoring to form endogenous security decision-making capabilities.

[0013] Therefore, there is an urgent need in this field for a novel distributed trusted decision-making architecture that can achieve secure, reliable, efficient and flexible group decision-making in decentralized peer-to-peer networks through mechanisms such as distributed consensus, multi-dimensional anomaly detection, and security boundary constraints (evolutionary direction anchoring). This would enable security capabilities to be inherent in the distributed architecture itself and to continuously evolve with the system's operation, forming an endogenous security system that coexists with the decentralized peer-to-peer architecture. Summary of the Invention

[0014] I. Technical problems to be solved To address the aforementioned shortcomings of existing technologies, and drawing upon the core idea of ​​the multi-head attention mechanism in the Transformer architecture, this invention proposes a novel distributed trusted decision-making paradigm. The core technical problem it aims to solve is: how to construct a complete trusted decision-making mechanism in a decentralized, peer-to-peer multi-agent architecture where nodes continuously self-modify and iterate their strategies during operation. This mechanism addresses the issue of static trust assumptions at deployment time becoming invalid for running nodes due to continuous node self-evolution. Instead of granting trust to nodes statically, the trust state of each node is continuously re-verified based on its current behavior. This approach can resist various known and unknown attack patterns (including Byzantine attacks, slow drift attacks, evolutionary poisoning, etc.) while maintaining the high efficiency, high scalability, and autonomous evolutionary capabilities of the distributed system, achieving a positive cycle of "the more it is used, the more secure and trustworthy it becomes."

[0015] Break it down into four sub-problems: 1. How to achieve trusted group decision-making in a decentralized architecture and prevent malicious nodes from dominating the decision-making results? 2. How to detect slow drift-type progressive attacks and distinguish between normal self-evolutionary behavior of nodes and malicious drift behavior? 3. How to ensure that the system's self-evolution direction does not deviate from the safety baseline and prevent the evolution from being poisoned? 4. How to establish a multi-layered identity trust system to address the risks of legitimate identity hijacking and static trust failure caused by the continuous self-evolution of nodes? Technical solution

[0016] To address the aforementioned technical challenges, this invention proposes a three-layer collaborative trusted decision-making architecture of "distributed arbitration + multi-window detection + iterative anchoring." The core idea is that trust is not pre-assigned but dynamically established through continuous behavioral verification and group consensus; security is not a static rule but a dynamic balance maintained through security boundary constraint mechanisms during dynamic evolution; and security capabilities are inherent in the distributed architecture itself, forming an endogenous security system that coexists with the decentralized peer-to-peer architecture.

[0017] The core architecture consists of four parts: 1. Distributed intelligent agent cluster (infrastructure layer) It consists of multiple intelligent agent nodes with independent decision-making capabilities, continuously self-modifying and iterating their strategies during operation, with each node having a built-in security kernel module. The cluster adopts a decentralized peer-to-peer architecture, with no single global control node; the failure of a single node does not paralyze the entire system. To address the issue of static trust assumptions at deployment time becoming invalid due to the continuous self-evolution of nodes during operation, the trust state of each node is continuously re-verified based on its current behavior, rather than being statically granted all at once.

[0018] Each agent's identity is formed by cross-validation of three independent anchor points: (1) Hardware anchor: Physical signature from the security chip, generated based on Physically Unclonable Function (PUF), to ensure the physical unforgeability of the identity; the hardware anchor is provided by an independent hardware security chip, which provides each intelligent agent with an immutable physical identity root and a secure execution environment. The key parts of the kernel security policy are executed within the security chip, ensuring that even if the operating system layer is compromised, the security core can still operate normally, and the security protection strength is improved by orders of magnitude. (2) Behavioral anchors: Behavioral fingerprints generated from historical interaction sequences, including decision-making patterns, response delay characteristics, information processing preferences, etc., to characterize the behavioral patterns of nodes; (3) Iteration anchor: The feature signature evolved from the kernel security gene is updated synchronously with the gene iteration, reflecting the security evolution trajectory of the node.

[0019] The three types of anchor points are independent of each other yet mutually corroborate each other. When a single anchor point fails, the remaining anchor points can still maintain the downgraded judgment of identity credibility, thus avoiding a complete loss of security due to the breach of a single identity credential.

[0020] Each agent node has a kernel security gene, which is stored in a one-way hash chain structure: each iteration is based on the hash value of the previous version, and any tampering with the historical security gene will cause the hash chain to break, thus being identified and rejected by all nodes in the network.

[0021] Each agent node adopts a progressive activation mechanism that prioritizes a secure kernel: when a node starts up, it first loads a secure kernel module to complete hardware-level security verification and genetic integrity verification. Only after the security verification is passed can the functional modules and decision modules be activated. If the secure kernel fails to load or the verification fails, the entire node enters a restricted security mode, retaining only the most basic secure communication and fault reporting functions, and does not perform any decision-making or task operations, thus preventing tampered nodes from entering the cluster from the source of startup.

[0022] 2. Distributed arbitration mechanism (decision-making level) Embedded in each agent node, it is used to reach a group decision through a dynamic weighted voting consensus algorithm when multiple agents have differing opinions. The voting weight of each node is dynamically calculated based on its historical decision accuracy, security reputation score, and task domain matching degree, and is updated in real time as the node's behavior is continuously re-verified.

[0023] The distributed arbitration mechanism adopts a zero-trust identity assumption: even if the identity is verified, the output of the agent is not directly accepted. Instead, a weighted judgment is made based on its historical reputation, current context, and multi-source cross-validation results. When the identity credibility is lower than the threshold, the decision weight of the agent is automatically reduced, thereby achieving continuous re-verification of trust in the continuously self-evolving nodes in operation.

[0024] The key mechanisms of dynamic weighted voting include: (1) Weight cap threshold: The weight of any single node must not exceed the predetermined proportion of the total weight of the cluster to prevent excessive power of a single node from causing security risks; (2) Parallel hardware computation of weights: The weight computation process is executed in parallel within the hardware security chip of each node. Each node independently calculates its own weight and the weight distribution of the entire network. The computation results are publicly available and verifiable in real time. All nodes can independently verify the rationality of the weights, preventing weight manipulation and single-point computation bottlenecks. (3) Confidence-weighted decision-making mode: Each agent attaches a confidence evaluation to its output decision options. The final decision of the cluster is weighted and fused according to the weight of each node and its corresponding confidence of the option. The confidence is monotonically correlated with the selection probability. (4) Nondeterministic selection mechanism: A decision probability distribution is formed based on the weighted fusion results, and the final decision is randomly selected from the probability distribution. Low-confidence options retain non-zero probabilities to maintain the system's exploratory ability and evolutionary possibility, and avoid system rigidity caused by excessive conservatism.

[0025] The distributed arbitration mechanism is also equipped with an anti-reputation cartel mechanism (node ​​collusion detection mechanism): it periodically checks the correlation of reputation scores between nodes. If a group of nodes is found to have an abnormally high rating pattern for each other, the reputation credibility of all nodes in the group is automatically reduced and an independent audit process is triggered.

[0026] 3. Global Behavior Pattern Monitoring Module (Monitoring Layer) It is used to collect overall behavior data at the cluster level in real time, identify systemic security risks by calculating the deviation of macro indicators in multiple time windows and dimensions, and distinguish between normal self-evolutionary behavior and malicious drift behavior of nodes.

[0027] Macro indicators include: cluster decision deviation, information propagation map, resource consumption pattern, cross-node behavior synchronization, and other group-level behavioral characteristics.

