Exploring GCN, GAT, and GIN Fusion for Illicit Transaction Classification in Cryptocurrency Networks

Abstract

The rise of blockchain and cryptocurrency networks has fueled financial innovation, yet it also presents new opportunities for illicit activities such as fraud and money laundering. Traditional detection approaches struggle with the non-Euclidean, large-scale nature of cryptocurrency transaction networks. Graph Neural Networks offer a promising solution for capturing complex relational data. This paper proposes four fusion architectures—Triple Parallel Layer, Hierarchical Staging, Attention-Weighted Residual Fusion, and Multi-View Feature Aggregation—combining Graph Convolutional Network, Graph Attention Network, and Graph Isomorphism Network to enhance classification of illicit transactions. Experiments on the Elliptic Bitcoin dataset show that the proposed models achieve classification accuracies up to 97.17%, significantly outperforming standalone Graph Convolutional Network, Graph Attention Network, and Graph Isomorphism Network models. These results underscore the superior performance and robustness of the fusion architectures, with improvements in accuracy ranging from 1.1% to 2.9% over individual models, marking a step forward in financial crime detection within decentralized networks.