Zero-trust security agent information communication network control and collaboration method and system based on skill chain

Abstract

The invention belongs to the technical field of artificial intelligence and block chain fusion, and discloses a skill chain-based zero-trust security agent information communication network control and cooperation method and system, and a TrustAgentNet establishes a powerful and audible unified trusted agent identity recognition system. An intelligent agent is required to register its identity through a certificate issuing mechanism on an alliance block chain, and each participant can obtain an encrypted and secure digital certificate. This inherent trust basis can reduce the risk associated with malicious or infected agents, thereby creating a safer and more reliable collaborative environment. The TrustAgentNet provides a scheme which is easier to manage the information of this kind. The skill combinations and related tags can be permanently recorded on the block chain, thereby creating a shared transparent registry. The smart contract can compulsively use standardized formats and ontologies to describe these capabilities, ensure consistency, and promote accurate and efficient discovery of agents with specific professional knowledge.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and blockchain integration technology, and in particular relates to a zero-trust secure intelligent agent information communication network management and collaboration method and system based on skill chain. Background Technology

[0002] With the profound development of artificial intelligence technology, the role of AI models as the "nerve center" of networks is becoming increasingly clear. Networks will no longer be merely data transmission platforms, but intelligent infrastructures capable of autonomous learning, cognition, and decision-making. In particular, within the AgenticAI architecture, task-oriented AI agents achieving complex goals through interactive collaboration has become a core paradigm. Google's Agent2Agent protocol and industry attention to standardized multi-agent collaboration are further driving the generalized deployment of intelligent agents in networks.

[0003] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0004] (1) Zero-trust identity management and verification for distributed heterogeneous intelligent agents. The core challenge of AgentNet deployment lies in managing and verifying the identities of heterogeneous intelligent agents. The system consists of physical intelligent agents such as robots and drones, as well as virtual intelligent agents based on large language models (LLM) or large visual models (LVM). These entities differ fundamentally in physical form, operating protocols, and underlying architecture, urgently requiring the construction of a unified identity management framework. This framework aims to provide cohesive identity allocation, maintenance, and association mechanisms for intelligent agents with different skill sets, thereby realizing a zero-trust security architecture in the network system and ensuring smooth discovery, interaction, and collaboration among intelligent agents.

[0005] (2) Zero-trust auditing and verification of agents' claimed skill sets. To ensure efficient multi-agent planning and task allocation, it is essential to guarantee that the skills claimed by agents match their actual capabilities. Verifying whether potential participants possess specific skills is crucial when making decisions in controllers or collaborative clusters. Since agents operate in a decentralized environment, according to the zero-trust principle, a robust system must be established to verify and trace the performance of agents in various real-world environments. Implementing this framework is a necessary prerequisite for establishing a foundation of system trust and ensuring the reliable operation of multi-agent interaction and cooperation within the network.

[0006] (3) Privacy Protection and Secure Sharing in Multi-Agent Collaborative Learning and Inference. Ensuring data privacy while supporting the continuous evolution of agent skills is another core challenge. Unlike traditional static models, AgentNet emphasizes lifelong learning, dynamic adaptation, and collaboration among diverse agents, with agents continuously optimizing their capabilities through knowledge exchange. Therefore, developing a new collaborative learning paradigm is crucial. This paradigm needs to promote decentralized coordination among multi-agent agents across domains while strictly protecting the private data and sensitive information of each agent, balancing the benefits of knowledge sharing with the risks of privacy leakage, and ensuring the security of the AgentNet ecosystem. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method and system for the management and collaboration of zero-trust secure intelligent agent information communication networks based on skill chains.

[0008] This invention is implemented as follows: A zero-trust secure intelligent agent information communication network management and collaboration method based on skill chain includes:

[0009] S1, Identity Authentication: The intelligent agent submits an identity request to the certificate authority and receives a digital certificate.

[0010] S2, Task Mapping: The controller determines the target skill combination based on the task identifier carried by the agent.

[0011] S3, Skill Acquisition: The controller submits a transaction to the skill chain, acquires the skill combination tag from peer nodes, and downloads the associated skill combination from the skill set repository;

[0012] S4, Local Training: The agent uses private data to train the model in the skill set and generates updated parameters.

[0013] S5, parameter submission: The agent submits the updated parameters as a transaction to the collaboration chain corresponding to the skill combination;

[0014] S6, Aggregated Release: The collaborative chain uses smart contracts to aggregate update parameters of multiple smart agents to generate a global model and write it into the Skill Chain public dataset.

[0015] Furthermore, the identity authentication uses public key infrastructure and records the digital certificate status in a certificate revocation list.

[0016] Furthermore, the skill chain is an agent consortium chain architecture and stores tag metadata for each skill item, including basic information, capability description, evaluation metrics, file hash value, and skill set repository CID.

[0017] Furthermore, the collaborative chain adopts a Byzantine fault-tolerant consensus mechanism and performs signature verification and timing checks on submitted transactions before aggregation to prevent tampering.

[0018] Furthermore, the aggregation adopts a federated averaging algorithm and generates a hash value after aggregation, which is then written into the block header.

[0019] Another objective of this invention is to provide a zero-trust secure intelligent agent information communication network management and collaboration system based on a skill chain, comprising:

[0020] The identity authentication module is used to parse digital certificates to determine the identity of the intelligent agent;

[0021] The task mapping module is used to query the mapping table based on the task identifier and output the skill combination.

[0022] The skill acquisition module is used to send transactions to skill chain nodes and obtain the skill combination tags;

[0023] The skills download module is used to download associated skill sets from the skills group repository based on tags.

[0024] A local training module is used to train the skill combination based on private data and generate updated parameters.

[0025] The parameter submission module is used to encapsulate the updated parameters into a transaction and submit it to the corresponding collaborative chain;

[0026] The aggregation and release module is used to aggregate the updated parameters within the collaboration chain and write the global model back to the skill chain.