[0028] The multi-time window mechanism includes: (1) Short-term sudden window: detect sudden behavioral changes in a short period of time and identify rapid attacks, such as a collective attack by Byzantine nodes; (2) Long-term cumulative window: detect slow drift of behavior over a longer time scale and identify gradual attacks like boiling a frog in lukewarm water.

[0029] The two windows are set with independent anomaly thresholds and alarm triggering logic to achieve full-coverage detection that "acute attacks can be caught and chronic attacks cannot escape".

[0030] The long-term cumulative window employs a sliding baseline mechanism: the behavioral baseline updates slowly with the system's normal evolution, but the baseline movement speed has an upper limit. If the detected behavioral drift speed exceeds the normal evolution speed threshold, a slow drift attack alarm is triggered. This mechanism can effectively distinguish between "behavioral changes caused by normal system self-evolution" and "behavioral drift caused by malicious poisoning"—the former is slow and matches the evolution speed, while the latter is fast and abnormally accelerated.

[0031] The global behavior pattern monitoring module adopts an anomaly scoring mechanism that integrates multiple anomaly detection algorithms: it combines the detection results of multiple anomaly detection algorithms (such as isolated forest, entropy analysis, graph neural network, and at least two of time series anomaly detection) to give a comprehensive score to the cluster's abnormal state, thereby improving the detection accuracy and reducing the false alarm rate.

[0032] 4. Strategy Iteration and Optimization Module (Evolutionary Layer, including safety boundary constraint mechanism) Based on the results of security event processing, the strategies of each node are iteratively optimized in reverse, using a three-layer iterative architecture: (1) Kernel security gene layer: extremely low frequency of changes and high threshold approval, defining the most fundamental security bottom line of the system; changes at this layer require multiple authorizations in a physically isolated environment and must be verified for consistency across the entire cluster; (2) Core framework layer: The core framework that defines security mechanisms is subject to mid-frequency changes and requires multi-party verification; changes at this layer require joint signature confirmation from a certain proportion of trusted nodes. (3) Scenario strategy layer: High frequency of change, quick adaptation, security strategy for specific scenarios; this layer can be iterated autonomously by a single node, but needs to be broadcast to the whole network and be subject to supervision.

[0033] The change triggering conditions, approval processes, and verification mechanisms vary at each level, with higher thresholds for changes at lower levels, forming a gradient evolutionary architecture where "the upper level is flexible and adaptable, while the lower level is rock-solid."

[0034] The safety boundary constraint mechanism (i.e., the evolutionary direction anchoring mechanism) is one of the core innovations of this invention, used to ensure that system evolution does not deviate from the safety baseline. The safety boundary constraint mechanism adopts a dynamic, elastic boundary: the safety boundary is not a fixed threshold, but rather an elastic boundary that is dynamically adjusted according to the system operating environment, task requirements, and current safety margin under the rigid constraints of the kernel safety genes. The safety boundary constraint mechanism includes: (1) Core anchoring factor: Defines the inviolable security bottom line, and all strategy iterations must not conflict with it. This is the system's "security genetic code", the starting point and destination of all evolution; modifications to the core anchoring factor must meet multiple authorization verification under physical isolation conditions; (2) Event Credibility Assessment: The credibility of security events that trigger iterations is rated, and the credibility weight is calculated based on factors such as the diversity of event sources, the degree of cross-validation, and the scope of impact. A multi-source cross-validation mechanism is adopted: security events from a single source have the lowest credibility, and the credibility of the same event detected by multiple independent nodes increases progressively. Only security events with credibility reaching the threshold can trigger ecosystem-level policy iterations to prevent the poisoning of evolutionary direction through false events; (3) Iteration direction verification: Before and after each iteration, the consistency between the strategy change direction and the core anchoring factor is verified. If the deviation exceeds the threshold, the iteration is rejected to ensure that the evolution always proceeds within a safe track.

[0035] The system also features a four-layer security defense system, forming a defense-in-depth structure: Layer 1: Admission Reputation Layer – New nodes must undergo security review and reputation initialization before joining the cluster; nodes with low reputation have limited permissions. A progressive permission release mechanism is adopted: new nodes initially have only the minimum permissions, and gradually gain higher decision-making weight and more system access permissions as their continuous safe operation time and contributions increase.

[0036] The second layer: Byzantine fault tolerance layer—the decision-making process tolerates a certain percentage of malicious nodes outputting erroneous information, and the dynamic weight mechanism naturally has the ability to resist Byzantine attacks.

[0037] The third layer: Anomaly detection layer—real-time monitoring of the behavioral pattern deviation of each node, identifying potential hijacked nodes, and even nodes with legitimate identities can be detected for anomalies in a timely manner.

[0038] The fourth layer: Iterative constraint layer—the kernel security gene provides an immutable anchor point for the entire system. Even if the upper-layer policy is contaminated, it can still be rolled back to a safe state through the anchor point.

[0039] A three-tiered fault response mechanism ensures the system's survivability under extreme conditions: Level 1: Degradation self-preservation mode - When the cluster consensus mechanism fails, each node switches to independent operation mode to maintain minimum security protection, "dispersed but not collapsed"; Level Two: Dynamic Election Mechanism – When the cluster needs a coordinating node, the remaining healthy nodes elect a temporary coordinating role through a distributed consensus algorithm to restore the cluster's collaborative capabilities. The election uses a reputation-weighted algorithm, where a node's electability is positively correlated with its historical security reputation score, and the entire election process is traceable. Level 3: Decentralized Self-Organizing Mode – When more than half of the nodes fail, the remaining nodes automatically enter a completely decentralized self-organizing state, maintaining basic functions through probabilistic pairwise verification: each node periodically and randomly selects other nodes for cross-verification of security status, rather than full synchronization, ensuring security coverage while keeping communication overhead within an acceptable range, waiting for the system to gradually recover, ensuring "disconnection without death".

[0040] The system's security rules adopt a three-tiered governance system adapted to a decentralized peer-to-peer architecture: (1) Bottom layer: Kernel security gene layer - cannot be directly modified, physically anchored by hardware security chip, unique across the entire network and cannot be forked; (2) Middle layer: Security principle framework layer - It can only be modified if a certain proportion of trusted nodes jointly sign and confirm it. It defines the core security boundary and uses majority consensus to update. (3) Top layer: Scene behavior rules layer - specific behavioral constraints that can be autonomously iterated and adjusted by a single node, broadcast to the whole network and subject to supervision.

[0041] There is a constraint verification mechanism between the upper and lower layers. The rules of the lower layer must not conflict with the principles of the upper layer, and each node independently verifies and autonomously accepts the rules.

[0042] The security rule updates employ a peer-to-peer reputation-weighted distributed, hierarchical diffusion mechanism (ripple diffusion): updates first take effect on a single or a small number of nodes and enter an observation period; after successful observation, the propagation scope gradually expands; updates are propagated from the initiating node to its trusted neighbor nodes, and receiving nodes independently verify the updates using their own kernel security features before accepting them and continuing to propagate them outwards; the propagation path is naturally formed by the reputation relationship network between nodes, rather than a preset hierarchy; high-reputation nodes propagate faster and cover a wider area; each step of the diffusion requires security verification, and if an anomaly occurs at any stage, the diffusion is immediately stopped and automatically rolled back to the previous stable version. This "small-scale trial first, then gradual rollout" update model effectively controls the impact of defective rules and avoids the systemic risks caused by a one-time full update.

[0043] Each agent node has independent verification and autonomous rejection rights when receiving security rule updates: the node uses its own kernel security genes to perform consistency verification on the new rules. If a conflict is found, it has the right to reject the update and broadcast an anomaly alarm to its directly connected neighboring nodes, naturally forming a propagation blockage; the rejected update will not continue to spread outward at that node, ensuring that the update of security rules will not be used by attackers to mass-deploy malicious rules.

