Neighborhood Subgraph-based Illicit Transaction Detection in Cryptocurrency Networks

With the widespread adoption of cryptocurrencies, associated illicit activities such as money laundering, fraud, extortion, and Ponzi schemes have garnered significant attention. Traditional studies have extensively employed graph neural network technologies to identify illicit transactions, typically treating the entire transaction network as input to classify its nodes. Although these methods perform well on small-scale networks, they face performance limitations when processing large-scale blockchain transaction data. To address this issue, we propose a neighborhood subgraph-based method for detecting illicit transactions, integrating GCN and LSTM. In this approach, the GCN captures and integrates information from adjacent nodes for each transaction, facilitating a deep understanding of network structures. The LSTM is tasked with capturing the sequence and variations in fund flows, effectively tracking the direction and temporal characteristics of the money flow. Experimental results indicate that when using 3-hop neighborhood subgraphs, our model's performance surpasses other GCN-based methods. Compared to previous models that require considering the entire transaction network, our approach only needs to utilize information from an average of 80 nodes, significantly improving efficiency.