HSTA: Ethereum Phishing Fraud Detection Model Based on Dynamic Graph Hybrid Spatio-Temporal Attention Mechanism

With the popularization of blockchain technology, phishing and fraud activities on platforms such as Ethereum have become increasingly rampant, posing a serious threat to the transaction security of the blockchain ecosystem. Existing detection methods struggle to capture the dynamic evolution of transaction networks and effectively integrate spatio-temporal features to uncovering complex fraud patterns. Therefore, in this paper, we propose a novel phishing fraud detection method using a Hybrid Spatio-Temporal Attention network (HSTA) to address this increasing problem. Specifically, we first construct the Ethereum transaction records as a continuous-time dynamic graph. Then, we introduce a hybrid spatio-temporal attention mechanism to dynamically capture complex network dependencies. Finally, we aggregate neighborhood information via multi-layer Transformer encoders to generate node embeddings with strong representation capabilities for the downstream phishing transaction chain prediction task. The experimental results on the real Ethereum phishing fraud dataset and two public datasets show that HSTA outperforms the existing baseline models in both transduction and inductive learning scenarios. These findings demonstrate the practical value and effectiveness of our proposed HSTA model in combating phishing fraud activities in blockchain environments. Our code can be obtained at https://​github.​com/​cmx12138/​HSTA.