[0044] The system is also configured with an interface for collaboration with an external security status monitoring and control system: this interface is used to receive cluster security status change warning signals issued by the external security status monitoring and control system; in response to the change warning signal, the distributed arbitration module triggers a recalculation of the security reputation score of the associated nodes, and reduces the voting weight of the node or suspends its qualification to participate in consensus according to the recalculation result, until the security status of the node is restored after reverification, thereby forming a cross-system security closed loop linking external monitoring and warning with internal consensus arbitration. This interface is connected to the "detection-prediction-intervention-verification" closed loop of the external security status monitoring and control system: the external system is responsible for the detection and prediction of security phase transitions and outputs change warnings, while this system is responsible for intervention and verification based on consensus arbitration. The collaboration between the two sides constitutes a complete cross-system security closed loop (see Example 4 for specific implementation).

[0045] This invention also provides a self-evolving multi-agent cluster distributed trusted decision-making method, comprising the following steps: S1: Construct a decentralized peer-to-peer distributed cluster consisting of multiple intelligent agent nodes with independent decision-making capabilities that continuously modify themselves and iterate their strategies during operation, with each node having a built-in security kernel module; S2: When multiple agents disagree on a decision, a dynamic weighted voting consensus algorithm is used through a distributed arbitration module to reach a group decision. The voting weight of each node is dynamically calculated based on its historical decision accuracy, security reputation score, and task domain matching degree, and is updated in real time as the node behavior is continuously re-verified. Furthermore, each agent attaches a confidence assessment to its output decision options. The final cluster decision is weighted and fused based on the weight of each node and its corresponding confidence of the option. A nondeterministic selection mechanism is used to extract the final decision from the probability distribution. Low-confidence options are retained with non-zero probability to maintain the system's exploration capability and iteration possibility. S3: Collects overall behavior data at the cluster level in real time, identifies systemic security risks by calculating the deviation of macro indicators in multiple time windows and dimensions, and distinguishes between normal self-evolutionary behavior and malicious drift behavior of nodes; furthermore, it detects short-term sudden anomalies and long-term cumulative drift at the same time, and sets independent detection thresholds and alarm logic for each. Long-term cumulative detection adopts a sliding baseline mechanism, and the baseline movement speed has an upper limit. If the upper limit is exceeded, a slow drift attack alarm is triggered. S4: Based on the processing results of security events, a three-layer iterative architecture is adopted to iteratively optimize the strategies of each node. The security boundary constraint mechanism ensures that the self-evolutionary iteration does not deviate from the security bottom line. Furthermore, the security events that trigger the iteration are rated for credibility. A multi-source cross-validation mechanism is used to evaluate the credibility of the events. Only security events with credibility reaching the threshold can trigger ecosystem-level strategy iteration.

[0046] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described self-evolving multi-agent cluster distributed trusted decision-making method.

[0047] This invention can be applied to at least one of the following scenarios: multi-agent collaborative platform, robot swarm, autonomous driving fleet, distributed computing network, multi-model fusion decision system, intelligent recommendation system, self-evolving agent swarm, world model driven robot system, and multi-agent system safety supervision and governance platform.

[0048] The beneficial effects of this invention can be summarized as "a breakthrough in architectural paradigm + four core technological innovations + one long-term strategic value".

[0049] A paradigm shift in architecture: from "centralized control" to "distributed autonomy"—a security paradigm truly compatible with distributed, self-evolving architectures. Traditional distributed security solutions are essentially an extension of the "centralized control" mindset—either having a physical central control node or a set of globally unified static rules. This model contradicts the inherent characteristics of distributed systems—"decentralized, self-organizing, and evolvable"—and is even less able to address the triple failure scenarios of self-evolving systems: "verification window failure, feedback delay failure, and the exploration-security paradox."

[0050] The distributed trusted decision-making architecture proposed in this invention is a security paradigm truly designed based on the first principles of distributed systems: trust is dynamically established through group consensus and continuous re-verification; security is guaranteed through multi-dimensional cross-verification; and evolution is guided directionally by security boundary constraint mechanisms. Security is inherent in the distributed architecture itself, rather than being imposed from the outside, forming an endogenous security system that coexists with the decentralized peer-to-peer architecture. This paradigm breakthrough enables security mechanisms and system architecture to move from "opposition" to "symbiosis."

[0051] Four core technological innovations: 1. Dynamic Weighted Distributed Arbitration Mechanism: Overcoming the Security Dilemma of "Most Powers Obeying the Minority" The traditional "one person, one vote" voting mechanism allows attackers to manipulate the entire system by controlling just over half of the nodes. The dynamic weighting mechanism of this invention dynamically binds voting rights to a node's historical performance, security reputation, and domain matching degree, and updates it in real time as the node's behavior is continuously re-verified. This forces attackers to compromise a large number of high-reputation nodes simultaneously to influence the decision-making outcome, resulting in an exponential increase in attack costs.

[0052] More importantly, the weight calculation is public and verifiable, and each node can independently verify it, thus eliminating the possibility of weight manipulation from a fundamental mechanism. The nondeterministic selection mechanism, while ensuring security, preserves the system's exploratory capabilities and avoids system rigidity caused by "tyranny of the majority."

[0053] 2. Multi-time-window sliding baseline detection: leaving no room for slow-drift attacks to hide. This is the first systematic detection scheme in the field specifically designed to target slow drift attacks. Traditional security detection is "snapshot-based"—comparing the current state to a fixed baseline and issuing an alert if a threshold is exceeded. This approach is effective against sudden attacks, but completely ineffective against slow, gradual drifts—because the baseline also changes slowly over time, and by the time an anomaly is detected, the system may already be completely compromised.

[0054] The sliding baseline mechanism of this invention solves this problem by "speed limiting": the baseline can move, but the moving speed cannot exceed the upper limit of normal evolutionary speed. Like a speed camera, it takes a picture if you exceed the speed limit, no matter where you end up. This detection approach breaks away from the traditional framework of "comparing to a fixed standard" and shifts to a new paradigm of "comparing to the rate of change," effectively distinguishing between normal self-evolutionary behavior and malicious drifting behavior of nodes. It has universal value in detecting various covert, progressive attacks.

[0055] 3. Safety boundary constraint mechanism (evolutionary direction anchoring): Equipping self-evolving systems with a "safety gyroscope". As AI systems acquire increasingly strong self-learning and self-evolution capabilities, "evolutionary deviation" is becoming a real and serious security challenge. Existing security technologies are designed for static systems and have never considered scenarios where the system itself evolves or can be targeted with malware.

[0056] The safety boundary constraint mechanism (evolutionary direction anchoring mechanism) of this invention is equivalent to equipping a self-evolving system with a gyroscope—no matter how the system evolves or how fast it evolves, the core anchoring factor always defines a safe "true north" direction. Iterative direction verification ensures that each iteration does not deviate from this direction, while event credibility assessment filters out false evolutionary triggers. The dynamic elastic boundary, under the rigid constraint of the kernel safety gene, flexibly adjusts the safety boundary according to the operating environment, task requirements, and safety margin, avoiding the "one-size-fits-all" fixed threshold from stifling the system's exploration capabilities and responding to the "exploration-safety paradox" of self-evolving systems. Multi-layered, three-dimensional protection ensures that the system always stays on a safe track while evolving rapidly.

[0057] 4. Three-type anchor point cross-identity authentication: breaking the equation "identity = trust". In traditional security systems, successful identity verification is equivalent to gaining trust—"If you are you, I will believe what you say." However, in real-world scenarios where intelligent agents can be hijacked and identities can be stolen, this equation is dangerous.

