Detecção de mixers na Ethereum

Abstract

This work proposes a machine learning-based approach for the automated detection of mixer-type smart contracts on the Ethereum blockchain, focusing on the Tornado Cash protocol. Two supervised algorithms — Random Forest and Support Vector Machine (SVM) — were trained with on-chain data extracted from the Blockchair platform. The selected metrics included entropy of transferred values, entropy of gas fees, internal calls, and transaction pattern variations. The training set was composed exclusively of interactions with the 10 ETH version of Tornado Cash, in order to assess the models’ ability to generalize to other protocol denominations. The evaluation was conducted using a real scenario from October 29, 2020, when more than 1.1 million transactions were recorded on the Ethereum network, of which only 239 were associated with Tornado Cash. The Random Forest model achieved better performance, correctly identifying the active Tornado Cash contracts, while the SVM presented a higher number of false positives. The results indicate that it is possible to detect suspicious behaviors on the Ethereum network using supervised learning, although improvements such as filtering of legitimate contracts and the incorporation of off-chain data are necessary to reduce inaccuracies.