[0027] Furthermore, the identity authentication module includes a certificate verification unit and a certificate revocation management unit.

[0028] Furthermore, the skill acquisition module has a timed triggering mechanism to achieve periodic skill updates.

[0029] Another objective of this invention is to provide an intelligent agent device, comprising a communication unit, a storage unit, a training unit, and a control unit; the communication unit is used to exchange blocks and transactions with skill chain nodes and collaboration chain nodes; the storage unit is used to store downloaded skill combinations and local data; the training unit is used to perform private data training and protect gradients using homomorphic encryption; the control unit is used to invoke the communication unit, storage unit, and training unit and execute the aforementioned zero-trust secure intelligent agent information communication network management and collaboration method based on the skill chain.

[0030] Another object of the present invention is to provide a computer-readable storage medium having instructions stored thereon, which, when executed by a processor, cause the processor to perform the described method for managing and coordinating zero-trust secure intelligent agent information communication networks based on skill chains.

[0031] The advantages and positive effects of the technical solution to be protected by the invention are as follows:

[0032] Current agent communication networks typically rely on centralized or opaque agent identification and verification mechanisms. TrustAgentNet establishes a robust and auditable zero-trust trusted agent identification system. By requiring agents to register their identities with a certificate authority on a consortium blockchain, each participant receives a cryptographically secure digital certificate. This inherent trust foundation reduces the risks associated with malicious or infected agents, thereby creating a more secure and reliable collaborative environment.

[0033] A dual-layer storage architecture is introduced, separating the lightweight, immutable skill tag registry (STR) on-chain from the large off-chain model files hosted in decentralized storage. By recording only cryptographic content identifiers (CIDs) on CoS, the integrity of content is ensured when agents retrieve models. Authorized agents can upload or download skills on CoS via blockchain transactions, thereby building a secure agent AI marketplace.

[0034] Existing methods for describing and organizing agent skill combinations and labels may suffer from inconsistencies, lack of standardization, and limited search capabilities. TrustAgentNet proposes a CoS smart contract mechanism that uses a Proof-of-Performance (PoP) challenge to audit the initial skill performance claims submitted by agents in the context of real-world performance. This mechanism can dynamically adjust the agent's trust score, providing a verifiable and objective metric for capability integrity management.

[0035] Traditional multi-agent communication networks typically centralize the learning process within a single agent or rely on centralized servers for model aggregation, such as the classic federated learning (FL). TrustAgentNet achieves decentralized lifelong learning by facilitating collaborative training and inference among multiple agents (without direct data sharing). A Collaboration Chain (CoC) module was developed within CoS to support on-chain multi-agent collaborative model training and inference. This module utilizes transaction and aggregation smart contracts to support federated learning (such as...). Figure 4 As shown), distributed hybrid expert models (MoE, such as...) Figure 5 (as shown) and multi-agent model partitioning and sharing (MoPS) for large models, such as Figure 6 Various collaborative learning frameworks (as shown) can be used to achieve collective intelligence among multiple agents while ensuring data privacy.

[0036] The expected benefits and commercial value of the technical solution of this invention after transformation are as follows: This invention can be widely applied to intelligent network scenarios such as 6G, IoT, and edge computing. It achieves AI agent identity authentication, skill verifiability, and privacy-focused collaborative training and inference through blockchain, possessing high security and scalability. It has the commercial potential to be deployed in government and enterprise intelligent systems, industry AI platforms, and other similar applications.

[0037] This invention is the first to propose using blockchain to support skill management and collaborative training in AgenticAI networks, establishing a dual-chain structure of "skill chain + collaboration chain". This solves long-standing problems such as untrustworthy AI agent identities, unverifiable capabilities, and insecure collaborative learning. There is no similar systematic solution globally.

[0038] This solution effectively addresses key technical challenges such as heterogeneous identity authentication, capability verification, and collaborative learning for data privacy among multi-agent systems. In particular, it enables trusted training even without a central coordination node, breaking through the security and scalability bottlenecks of traditional AI systems and federated learning.

[0039] Wide range of applications: This invention can be widely applied to intelligent network scenarios such as 6G networks, large-scale Internet of Things, and edge computing, enabling AI agents to securely access and evolve their capabilities in these environments.

[0040] Reduce operating costs and risks: By establishing a trusted identity and skills verification mechanism, the security risks associated with malicious or infected intelligent agents are significantly reduced, and economic losses caused by security vulnerabilities are reduced.

[0041] The technical solution of this invention solves a technical problem that people have long desired to solve but have been unable to achieve:

[0042] Breaking the "Information Silos" and Trust Problems: Effectively Solving the Issues of Inconsistency in Industrial Multi-Agent Deployment

[0043] The problem of "information silos" caused by data sensitivity among algorithm models in the same position, and the gap in identity trust between heterogeneous intelligent agents.

[0044] Achieving highly reliable distributed collaboration: Solving the problems of traditional federated learning in terms of model traceability, etc.

[0045] The shortcomings in the new tracing mechanism ensure that the system can still converge and continue to operate even when some nodes are poisoned (such as 30% of nodes being attacked) or the coordinator fails at a single point of failure.

[0046] Balancing privacy and collective intelligence: This addresses the challenge of achieving this without directly sharing private data.

[0047] How to achieve collaborative improvement of collective intelligence among multiple agents across different domains?

[0048] The technical solution of this invention overcomes technical bias: it changes the perception that "zero trust equals high overhead": through vectorized semantic indexing and low-rank decomposition with equal weighted compression techniques, it proves that while achieving zero-trust security auditing, it is possible to achieve it without increasing or even significantly optimizing the initialization overhead and operational burden of the intelligent agent. Attached Figure Description

[0049] Figure 1 This is a flowchart of the zero-trust secure intelligent agent information communication network management and collaboration method based on skill chain provided in the embodiments of the present invention.

