Revealing unexplored deanonymization risks over the ethereum blockchain

Blockchain technology, especially Ethereum, has enabled major advances in areas like DeFi, digital identity, and governance, as its pseudonymity boosts privacy while hindering attribution and accountability. High-profile cases, such as the sanctioning of Tornado Cash for its role in laundering illicit funds, highlight the urgent need for effective mechanisms to trace malicious fund activity on the blockchain environment. In this dissertation, we propose a novel solution called EtherDiffLinkage (EDL), a behavioral-difference–based framework designed to implement account owner linkage and improve the traceability of malicious fund movements on Ethereum. Technically, we propose a model-based supervised linkage strategy that addresses the limitations of traditional distance metrics. In this framework, the integration of explainable machine learning techniques guarantees transparency and interpretability of the decision. Specifically, we first profile accounts based on behavioral characteristics; enabling the identification of accounts controlled by the same owner, even in cases where significant behavioral variations are present. To ensure transparency in its decision-making process, we leverage an inherently explainable AI (xAI)-based model, providing blockchain investigators with robust and interpretable insights into the mechanisms of deanonymization. Extensive experiments on a real-world dataset demonstrate that EDL outperforms state-of-the-art methods across diverse empirical conditions, effectively addressing key challenges in the deanonymization of malicious accounts on the Ethereum blockchain.