[0058] This invention's three-type anchor cross-verification mechanism upgrades identity from a "single credential" to a "multi-dimensional feature combination"—hardware anchors ensure physical unforgeability, behavioral anchors detect dynamic behavioral anomalies, and iterative anchors track long-term security trajectories. These three elements mutually reinforce each other, meaning that even if an attacker breaches one dimension, they cannot simultaneously impersonate others, significantly raising the barrier to identity impersonation. The zero-trust assumption goes a step further—even if identity verification is complete, the output is not directly accepted; instead, it is continuously evaluated based on reputation and context. Even if a node continuously modifies itself during operation, its trust state is always dynamically updated based on its current behavior, truly achieving "never trust, always verify."

[0059] A long-term strategic value: providing a secure architectural foundation for the deployment of large-scale autonomous intelligent agent systems. As AI agents evolve from single-agent to multi-agent collaboration, from closed to open environments, and from auxiliary tools to autonomous decision-making, distributed and trustworthy decision-making will become an indispensable core infrastructure. The architecture proposed in this invention can support large-scale distributed systems ranging from a few robots to hundreds or thousands of agents, covering traditional scenarios such as robot swarms, autonomous vehicle fleets, distributed computing networks, and multi-model fusion decision-making, as well as emerging scenarios such as self-evolving agent swarms, world-model-driven robot systems, and multi-agent system security monitoring and governance platforms, thus removing security obstacles for the large-scale application of autonomous agents. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall system architecture of the present invention, showing the composition and data flow of the four core modules: distributed intelligent agent cluster, distributed arbitration module, global behavior pattern monitoring module, and strategy iteration optimization module. Figure 2 This diagram illustrates the cross-validation mechanism for three types of anchors, showing how three independent anchors—hardware anchors, behavioral anchors, and iterative anchors—jointly constitute the identity trust system, as well as the degradation judgment logic when a single anchor fails. Figure 3 This is a schematic diagram of the multi-time window detection principle, illustrating the dual-time-axis detection mechanism of short-term burst windows and long-term cumulative windows, as well as the speed limiting principle of the sliding baseline; Figure 4 This diagram illustrates the three-layer iterative architecture and security boundary constraint mechanism, showcasing the three-level iterative architecture of kernel security gene layer, core framework layer, and scenario strategy layer, as well as the three layers of protection (core anchoring factor, event credibility assessment, and iterative direction verification) and dynamic elastic boundary of security boundary constraints (evolutionary direction anchoring). Figure 5 This is a schematic diagram of a four-layer security defense system, illustrating the defense-in-depth architecture consisting of an access reputation layer, a Byzantine fault tolerance layer, an anomaly detection layer, and an iterative constraint layer. Figure 6 This diagram illustrates a distributed, hierarchical diffusion mechanism for reputation weighting in peer-to-peer networks. It shows how security rule updates gradually spread outward from the initiating node along the reputation relationship network between nodes, with each step undergoing independent verification and automatic blocking in case of anomalies. Figure 7 This is a schematic diagram of the cross-system security closed-loop collaborative interface, illustrating the collaborative process and data flow between this system and the external security status monitoring and control system, from "receiving sudden warnings—recalculating security reputation scores—consensus downgrading or pausing—re-verification and recovery". Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0062] Example 1: Multi-robot collaborative search and rescue system (self-evolving cluster scenario) This embodiment demonstrates the complete workflow of the present invention in a multi-robot collaborative search and rescue scenario at a disaster site. The robots participating in the search and rescue continuously update their search and rescue strategies and collaboration modes based on online learning during mission execution, representing a typical self-evolving intelligent agent cluster.

[0063] System Configuration: (1) Robot cluster: 8 search and rescue robots, each equipped with lidar, thermal imaging camera, gas sensor and robotic arm. (2) Communication method: self-organizing mesh network, no central base station, supports multi-hop communication (3) Mission objectives: Search for survivors in the earthquake rubble, mark dangerous areas, and coordinate the removal of obstacles. (4) Security Kernel: Each robot has a built-in security kernel gene, which includes core anchoring factors such as "do not enter structurally unstable areas", "do not touch suspected explosives" and "prioritize human safety". The security kernel gene is stored in a one-way hash chain structure. Any tampering with the historical security gene will cause the hash chain to break, thereby being identified and rejected by all nodes in the network. (5) Initial reputation: All robots have an initial reputation value of 0.5 and a weight cap of 0.15 (the weight of a single node shall not exceed 15% of the total weight). (6) Time window configuration: short window 5 minutes, long window 24 hours; baseline movement speed limit is 5% per day. Typical workflow: Step 1: Task Allocation and Distributed Decision Making After receiving the mission, each robot independently assesses the sub-tasks (such as searching, marking, and moving) that it is suitable to perform based on its own location, sensor status, remaining battery power, and other factors, and broadcasts its mission preferences and confidence levels to the entire network.

[0064] A distributed arbitration mechanism is activated: each node independently calculates the dynamic weights of all other nodes—considering factors such as historical task completion accuracy, safety records, and current task domain matching, and updates them in real time as the behavior of each node is continuously re-verified. For example, a robot equipped with a gas sensor has a higher weight in chemical leak detection tasks, while a robot equipped with a robotic arm has a higher weight in handling tasks.

[0065] Each node independently calculates the final group task allocation scheme based on weights and confidence levels. Because each node uses the same computational logic and input information, it can independently produce a consistent allocation result without the need for a central node for scheduling. If disagreements arise, a weighted random arbitration is used to determine the final scheme.

[0066] Step 2: Collaborative Execution and Anomaly Detection The robots execute their assigned tasks separately. During execution, the global behavior pattern monitoring module continuously monitors the overall cluster status and distinguishes between normal self-evolutionary behavior and malicious drifting behavior of the robots. (1) Short-term window: Monitor whether nodes suddenly output abnormal commands, whether abnormal behavior synchronization occurs (which may mean that nodes are being hijacked in batches), and whether abnormal cascading occurs in information propagation. (2) Long-term window: Monitor slowly changing indicators such as changes in the cluster's task preferences, risk tolerance levels, and decision-making speed, and compare them with the sliding baseline. If the proportion of robots entering dangerous areas suddenly increases by 20% compared to usual on a certain day, and the drift speed exceeds the threshold of 5% / day, a slow drift alarm will be triggered.

[0067] The detection employs a fusion of multiple anomaly detection algorithms: for example, using isolated forest to detect individual anomalies, using entropy analysis to detect changes in system disorder, using graph neural networks to analyze changes in node interaction patterns, and using time series anomaly detection to detect trend shifts. The results of these multiple algorithms are weighted and fused to obtain the final anomaly score.

[0068] Step 3: Event Feedback and Strategy Evolution When a safety incident occurs (such as a robot nearly falling into a trap or an area being marked as high-risk), the relevant information is disseminated throughout the cluster via the Mesh network.

[0069] Each node independently evaluates the credibility of the event: (1) If only one robot reports the event, the confidence level is 0.3 (low). (2) If two or more robots with independent paths report similar events, the confidence level is 0.7 (medium). (3) If four or more robots in different locations verify the event from different angles, the credibility is 0.9 (high). Only events with a confidence level exceeding 0.6 can trigger policy updates at the core framework layer; only events with a confidence level exceeding 0.8 can affect the fine-tuning of the kernel security gene layer (note that this is fine-tuning, not modification of the core anchoring factor).