[0050] Figure 2 This is a flowchart illustrating the skill set acquisition process implemented in this invention.

[0051] Figure 3 This is a schematic diagram of the skill tags of the present invention.

[0052] Figure 4 This is a flowchart illustrating the collaborative chain establishment process implemented in this invention.

[0053] Figure 5 This is a schematic diagram of parameter aggregation (FL mode) implemented by the present invention.

[0054] Figure 6 This is a schematic diagram of the parameter aggregation (MOE mode) implemented in this invention.

[0055] Figure 7 This is a schematic diagram of the parameter aggregation (MOPS mode) implemented in this invention.

[0056] Figure 8 This is a flowchart of the process for adding a new intelligent agent to the intelligent agent alliance chain and updating the skill set, provided by an embodiment of the present invention.

[0057] Figure 9 This is a schematic diagram of the prototype of the present invention.

[0058] Figure 10 This is a flowchart illustrating how TrustAgentNet skills, once on-chain, effectively resist various malicious attacks, as provided in this embodiment of the invention.

[0059] Figure 11 yes Figure 10 The specific performance of TrustAgentNet compared to traditional federated learning when encountering a single point of failure attack on the coordinator in the corresponding embodiment.

[0060] Figure 12 yes Figure 10 The corresponding implementation shows the specific performance of TrustAgentNet compared to traditional federated learning when 30% of the agents are poisoned.

[0061] Figure 13This is a flowchart illustrating how a secure intelligent agent intelligently switches its on-chain skill mode after a user is attacked, as provided in this embodiment of the invention.

[0062] Figure 14 This is a block diagram of the zero-trust secure intelligent agent information communication network management and collaboration system based on skill chain provided in the embodiments of the present invention.

[0063] Figure 15 This is a performance comparison chart of the TrustNet system under different consensus node ratios provided in the embodiments of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0065] like Figure 1 As shown in the figure, the zero-trust secure intelligent agent information communication network management and collaboration method based on skill chain provided by this invention includes the following steps:

[0066] S1, Identity Authentication: The intelligent agent submits an identity request to the certificate authority and receives a digital certificate.

[0067] S2, Task Mapping: The controller determines the target skill combination based on the task identifier carried by the agent.

[0068] S3, Skill Acquisition: The controller submits a transaction to the skill chain, acquires the skill combination tag from peer nodes, and downloads the associated skill combination from the skill set repository;

[0069] S4, Local Training: The agent uses private data to train the model in the skill set and generates updated parameters.

[0070] S5, parameter submission: The agent submits the updated parameters as a transaction to the collaboration chain corresponding to the skill combination;

[0071] S6, Aggregated Release: The collaborative chain uses smart contracts to aggregate update parameters of multiple smart agents to generate a global model and write it into the Skill Chain public dataset.

[0072] The identity authentication provided in this embodiment of the invention uses public key infrastructure and records the status of digital certificates in a certificate revocation list.

[0073] The skill chain provided in this embodiment of the invention is an agent consortium chain architecture, and each skill item stores tag metadata containing basic information, capability description, evaluation indicators, file hash value, and skill set repository CID.

[0074] The collaborative chain provided in this embodiment of the invention adopts a Byzantine fault-tolerant consensus mechanism and performs signature verification and timing checks on submitted transactions before aggregation to prevent tampering.

[0075] The aggregation provided in this embodiment of the invention uses a federated averaging algorithm and generates a hash value after aggregation, which is then written into the block header.

[0076] The intelligent agent identity authentication provided in this embodiment of the invention: When an intelligent agent participates in an intelligent agent consortium blockchain, it first needs to apply for a unique identity from the Certificate Authority (CA) of the intelligent agent consortium blockchain; the CA will issue a digital certificate to the intelligent agent, which contains a public key, a certificate signed by the CA and a corresponding private key, following the standard public key infrastructure (PKI) system.

[0077] The intelligent agent provided in this embodiment of the invention actively acquires and downloads skill sets from the skill chain; each available skill in the network is tagged with descriptive functions and lifecycle traceability information, and is stored on peer nodes of the skill set chain database. The mapping agent actively acquires relevant skill sets (such as...) by submitting query transactions to the skill chain (CoS). Figure 2 As shown), this process is permanently stored by the query record module. The CoS smart contract returns the corresponding skill tag (e.g., Figure 3 As shown, the unique Content Identifier (CID) contained within the skill file serves as a cryptographic pointer to the actual skill file in the decentralized storage network. After downloading the skill component, the agent independently recalculates the file's cryptographic hash and verifies its complete consistency with the tamper-proof CID in the CoS tag. Since the CID is a cryptographic signature of the file content, a matching CID ensures that the file acquired by the agent is indeed the audited original registered on CoS, effectively preventing tampering or replacement during transmission over the storage network. Once cryptographic verification is successful, the agent card (usually a JSON metadata file) is synchronously updated to reflect the new capabilities. Subsequently, the updated agent card is broadcast to the agent controller, enabling the agent to immediately accept further task assignments based on the newly acquired skills. It is worth noting that the skill acquisition process is not limited to the initialization phase but can continue throughout the agent's entire lifecycle, adapting to changes in task requirements.

[0078] This invention provides a method for downloading associated skill sets from a skill set repository based on tags and uploading them to a collaborative chain. After downloading skill set s from the chain, an agent can use its private dataset to train a base model locally. For each skill set s, a dedicated collaborative chain is established, which can only be joined by authorized agents for private and confidential model training. An agent possessing skill set i must be authenticated and authorized to join collaborative chain i, after which it can submit its locally trained parameters as a transaction to the collaborative chain.

