An Empirical Study of Two Bitcoin Artifacts Through Deep Learning

Abstract

Human artifacts like technical papers and computer programs often carry the individual styles of their creators. If retrieved properly, such style information from the artifacts can be used to categorize the artifacts, compare the relative “similarities” among artifacts, and may even be used for tracing the authorship of a new artifact. Bitcoin is a peer-to-peer cryptocurrency and its author(s) goes/go by the pseudonym of Satoshi Nakamoto. In this article, we use deep learning to study the styles of two Bitcoin artifacts: the first version of Bitcoin’s source code, v0.1.0, which was released in early 2009, and the original Bitcoin white paper, which is dated Oct. 2008. Both studies use the deep learning technique, which first utilizes extensive computing power to generate a neural network model from labelled training data and then uses the model to predict the authorship of unknown data. For the Bitcoin source code artifact, the data set is a set of cryptography software that were built around 2008/2009 and it has 16 known labels. Our model achieves 89.1 % validation accuracy and our prediction results show that the Bitcoin source code is likely produced by multiple authors and Hal Finney is not one of them. For the Bitcoin white paper, we compiled a second data set of financial cryptography papers that are in the same knowledge domain. This data set has 436 known labels. Our model achieves 55.1 % validation accuracy and it has identified four technical papers that are “similar” to the Bitcoin white paper.