Intelligent model credible sharing method and system based on block chain and computing power network

Abstract

The invention discloses an intelligent model credible sharing method and system based on a block chain and a computing power network, and belongs to the technical field of cross of block chains, computing power networks and artificial intelligence. The method comprises the following steps: obtaining a verifiable evolutionary fingerprint of a to-be-shared intelligent model and a performance reference thereof, and respectively storing the verifiable evolutionary fingerprint and the performance reference in a main chain and a side chain of a block chain; constructing a semantic ability portrait vector of the computing power node; receiving a model calling request verified by an attribute-based encryption mechanism; based on the capability portrait vector, taking the minimum comprehensive cost as a target, dynamically selecting a target computing power node from nodes meeting preset security and timeliness constraints to execute a model reasoning task, the comprehensive cost being related to estimated energy consumption, time delay and security reputation of the nodes; and performing multi-party verification on the task execution result, and updating the excitation state of the model contributor and the computing power node according to the verification result and the node performance. According to the invention, credible, efficient and green sharing of the intelligent model in a distributed environment is realized.

Description

Technical Field

[0001] This invention relates to the field of blockchain, computing power networks and artificial intelligence, and in particular to a trusted sharing method and system for intelligent models based on blockchain and computing power networks. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, intelligent models such as deep learning and machine learning models have been widely applied in fields such as intelligent perception, intelligent decision-making, data analysis, and automatic control. The application of intelligent models has evolved from local deployment in single scenarios to a shared service model across industries and regions. Model reuse and collaborative invocation have become key means to improve the overall efficiency of AI systems. However, the training and inference processes of intelligent models are highly dependent on high-performance computing resources, and the demand for computing power continues to rise as the model size increases. In actual deployment environments, computing resources exhibit decentralized characteristics across clouds, edges, and regions, forming a heterogeneous computing power network composed of cloud computing nodes, edge computing nodes, and dedicated acceleration devices (such as GPUs and TPUs). Against this backdrop, how to achieve efficient deployment, flexible scheduling, and secure invocation of intelligent models within complex computing power networks has become a core challenge facing current AI systems.

[0003] Currently, the mainstream methods for sharing intelligent models mainly rely on centralized model management platforms or third-party model service marketplaces. These platforms are typically managed by a centralized organization that is responsible for the storage, distribution, version management, and API provision of models, which lowers the barrier to entry for model use to some extent. However, this model has several key flaws: First, the model's source, training data, training process, and version information are all maintained unilaterally by the platform, lacking transparency and verifiability. Model files are easily tampered with or replaced during storage and transmission, making it difficult for model users to verify their authenticity and integrity. Second, key behaviors such as model release, updates, call records, and revenue distribution lack a unified, immutable, and reliable evidence storage mechanism. In the event of model infringement, performance disputes, or security incidents, effective auditing and accountability are difficult to conduct. Third, existing platforms typically statically bind model inference tasks to specific computing power nodes, failing to dynamically coordinate scheduling based on the real-time load, geographical location, or resource type of each node in the computing power network. This results in uneven utilization of computing resources, with some nodes overloaded while others are idle, severely impacting overall execution efficiency. Finally, the lack of effective security protection and rights safeguards during model sharing makes models vulnerable to illegal copying, reverse engineering, or unauthorized abuse. The intellectual property rights and economic interests of model providers are difficult to protect effectively, and the lack of reasonable incentive mechanisms hinders the sustainable development of a high-quality model ecosystem.

[0004] Blockchain technology, with its core characteristics of decentralization, immutability, traceability, and verifiability, provides a new technological path for building a trusted data and resource sharing mechanism. By writing key metadata of intelligent models (such as hash fingerprints, training configurations, version identifiers, and provider identities) into the blockchain ledger, the reliability of model origins, verifiability of versions, and traceability of calls can be effectively guaranteed. Meanwhile, computing power networks, through unified abstraction, modeling, and scheduling of distributed heterogeneous computing power resources, lay the foundation for cross-domain collaborative execution of model inference tasks. However, current research on blockchain and computing power networks largely focuses on partial functions such as computing power resource trading, data notarization, or model markets, and has not yet formed a systematic solution for sharing intelligent models throughout their entire lifecycle. Especially in key areas such as trusted model registration, dynamic collaborative scheduling of computing power, automatic constraints on shared behavior, and rights protection, a deep-integrated collaborative mechanism is lacking, making it difficult to simultaneously ensure model trustworthiness, security, and computing power utilization efficiency. Summary of the Invention

[0005] This invention aims to solve the technical problems of low model credibility, rigid computing power scheduling, and lack of effective constraints in the sharing process in the prior art, and proposes a method and system for intelligent model trusted sharing based on blockchain and computing power network.

