Ensuring Trustworthy Neural Network Training via Blockchain
Document Type
Conference Proceeding
Publication Date
2023
Abstract
As Artificial Intelligence prevalence grows, it highlights the risk in relying on compromised models, thereby fueling a growing need to ensure the integrity of trained AI models. In this paper, we present a novel blockchain-based system, designed to authenticate the integrity of trained neural network models. The system addresses the risk of manipulation of a model by strategically re-computing intervals of the training process. Further, the blockchain network provides a traceable, immutable, trusted ledger for cataloging the intricate processes of training and validation. We consider two primary entities involved: ‘submitters’, who submit trained models, and ‘verifiers’, who re-train distinct sections of the submitted models to validate their integrity. The design of the blockchain system emphasizes efficiency by selectively targeting a portion of all training intervals. This is made possible through the use of an innovative weight-analysis algorithm, which applies an Absolute Change approach to identify outliers. We implement our solution to demonstrate that the proposed blockchain system is robust, and the weight-analysis algorithm is accurate and scalable.
Publication Information
Navarro, Edgar; Standing, Kyle J.; Dagher, Gaby G.; and Andersen, Tim. (2023). "Ensuring Trustworthy Neural Network Training via Blockchain". In 2023 IEEE 5th International Conference on Cognitive Machine Intelligence (CogMI) (31-40). IEEE. https://doi.org/10.1109/CogMI58952.2023.00015