HDFG: Ethereum Smart Contract Honeypot Detection Based on Pre-Training Techniques

Abstract

In recent years, a new fraud method, namely smart contract honeypots, has emerged on the famous blockchain platform Ethereum. The difference from smart contract vulnerabilities is that the contract honeypot essentially has no vulnerabilities, luring victims to call in a seemingly vulnerable form. However, the victims ultimately cannot obtain the desired benefits and will lose certain funds. Deep learning algorithms are preferred among current contract honeypot detection methods because they can learn more general characteristics and do not rely on expert experience. Most previous works use natural language models to learn the opcodes of contract honeypots but overlook the relevant structural features of the source code. We propose a novel method called the Smart Contract Honey-pot Data Flow Graph, which utilizes a data flow graph to extract the calling relationships of critical source code within contract honeypots and employs a pre-trained model for representation learning. First, contract honeypots generally have a code that transfers money to the calling address, which is critical information for constructing a source code data flow graph. Then, the pre-trained model is used to learn the source code representation and perform downstream classification tasks. The F1-score of our model significantly outperforms the state-of-the-art approaches in the contract honeypot classification task and is close to the highest performance in the detection task. In addition, this model is an end-to-end model that can detect unknown-type contract honeypots.