[0006] In a first aspect, embodiments of the present invention provide a trusted sharing method for intelligent models based on blockchain and computing power networks, including: Obtain the verifiable evolutionary fingerprint and inference performance benchmark of the target model to be shared, and store them separately on the main chain and side chain of the blockchain; In response to a model invocation request that has been verified by attributes, a target computing power node is dynamically selected to perform the inference task based on the capability profile vectors of multiple computing power nodes, with the goal of minimizing the overall cost; wherein, the overall cost includes the estimated energy consumption, estimated latency, and security reputation information of the target computing power node; The execution results of the inference task are verified, and the incentive states associated with the target model and the target computing node are updated based on the verification results and task execution performance.

[0007] The technical effects of the intelligent model trusted sharing method based on blockchain and computing power network disclosed in this invention are as follows: by introducing verifiable model evolution trajectory, heterogeneous computing power semantic matching mechanism and dynamic scheduling strategy based on carbon efficiency-security dual constraints, the intelligent model can achieve trusted circulation throughout its entire life cycle, privacy-protected invocation and green and efficient execution in cross-domain heterogeneous environment.

[0008] Furthermore, the verifiable evolutionary fingerprint is generated by concatenating and hashing the structural evolution features of the target model, the hash digest of its training dataset, the training round information, and the hardware credentials of the provider's device.

[0009] Furthermore, storing the two separately on the main chain and side chain of the blockchain specifically includes: writing the verifiable evolution fingerprint into the main chain, storing the inference performance benchmark into the side chain, and maintaining the consistency verification relationship between the two through a cross-chain state channel.

[0010] Furthermore, the capability profile vector includes: runtime compatibility score with known AI frameworks, carbon emission intensity per unit of computational task, and an identifier indicating whether the model supports cryptographic inference.

[0011] Furthermore, the attribute-verified model call request refers to a request that encapsulates an authorized set of user attributes, and that the set has passed the smart contract's verification of the preset access control policy. The verification process does not disclose the complete identity information of the requester.

[0012] Furthermore, the optimization objective function upon which the dynamic selection of target computing nodes is based is: ; Where Ω represents the set of candidate nodes that satisfy the preset security conditions and delay constraints, and x i Let E be the decision variable, representing whether to select node i. i and T i S represents the estimated energy consumption and estimated end-to-end delay when node i is selected. i Assign a security reputation score to node i, where α, β, and γ are configurable weighting coefficients, and E max and T deadline This represents the maximum energy consumption allowed by the system and the task deadline.

[0013] Furthermore, the security reputation score S of node i i It consists of the product of the identifier of whether node i supports cryptographic inference and the integrity metric of the trusted execution environment.

[0014] Furthermore, the verification of the execution results adopts a consistency verification protocol based on differential privacy: multiple verification nodes add independent noise to the same input and then infer separately, and the validity of the result is determined based on the consistency of the majority of output results.

[0015] Furthermore, the updated incentive state associated with the target model is calculated using a weighted iterative method based on the accuracy gain score and carbon efficiency compliance score of this inference task.

[0016] Secondly, embodiments of the present invention provide a trusted sharing system for intelligent models based on blockchain and computing power networks. The system is used to implement the trusted sharing method for intelligent models based on blockchain and computing power networks, including: The fingerprint evidence storage module is configured to obtain the verifiable evolutionary fingerprint and inference performance benchmark of the target model to be shared, and store the two on the main chain and side chain of the blockchain respectively. The inference task execution module is configured to respond to a model call request that has been verified by attributes, and dynamically select a target computing power node to execute the inference task based on the capability profile vectors of multiple computing power nodes, with the goal of minimizing the overall cost; wherein, the overall cost includes the estimated energy consumption, estimated latency and security reputation information of the target computing power node; The inference verification module is configured to verify the execution results of the inference task and update the incentive state associated with the target model and the target computing node based on the verification results and task execution performance.

