Across-Platform Detection of Malicious Cryptocurrency Transactions via Account Interaction Learning

TL;DR

This paper introduces ShadowEyes, a graph contrastive learning approach for detecting malicious cryptocurrency transactions across platforms, effectively leveraging unlabeled data and transaction evolution simulation to improve generalization.

Contribution

The paper proposes a novel graph-based framework with data augmentation and contrastive learning to enhance cross-platform malicious transaction detection with limited labeled data.

Findings

Outperforms SOTA in zero-shot gambling transaction detection with 76.98% F1 score.

Achieves around 90% F1 score in across-platform detection, 10% higher than existing methods.

Effective in real-world scenarios with limited labeled data and evolving malicious behaviors.

Abstract

With the rapid evolution of Web3.0, cryptocurrency has become a cornerstone of decentralized finance. While these digital assets enable efficient and borderless financial transactions, their pseudonymous nature has also attracted malicious activities such as money laundering, fraud, and other financial crimes. Effective detection of malicious transactions is crucial to maintaining the security and integrity of the Web 3.0 ecosystem. Existing malicious transaction detection methods rely on large amounts of labeled data and suffer from low generalization. Label-efficient and generalizable malicious transaction detection remains a challenging task. In this paper, we propose ShadowEyes, a novel malicious transaction detection method. Specifically, we first propose a generalized graph structure named TxGraph as a representation of malicious transaction, which captures the interaction features of each malicious account and its neighbors. Then we carefully design a data augmentation method tailored to simulate the evolution of malicious transactions to generate positive pairs. To alleviate account label scarcity, we further design a graph contrastive mechanism, which enables ShadowEyes to learn discriminative features effectively from unlabeled data, thereby enhancing its detection capabilities in real-world scenarios. We conduct extensive experiments using public datasets to evaluate the performance of ShadowEyes. The results demonstrate that it outperforms state-of-the-art (SOTA) methods in four typical scenarios. Specifically, in the zero-shot learning scenario, it can achieve an F1 score of 76.98% for identifying gambling transactions, surpassing the SOTA method by12.05%. In the scenario of across-platform malicious transaction detection, ShadowEyes maintains an F1 score of around 90%, which is 10% higher than the SOTA method.