[0079] The skill sets provided in this embodiment of the invention are actively updated and evolved on the collaboration chain (e.g. Figure 4 (As shown in the diagram), for each collaboration chain i, after receiving updated model parameters from multiple authorized agents, the smart contract aggregates these updates from multiple agents (e.g., using FedAvg) to compute refined global model parameters; the updated global skill set model and labels are written to the public dataset of the collaboration chain and the skill set chain; agents that have joined the skill set chain can view and download updated skill set data information in the public dataset, including model parameters and skill set labels; therefore, the skill set on the chain can iteratively evolve through decentralized collaborative learning, utilizing the unique datasets of different agents without compromising their data privacy;

[0080] A Collaborative Chain (CoC) module was developed within CoS to support on-chain multi-agent collaborative model training and inference. This module utilizes transaction and aggregation smart contracts to support federated learning (such as...). Figure 5 As shown), distributed hybrid expert models (MoE, such as...) Figure 6 (as shown) and multi-agent model partitioning and sharing (MoPS) for large models, such as Figure 7 Various collaborative learning frameworks (as shown) can be used to achieve collective intelligence among multiple agents while ensuring data privacy.

[0081] Specific implementation of the present invention:

[0082] In typical industrial multi-agent deployments, the biggest obstacle is the "information silos" and trust gaps between algorithm models in different roles: edge nodes in the factory have access to real-world operating data but worry about uploading sensitive samples to the cloud; while cloud-based algorithm teams urgently need large-scale heterogeneous samples to improve prediction accuracy. While previously common federated learning frameworks alleviated data silos, they still fall short in terms of model traceability, update tracking, and cross-domain skill reuse, leading to long model iteration cycles, broken accountability chains, and difficulty in supporting the demanding production rhythms of industry settings.

[0083] To address this, this method first introduces a verifiable credential system based on X.509 extended fields on the identity side, replacing the traditional one-time "node-center" authentication with on-chain identities that are dynamically refreshed with each task. The programmable attestation embedded in the digital certificate works in conjunction with the national cryptographic standard SM2 signature, ensuring seamless credential renewal even during edge device reconfiguration, firmware hot-updates, or network switching, fundamentally solving the "trust drift" problem during agent migration.

[0084] Subsequently, the controller retrieves the optimal "skill combination" by parsing the task descriptor and using a vectorized skill semantic index. This process essentially involves semantic clustering and differentiated packaging of numerous micro-models within the distributed model repository (SkillChain): selecting base domain models using tensor space cosine similarity, and then re-weighting the skill labels using a Graph Attention Network to generate a "minimum feasible skill package" tailored to specific edge scenarios. In this way, industrial equipment no longer needs to pull redundant weights from the entire repository, reducing download volume and initialization latency to one-tenth of the original.

[0085] During the local training phase, an adaptive differential privacy (FedAdaDP) strategy is employed, performing self-noise injection and learnable pruning threshold scheduling before each gradient update to balance convergence speed and privacy budget. To avoid straining edge hardware computing power, the algorithm uses low-rank decomposition and quantization-aware fine-tuning (QAT) techniques to compress over 90% of the convolutional weights into 8-bit representations, while maintaining an inference accuracy drop of less than 1%. This process ensures that the updated parameters generated at the edge have publishing value without revealing local sample distribution characteristics.

[0086] The parameter submission chain utilizes a two-layer transaction channel: metadata is broadcast through the skill chain mainnet, and the actual gradients are stored in IPFS shards and written to the collaboration chain sidechain in Merkle proof format. This design ensures that every model update leaves a verifiable hash fingerprint; even if a rollback is required or a compliance audit occurs in the future, the source node and timestamp can be quickly located based on the on-chain logs, achieving "model-level traceable quality management".

[0087] Finally, the collaborative chain smart contract triggers asynchronous aggregation. When the specified number of updates or cumulative contribution thresholds are met, the contract calls the aggregation operator (supporting strategies such as FedAvg, FedNova, and SCAFFOLD) to generate global weights, which are then pushed back to the skill chain public area after consensus is reached. This forms a "skill-training-release" closed loop: edge devices always receive the latest verifiable models, the cloud can quantify the contribution of each node in real time and distribute computing power incentives as needed, thereby transforming complex cross-layer collaboration into a deployable, operable, and sustainably evolving intelligent agent ecosystem in industrial settings.

[0088] I. A Skill Chain-Based Zero-Trust Secure Intelligent Agent Information Communication Network Architecture Framework

[0089] (1) Infrastructure

[0090] This includes hardware infrastructure, such as cloud and edge computing and storage resources, and communication networks that connect agents and blockchains and grant them access to these resources, as well as software systems available to agents, including high-quality datasets, accumulated skill sets, and comprehensive world models.

[0091] (2) Skill chain

[0092] It includes a blockchain containing the following sub-functional modules: 1) Authentication: Providing authenticated digital identities for new agents joining the network; 2) Skill set tag storage: Storing all available and up-to-date skill set tags (generally JSON metadata), each tag being associated with a skill set to describe its functionality and lifecycle traceability; 3) Tag authorization: Providing trusted tags to agents based on their functional requests.

[0093] (3) Decentralized storage network for skill sets

[0094] Leveraging distributed storage resources across the network, TrustAgentNet provides persistent, content-addressed storage services for all skill sets using decentralized storage technologies such as the InterPlanetary File System (IPFS). When a skill set is stored, the system generates a unique Content Identifier (CID) (similar to the Content ID in IPFS). This CID is recorded in the skill tag and stored on the Skill Chain (CoS). This means that CoS does not directly store the massive file bodies, but rather stores cryptographic hash links pointing to these files. Any agent can use the CID recorded in the CoS tag to query or download a specific skill set in a trustless manner. Due to the characteristics of the CID, this ensures that the retrieved content is completely consistent with the data originally registered and audited on CoS. This separation of the decentralized ledger (CoS) and decentralized storage (skill storage) is the cornerstone of TrustAgentNet's scalability and trustworthiness.