[0017] The technical effects of the intelligent model trusted sharing system based on blockchain and computing power network disclosed in this invention are as follows: the modular design realizes functional division and collaboration; the standardized encapsulation and metadata management of the model management module reduce the sharing threshold of heterogeneous models; and the integration of the blockchain evidence storage module and the trusted verification module ensures the closed-loop trust of "evidence storage-verification", solving the problems of functional fragmentation and poor compatibility in traditional decentralized solutions. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the trusted sharing method for intelligent models based on blockchain and computing power networks provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the semantic capability profile vector construction and dynamic task scheduling mechanism for computing power nodes provided in this embodiment of the invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0020] refer to Figure 1 and Figure 2As shown, the trusted sharing method for intelligent models based on blockchain and computing power networks involves entities such as model providers, model users, main chains, side chains, computing power networks, and scheduling modules. The process begins with the generation of a verifiable evolutionary fingerprint by the model provider. This fingerprint, along with the model performance benchmark, is stored on the main chain and side chains respectively, forming a collaborative notarization between the main and side chains. Node information in the computing power network is abstracted into semantic computing power profiles. When a model user initiates a request, after verification via attribute-encrypted access control, the scheduling module assigns tasks to computing power nodes according to a carbon efficiency-security dual-constraint scheduling algorithm. After the task is completed, multi-dimensional verification and incentive feedback are performed, and the results are recorded on the blockchain, updating the incentive values ​​of relevant parties. Its core lies in constructing a complete trusted sharing closed loop through six collaborative steps, specifically including the following steps: Step S1: Model Registration and Verifiable Evolutionary Fingerprint Generation. Obtain the verifiable evolutionary fingerprint and inference performance benchmark of the target model to be shared, and store them separately on the main chain and side chain of the blockchain. This step aims to generate a unique and unforgeable "identity fingerprint" for each model version.

[0021] Specifically, the model provider first versions and packages the model (e.g., a Transformer medical image segmentation model). Next, it analyzes the model parameter set M, extracting structural evolution features such as "layer depth change rate" (the change rate increases from 6 to 8 layers by 2 / 6) and "parameter gradient sparsity" to construct the feature vector Ψ(M). Simultaneously, it calculates the training dataset D. train Hash digest of (e.g., 10,000 lung CT images) Then, combining the global round number e=150 at the end of training, and the hardware-level identity credential σ obtained from the provider's Trusted Platform Module (TPM), this information is concatenated and input into a cryptographic hash function (such as SHA3-256) to generate the final verifiable evolutionary fingerprint Fevol. This fingerprint deeply integrates the model's "structural genes," "data lineage," "training history," and "hardware origin," ensuring the credibility of the model's origin and the traceability of its evolutionary path.

[0022] The fingerprint generation method is as follows: ; Where Ψ(M) represents the structural evolution feature vector extracted from the model parameter set M (such as the rate of change of layer depth, parameter gradient sparsity), ΘD train For training dataset D train The tamper-resistant hash digest, e is the global round number when training is complete, σ is the identity credential issued by the Trusted Platform Module (TPM) of the provider device, and || denotes vector concatenation. This is a cryptographically secure hash function. The fingerprint Fevol uniquely identifies the evolutionary trajectory of this model version, possessing unforgeable and traceable characteristics.

[0023] This step resolves the issues of unclear model origins, inconsistent versions, and susceptibility to tampering, laying the foundation for subsequent reliable evidence preservation and traceability.

[0024] Step S2: Main-sidechain collaborative evidence storage and evolution anchoring: In response to the attribute-verified model call request, based on the capability profile vectors of multiple computing power nodes, and with the goal of minimizing the overall cost, a target computing power node is dynamically selected to execute the inference task; wherein, the overall cost includes the estimated energy consumption, estimated latency, and security reputation information of the target computing power node.

[0025] To balance efficiency and richness in data storage, this invention employs a main-sidechain architecture. The "evolutionary fingerprint" Fevol generated in step S1 is written to the main chain as the core credential for the model's lineage. Simultaneously, the model's performance benchmark (e.g., FLOPs=456, latency 28ms, accuracy mIoU=89.2%) is modeled using the Pareto front and stored in a performance sidechain. A lightweight cross-chain state channel establishes a bidirectional verification relationship between the main chain fingerprint and the sidechain performance summary. When a user invokes the model, they can trace the model's origin through the main chain and verify whether the current deployment meets the performance commitment through the sidechain. This involves tracing back to the original training context based on the Fevol on the main chain and verifying whether the current deployment meets the promised performance boundaries through the sidechain. This dual-chain mechanism ensures end-to-end trustworthiness of the model from training to inference.

