Directed dynamic attribute graph anomaly detection based on evolved graph attention for blockchain

Abstract

Blockchain is gradually becoming an important data storage platform for Internet digital copyright confirmation, electronic deposit, and data sharing. Anomaly detection on the blockchain has received extensive attention as the foundation for securing blockchain-based digital applications. However, the current blockchain anomaly detection for obtaining network nodes' depth and dynamic change features still needs improvement. In this paper, we propose a blockchain anomaly detection method based on evolved graph attention. Different from general blockchain network modeling methods, we first adopt a dynamic attribute graph network construction method to model each transaction using edges to provide more learnable transaction attribute information for graph representation learning in blockchain networks. Then, we propose an evolved graph attention model structure, which learns the time evolution features of the blockchain network and dynamically updates the learning weights of the subnetwork nodes with different time steps while fully extracting the deep features of the blockchain nodes and avoiding the influence of noisy data on the classification results. In order to solve the dataset imbalance problem, we also apply the SMOTE method for graph-structured data on blockchain for the first time. Finally, we identify node labels in blockchain networks using a binary classification method and verify our proposed scheme through multiple rounds of experiments.