Illegal Accounts Detection on Ethereum Using Heterogeneous Graph Transformer Networks

Abstract

Numerous applications based on Ethereum have been utilized in a variety of scenarios, such as financial services. However, due to the lack of effective regulation in the blockchain, a significant number of illegal users cash in on the anonymity of blockchain accounts, which has an extremely negative impact. Existing illegal account detection methods employ machine learning techniques to train fundamental account characteristics and fail to extract efficient high-order features by graph structures, leading to inaccuracies in account detection. To address this issue, we propose a novel illegal account identification method based on a heterogeneous transformer network. Specifically, we design an account-centric heterogeneous information network model to express real transaction data on Ethereum for the first time. This model can describe the network structure information more comprehensively. Additionally, we propose to apply the graph transformer network to automatically learn the multi-hop metapath and obtain high-order node information and links. These features, in turn, improve the quality and performance of our model. Finally, we employ the graph convolutional network to classify nodes and complete the account identification task and ensure the security of the Ethereum system. Furthermore, we compare our method with other existing detection models. Our experiments demonstrate that the proposed approach achieves an accuracy of 95.57%, which surpasses that of traditional machine learning models and existing detection schemes.