Abstract
Machine learning (ML) models have rapidly become foundational elements of the modern world, with numerous organizations integrating ML systems into both major and minor tasks. However, users are often required to place blind trust in these ML models hosted by organizations, as there is no mechanism for independent verification of model correctness or detection of hidden biases. To address this challenge, we propose the integration of blockchain technology, known for its inherent security and trustworthiness, with the storage of ML-based predictive models. Blockchain, as a decentralized ledger comprised of blocks, ensures that peers in the network maintain an identical copy of the ledger, with new blocks added through a consensus mechanism. In this thesis, we utilize neural networks as predictive models and employ smart contracts to store the models in a public blockchain network. Smart contracts are self-executing protocols that can be encoded to store various parameters of a predictive model, including a neural network’s weights and other relevant factors required for model recreation. We design algorithms to efficiently upload the parameters of a predictive model stored in a smart contract to the blockchain and retrieve them by regular users. By storing the models on a public blockchain network, all users in the network can test, validate and review them. To enhance efficiency, we further introduce a meta-block at the beginning of the blockchain containing indexing information about the predictive models and their review comments. By organizing and structuring this information, we streamline the retrieval process for predictive models. Finally, through case studies conducted using an Ethereum blockchain simulation, this proposed approach demonstrates its efficacy in establishing security, immutability, and transparency for ML-based predictive models.