CN-121996419-A - Intelligent model trusted sharing method and system based on blockchain and computational power network

CN121996419ACN 121996419 ACN121996419 ACN 121996419ACN-121996419-A

Abstract

The invention discloses an intelligent model trusted sharing method and system based on a blockchain and an algorithm network, and belongs to the technical field of blockchain, algorithm network and artificial intelligence intersection. The method comprises the steps of obtaining verifiable evolution fingerprints and performance references of an intelligent model to be shared and respectively storing the verifiable evolution fingerprints in a main chain and a side chain of a blockchain, constructing semantic capability portrait vectors of computing nodes, receiving model calling requests verified by an attribute-based encryption mechanism, dynamically selecting target computing nodes from nodes meeting preset safety and aging constraints based on the capability portrait vectors with minimum comprehensive cost as a target, executing model reasoning tasks, wherein the comprehensive cost is related to estimated energy consumption, time delay and safety reputation of the nodes, performing multiparty verification on task execution results, and updating excitation states of model contributors and computing nodes according to verification results and node performances. The invention realizes the credible, efficient and green sharing of the intelligent model in the distributed environment.

Inventors

ZHOU WEN
YANG JIA
TAN MIN
CHEN YANSHU
YE ZHIWEI
He Qidai
CAI TING
WANG MINGWEI
ZHANG JIXIN
XIANG CHUNLI
LEI MENGYA

Assignees

湖北工业大学

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (10)

1. The intelligent model trusted sharing method based on the blockchain and the computing power network is characterized by comprising the following steps of: Obtaining verifiable evolution fingerprints and reasoning performance references of a target model to be shared, and respectively verifying the verifiable evolution fingerprints and reasoning performance references in a main chain and a side chain of a blockchain; Responding to a model call request subjected to attribute verification, and dynamically selecting a target computing node to execute an reasoning task based on capability image vectors of a plurality of computing nodes with the aim of minimizing comprehensive cost, wherein the comprehensive cost comprises estimated energy consumption, estimated time delay and safety reputation information of the target computing node; and verifying the execution result of the reasoning task, and updating the excitation states associated with the target model and the target computing power node according to the verification result and the task execution performance.
2. The method of claim 1, wherein the verifiable evolution fingerprint is generated by stitching and hashing structural evolution features of the target model, a hash digest of its training dataset, training round information, and provider device hardware credentials.
3. The method of claim 1, wherein the verifying the two separately in the backbone and the side chain of the blockchain specifically comprises writing a verifiable evolution fingerprint to the backbone, storing an inference performance benchmark in the side chain, and maintaining a consistency verification relationship between the two through a cross-chain state channel.
4. The method of claim 1, wherein the capability image vector includes an identification of a runtime compatibility score for a known AI framework, a carbon emission intensity for a unit calculation task, and whether model encryption reasoning is supported.
5. The method of claim 1, wherein the attribute-verified model calls a request that an authorized user attribute set is encapsulated in the request and passes verification of a preset access control policy by an intelligent contract, and the verification process does not reveal complete identity information of the requesting party.
6. The method of claim 1, wherein the optimization objective function upon which the dynamically selected target computing node is based is: ; Wherein Ω is a candidate node set satisfying preset security conditions and time delay constraints, x i is a decision variable, indicating whether to select node i, E i and T i are respectively estimated energy consumption and estimated end-to-end time delay when node i is selected, S i is a security reputation score of node i, α, β, γ are configurable weight coefficients, and E max and T deadline are maximum energy consumption values and task deadlines allowed by the system.
7. The method of claim 6, wherein the security reputation score S i for node i consists of a product of an identification of whether node i supports cryptographic reasoning and an integrity metric value for a trusted execution environment.
8. The method of claim 1, wherein the verifying the execution result uses a consistency check protocol based on differential privacy, wherein a plurality of verification nodes respectively infer after adding independent noise to the same input, and determine the validity of the result according to the consistency of a plurality of output results.
9. The method of claim 1, wherein the updating the excitation state associated with the target model is performed by a weighted iterative calculation based on the accuracy gain score and the carbon efficiency compliance score of the current inference task.
10. A system for trusted sharing of intelligent models based on blockchain and a computing power network, wherein the system is configured to implement the method for trusted sharing of intelligent models based on blockchain and a computing power network as claimed in any one of claims 1 to 9, comprising: The fingerprint certification module is configured to acquire verifiable evolution fingerprints and reasoning performance references of the target model to be shared, and certify the verifiable evolution fingerprints and reasoning performance references on a main chain and a side chain of the blockchain respectively; The reasoning task execution module is configured to respond to a model call request verified by attributes, dynamically select a target computing node to execute a reasoning task based on capability image vectors of a plurality of computing nodes with the aim of minimizing comprehensive cost, wherein the comprehensive cost comprises estimated energy consumption, estimated time delay and safety reputation information of the target computing node; And the reasoning verification module is configured to verify the execution result of the reasoning task and update the excitation state associated with the target model and the target computing node according to the verification result and the task execution performance.