[0095] (4) Task-specific collaboration chains (CoC)

[0096] The CoS aims to facilitate privacy-preserving coordination among multiple agents to collaboratively complete training or inference tasks within a CoS (Co-Agent Chain). Task initiators launch tasks by submitting transactions to the CoS, thereby establishing corresponding CoCs (Co-Agent Chains). Agents can then voluntarily apply to join the chain or be invited to do so. Joining an existing collaborative chain requires endorsement, typically requiring a certain percentage of consensus among currently participating agents. Once endorsement is granted, each agent can submit its local outputs to the blockchain as transactions for verification and aggregation. CoS supports massive parallel CoC execution and is compatible with various multi-agent decentralized collaboration frameworks. Each chain focuses on collaborative learning and capability evolution to meet specific task requirements.

[0097] (5) Intelligent agent

[0098] This includes various types of task-oriented agents. Each type of agent possesses the same capabilities. Each agent implemented in a specific environment has a unique set of local observation datasets and can have multiple skill combinations to complete a specific task. The agent's skill combinations can be further evolved based on local training.

[0099] Two key intelligent agents:

[0100] Firstly, the task-skill mapping agent (M-agent). In agent communication networks, agents are tailored to specific tasks. Therefore, the task-skill mapping agent is crucial for efficient operation, acting as a central hub that associates a specific task with the necessary skill set required to complete it. By continuously learning and accumulating knowledge in its knowledge base, the mapping agent gradually improves its understanding of the mapping function. When an authenticated agent needs to perform a task, it submits its request to the mapping agent, which then replies with a list of corresponding skill combinations. This mechanism ensures an effective match between task objectives and agents with appropriate capabilities within the network.

[0101] Secondly, the S-agent. This agent can assess the safety of other agents. It monitors the operation of other agents in real time (e.g., accuracy, loss, gradient). When anomalies occur (e.g., abnormal oscillations in accuracy and loss, gradient shift), the S-agent assesses the safety of the agent and isolates or updates the agent's skills based on the assessment results.

[0102] (6) Skill set / agent tag

[0103] A smart agent label is a trusted digital certificate backed by blockchain technology used to demonstrate the capabilities of a smart agent. This allows the smart agent controller to find the most suitable smart agent for a given subtask. Typically, a smart agent label can be a JSON metadata file containing labels for its entire skill set.

[0104] (7) Intelligent agent controller

[0105] Coordinate the collaboration and task cognition behaviors between agents as defined in (4). It includes the following sub-functional modules: 1) Task cognition: formulate corresponding tasks according to the user's semantics; 2) Task separation: decompose the task into multiple sub-tasks that need to be executed to achieve the user's semantic goals; 3) Agent adaptation: assign each sub-task to a specific agent according to the agent's label.

[0106] like Figure 8 As shown, this flowchart illustrates the complete lifecycle of a "new agent" applying to join the agent consortium blockchain and acquiring a skillset. The entire process begins at the top with the "Start: New Agent Applying to Join" node. First, the Certificate Authority (CA) of the agent consortium blockchain determines at node S1 whether the agent is qualified for entry. If the CA determines "not allowed," the process terminates directly along the red dotted line on the right; if "allowed," it proceeds to the subsequent task collaboration and skills management stages.

[0107] (1) After CA certification is passed, the agent actively submits its task requirements; the system matches its requirements through the "task-skill mapping" module (M-agent) and returns a list of candidate skills. Subsequently, the agent applies to the skill chain at the S3 node to obtain the required skill set tag information. After obtaining the skills, the system will synchronously update the agent's digital certificate tag and notify the scheduler so that its new capabilities can be identified in subsequent training scheduling. This step can be executed multiple times to ensure that skills and task requirements are continuously aligned.

[0108] (2) Next, in the S4 stage: the agent actively downloads the required skill set from the skill set repository to its local machine based on the authorized skill set tags it possesses, and conducts self-training in a sandbox or edge environment. After completing local training, the agent needs to be certified and authorized by the MSP (Member Service Provider) before it can officially join the collaborative chain and upload its local training parameters. Through this "on-chain-off-chain" collaborative training architecture, the system protects model privacy and data sovereignty while ensuring incremental updates of the global model.

[0109] At the S5 node, the on-chain smart contract aggregates the training parameters submitted by each agent, updates the global model, and recalculates the aggregation coefficients. Subsequently, the SkillChain public dataset releases the latest skill set version. All agents already on the chain can download and apply this latest skill in real time, forming a continuously iterating closed loop. If the CA's initial review fails or certification fails midway, the process will return to the terminal node along the red dotted line path, ensuring the security and compliance of the agent consortium blockchain.

[0110] The present invention provides a method for managing and coordinating a zero-trust secure intelligent agent model based on a skill chain, comprising the following steps:

[0111] Step S1: Agent identity authentication.

[0112] For an intelligent agent to participate in an intelligent agent consortium blockchain, it first needs to apply for an identity from the blockchain's Certificate Authority (CA). The CA will issue a digital certificate to the intelligent agent, which contains a public key, a certificate signed by the CA, and a corresponding private key, following a standard Public Key Infrastructure (PKI) system.

[0113] Step S2: Task-Skill Set Mapping.

[0114] Based on the task carried by the agent, access the task-skill set mapping to obtain the list of skill combinations required to complete the task.

[0115] Step S3: Obtain skill set tags from the skill chain.

[0116] Each available skill in the network, tagged with descriptive functionality and lifecycle traceability information, is stored on peer nodes in the skill set chain database. Authenticated agents can acquire relevant skill set tags by submitting proposals to the skill set chain. Notably, skill set acquisition is not limited to initialization but can also occur periodically throughout the agent's operations. Once an agent obtains the necessary skill set tags, its authentication tag (which can be stored as a JSON metadata file) is updated and published to the agent controller for further task allocation.

