Detecting Rug Pull Scams on Blockchain via Feature Fused Graph Classification

The development of decentralized finance (DeFi) is largely motivated by the blockchain technology. As one of the core applications of DeFi, decentralized exchanges (DEXes) have become a popular choice to trade digital assets, but they also suffer from various scams. Rug pull is a fastly increasing scam, which commits fraud by issuing meaningless tokens. To detect rug pull scams, existing methods usually employ machine learning models based on manually extracted features, while they ignore the learning of the topology structure of token transactions, which is crucial for detecting rug pull scams. In our work, we propose a feature fused graph classification model for rug pull scams detection. In addition to manually extracting token features, we consider the implicit structural information of token transactions. After constructing a graph for transactions of each token, we utilize graph pooling to learn the embedding of the graph, and concatenate it with the extracted features so as to form a complete token representation for classification. Moreover, we propose two new temporal features in feature extraction. The experimental results indicate that our detection model outperforms other methods in detection performance.