As the second largest blockchain platform, Ethereum has spawned many decentralized applications and significantly boosted the development of blockchain finance. However, Ethereum’s anonymity has led to a spike in illegal activities, threatening blockchain financial security. Identity inference can reveal identities behind anonymous addresses and detect illicit behaviors. However, existing methods face issues such as a lack of large-scale, multi-identity transaction datasets; no effective subgraph sampling; and weak node representations. In this paper, we propose a systematic identity inference scheme to address these issues. Specifically, as there is no publicly available Ethereum multi-identity inference dataset, we first construct a large-scale transaction dataset. To the best of our knowledge, this is the largest labeled transaction dataset for Ethereum identity inference. Then, we systematically investigate the impact of different subgraph sampling strategies and propose a joint sampling strategy that well preserves structural information and behavioral patterns of nodes. Finally, we design a graph transformer-based identity inference model called \(I^2GT\)  that leverages the strong representation learning capabilities of the transformer architecture to produce more effective node representations. Extensive experiments on real-world Ethereum datasets demonstrate that our method(90.30% F1-score, and 98.57% AUC) achieves the state-of-the-art performance.