Identifying, Defending, and Predicting MEs in Ethereum via Graph Neural Networks

Abstract

Maximal Extractable Value (MEV) has been a longstanding unfairness and volatility in Ethereum's final execution, as there are opportunities for transaction ordering to allow for private gains that precede the observation of ordinary users. This study builds a framework for MEV identification, adaptive defense, and short-horizon prediction based on graphs. Records of transactions from MEV labels, bundle-level observations, and Ethereum on-chain data are organized into a heterogeneous transaction graph. A Relational Graph Convolutional Network (RGCN) is employed to learn representations of accounts and transactions that are aware of their relations, and the learned representations are integrated with engineered transaction features in an eXtreme Gradient Boosting (XGBoost) classifier. The defense module applies incremental updates with contrastive self- supervision in order to deal with the evolving nature of attacks. Additionally, a Temporal Graph Neural Network estimates the near-future MEV risk based on the historical graph states. Experimental results show that the graph-based design outperforms traditional classifiers, with the highest accuracy of 91.6% and the highest F1 score of 89.8%; while the temporal modelling gives better and more stable early-warning accuracy as the prediction horizon grows, with the best accuracy of 88.7% at the 10th horizon.