Block-chain Abnormal Transaction Detection Method Based on Auto-encoder and Attention Mechanism

Abstract The significant changes brought by block-chain technology have posed many challenges to financial services, ecological security, and privacy protection. Therefore, in order to achieve intelligent block-chain supervision and assess the risk of potential money laundering, terrorist financing, and other financial crimes of customers, anomaly detection of blockchain networks is required. Structurally, blockchain data is essentially represented by a graph, where nodes represent addresses and edges represent behaviors such as transactions, and the model after constructing the transaction graph can extract high-dimensional features in the graph structure relationships. Existing anomaly detection methods ignore the interaction information between network structure and node attributes and have limited ability to detect anomalies. Based on this, this paper proposes GraphAEAtt, a deep learning framework based on self-encoder and attention mechanism, which consists of a structural auto-encoder and an attribute auto-encoder to jointly learn node and attribute feature vector representations, and in addition, introduces an attention mechanism to learn the correlation between nodes and their neighboring nodes. First the structural encoder converts the observed raw node attributes into a vector representation of the low-dimensional potential space, and then the shared attention mechanism is used to aggregate the embeddings of all neighboring nodes to finally generate node embedding. The attribute encoder uses a multi-layer perceptron to map the observed attribute data into a potential attribute embedding representation. Then, a structure decoder is used to reconstruct the adjacency matrix and an attribute decoder to reconstruct the attribute matrix, and the reconstruction error of the nodes is measured from both structure and attribute perspectives as the objective function for neural network training. Then anomaly detection is implemented based on the reconstruction error of the nodes measured from both structure and attribute perspectives. Finally, a large number of experiments are conducted to verify the effectiveness of the proposed method in real datasets.