Indy528 — Federated Learning Model Tokenization with Non-Fungible Tokens(NFT) and Model Cards

Abstract

Non-Fungible Tokens (NFTs) are used to represent ownership of unique items. The tokens provide means to validate integrity. That is, the item is unique and cannot be replaced with something else. NFTs are securely stored in the blockchain-based decentralized storage in the format of ERC721 tokens, so it is technically difficult to modify the record of ownership or duplicate a new NFT token. Currently, NFTs are used to tokenise objects such as arts, collectables, and even real estate. In this research, we propose an open platform “Indy528” and utilize NFT tokens to represent federated machine learning (FML) models. To the best of our knowledge, this is the very first research that tries to represent machine learning models as NFTs. Different peers in the blockchain network can build FML models using a blockchain-enabled coordinator-less FML system. The data provenance information of the FML models, which records the model data changes from the origin, is recorded as Model Cards. The FML model ownership, storage locations of the FML models and data provenance information(with Model Cards) are encoded into NFT tokens and stored in the decentralized ledger. We design a novel NFT token schema i528 (which is an extension of ERC721) to represent machine learning models as NFTs. As a use-case of Indy528, we propose a concept of the NFT-based decentralized FML model marketplace. Model creators can train the FML models and publish them in the marketplace as NFTs. Parties who want to use the models can buy them from the model creators as NFTs. The buyers can view model information, accuracy, training process etc via the model card objects stored with the NFT. This design methodology provides enhanced transparency and auditability to the federated learning process while providing an open platform to share and trade machine learning models in the marketplace.