[0117] Step S4: Download the associated skill set from the skill set repository and upload it to the collaboration chain.

[0118] After an agent downloads its associated skill set *s* from the skill set repository based on its authorized skill set tags, it can use its private dataset to train the base model locally. For each skill set *s*, a dedicated collaborative chain is established, which only authorized agents can join for private and confidential model training. An agent possessing skill set *i* must be authenticated and authorized to join collaborative chain *i*, after which it can submit its locally trained parameters as transactions to the collaborative chain.

[0119] Step S5: Skill sets continue to evolve along the collaboration chain.

[0120] For each collaborative chain i, after receiving updated model parameters from multiple authorized agents, the smart contract aggregates these updates from multiple agents (e.g., using FedAvg) to compute refined global model parameters. The updated global skill set model and labels are written to the public dataset of both the collaborative chain and the skill set chain. Agents already on the skill set chain can view and download updated skill set data information, including model parameters and skill set labels, from the public dataset. Therefore, on-chain skill sets can iteratively evolve through decentralized collaborative learning, leveraging the unique datasets of different agents without compromising their data privacy.

[0121] Specific implementation of the present invention:

[0122] Example 1

[0123] like Figure 10 The diagram illustrates the security improvements of TrustAgentNet compared to federated learning. Each step is described in detail below:

[0124] Prototype system deployment and network setup

[0125] A desktop server equipped with an Intel i5-10500T processor and 64GB of RAM was used to deploy the blockchain node and agent system. The system employed Docker technology to simulate multiple blockchain nodes and agents, constructing a test scenario closer to a real-world environment. In the network topology, the system was configured with 20 user nodes, each bound to one agent and possessing its own independent local dataset. To verify the collaborative learning performance, handwritten digit recognition was selected as the skill task, trained on the MNIST dataset using the LeNet-5 model. The agents aggregated the model using the FedAvg method.

[0126] Blockchain platform initialization

[0127] The system utilizes the open-source blockchain platform Hyperledger Fabric to implement two chain systems: the Chain of Skillsets and the Chain of Collaboration. The Chain of Skillsets consists of 3 orderer nodes and 7 organizations; each chain of collaborations consists of 3 orderer nodes and 5 organizations, with each organization configured with 1 peer node. The consensus mechanism adopts the Etcdraft protocol, and the algorithm uses Raft to ensure the stability and security of the blockchain system in a multi-node environment.

[0128] Verify security enhancements for skills training under malicious attacks

[0129] To evaluate the enhanced security of TrustAgentNet compared to traditional federated learning in distributed learning, we tested its resilience against two common attack types: (a) attacks targeting the coordinator server, resulting in a single point of failure; and (b) poisoning attacks, where 30% of the agents maliciously manipulate their uploaded local model parameters. Experiments showed that when the coordinator server shut down during the 5th iteration, classic federated learning ceased operation; however, TrustAgentNet, trained on-chain, continued training uninterrupted thanks to the distributed ledger across all servers. Furthermore, when 30% of the participating agents were subjected to a poisoning attack, classic federated learning failed to converge, while TrustAgentNet maintained convergence through consensus and endorsement mechanisms. The experimental results highlight the improved security of TrustAgentNet.

[0130] Example 2

[0131] like Figure 13 As shown, this example aims to propose a working mode in which federated learning and TrustAgentNet work together. Under normal circumstances, federated learning still runs, but when subjected to malicious attacks, it switches to TrustAgentNet working mode.

[0132] Step S1: Initialize the training scenario.

[0133] Participating users began training the handwritten digit recognition model using a traditional federated learning (off-chainFL) model.

[0134] Step S2: Simulate an attack on the central server.

[0135] The training process was disrupted when the central server was attacked, leading to performance degradation and a security threat. The security agent (S-agent) assessed the system's current resource availability (computing, storage, network, etc.) to determine whether it was feasible to switch to blockchain training mode.

[0136] Step S3: Switch training mode.

[0137] If resources are sufficient, S-agent coordinates and notifies all participating users to switch the training mode to the blockchain-enhanced TrustAgentNet mode. Blockchain mode training is initiated, and training parameters are verified through an on-chain consensus mechanism and recorded in a distributed ledger, ensuring data immutability and security. Simultaneously, the system quickly resumes the training task, resists attacks, and achieves continuous and secure training.

[0138] This invention has completed a prototype system based on the TrustAgentNet architecture. For details of the architecture, please refer to [link / reference needed]. Figure 8 .like Figure 9As shown, the prototype system has effectively completed the core "skill chain + collaboration chain" dual-chain system, and clearly and completely demonstrated the entire system architecture.

[0139] During the prototype system's startup phase, each edge agent first completes node-level identity verification and key derivation through a two-way handshake with the root certificate authority (CA). Subsequently, the AI ​​Agent Controller integrates it into a unified scheduling domain. This allows the control plane to perceive the trustworthiness of newly connected entities within milliseconds and bind a set of verifiable public key identities to subsequent on-chain transactions, thereby establishing a root of trust outside the physical boundary. The entire process is similar to... Figure 9 The “Identity Authentication” path on the left ensures that all on-chain operations have a traceable subject mapping, preventing network integrity from being compromised by the insertion of anonymous nodes.

[0140] The system's knowledge resources are carried by "Chain of Skillsets." This chain operates as an agent consortium blockchain, with each block containing immutable metadata tags for a single skillset: basic information, capability description, evaluation metrics, file hash value, and skillset repository CID are all written into an on-chain JSON structure. Figure 9 (For example, a tag), and it is fused with the previous block via Merkle hashing. Researchers can retrieve the optimal implementation of a specific capability at any point in time, and the scheduler only needs to verify the block header to confirm the uniqueness and integrity of the packet—this avoids the version drift and tampering risks common in traditional centralized model libraries, and also makes cross-domain reuse and traceability auditing possible.

