Abstract
As one key component of blockchain technology, a consensus protocol is used to ensure that all the participants agree on the ordering and validity of transactions in a decentralized network. Diverse types of consensus protocols have been put forward, as revealed by a state-of-the-art literature review conducted in this thesis research. However, the existing protocols often face scalability and throughput limitations. In this thesis, based on the Directed Acyclic Graph (DAG)architecture, we propose an improved consensus protocol using semi-supervised learning with Graph Convolutional Networks (GCNs), which can learn from both labeled and unlabeled data and process graph-structured data efficiently. Leveraging the power of GCNs, the proposed approach aims to provide high transaction throughput, fast block validation, and detection of malicious blocks on blockchain networks. The performance of the proposed approach is compared to that of an existing protocol named UL-BlockDAG, which applies unsupervised learning (UL) for block classification and uses spectral graph theory to analyze block connectivity patterns. Empirical studies using blockchain networks of different sizes show that the proposed GCN-based approach outperforms UL-BlockDAG significantly in terms of scalability, throughput, and security. In addition, the effects of block size, transaction count, network bandwidth and graph size are investigated.