Hyperedge classification method via graph-based deep learning for detecting illicit bitcoin activities

Abstract

Cryptocurrencies are used as a channel for dealing with illegal funds to criminals. In fact, more than billions-dollar of Bitcoin have been penetrating cryptocurrency exchange every year. Despite the Bitcoin financial forensics' urgency to investigate criminals, existing methods inhere fundamental limitations due to lack of consideration about graph data even though Bitcoin user graph is hypergraph data. To address the limitations of existing methods, we develop a novel Hyperedge Classification method by approximating structure-based edge similarity through graph neural net to detect illegal transactions, represented as hyperedge in the Bitcoin user graph. Moreover, we present a highly scalable graph neural net algorithm by utilizing clustering-based scalable graph neural net and random walk-based sampling method to handle large Bitcoin graph. Based on novel Hyperedge classification, we propose a framework called CENSor, which enables powerful and robust detection than legacy techniques about both illegal entity detection and illegal transaction detection. Compared to the previous graph-based method, our framework succeeds in improving the increment of F-1 score ten times better. We could represent the importance of utilizing proper graph information in Bitcoin analysis by visualizing the Bitcoin cluster graph and Hyperedge-Node switched graph.