KR-20260065598-A - Method and system for tokenizing artificial intelligence model weights linked to real-world assets

Abstract

The present invention relates to a method and system for converting weight data of a trained artificial intelligence model into real-world asset-linked tokens (RWA tokens). According to the present invention, an encrypted hash value of the weight data is generated, the encrypted weights are stored in an off-chain secure storage, weight metadata and RWA tokens are recorded on a blockchain, and access to the weights is allowed only within a Trusted Execution Environment (TEE) based on proof of token ownership. In particular, to accommodate the characteristics of dynamic weights that are continuously updated by fine-tuning, the invention provides snapshot version tokenization, differential data separation tokenization, governance approval-based updates, and a living asset token structure. Furthermore, inference execution rights, fine-tuning rights, and governance voting rights are issued separately by type, and automatic profit distribution and automatic payment of fine-tuning royalties are realized through smart contracts.

Inventors

• 안범주

Assignees

• 안범주

Dates

Publication Date

20260508

Application Date

20260421

Claims (1)

1. As a method for converting artificial intelligence model weights into real asset-linked tokens (RWA tokens), One or more processors, A step of generating an encrypted hash value for the weight data of a trained artificial intelligence model; A step of encrypting the weight data and storing it in a distributed storage or off-chain secure storage, and generating weight metadata including the encrypted hash value and the storage location identifier; The step of recording the above weight metadata on a blockchain network and issuing one or more RWA tokens representing ownership and usage rights for the above artificial intelligence model weights; and A step of allowing access to the corresponding weight data when the token holder proves ownership of the RWA token. An artificial intelligence model weight tokenization method including executing

Description

Method and system for tokenizing artificial intelligence model weights linked to real-world assets The present invention relates to a method and system for converting weight data of a trained artificial intelligence model into a blockchain-based Real World Asset Token (hereinafter RWA Token) to represent ownership, usage rights, profit claims, and governance voting rights as digital assets and enable trading. In particular, the invention relates to a method and system for real-world asset-linked tokenization of artificial intelligence model weights, comprising a version control mechanism that accommodates the characteristics of dynamic weights that are continuously updated through fine-tuning and reinforcement learning, a secure access control structure combining a Trusted Execution Environment (TEE) and a Zero-Knowledge Proof (ZKP), and a smart contract-based automatic profit distribution system. With the rapid advancement of artificial intelligence technology, high-performance AI models are being developed across various domains, including large-scale language models, image generation models, and speech recognition models. The core asset of these AI models is weight data, consisting of hundreds of millions to trillions of parameters. This weight data possesses substantial economic value as an optimized outcome achieved through long-term learning processes involving massive computational resources and vast training data. In particular, the weights of expert AI models specialized for specific domains often outperform general-purpose models in inference performance within those domains, serving as core intellectual property for companies and research institutions. However, despite their economic value, the ownership structure of conventional AI model weights as assets was unclear, and there was no system to transparently record and enforce the rights of each contributor regarding models developed through joint investment by multiple investors. Furthermore, the lack of technical means to automatically distribute revenue generated from inference services provided by AI models to multiple stakeholders according to their contribution ratios led to frequent contract disputes and settlement delays. Real-world asset-linked tokenization technology utilizing blockchain technology provides a method to represent and trade ownership and usage rights for traditional physical assets, such as real estate, precious metals, and bonds, as tokens on the blockchain. However, conventional real-world asset-linked tokenization technology has been designed for assets with fixed physical entities, such as real estate or precious metals, or digital content with immutable content. Since the content of these static assets does not change once they are tokenized, the encryption hash-based binding between the token and the asset maintains its validity. In contrast, the weights of artificial intelligence models are continuously updated through techniques such as fine-tuning, Reinforcement Learning from Human Feedback (RLHF), and continual learning. Even a single fine-tuning process can change billions of parameter values in the weight data, leading to a discrepancy between the cryptographic hash values originally recorded on the blockchain and the hash values of the updated weights. Consequently, technical problems arise where it becomes ambiguous which version of the weights an already issued token represents, and the attribution of token holders' access rights and profit claims becomes unclear. These characteristics unique to dynamic assets represent a fundamental technical challenge that cannot be resolved by conventional real-world asset-linked tokenization technologies. Furthermore, due to the massive amount of data (several gigabytes to several terabytes) of weights for artificial intelligence models, it is practically impossible to store them directly on a blockchain network. A security access control mechanism is required to store weight data off-chain while granting access rights to the weight data only to legitimate token holders and preventing unauthorized viewing of the weight contents. In particular, since there is a risk that weight information may be exposed externally if the weight data itself is decrypted in an off-chain storage, there is a technical necessity for decryption and inference operations to be performed only within a trusted execution environment that is inaccessible from the outside. Furthermore, in the case of artificial intelligence models with a Mixture of Experts (MoE) structure, since the entire model consists of a set of multiple expert module weights rather than a single set of weights, it is possible to implement independent ownership structures and revenue distribution systems for each expert module, and this is also economically advantageous. However, there were no technical means to realize this. FIG. 1 is a block diagram illustrating the overall architecture of an artificial intelligence model wei