[0070] Iteration Direction Verification: Before and after each strategy update, the consistency score between the new strategy and the core anchoring factor must be calculated. For example, if the new strategy suggests "lowering the structural stability assessment standard to improve search and rescue efficiency," which conflicts with the anchoring factor of "not entering structurally unstable areas," the iteration will be rejected. Dynamic Flexible Safety Boundary: Under the premise that the rigid bottom line of "not entering structurally unstable areas" is not breached, the specific stability assessment threshold can be dynamically adjusted according to the type of ruin structure, the urgency of the mission, and the current safety margin.

[0071] Step 4: Fault Handling and System Resilience Assuming an aftershock causes three robots to lose contact (more than one-third of the total), the system will enter a level two fault response phase: The remaining five health robots were elected through a reputation-weighted election, temporarily selecting one coordination robot to handle simple task scheduling, while the other robots continued the search and rescue mission. The coordination robot is only responsible for information aggregation and simple coordination; it does not possess any special power exceeding that of other nodes, and major decisions still require distributed consensus.

[0072] If the number of missing robots increases to 5 (more than half), the system will enter a Level 3 fault response phase. The remaining three robots entered a completely decentralized self-organizing mode, with each robot making independent decisions and acting independently. They periodically and randomly selected another robot for pairwise safety checks, rather than full synchronization, ensuring safety coverage while keeping communication overhead within acceptable limits. Although collaboration efficiency decreased, the system did not completely shut down and could continue to perform basic search tasks, awaiting external reinforcements. This demonstrates the system's resilience of being "dispersed but not collapsing, broken but not dead."

[0073] Example 2: Trusted Task Scheduling in Distributed Computing Networks This embodiment demonstrates the application of the present invention in a distributed computing network scenario and verifies its applicability in scenarios with large-scale nodes and high-frequency transactions.

[0074] Scene description: This distributed computing network comprises over 1000 computing nodes, which can freely join and leave. Task publishers break down AI training tasks into smaller chunks and distribute them to various nodes for execution. Each node's task execution strategy continuously evolves and updates based on task experience, exhibiting typical self-evolutionary characteristics. However, because nodes are anonymous or semi-anonymous, there is a risk of malicious nodes returning incorrect calculation results or poisoning training data. Therefore, a reliable task scheduling and result verification mechanism is needed to ensure the correctness of the calculation results without relying on a central node.

[0075] System Configuration: (1) Node size: 1024 computing nodes, dynamically added and removed. (2) Task type: AI model training sharding, data labeling, inference service (3) Reputation system: Each node has a multi-dimensional reputation score, including calculation accuracy, completion speed, and cheating records. (4) Weight cap: The weight of a single node shall not exceed 0.1% (one-thousandth) to prevent large nodes from monopolizing the market. (5) Detection window: short-term window 1 hour, long-term window 7 days; baseline drift rate upper limit is 2% per day. Workflow: 1. Task Allocation: The task publisher broadcasts the task to the network, and nodes compete for the task based on their computing power and reputation. A distributed arbitration mechanism dynamically calculates the weight of each node based on factors such as historical accuracy, task matching degree, and current load, and updates it in real time as the results of continuous re-verification of node behavior are updated. A weighted random selection process determines which nodes will execute the task. The same task may be assigned to multiple nodes for repeated execution, used for cross-validation of results.

[0076] 2. Result Verification: After multiple nodes return their calculation results, a weighted majority vote is used to determine the correct result. Voting weights are based on a node's historical accuracy reputation—nodes with higher past accuracy have greater voting weight. However, due to a weight cap, even the node with the highest reputation cannot determine the result alone.

[0077] The anti-reputation cartel mechanism operates periodically: it analyzes the voting consistency patterns among nodes, and if it finds that a certain group of nodes always votes the same way and deviates from the group average, it suspects that they may have formed a cartel (collusion and cheating), automatically reduces the reputation of this group of nodes, and triggers a more stringent audit.

[0078] 3. Slow Drift Detection: Long-term window monitoring of the accuracy trends of each node. Under normal circumstances, node accuracy should be relatively stable or increase slowly (because the model is optimizing). If the accuracy of a node increases abnormally rapidly in a short period of time, but at the same time the "uniqueness" of the returned results decreases (always consistent with the majority, without independent insights), it may be that the node has changed from "independent computation" to "free-riding" by "copying others' results". The sliding baseline mechanism can detect this abnormal change—the rate of accuracy increase exceeds the upper limit of normal evolution, thus distinguishing the normal self-evolutionary behavior of a node from malicious drift behavior.

[0079] 4. Evolution and Adaptation: The system continuously updates the reputation score of each node based on the verification results. Honest and accurate nodes gain reputation and receive more tasks and higher rewards; cheating or making mistakes lead to a decline in reputation and fewer tasks. This is a market-driven evolutionary selection process—good nodes thrive, while poor nodes are marginalized, and the overall credibility of the system continuously improves over time.

[0080] The security boundary constraint mechanism ensures that this "survival of the fittest" will not go to extremes—for example, a few nodes will not monopolize all tasks, because the weight cap mechanism and the anti-reputation cartel mechanism always maintain the diversity and decentralization of the system.

[0081] Example 3: A Security Decision System Based on Multi-Model Fusion This embodiment demonstrates the application of the present invention in a multi-AI model fusion decision-making scenario, and verifies its versatility in heterogeneous intelligent agent environments.

[0082] Scene description: Financial risk control systems require the integration of outputs from multiple AI models to make final decisions. Different models have different training data, algorithmic architectures, and biases, and each model is continuously fine-tuned and updated based on new data, with decision-making behavior evolving over iteration. Some models are better at identifying fraud, some are better at assessing credit risk, and some are more sensitive to new types of attacks. The core issue is how to integrate the outputs of these heterogeneous models, leveraging their respective strengths while preventing any single model from being compromised and contaminating the final decision.

[0083] System Configuration: (1) Number of models: 12 AI models, including deep learning models, traditional machine learning models, and rule engines. (2) Decision type: Transaction risk score, outputting a risk value of 0-1. (3) Reputation dimension: historical accuracy, recall, and detection capability for different attack types of each model. (4) Three-level security rules: kernel layer (risk control red line, cannot be modified), framework layer (risk calculation logic, requires approval to modify), and scenario layer (specific threshold, can be dynamically adjusted). Workflow: 1. Multi-model distributed decision-making: Each model independently assesses the risk of a transaction, outputting a risk score and confidence level. A distributed arbitration mechanism integrates the outputs of all models, dynamically adjusts the weights based on each model's historical performance on this type of transaction, and calculates the final fused risk score.

[0084] For example, in online transaction fraud scenarios, models that excel at graph computation have higher weights; in offline credit assessment scenarios, models that excel at behavioral analysis have higher weights. These weights are not fixed but dynamically calculated based on the model's real-time performance in the corresponding domain.

[0085] 2. Anomaly Model Detection: The global behavior pattern monitoring module monitors the output patterns of each model in real time. If the output distribution of a model suddenly changes, or the deviation from the consensus results of other models continues to increase, it is possible that the model has been altered by an attacker through adversarial example attacks or data poisoning. Short-term windows detect mutation attacks, while long-term windows detect slow model drift.

[0086] For example, an attacker might gradually "deviate" from a model's decision boundary by continuously injecting small adversarial examples. These changes may seem minor on a daily basis, but can accumulate to a significant degree. A sliding baseline mechanism can detect anomalies in the speed of this drift, distinguishing it from the model's normal self-evolutionary updates based on new data, and providing timely alerts.

[0087] 3. Ripple Update of Rules: When risk control rules need to be adjusted (e.g., a new fraud pattern emerges), the update is not pushed out all at once. Instead, it is first piloted on 1-2 models. After observing for a period of time and confirming that there are no side effects, it is gradually promoted to more models along the reputation relationship network between nodes. If a problem is found at a certain stage—for example, the false positive rate of the new rule is too high—it is immediately rolled back to the previous version without affecting the overall system.