[0141] Each skill set automatically derives an independent "Chain of Collaboration" during continuous training. The collaboration chain solidifies events such as gradient submission, aggregation signature, and model snapshot into blocks, and ensures parameter correctness and timing consistency in the multi-party collaboration process through lightweight Byzantine fault-tolerant consensus. Figure 9 The purple and blue vertical links in the middle illustrate the controller pushing encrypted gradient streams to the collaborative chain nodes. When any node acts maliciously or uploads invalid updates, the on-chain auditor can locate and roll back the behavior within one or two rounds of consensus, preventing pollution of the global model. This "one skill, one chain" sidechain mechanism not only reduces the load on the main chain but also naturally provides chain-level boundaries for training privacy isolation.

[0142] The AI ​​Agent Controller at the bottom is responsible for dynamically loading and hot-updating multiple skill packages on the user interface (UE), and monitoring the usage of local inference and training resources in real time. Once the collaboration chain completes parameter convergence, the controller triggers the writing of the new version tag back to the skill chain, and then distributes the model weights, verified by the block header, to each associated terminal, achieving a second-level closed loop. In this way, the system maintains distributed transparency while possessing the low-latency operating characteristics required for industrial applications; the dual-chain synergy of "skill chain + collaboration chain" becomes the core tension of the entire architecture, truly enabling the continuous evolution and economical sharing of trustworthy AI.

[0143] Based on Example 1, the experimental results are shown below. Figure 11 and Figure 12 The figure clearly shows the difference in training accuracy between TrustAgentNet and traditional federated learning when TrustAgentNet encounters two common attacks.

[0144] Figure 11 The results show that when a single point of failure occurs in the coordinator, although both traditional federated learning and TrustAgentNet continue training, TrustAgentNet has a slight advantage in accuracy.

[0145] Figure 12 The results show that when 30% of the agents are poisoned, the accuracy of TrustAgentNet still steadily increases with the number of iterations, but traditional federated learning is greatly affected, with its accuracy plummeting and failing to recover to the pre-attack level even with the number of iterations.

[0146] like Figure 14 As shown in the figure, an embodiment of the present invention provides a zero-trust secure intelligent agent information communication network management and collaboration system based on a skill chain, comprising:

[0147] The identity authentication module is used to parse digital certificates to determine the identity of the intelligent agent;

[0148] The task mapping module is used to query the mapping table based on the task identifier and output the skill combination.

[0149] The skill acquisition module is used to send transactions to skill chain nodes and obtain the skill combination tags;

[0150] The skills download module is used to download associated skill sets from the skills group repository based on tags.

[0151] A local training module is used to train the skill combination based on private data and generate updated parameters.

[0152] The parameter submission module is used to encapsulate the updated parameters into a transaction and submit it to the corresponding collaborative chain;

[0153] The aggregation and release module is used to aggregate the updated parameters within the collaboration chain and write the global model back to the skill chain.

[0154] The identity authentication module provided in this embodiment of the invention includes a certificate verification unit and a certificate revocation management unit.

[0155] The skill acquisition module provided in this embodiment of the invention has a timed triggering mechanism to achieve periodic skill updates.

[0156] The zero-trust secure intelligent agent information communication network management and collaboration system based on skill chains provided by this invention takes "trustworthy identity, controllable capabilities, verifiable processes, and traceable results" as its core design concept. Through the deep integration of blockchain and intelligent agent training mechanisms, it realizes secure collaboration and capability evolution of multiple intelligent agents in a zero-trust network environment. Its overall working principle is as follows.

[0157] During system operation, the agent first connects to the network through the identity authentication module. This module, based on a digital certificate system, verifies the integrity and legality of the certificates submitted by the agent and, in conjunction with the certificate revocation management unit, determines whether the certificate is in a valid and trusted state. This completes zero-trust identity verification at the network boundary, preventing unauthorized or invalid nodes from participating in subsequent communication and collaboration. After successful identity authentication, the agent sends a request to the task mapping module carrying a task identifier. The task mapping module, based on a pre-built task-skill mapping table, parses the abstract task requirements into executable skill combinations, providing clear guidance for subsequent capability acquisition.

[0158] Subsequently, the skill acquisition module sends a transaction request to the skill chain node based on the skill combination description, obtaining the corresponding skill combination tag through the on-chain consensus mechanism. This process leverages the immutability and traceability of blockchain to ensure the authenticity and consistency of skill identifiers. Simultaneously, the skill acquisition module incorporates a timed trigger mechanism, enabling the agent to periodically detect changes in the skill chain state. When a skill is updated or optimized, the acquisition process is automatically triggered, thus achieving continuous capability evolution. After obtaining the skill combination tag, the skill download module downloads the associated specific skill model or parameter set from the skill group repository based on the tag index and stores it in a local secure environment.

[0159] After skill deployment, the local training module uses the downloaded skill set as the initial capability base and combines it with the agent's own private data to perform local training or fine-tuning operations. During training, the system encrypts gradients or model update parameters using privacy protection technologies such as homomorphic encryption, ensuring data privacy and business security from a mechanism perspective by eliminating the need to expose raw data or plaintext parameters during collaboration. After training is complete, the parameter submission module encapsulates the encrypted update parameters into a transaction and submits it to the corresponding collaborative chain node.

[0160] The aggregation and publishing module in the collaborative chain is responsible for verifying, decrypting, authorizing, and aggregating updated parameters from multiple agents to form a globally optimized skill model or parameter set, and then writing the aggregation results back to the skill chain. Through this on-chain publishing mechanism, the system achieves global sharing and consistent updating of skill capabilities while maintaining the privacy of local data for each agent. This forms a closed-loop operating mechanism of "identity verification—skill acquisition—local training—secure collaboration—capability write-back," enabling secure management and efficient collaboration of multi-agent communication networks under zero-trust conditions.

