Bitcoin Double-Spending Attack Detection using Graph Neural Network

Abstract

Bitcoin transactions include unspent transaction outputs (UTXOs) as their inputs and generate one or more newly owned UTXOs at specified addresses. Each U TXO can only be used as an input in a transaction once, and using it in two or more different transactions is referred to as a double-spending attack. Ultimately, due to the characteristics of the Bitcoin protocol, double-spending is impossible. However, problems may arise when a transaction is considered final even though i ts finality has not been fully guaranteed in order to achieve fast payment. In this paper, we propose an approach to detecting Bitcoin double-spending attacks using a graph neural network (GNN). This model predicts whether all nodes in the network contain a given payment transaction in their own memory pool (mempool) using information only obtained from some observer nodes in the network. Our experiment shows that the proposed model can detect double-spending with an accuracy of at least 0.95 when more than about 1% of the entire nodes in the network are observer nodes.