[0088] When a new rule is received, each model performs a consistency check using its own kernel security features. If the new rule violates the most basic risk control principles (such as "risk scores cannot be adjusted due to racial factors"), the model has the right to refuse to load the new rule and broadcast an anomaly alert to its directly connected neighboring nodes, naturally creating a propagation blockage; the rejected update will not continue to spread outward from that model.

[0089] This embodiment demonstrates that the distributed trusted decision-making architecture of the present invention is applicable not only to physical robot swarms but also to multi-model fusion systems at the software level. Its core ideas—dynamic weights, multi-window detection, security boundary constraints, and ripple-like progressive diffusion—have universal architectural value for any scenario requiring the assurance of security and trustworthiness in distributed, heterogeneous, and evolving systems.

[0090] Example 4: Cross-system security closed loop in coordination with external security status monitoring and control system This embodiment demonstrates the interface implementation scheme for the collaborative work between the system of the present invention and an external safety status monitoring and control system (such as an external monitoring system that employs global behavior deviation analysis, safety phase transition prediction, and minimum intervention control). It corresponds to the interface definition and arbitration linkage process of the system side in the cross-system "detection-prediction-intervention-verification" safety closed loop.

[0091] Scene description: A self-evolving intelligent agent swarm consisting of 200 warehouse robots continuously updates its path planning and collaboration strategies based on online reinforcement learning. An external safety status monitoring and control system performs third-party global monitoring and phase transition prediction for the swarm, while the system in this invention is responsible for trusted decision-making and consensus arbitration within the swarm. The two systems are connected through a standardized collaborative interface.

[0092] Interface configuration: (1) Interface type: bidirectional message interface, using a signature message mechanism. The warning message fields include: warning identifier, set of associated node identifiers, risk type, risk confidence level (value 0-1), suggested handling level (observation / downgrading / suspension consensus), warning validity period, and external system signature; (2) Triggering conditions: When the probability of a safe phase transition in the cluster or a local subgroup calculated by the external system exceeds its warning threshold, a cluster safety status change warning signal is sent to this system; (3) Security verification: After receiving the warning signal, each node of this system first uses the kernel security gene to verify the consistency of the message source certificate and signature to prevent the forgery of warnings; warnings that fail the verification are directly discarded and recorded in the exception log.

[0093] Workflow: Step 1: Warning Receiving and Verification An external system detected an abnormally high correlation in the behavior of a local subgroup (12 robots) and a decreased recovery rate after disturbance (critical slowing down). The probability of a gradual phase transition occurring in this region was determined to be 0.72, and a sudden change warning signal was issued: associated node set R (12 robots), risk type "slow drift / gradual phase transition", confidence level 0.72, recommended handling level "downgrading", validity period 30 minutes. This warning is processed and signed by each node in the system.

[0094] Step 2: Recalculate the safety reputation score In response to a mutation warning signal, the distributed arbitration module triggers a recalculation of the security reputation score of nodes within set R. This involves incorporating the warning confidence level as prior validation data into the reputation model, while simultaneously retrieving the behavioral drift records of nodes within R over a long-term cumulative window, recent decision accuracy, and multi-source cross-validation results to comprehensively recalculate the security reputation score. For example, the reputation score of a node within R may decrease from 0.81 to 0.52.

[0095] Step 3: Consensus Demotion and Qualification Handling Based on the recalculation results, tiered measures are implemented: Nodes with reputation scores below the de-weighting threshold (e.g., 0.6) have their voting weight reduced accordingly based on the reputation reduction, while still being subject to the existing weight cap mechanism; Nodes with reputation scores below the suspension threshold (e.g., 0.4) or whose warning suggestion level is "suspension of consensus" are suspended from participating in consensus, and their unfinished tasks are taken over by healthy nodes. De-weighting and suspension decisions take effect across the entire network after confirmation by dynamic weighted voting consensus, and the action records are written to each node's local log, ensuring full traceability.

[0096] Step 4: Re-verify and restore The node being processed enters a re-verification process: three types of anchors are re-cross-verified (physical signature verification of hardware anchors, re-accumulation of behavioral anchors within the observation window, and consistency verification of iterative anchors and kernel security gene hash chain), and its behavioral drift speed is continuously monitored during the observation period (e.g., 24 hours); after the re-verification is passed, its voting weight and consensus qualification are gradually restored according to the progressive permission release mechanism to avoid the risks of a one-time restoration.

[0097] Step 5: Closed-loop feedback This system feeds back the handling results and recovery progress to the external safety status monitoring and control system; the external system continues to monitor changes in phase transition probabilities: several prognostic phase transition probabilities fall below 0.3 and remain stable, confirming the effectiveness of this closed-loop process. This system records the entire "early warning-handling-recovery" process to update the prior parameters and long-term window baseline of the reputation model; the external system incorporates the intervention effect into the update of its detection baseline and prediction model parameters, and the closed loop continues to operate.

[0098] In this embodiment, the external security status monitoring and control system is responsible for the "detection-prediction" link and outputs a sudden change warning to this system; this system is responsible for the consensus arbitration part in the "intervention-verification" link: the external sudden change warning triggers the reputation recalculation and consensus downgrading of this system (intervention linkage), and the intervention effect is fed back by the consensus verification of this system and updates the detection baseline and prediction model parameters of the external system in reverse (verification linkage), thereby forming a complete cross-system security closed loop of "detection-prediction-intervention-verification".

[0099] Application Scenario Extension Description In addition to the above embodiments, the present invention can also be applied to the following scenarios: (1) Multi-agent collaboration platform and intelligent recommendation system: Each agent in the platform continuously learns online. The trust continuous re-verification and dynamic weight arbitration mechanism of this invention prevents hijacked or poisoned nodes from dominating platform decision-making. (2) Self-evolving intelligent agent clusters: Various intelligent agent clusters with online learning and self-modification capabilities are provided with endogenous security guarantees throughout their entire life cycle by the continuous re-verification, slow drift detection and security boundary constraint mechanisms of this invention; (3) World model driven robot system: The robot performs simulation prediction and strategy pre-playing based on the internal world model. The strategy evolves continuously as the world model is updated. The safety boundary constraint mechanism of this invention ensures that the strategy iteration driven by the world model does not break through the core anchoring factor. The cross-validation of three types of anchor points ensures that the node identity is always credible in the continuous evolution. (4) Multi-agent system security supervision and governance platform: The reputation system, three-level hierarchical rule governance and cross-system collaboration interface of the present invention can serve as a platform-level security governance infrastructure, providing regulators with credible decision auditing, risk warning and collaborative handling capabilities.

