Phishing Detection on Ethereum via Learning Representation of Transaction Subgraphs

With the widespread application of blockchain in the financial field, blockchain security also faces huge challenges brought by cybercrimes, especially phishing scams. It forces us to explore more efficient countermeasures and perspectives for better solution. Since graph modeling provides rich information for possible downstream tasks, we use a surrounding graph to model the transaction data of a target address, aiming to analyze the identity of an address by defining its transaction pattern on a high-level structure. In this paper, we propose a graph-based classification framework on Ethereum. Firstly we collect the transaction records of some verified phishing addresses and the same number of normal addresses. Secondly we form a set of subgraphs, each of which contains a target address and its surrounding transaction network in order to represent the original address on graph-level. Lastly, based on the analysis of the Ether flow of the phishing scam cycle, we propose an improved Graph2Vec, and make classification prediction on the subgraphs we built. The experimental results show that our framework has achieved a great competitiveness in the final classification task, which also indicate the potential value of phishing detection on Ethereum via learning the representation of transaction network.