Phishing Detection in Ethereum via Transaction Graph Embedding

As blockchain technology advances at an unprecedented pace, phishing scams increasingly exploit vulnerabilities in Ethereum transactions. These attacks typically involve fraudulent addresses that deceive users and illicitly expropriate digital assets, posing significant threats to the security and integrity of the blockchain ecosystem. In this work, we propose txnet2vec, a novel framework for detecting phishing addresses based on transaction graph analysis. Our approach begins by collecting labeled Ethereum transaction data and constructing a directed, weighted transaction graph, where nodes represent addresses and edges denote transactions. To capture both structural and transactional characteristics, we employ network embedding techniques to learn low-dimensional representations of addresses. To further improve detection accuracy, we design an attention-based feature fusion mechanism that integrates multiple random-walk-based sampling strategies, incorporating transaction amounts, temporal features, and market-driven behaviors. The learned embeddings are then fed into a Support Vector Machine classifier to distinguish phishing from benign addresses. Extensive experimental results demonstrate that txnet2vec achieves superior performance compared to existing baselines in phishing detection within Ethereum's transaction network.