Research On Bitcoin Abnormal Transaction Detection Based On Graph Neural Network

Since the concept of digital currency was put forward in 1983,with the continuous development of blockchain technology,it has received widespread attention from various countries.The birth of Bitcoin in 2009 triggered a wave of encrypted digital currencies.Bitcoin is the most typical,most valuable,and most concerned digital currency in the blockchain.Many criminals use its point-to-point decentralization feature to hide their true identity and provide currency transactions for some illegal goods and services in the darknet market.Financial security issues have major hidden dangers,and timely detection of abnormal Bitcoin transactions and responses can minimize and avoid major economic losses.Therefore,abnormal transaction detection for Bitcoin is particularly important in the field of digital currency security.The current abnormal transaction detection methods for Bitcoin are still in the initial exploration and development.Traditional machine learning methods usually only focus on the feature representation of nodes and ignore the topology of the graph and the dynamics of Bitcoin transactions.In response to the above problems,this thesis combines the graph neural network and the variation of the recurrent neural network to propose a scheme for detecting abnormal Bitcoin transactions.The main work is as follows:(1)Modeling Bitcoin transactions as dynamic graphs.The data structure of Bitcoin transactions is irregular.According to the characteristics of the Bitcoin transaction chain,it can be mapped to an undirected graph,and the maps of Bitcoin transaction mapping at different times may be completely different.Further analyze its dynamics in the time series.And model it as a dynamic graph according to the time dimension,which is more in line with real application.(2)A graph convolutional network model SA-GCN with an adaptive neighborhood aggregation method is proposed for feature extraction of Bitcoin transactions.Input the undirected graph of Bitcoin transaction mapping into the network,the model can learn the topological structure and node characteristic information of the graph at the same time.By adding a gating unit,it can calculate the neighborhood entropy and central neighborhood similarity of the metrics based on information theory.Control the degree of neighborhood aggregation of each node.At the same time,the residual error mechanism is introduced to update the network layer,which realizes an adaptive neighborhood aggregation method and forms a better feature representation.(3)A SA-GCN-GRU model combined with a gated recurrent unit GRU is proposed to detect abnormal Bitcoin transactions.Bitcoin transaction data is dynamically changing.As new accounts are generated and new transactions occur,a constantly changing graph structure will be generated.Bitcoin transactions are transformed into time series data with spatial characteristics,combined with gated loop units,and at the same time from Time dimension and space dimension model the Bitcoin transaction data,and update the graph convolutional network model along the time dimension to better capture the dynamics of Bitcoin transaction data.This thesis conducts experiments on three public Bitcoin transaction data sets Elliptic/Bitcoin OTC/Bitcoin Alpha,and the experimental results prove the effectiveness of the scheme.