Claims

1. A self-evolving multi-agent cluster distributed trusted decision-making system implemented using computer technology, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the following modules: The distributed intelligent agent cluster consists of multiple intelligent agent nodes with independent decision-making capabilities that continuously modify themselves and iterate their strategies during operation. Each node has a built-in security kernel module. The cluster adopts a decentralized peer-to-peer architecture and there is no single global control node. To address the issue that the static trust assumptions at the time of deployment may fail for running nodes due to the continuous self-evolution of nodes during operation, the trust state of each node is continuously re-verified based on its current behavior, rather than being statically granted all at once. The distributed arbitration module, embedded in each agent node, is used to reach a group decision through a dynamic weighted voting consensus algorithm when multiple agents have disagreements on decisions. The voting weight of each node is dynamically calculated based on its historical decision accuracy, security reputation score, and task domain matching degree, and is updated in real time with the results of continuous re-verification. The global behavior pattern monitoring module is used to collect overall behavior data at the cluster level in real time. It identifies systemic security risks by calculating the deviation of macro indicators through multiple time windows and dimensions, and distinguishes between the normal self-evolutionary behavior and malicious drift behavior of nodes. The macro indicators include cluster decision deviation, information propagation graph, resource consumption pattern and cross-node behavior synchronization. The time windows include short-term burst windows and long-term cumulative windows to detect mutation attacks and slow drift attacks, respectively. The strategy iteration and optimization module iteratively optimizes the strategies of each node based on the processing results of security events. It adopts a three-layer iterative architecture: kernel security gene layer (extremely low frequency, high threshold change), core framework layer (medium frequency, requires multi-party verification), and scenario strategy layer (high frequency, can be quickly adapted). The strategy iteration and optimization module is equipped with a security boundary constraint mechanism to ensure that the self-evolution iteration of each node does not deviate from the preset security bottom line, so that the security capability is endogenous to the distributed architecture itself, forming an endogenous security system that coexists with the decentralized peer-to-peer architecture.

2. A self-evolving multi-agent cluster distributed trusted decision-making system implemented using computer technology, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the following modules: The distributed intelligent agent cluster consists of multiple intelligent agent nodes with independent decision-making capabilities that continuously modify themselves and iterate their strategies during operation. Each node has a built-in security kernel module. The cluster adopts a decentralized peer-to-peer architecture and there is no single global control node. To address the issue that the static trust assumptions at the time of deployment may fail for running nodes due to the continuous self-evolution of nodes during operation, the trust state of each node is continuously re-verified based on its current behavior, rather than being statically granted all at once. The distributed arbitration module, embedded in each agent node, is used to reach a group decision through a dynamic weighted voting consensus algorithm when multiple agents have disagreements on decisions. The voting weight of each node is dynamically calculated based on its historical decision accuracy, security reputation score, and task domain matching degree, and is updated in real time with the results of continuous re-verification. The global behavior pattern monitoring module is used to collect overall behavior data at the cluster level in real time. It identifies systemic security risks by calculating the deviation of macro indicators through multiple time windows and dimensions, and distinguishes between the normal self-evolutionary behavior and malicious drift behavior of nodes. The macro indicators include cluster decision deviation, information propagation graph, resource consumption pattern and cross-node behavior synchronization. The time windows include short-term burst windows and long-term cumulative windows to detect mutation attacks and slow drift attacks, respectively. The strategy iteration and optimization module iteratively optimizes the strategies of each node based on the processing results of security events. It adopts a three-layer iterative architecture: kernel security gene layer (extremely low frequency, high threshold change), core framework layer (medium frequency, requires multi-party verification), and scenario strategy layer (high frequency, can be quickly adapted). The strategy iteration and optimization module is equipped with a security boundary constraint mechanism to ensure that the self-evolution iteration of each node does not deviate from the preset security bottom line, so that the security capability is endogenous to the distributed architecture itself, forming an endogenous security system that coexists with the decentralized peer-to-peer architecture.

3. The system according to claim 2, characterized in that, The kernel security framework includes an unmodifiable core anchoring factor, which defines the lowest-level security boundary of the system. Policy iterations at any level must not conflict with the core anchoring factor. Modifications to the core anchoring factor must satisfy multi-authorization verification under physical isolation conditions.

4. The system according to claim 1, characterized in that, Each agent's identity is formed by cross-validation of at least three independent anchor points: Hardware anchor: Physical signature from the security chip, generated based on a physically unclonable function (PUF); Behavioral anchors: Behavioral fingerprints generated from historical interaction sequences, including decision-making patterns, response latency characteristics, and information processing preferences; Iterative anchors: Feature signatures evolved from kernel security genes, updated synchronously with gene iteration; When a single anchor point fails, the remaining anchor points can still maintain the downgraded assessment of identity credibility.

5. The system according to claim 4, characterized in that, The distributed arbitration module adopts a zero-trust identity assumption—even if the identity is verified, the output of the agent is not directly accepted. Instead, a weighted judgment is made based on its historical reputation, current context, and multi-source cross-validation results. When the identity credibility is lower than the threshold, the decision weight of the agent is automatically reduced, thereby achieving continuous re-verification of trust in the continuously self-evolving nodes in operation.

6. The system according to claim 1, characterized in that, The dynamic weighted voting consensus algorithm has a weight upper limit threshold, and the weight of any single node must not exceed a predetermined proportion of the total weight of the cluster. The weight calculation process is executed in parallel within the hardware security chip of each node, and the calculation results are publicly verifiable across the entire network in real time. All nodes can independently verify the rationality of the weights.

7. The system according to claim 1, characterized in that, The distributed arbitration module adopts a confidence-weighted decision-making mode—each agent attaches a confidence assessment to its output decision options, and the final decision of the cluster is a weighted fusion based on the weight of each node and its corresponding confidence for the option. The confidence level and the selection probability are monotonically correlated.

8. The system according to claim 7, characterized in that, The cluster's final decision-making adopts a nondeterministic selection mechanism—a decision probability distribution is formed based on the weighted fusion results, and the final decision is randomly selected from the probability distribution; low-confidence options retain non-zero probabilities to maintain the system's exploratory capability and iterative possibility.

9. The system according to claim 1, characterized in that, The system has a node collusion detection mechanism—it periodically checks the correlation of reputation scores between nodes. If a group of nodes is found to have an abnormally high rating pattern for each other, the system automatically lowers the reputation credibility of all nodes in that group and triggers an independent audit process.

10. The system according to claim 1, characterized in that, The system has a three-level fault response mechanism: Level 1: Degradation self-preservation mode – When the cluster consensus mechanism fails, each node switches to independent operation mode to maintain minimum security protection; Level 2: Dynamic election mechanism - When the cluster needs a coordinating node, the remaining healthy nodes elect a temporary coordinating role through a distributed consensus algorithm to restore the cluster's collaborative capabilities. Level 3: Decentralized Self-Organizing Mode – When more than half of the nodes fail, the remaining nodes automatically enter a completely decentralized self-organizing state, maintaining basic functions through probabilistic pairwise verification, and waiting for the system to gradually recover.

11. The system according to claim 10, characterized in that, The dynamic election mechanism adopts a reputation-weighted election algorithm—a node's right to be elected is positively correlated with its historical security reputation score, and the entire election process is traceable, preventing malicious nodes from controlling the cluster by manipulating the election.

12. The system according to claim 10, characterized in that, In the probabilistic pairwise verification mechanism, each node periodically and randomly selects other nodes to perform cross-verification of security status, rather than full synchronization, thus ensuring security coverage while keeping communication overhead within an acceptable range.

13. The system according to claim 1, characterized in that, The multi-time-window detection mechanism of the global behavior pattern monitoring module includes: Short-term burst window: Detects sudden behavioral changes within a short period of time and identifies rapid attacks; Long-term cumulative window: detects slow behavioral drift over a longer timescale and identifies progressive attacks; The two windows each have independent exception thresholds and alarm triggering logic.

14. The system according to claim 13, characterized in that, The long-term cumulative window adopts a sliding baseline mechanism—the behavior baseline is slowly updated with the normal iteration of the system, but the baseline movement speed is capped; if the behavior drift speed is detected to exceed the normal iteration speed threshold, a slow drift attack alarm is triggered.

15. The system according to claim 1, characterized in that, The global behavior pattern monitoring module adopts an anomaly scoring mechanism that integrates multiple anomaly detection algorithms—combining the detection results of multiple anomaly detection algorithms to comprehensively score the abnormal state of the cluster, thereby improving the detection accuracy and reducing the false alarm rate; the multiple anomaly detection algorithms include at least two of the following: isolated forest, entropy analysis, graph neural network, and time series anomaly detection.

