Expose Your Mask: Smart Ponzi Schemes Detection on Blockchain
Shuhui Fan, Shaojing Fu, Haoran Xu, Chengzhang Zhu
Abstract
The anonymity of blockchain has caused Ponzi schemes to be transferred to smart contract platforms by scammers. These Ponzi schemes wearing the mask of smart contracts caused huge losses to people, which makes the detection of smart Ponzi schemes attract people's attention. Recent methods mainly focus on machine learning technology to enable automatic detection for smart Poniz schemes. However, there are some problems with their methods. Firstly, the gradient boosting algorithm in machine learning they used have the problem of prediction shift due to target leakage when processing category features and calculating gradient estimates. Secondly, they ignored the imbalance and repetitiveness of Ponzi schemes on smart contract platforms. These problems can directly lead to model overfitting and affect the generalization ability of trained models. This paper proposes a novel Ponzi schemes detection method on smart contract platform for blockchain. Our method addresses the above issues with the following strategies. Firstly, we leverage ordered target statistic (TS) to process the category features of smart contract. Secondly, we solve the imbalance of dataset through a data augmentation method. Thirdly, with the idea of ordered boosting algorithm, we train a PonziTect model to fight prediction shift caused by target leakage. Based on the above ideas, the experimental results fully manifest the effectiveness and reliability of our model in detecting smart Ponzi schemes on the blockchain. Specifically, our model achieves 98% F-score on the real-world dataset, which significantly outperforms the existing methods. Using our method, we estimate that there are about 532 Ponzi schemes on Ethereum.