Translets: Toward Explainable Phishing Fraud Detection in Ethereum

Xikang Yang, Biyu Zhou, Xuehai Tang, Xiaodan Zhang, Jizhong Han, Songlin Hu

Abstract: In recent years, the cryptocurrency platform becomes a prime target of various cybercrimes. Criminals use phishing fraud to commit massive scams on Ethereum (one of the most widely used cryptocurrency platforms), which poses a significant threat to the security of the cryptocurrency ecosystem. In this context, the use of Ethereum transaction information to detect and identify phishing fraud accounts is essential to ensure a secure and regulated trading platform. However, the previous proposals do not explore the behavior patterns of phishing accounts in depth, and also lack interpretability. To address this problem, we propose a novel and interpretable Ethereum phishing fraud detection method by extracting more fine-grained and interpretable account transaction features. The key idea is to extract both the spatial structure and temporal behavior patterns of the Ethereum transaction network as the “Translets” features via the highly interpretable attribute sub graphs and subsequences. Based on these features, a classifier with good interpretability is adopted, and the proposed interpreter is combined to interpret the detection outputs. The experimental results on real-world Ethereum phishing fraud account datasets demonstrate that our method not only has advantages in precision, recall, and F1 score but also provides interpretability in recognizing Ethereum phishing accounts.