Leveraging Data-intensive Graph Attention Networks to Quantify the Bitcoin Blockchain True Decentralization

Juan Muñoz, Benito Rojas

Abstract


The decentralized nature of the Bitcoin blockchain is a cornerstone of its success, yet measuring true decentralization remains a significant challenge. Moreover, with the institutional participants entering the space, there is an increasing necessity for compliance with global financial regulations, to perform forensic analysis for fraud-activity, etc., and yet, ensuring the integrity of a transaction with high probability of it being validated in a truly decentralized way is not as straight forward as the media and industry reports communicate. Our approach consists of two key components. First, we present an open-source Extract, Transform, Load (ETL) process tailored for blockchain data, which fosters improved analytical transparency and aids regulatory and forensic analysis. Secondly, we apply a graph-based data structure approach that employs GATs to classify Bitcoin addresses. This method is chosen for its robustness in handling relational data and its ability to focus selectively on informative parts of the transaction graph. The efficacy of our approach is demonstrated through controlled experiments, yielding: 10 years of historical transactions, narrowed to a subset of the year 2022 thanks to temporal and informational-based comparisons, producing a classification of wallet addresses (Exchange, Gambling, Mining, Service) with accuracy of 92.87%, precision of 89.35%, and recall of 92.87%. As well the accompanying centrality metrics. Our findings have significant implications for financial technology, enabling more informed policy decisions and driving innovation in the field.

Keywords


Graph attention networks, graph machine learning, blockchain

Full Text: PDF