Predicting functional roles of Ethereum blockchain addresses

Ethereum is one of the largest blockchain programming platforms. Users in Ethereum are identified using public-private key addresses, which are difficult to connect to real-world identities. This has led to a variety of illegal activities being encouraged. However, based on their transactions’ functional roles, these addresses can be linked and identified. In this paper, we proposed a methodology for predicting the functional roles of Ethereum addresses using machine learning. We build machine learning models to predict the functional role of an address based on various features derived from the transactional history over varying window sizes. We have used labeled dataset of 300 million transactions that are publicly available on the Ethereum blockchain. The test data results show that the XGBoost classifier with eleven features vector and 200 window sizes can predict the role of an unseen address with the best achievable accuracy of 73%. We have also trained and tested the deep learning models on the dataset, CNN model predicted the labels with 86% accuracy. Using machine learning models, we have also devised a measure of anonymity and compared it for unlabelled addresses. Further, to qualitatively validate our prediction, we also discovered Ethereum addresses used on the dark web pages and predicted their functional roles with our trained models. Most of these addresses were behaving like Wallet_app , Shapeshift , and Mining and this prediction was aligned with the background information extracted from the context of address usage on the dark web page.