Blockchain provides a decentralized environment for applications and information systems in various fields. It is an innovative revolution for the traditional Internet. However, without proper regulatory mechanisms, the blockchain technology has gradually become a hotbed of criminal activities, such as Ponzi scheme that brings huge economic losses to people. To maintain the security of the blockchain system, the machine learning technique, which can detect smart Ponzi schemes automatically has recently received extensive attention. However, the existing method has potential target leakage and prediction shift problems when dealing with category features and calculating gradient estimates. Besides, they also ignore the imbalance and repeatability of smart contracts, which often causes the model to overfit. In this paper, we introduce a novel method for detecting smart Ponzi schemes in blockchain. Specifically, we first expand the dataset of smart Ponzi schemes and eliminate the unbalanced dataset via data enhancement. Then, we leverage ordered target statistics (TS) to handle the category features of smart contract without target leakage. Finally, we propose an anti-leakage smart Ponzi schemes detection (Al-SPSD) model based on the idea of ordered boosting. Experimental results show that our proposal outperforms the competitive methods and is effective and reliable in detecting smart Ponzi schemes. Al-SPSD achieves 96% F-score and detects about 1,621 active smart Ponzi schemes in Ethereum.
Read full abstract