Block-Chain Abnormal Transaction Detection Method Based on Dynamic Graph Representation.

The advent of cryptocurrency introduced by Bitcoin ignited an explosion of technological and entrepreneurial interest in payment processing. The user scale of Bitcoin is dynamic, and the participating identities are anonymous, which will lead to more hidden, sophisticated and intelligent money laundering crimes. Therefore, in order to realize intelligent anti-money laundering, it is necessary to accurately detect abnormal transactions. Recently, graph representation learning has shown strong advantages in the field of machine learning, and the current blockchain anomaly detection models based on graph representation learning are mainly designed for static graphs, however, real-world graphs evolve over time. Based on this, this paper proposes a block-chain abnormal transaction detection model DynAEGCN based on dynamic graph representation learning. This model uses the autoencoder as the framework. Firstly, the encoder uses the graph convolutional neural networks to gather neighborhood information to obtain lowdimensional feature vectors. Then, considering the dynamics of graphs, the GRU network is used to evolve the graph model itself over time. Finally, the decoder reconstructs the adjacency matrix and compares it with the real graph to construct the loss. Extensive experiments on the Bitcoin transaction dataset for edge classification tasks against financial crimes show that DynAEGCN model has better performance compared with related approaches.