A time-aware LSTM model for detecting criminal activities in blockchain transactions

Abstract

This paper introduces the Time-Aware LSTM (T-LSTM) model to identify criminal activities involving USDT on wallet addresses within the blockchain ecosystem. The model utilizes a time-aware LSTM architecture to learn the continuous variations in node address features over different transaction time intervals. Additionally, a gating mechanism filters the influence intensity of neighboring transaction node addresses on the central node. The gating mechanism accounts for the transactional correlation strength between node addresses. Finally, a self-attention mechanism is employed to integrate node address features across various transaction timestamps, producing a comprehensive feature representation for the addresses. Experimental results demonstrate that the T-LSTM model effectively captures the dynamic feature changes of node addresses over irregular transaction intervals, outperforming traditional detection models regarding precision, recall, and F1 score on the test set.