[0026] Step S3: Semantic Capability Modeling of Computing Power Network. For accurate scheduling, a unified capability profile of heterogeneous computing power nodes (such as cloud GPU servers and edge computing boxes) is required. This step constructs a four-dimensional semantic capability profile vector for each node i. .

[0027] Among them, κ i (Compatibility Score): The score is determined by evaluating the model loading success rate and inference stability by running the probe program in target frameworks such as PyTorch / TensorFlow (0-1 points).

[0028] ρ i (Carbon emission intensity): Calculated based on the grid carbon emission factor of the node's location (e.g., 785gCO2 / kWh for the North China power grid) and the node's hardware energy efficiency ratio.

[0029] ϕ i (Memory efficiency ratio): The bandwidth efficiency ratio of data transfer between memory and video memory is obtained through benchmark tests such as STREAM.

[0030] ξ i(Encryption Support Identifier): This is determined by detecting whether the CPU supports Intel SGX or whether the GPU has NVIDIA Confidential Computing enabled (1 for support, 0 otherwise), i.e. whether it has the capability to support encrypted model inference (such as Intel SGX or NVIDIA Confidential Computing).

[0031] For example, a node located in a hydropower-rich area of ​​Yunnan, equipped with an A100 GPU, might have a profile of [0.95, 120, 0.88, 1].

[0032] By abstracting heterogeneous computing resources into standardized semantic vectors, precise data input is provided for subsequent intelligent matching and optimized scheduling.

[0033] Step S4: Fine-grained Access Control Based on Attribute Encryption. To protect user privacy and achieve fine-grained authorization, this step employs Attribute-Based Encryption (CP-ABE) based on ciphertext policy. Model users do not need to expose their real identities; they only need to construct an attribute set (e.g., {Task Type: Medical Image Segmentation, Maximum GPU Memory: 16GB, Institution Type: Tertiary Hospital}) and submit an encryption request. The model provider or platform administrator pre-sets the access policy in the smart contract (e.g., "Medical Image Segmentation AND (GPU Memory < 16GB OR Encryption Support = TRUE)"). The smart contract directly verifies on-chain whether the attributes in the encryption request meet the policy logic, without decrypting the user's identity throughout the process. After successful verification, a one-time encrypted access credential is generated and stored on the blockchain. This credential, along with the policy hash Hpolicy=Hash(policy expression), is written to the main chain, achieving minimal permissions and auditable behavior.

[0034] It implements a privacy protection paradigm of "controlling access without knowing the identity", ensures the principle of least privilege for model calls, and all access attempts are auditable.

[0035] Step S5: Dynamic task scheduling with carbon efficiency and security dual constraints. This step is the core of resource scheduling. After receiving a verified access request, the scheduling module (6) integrates the profile vectors C of all available computing power nodes. i Based on current task requirements (such as maximum allowable delay d) max =50 ms, maximum carbon emissions =5gCO2), solve the following optimization problem. The scheduling module optimizes with the objective of "carbon efficiency-safety dual constraints". The scheduling objective function is to minimize the overall cost: ; Where Ω represents the set of candidate nodes that satisfy the preset security conditions and delay constraints, and x iLet E be the decision variable, representing whether to select node i. i and T i These represent the estimated energy consumption and estimated end-to-end latency when selecting node i, including network transmission, queuing, and computation delays, S. i Assign a security reputation score to node i, where α, β, and γ are configurable weighting coefficients (e.g., 0.4, 0.3, 0.3), and E max and T deadline This represents the maximum energy consumption allowed by the system and the task deadline.

[0036] S i =ξi×IntegrityScore, where IntegrityScore comes from the remote proof protocol. The scheduler will meet the task deadline T. deadline and security requirements (such as policy requirements) i Given that (=1), solve this optimization problem and assign tasks to the node with the lowest overall cost. The scheduler uses a heuristic algorithm (such as the improved Hungarian algorithm) to determine the optimal task allocation method based on the given condition. (If the strategy requires safe reasoning) select the node that minimizes the objective function. For example, although a certain high-carbon node has lower latency, due to ρ i The score was too high and there was no TEE support. Its overall score was worse than that of Yunnan's low-carbon safety node, so the task was assigned to the latter.

[0037] Based on traditional considerations of latency and cost, this approach innovatively introduces carbon emissions and secure, reliable execution as core scheduling factors, achieving a three-in-one intelligent scheduling system that is green, efficient, and safe.

