LLM-Guided Temporal Graph Neural Network for Illicit Transaction Detection in Blockchain Networks

Abstract

Illicit transaction detection in blockchain networks is important for cryptocurrency anti-money laundering and other on-chain financial applications. Detecting suspicious transactions is challenging because illicit behaviors are often reflected through cross-entity fund-flow dependencies, temporal transaction evolution, and heterogeneous semantic attributes rather than isolated transaction features. This paper proposes an LLM-guided temporal graph neural network for illicit transaction detection in blockchain networks. The LLM is used as a schema-guided semantic interface rather than a final decision maker. It normalizes deterministic non-label transaction summaries, structural proxy tags, temporal behavior indicators, and label-free behavioral pattern tags into graph-compatible semantic attributes. A GraphSAGE encoder learns relational representations from blockchain transaction graphs, while an LSTM module captures temporal evolution across observation windows. Experiments on the Elliptic Bitcoin Dataset show that the proposed LLM-GNN-LSTM framework improves recall, F1-score, ROC-AUC, and PR-AUC over strong feature-based, static-graph, and temporal-graph baselines, including XGBoost, LightGBM, GAT, GraphSAINT, EvolveGCN, and TGN.