Detecting Bot Wallets on the Ethereum Blockchain

Abstract

Bots are increasingly relevant agents on the Ethereum blockchain that enable high degrees of automation and efficiency. On the other hand, they can pose adversarial danger to users and protocols. Therefore, it is crucial to detect bots and investigate their behavior in order to inform policy and human users of the associated risks. To better understand the bot landscape on Ethereum, we propose a categorization of bots integrating existing research with findings of new bot classes documented by code or on-chain data. This categorization consists of 7 main categories and 26 subcategories of bots. Existing bot detection systems are predominantly based on predefined rules and highly specific to certain types of bots, making them inflexible and unable to detect unanticipated changes in bot behavior. In order to explore a more data-driven approach, we investigate to what extent bots on Ethereum can be detected using supervised and unsupervised machine learning. As a benchmark, we use rule-based heuristic methods to distinguish bots from humans. Our findings suggest that bots can be successfully distinguished using supervised learning and unsupervised methods can find clusters exhibiting a high purity of bots. Supervised methods using only a small dataset showed better results than clustering using much more data, while both methods beat the rule-based heuristic benchmark models by a large margin. Finally, we define five time frames on the Ethereum blockchain and use our best-performing classifier to detect bots in each of them. Comparing the time windows, we observe that bots are more prevalent in hype phases of Ethereum.