[0038] Step S6: Multi-dimensional verification of reasoning results and dual-track incentive feedback. After the task is completed, the credibility of the results needs to be verified and participants should be fairly incentivized.

[0039] Result Verification: A lightweight verification protocol based on differential privacy is employed. Multiple verification nodes add Laplacian noise (Lap(0,Δf / ε)) to the same input and then independently infer. If the majority of outputs are consistent within the IoU threshold (e.g., >0.85), the result is deemed valid. Verification evidence is stored on the blockchain. The model output (e.g., segmentation mask) is first verified via the lightweight verification protocol: multiple randomly selected verification nodes apply a small differential privacy perturbation (e.g., adding Laplacian noise Lap(0,Δf / ε), where Δf is the sensitivity and ε=1.0) to the same input. If the majority of outputs are consistent within the IoU threshold (e.g., >0.85), the result is deemed valid. Verification evidence (including perturbation parameters, consistency score, and participating node signatures) and the result digest (e.g., hash value) are jointly stored on the main chain. Subsequently, the system updates the dual-track incentive value. For the model provider, its incentive value iterates according to the following rules.

[0040] Dual-track incentive: (1) Model provider incentives: The incentive values ​​are dynamically updated based on the actual performance of the model in this study. ; Among them, the accuracy gain score A t =max(0,mIoUt−baseline) / (1−baseline), where the baseline is set to 85%, i.e. Carbon efficiency compliance score C t When the model supports quantization for energy saving, i.e., if the model supports quantization compression or sparsification to reduce... , then Otherwise, it is 0; the attenuation factor η = 0.1, weight For example, if mIoU = 89.2% this time, then: A t =(89.2−85) / (100−85)=0.28; If the model is quantized using INT8, The final incentive increment is 0.1 × (0.7 × 0.28 + 0.3 × 0.36) = 0.0304. The incentive for computing nodes is calculated independently based on their actual carbon emissions, latency compliance rate, and security compliance records, and is used for subsequent scheduling priority ranking. This encourages providers to continuously optimize model performance and energy efficiency.

[0041] (2) Computing node incentives: Independent calculation and updates are performed based on the carbon emissions, latency compliance rate and security compliance records of the actual task execution, which will affect the future scheduling priority.

[0042] The credibility of the output results is ensured through multi-party verification; a dual-track incentive mechanism that links model performance, green indicators and node execution performance drives all participants to provide high-quality, efficient and low-carbon services from an economic perspective, forming a positive ecological cycle.

[0043] This method achieves the separation and trusted storage of model "identity" and "performance" data, which not only ensures the high-security level of evidence storage of key lineage information, but also meets the need for high-frequency updates of performance data, and ensures data consistency through cross-chain verification.

[0044] The following is a specific example: A tertiary hospital (the user) needs to call an AI model trained by an external research institution (the provider) for lung CT segmentation.

[0045] The provider encapsulates model v2.0, extracts the layer depth change rate from v1.0 (6-layer encoder) to v2.0 (8-layer encoder) = 2 / 6, calculates the dataset hash, and generates the fingerprint Fevol by combining the training round e = 150 and the TPM credential σ. Fevol is written to the main chain. The model's performance benchmark on the test set (mIoU = 89.2%, latency 28ms) is stored in the sidechain. The two chains are anchored via state channels. Two available nodes are discovered: Node A (cloud data center, profile [0.95, 650, 0.9, 1]) and Node B (edge ​​node, profile [0.7, 120, 0.5, 0]). The hospital constructs an attribute set {domain: medical, task: segmentation, memory ≤ 16GB} and initiates an encryption request. The smart contract verifies that it meets the policy "medical AND segmentation", and authorization is granted. The scheduling module receives the task (T deadline =50ms, encrypted inference required ξ i =1). Calculate the objective function values ​​for node A (high carbon but safe) and node B (low carbon but unsafe). Due to the policy requirement ξ... i =1, node B is excluded. Node A is selected under the time delay and security constraints. Node A performs inference. The three verification nodes add noise to the input CT image for verification; the output IoU is greater than 0.86 for all three nodes, indicating the result is valid. After evidence storage, the provider incentive is calculated: A t =(89.2−85) / (100−85)≈0.28. Since the model did not undergo special energy-saving optimization, C t =0. Let η=0.1, ω1=1, then the excitation value increases by 0.1*0.28=0.028. Node A updates its node excitation value based on its actual energy consumption and latency.

