TSFF: A Triple-Stream Feature Fusion Method for Ethereum Phishing Scam Detection

As a representative of the public blockchain, Ethereum has been applied in various industries. However, the vast number of transactions on the platform has also brought a number of illegal activities, such as phishing scams, which have caused significant damage to the Ethereum ecosystem. Due to anonymity of the blockchain, it is difficult for detectors to extract features that can be directly applied to phishing scams detection. Existing studies mainly model Ethereum transaction records as a network and mine key information from them to identify phishing addresses. However, these methods usually employ traditional feature engineering or network embedding, ignoring the fine-grained features in the transaction network. In addition, since the original network is too large to make learning difficult, existing work usually uses random walk (RW) to sample a part of nodes for training, thus ignoring the multiplicity of the network. To address these issues, in this article, we propose a three-stream feature fusion (TSFF) approach to enhance the feature representation of nodes. Specifically, we construct node states to guide RW sampling, and manually extracted 8-D features from the resulting dataset as basic features. Temporal features are jointly learned through long short-term memory network and contrastive learning. We combine residual blocks and graph convolutional network to extract fine-grained structural features from transactional networks. Finally, we fuse these three types of features and input them into a downstream classifier. Experiments show that our TSFF (85.3% Precision) outperforms the state-of-the-art methods, and the effectiveness of each feature is demonstrated.