Description

Intelligent model trusted sharing method and system based on blockchain and computational power network Technical Field The invention relates to the technical field of intersection of a blockchain, a computational power network and artificial intelligence, in particular to an intelligent model trusted sharing method and system based on the blockchain and the computational power network. Background With the rapid development of artificial intelligence technology, intelligent models such as deep learning models and machine learning models are widely applied in the fields of intelligent perception, intelligent decision making, data analysis, automatic control and the like. The application form of the intelligent model gradually evolves from local deployment in a single scene to a cross-industry and cross-region shared service mode, and multiplexing and collaborative calling of the model become key means for improving the overall efficiency of the artificial intelligent system. However, the training and reasoning process of smart models is highly dependent on high performance computing resources, and as model scale continues to increase, the need for computing power continues to rise. In an actual deployment environment, computing power resources show cross-cloud, cross-edge and cross-region decentralized characteristics, and a heterogeneous computing power network which is formed by cloud computing nodes, edge computing nodes and special acceleration equipment (such as GPU (graphics processing Unit), TPU (thermoplastic polyurethane) and the like) is formed. In this context, how to implement efficient deployment, flexible scheduling, and secure invocation of intelligent models in complex computational power networks has become a core challenge faced by current artificial intelligence systems. Currently, the mainstream intelligent model sharing mode mainly relies on a centralized model management platform or a third party model service market. Such platforms are typically provided by a centralized authority that is responsible for the storage, distribution, version management, and invocation interfaces of the models in a unified manner, reducing the thresholds used by the models to some extent. However, the model has a plurality of key defects that firstly, the source, training data, training process and version information of the model are maintained by a single side of a platform, transparency and verifiability are lacked, model files are easy to tamper with or replace in the storage and transmission process, so that the authenticity and integrity of the model are difficult to confirm by a model user, secondly, key behaviors such as issuing, updating, calling records and profit distribution of the model lack a unified and untouchable trusted evidence storage mechanism, once model infringement, performance disputes or security events occur, effective audit and responsibility tracing are difficult to carry out, and thirdly, the existing platform generally binds model reasoning tasks to specific computational nodes in a static mode, dynamic collaborative scheduling cannot be carried out according to real-time loads, geographic positions or resource types of all nodes in a computational network, so that computational resources are unbalanced, partial nodes are overloaded, the whole execution efficiency is seriously affected, finally, the model lacks effective security protection and equitable property protection mechanisms in the sharing process, the model is easy to suffer illegal copying, reverse engineering or unauthorized abuses, knowledge and economic benefit is difficult to be effectively maintained, and the sustainable mechanism of the model is reasonably and has high quality and sustainable development of the model is difficult to be guaranteed. The block chain technology provides a new technical path for constructing a trusted data and resource sharing mechanism by virtue of the core characteristics of decentralization, non-falsification, traceability, verifiability and the like. By writing key metadata (such as hash fingerprints, training configuration, version identification, provider identity and the like) of the intelligent model into the blockchain ledger, the source credibility, verifiable version and traceability of calling of the model can be effectively ensured. Meanwhile, the computing power network establishes a foundation for cross-domain collaborative execution of model reasoning tasks by carrying out unified abstraction, modeling and scheduling on distributed heterogeneous computing power resources. However, current research on blockchains and power networks focuses on local functions of power resource transactions, data storage certificates, or model markets, and no systematic solution for intelligent model full lifecycle sharing has been formed. Especially in key links such as the trusted registration of the model, dynamic collaborative scheduling of computing power, automatic constraint o