[0046] Based on the same inventive concept, embodiments of the present invention provide a trusted sharing system for intelligent models based on blockchain and computing power networks. The system is used to implement the trusted sharing method for intelligent models based on blockchain and computing power networks, including: The fingerprint evidence storage module is configured to obtain the verifiable evolutionary fingerprint and inference performance benchmark of the target model to be shared, and store the two on the main chain and side chain of the blockchain respectively. The inference task execution module is configured to respond to a model call request that has been verified by attributes, and dynamically select a target computing power node to execute the inference task based on the capability profile vectors of multiple computing power nodes, with the goal of minimizing the overall cost; wherein, the overall cost includes the estimated energy consumption, estimated latency and security reputation information of the target computing power node; The inference verification module is configured to verify the execution results of the inference task and update the incentive state associated with the target model and the target computing node based on the verification results and task execution performance.

[0047] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A trusted sharing method for intelligent models based on blockchain and computing power networks, characterized in that, include: Obtain the verifiable evolutionary fingerprint and inference performance benchmark of the target model to be shared, and store them separately on the main chain and side chain of the blockchain; In response to a model invocation request that has been verified by attributes, a target computing power node is dynamically selected to perform the inference task based on the capability profile vectors of multiple computing power nodes, with the goal of minimizing the overall cost; wherein, the overall cost includes the estimated energy consumption, estimated latency, and security reputation information of the target computing power node; The execution results of the inference task are verified, and the incentive states associated with the target model and the target computing node are updated based on the verification results and task execution performance.

2. The method as described in claim 1, characterized in that, The verifiable evolutionary fingerprint is generated by concatenating and hashing the structural evolution features of the target model, the hash digest of its training dataset, the training round information, and the hardware credentials of the provider's device.

3. The method as described in claim 1, characterized in that, The method of storing the two separately on the main chain and side chain of the blockchain specifically includes: writing the verifiable evolution fingerprint into the main chain, storing the inference performance benchmark into the side chain, and maintaining the consistency verification relationship between the two through the cross-chain state channel.

4. The method as described in claim 1, characterized in that, The capability profile vector includes: runtime compatibility score with known AI frameworks, carbon emission intensity per unit of computational task, and an identifier indicating whether the model supports cryptographic inference.

5. The method as described in claim 1, characterized in that, The attribute-verified model call request refers to a request that encapsulates an authorized set of user attributes, and that the set has passed the smart contract's verification of the preset access control policy. The verification process does not disclose the complete identity information of the requester.

6. The method as described in claim 1, characterized in that, The optimization objective function upon which the dynamic selection of target computing power nodes is based is: ； Where Ω represents the set of candidate nodes that satisfy the preset security conditions and delay constraints, and x i Let E be the decision variable, representing whether to select node i. i and T i S represents the estimated energy consumption and estimated end-to-end delay when node i is selected. i Assign a security reputation score to node i, where α, β, and γ are configurable weighting coefficients, and E max and T deadline This represents the maximum energy consumption allowed by the system and the task deadline.

7. The method as described in claim 6, characterized in that, The security and reputation score S of node i i It consists of the product of the identifier of whether node i supports cryptographic inference and the integrity metric of the trusted execution environment.

8. The method as described in claim 1, characterized in that, The verification of the execution results adopts a consistency verification protocol based on differential privacy: multiple verification nodes add independent noise to the same input and then infer separately, and the validity of the result is determined based on the consistency of the majority of output results.

9. The method as described in claim 1, characterized in that, The updated incentive state associated with the target model is calculated using a weighted iterative method based on the accuracy gain score and carbon efficiency compliance score of this inference task.

10. A trusted sharing system for intelligent models based on blockchain and computing power networks, characterized in that: The system is used to implement the trusted sharing method for intelligent models based on blockchain and computing power networks as described in any one of claims 1 to 9, including: The fingerprint evidence storage module is configured to acquire the verifiable evolutionary fingerprint and inference performance benchmark of the target model to be shared, and store the two on the main chain and side chain of the blockchain respectively. The inference task execution module is configured to respond to a model call request that has been verified by attributes, and dynamically select a target computing power node to execute the inference task based on the capability profile vectors of multiple computing power nodes, with the goal of minimizing the overall cost; wherein, the overall cost includes the estimated energy consumption, estimated latency and security reputation information of the target computing power node; The inference verification module is configured to verify the execution results of the inference task and update the incentive state associated with the target model and the target computing node based on the verification results and task execution performance.

Images