[0161] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0162] To verify the effectiveness of the zero-trust secure intelligent agent management system based on the skill chain proposed in this invention, this embodiment conducted multiple performance evaluations and comparative tests based on a hardware prototype. Specific evidence is as follows:

[0163] 1. Evidence of robustness against malicious attacks

[0164] Experimental setup: In a Collaborative Chain (CoC) environment, 30% of malicious nodes (MaliciousAgents) are artificially introduced to execute poisoning attacks, interfering with the aggregation of the global model.

[0165] Evidence data: Experimental results show (reference) Figure 10 , Figure 11 Even under attack, the global model's test accuracy can still be maintained above 85% using the distributed collaboration framework of this invention, and the convergence speed is minimally affected by interference.

[0166] Technical Conclusion: Compared with traditional centralized control solutions, this invention achieves the system's self-healing capability through blockchain and zero-trust verification mechanisms, effectively isolating the negative impact of malicious nodes.

[0167] 2. Evidence of system scalability

[0168] Experimental Setup: Configure a TrustNet system with 100 consensus nodes, setting the node ratio required for consensus to be 20%, 40%, 80%, and 100% respectively. Compare the runtime, actual measured CPU usage time, and bandwidth usage under different node ratios of consensus endorsement (see reference). Figure 15 ).

[0169] Evidence data: Experiments have shown that as the number of consensus nodes increases, the blockchain consensus time grows linearly rather than exponentially, and the CID (Content Identifier) ​​indexing mechanism ensures that the model retrieval speed is not affected by the node size.

[0170] Technical Conclusion: This solution possesses strong horizontal scalability and is suitable for large-scale heterogeneous agent networks.

[0171] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0172] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for managing and coordinating zero-trust secure intelligent agent information communication networks based on skill chains, characterized in that, The zero-trust secure intelligent agent information communication network management and collaboration method based on skill chain includes the following steps: S1, Identity Authentication: The intelligent agent submits an identity request to the certificate authority and receives a digital certificate. S2, Task Mapping: The controller determines the target skill combination based on the task identifier carried by the agent. S3, Skill Acquisition: The controller submits a transaction to the skill chain, acquires the skill combination tag from peer nodes, and downloads the associated skill combination from the skill set repository; S4, Local Training: The agent uses private data to train the model in the skill set and generates updated parameters. S5, parameter submission: The agent submits the updated parameters as a transaction to the collaboration chain corresponding to the skill combination; S6, Aggregated Release: The collaborative chain uses smart contracts to aggregate update parameters of multiple smart agents to generate a global model and write it into the Skill Chain public dataset.

2. The method for managing and coordinating zero-trust secure intelligent agent information communication networks based on skill chains as described in claim 1, characterized in that, The identity authentication uses public key infrastructure and records the status of digital certificates in a certificate revocation list.

3. The method for managing and coordinating zero-trust secure intelligent agent information communication networks based on skill chains as described in claim 1, characterized in that, The skill chain is an agent consortium chain architecture, and each skill item stores tag metadata containing basic information, capability description, evaluation indicators, file hash value, and skill set repository CID.

4. The method for managing and coordinating zero-trust secure intelligent agent information communication networks based on skill chains as described in claim 1, characterized in that, The collaborative chain adopts a Byzantine fault-tolerant consensus mechanism and performs signature verification and timing checks on submitted transactions before aggregation to prevent tampering.

5. The method for managing and coordinating zero-trust secure intelligent agent information communication networks based on skill chains as described in claim 1, characterized in that, The aggregation uses a federated averaging algorithm, and after aggregation is completed, a hash value is generated and written into the block header.

6. A zero-trust secure intelligent agent information communication network management and collaboration system based on a skill chain, implementing the skill chain-based zero-trust secure intelligent agent information communication network management and collaboration method as described in any one of claims 1-6, characterized in that, include: The identity authentication module is used to parse digital certificates to determine the identity of the intelligent agent; The task mapping module is used to query the mapping table based on the task identifier and output the skill combination. The skill acquisition module is used to send transactions to skill chain nodes and obtain the skill combination tags; The skill download module is used to download associated skill sets from the skill set repository based on tags; A local training module is used to train the skill combination based on private data and generate updated parameters. The parameter submission module is used to encapsulate the updated parameters into a transaction and submit it to the corresponding collaborative chain; The aggregation and release module is used to aggregate the updated parameters within the collaboration chain and write the global model back to the skill chain.

7. The zero-trust secure intelligent agent information communication network management and collaboration system based on skill chain as described in claim 6, characterized in that, The identity authentication module includes a certificate verification unit and a certificate revocation management unit.

8. The zero-trust secure intelligent agent information communication network management and collaboration system based on skill chain as described in claim 6, characterized in that, The skill acquisition module has a timed triggering mechanism to achieve periodic skill updates.

9. An intelligent agent device, characterized in that, The system includes a communication unit, a storage unit, a training unit, and a control unit; the communication unit is used to exchange blocks and transactions with skill chain nodes and collaboration chain nodes; the storage unit is used to store downloaded skill combinations and local data; the training unit is used to perform private data training and protect gradients using homomorphic encryption; the control unit is used to invoke the communication unit, storage unit, and training unit and execute the zero-trust secure intelligent agent information communication network management and collaboration method based on the skill chain as described in any one of claims 1 to 5.

10. A computer-readable storage medium having instructions stored thereon that, when executed by a processor, cause the processor to perform the zero-trust secure intelligent agent information communication network management and coordination method based on the skill chain as described in any one of claims 1 to 5.

Images