16. The system according to claim 1, characterized in that, The system is equipped with a four-layer security defense system: First layer: Admission reputation layer - New nodes must undergo security review and reputation initialization before joining the cluster, and nodes with low reputation have restricted permissions; The second layer: Byzantine fault tolerance layer – tolerating a certain percentage of malicious nodes outputting erroneous information during the decision-making process; The third layer: Anomaly detection layer—real-time monitoring of behavioral pattern shifts in each node to identify potentially hijacked nodes; The fourth layer: Iterative constraint layer—the kernel security gene provides an immutable anchor point for the entire system. Even if the upper-layer policy is contaminated, it can still be rolled back to a safe state through the anchor point.

17. The system according to claim 16, characterized in that, The admission reputation layer adopts a progressive permission release mechanism—new nodes initially have only the lowest permissions, and as the continuous safe operation time increases and contributions accumulate, they gradually gain higher decision-making weight and more system access permissions.

18. The system according to claim 1, characterized in that, The security boundary constraint mechanism adopts a dynamic elastic boundary—the security boundary is not a fixed threshold, but an elastic boundary that is dynamically adjusted according to the system operating environment, task requirements, and current security margin under the rigid constraints of the kernel security genes; the security boundary constraint mechanism includes: Core anchoring factor: Defines an inviolable safety baseline, and all strategy iterations must not conflict with it; Event credibility assessment: The credibility of security events that trigger iterations is rated, and the credibility weight is calculated based on factors such as the diversity of event sources, the degree of cross-validation, and the scope of impact. Iteration direction verification: Before and after each iteration, the consistency between the strategy change direction and the core anchoring factor is verified. If the deviation exceeds the threshold, the iteration is rejected.

19. The system according to claim 18, characterized in that, The event credibility assessment adopts a multi-source cross-validation mechanism—security events from a single source have the lowest credibility, while the credibility of the same event detected by multiple independent nodes increases progressively; only security events with credibility reaching the threshold can trigger ecosystem-level policy iteration.

20. The system according to claim 1, characterized in that, In the three-layer iterative architecture of the strategy iteration optimization module, the change triggering conditions, approval processes, and verification mechanisms are different at each level: Kernel security layer: Requires multiple authorizations in a physically isolated environment and must undergo full cluster consistency verification; Core framework layer: requires joint signature confirmation from a certain proportion of trusted nodes; Scenario strategy layer: It can be iterated autonomously by a single node, but it needs to be broadcast to the entire network and be subject to supervision.

21. The system according to claim 1, characterized in that, The system's security rules adopt a three-tiered governance system adapted to a decentralized peer-to-peer architecture: The lowest layer: the kernel security gene layer—cannot be directly modified, is physically anchored by a hardware security chip, is unique across the entire network, and cannot be forked; Middle layer: Security principle framework layer—modification requires joint signature confirmation from a certain proportion of trusted nodes, defines the core security boundary, and adopts majority consensus for updates; Top layer: Scene behavior rules layer—specific behavioral constraints that can be autonomously iterated and adjusted by a single node, broadcast to the entire network and subject to supervision; There is a constraint verification mechanism between the upper and lower layers. The rules of the lower layer must not conflict with the principles of the upper layer, and each node independently verifies and autonomously accepts the rules.

22. The system according to claim 21, characterized in that, When security rules need to be updated, a distributed, hierarchical diffusion mechanism with peer-to-peer network reputation weighting is adopted: the update is first propagated by the initiating node to its trusted neighbor nodes, and the receiving node independently verifies it using its own kernel security genes before accepting it and continuing to propagate it outward; the propagation path is naturally formed by the reputation relationship network between nodes, rather than a preset hierarchy; nodes with high reputation propagate faster and have a wider coverage; if a node finds a conflict during verification, it rejects the update and broadcasts an anomaly alarm to its neighbors, naturally creating a propagation blockage.

23. The system according to claim 21, characterized in that, When receiving security rule updates, each agent node has the right to independently verify and autonomously reject them. The node uses its own kernel security genes to verify the consistency of the new rules. If a conflict is found, it has the right to reject the update and broadcast an anomaly alarm to its directly connected neighboring nodes. The rejected update will not continue to spread outward from that node.

24. The system according to claim 4, characterized in that, The hardware anchor is provided by an independent hardware security chip, which provides each intelligent agent with an immutable physical identity root and a secure execution environment. The key parts of the kernel security policy are executed within the security chip, ensuring that even if the operating system layer is compromised, the security core can still operate normally, and the security protection strength is improved by orders of magnitude.

25. The system according to claim 1, characterized in that, The system is configured with an interface for collaboration with an external security status monitoring and control system: the interface is used to receive early warning signals of sudden changes in cluster security status issued by the external security status monitoring and control system; In response to the mutation warning signal, the distributed arbitration module triggers a recalculation of the security reputation score of the associated node, and reduces the voting weight of the node or suspends its qualification to participate in consensus according to the recalculation result, until the security status of the node is restored after re-verification, thereby forming a cross-system security closed loop that links external monitoring and early warning with internal consensus arbitration.

26. The system according to claim 1, characterized in that, The system is applied to at least one of the following scenarios: multi-agent collaboration platform, robot swarm, autonomous driving fleet, distributed computing network, multi-model fusion decision system, intelligent recommendation system, self-evolving agent swarm, world model driven robot system, and multi-agent system safety supervision and governance platform.

27. A self-evolving multi-agent cluster distributed trusted decision-making method, characterized in that, Includes the following steps: S1: Construct a decentralized peer-to-peer distributed cluster consisting of multiple intelligent agent nodes with independent decision-making capabilities that continuously modify themselves and iterate their strategies during operation, with each node having a built-in security kernel module; S2: When multiple agents disagree on a decision, a group decision is reached through a distributed arbitration module using a dynamic weighted voting consensus algorithm. The voting weight of each node is dynamically calculated based on its historical decision accuracy, security reputation score, and task domain matching degree, and is updated in real time as the node behavior is continuously re-verified. S3: Collects overall behavior data at the cluster level in real time, identifies systemic security risks by calculating the deviation of macro indicators in multiple time windows and dimensions, and distinguishes between normal self-evolutionary behavior and malicious drift behavior of nodes. S4: Based on the results of security event processing, a three-layer iterative architecture is adopted to iteratively optimize the strategies of each node, and a security boundary constraint mechanism is used to ensure that the self-evolution iteration does not deviate from the security bottom line.

28. The method according to claim 27, characterized in that, Step S2 further includes: each agent attaches a confidence assessment to its output decision options, the cluster's final decision is weighted and fused according to the weight of each node and its corresponding confidence of the option, and a nondeterministic selection mechanism is used to extract the final decision from the probability distribution, with low-confidence options retaining non-zero probabilities.

29. The method according to claim 27, characterized in that, Step S3 further includes: simultaneously detecting short-term sudden anomalies and long-term cumulative drift, setting independent detection thresholds and alarm logic for each; long-term cumulative detection adopts a sliding baseline mechanism, with an upper limit on the baseline movement speed, and triggering a slow drift attack alarm if the upper limit is exceeded.

30. The method according to claim 27, characterized in that, Step S4 further includes: rating the credibility of security events that trigger iterations, using a multi-source cross-validation mechanism to assess the credibility of events, and only security events with credibility reaching the threshold can trigger ecosystem-level policy iterations.

31. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the self-evolving multi-agent cluster distributed trusted decision-making method as described in any one of claims 27-30.