Blockchain Abnormal Transaction Detection Based On Graph Neural Network

In recent years,the cryptocurrency market has boomed and blockchain has attracted much attention as its key technology.However,the decentralized and anonymous characteristics of blockchain make it difficult to detect and track financial crimes of cryptocurrencies,which not only endangers financial security but also hinders the promotion of blockchain technology.With the expanding and diversifying applications of blockchain technology,such as digital identity authentication,supply chain management,Internet of Things,etc.,how to supervise abnormal transactions on blockchain has become an urgent problem to be solved.Machine learning,with its advantages of self-learning and selfadaptation,has been widely used in blockchain abnormal transaction detection with excellent results.However,the existing methods have the following technical challenges:(1)The existing blockchain abnormal transaction detection methods are difficult to effectively combine the attribute information and topology information of blockchain nodes,resulting in the inability to mine the potential transaction information and affecting the detection results.Meanwhile,the blockchain transaction data is huge and extremely unbalanced,which greatly affects the detection results of abnormal transactions.(2)Most of the existing studies do not consider that the blockchain transaction network changes at all times,which makes it difficult to effectively use the temporal information of blockchain transactions.Therefore,this paper proposes a blockchain anomaly transaction detection method based on graph neural network from both static and dynamic perspectives,which mainly includes:(1)To address the problem that existing methods are difficult to combine blockchain node attribute information resulting in low detection accuracy,this paper proposes a static blockchain abnormal transaction detection method based on random wandering.Firstly,a random wandering strategy for transaction networks is proposed,and the information of transaction nodes can be combined with biased sampling in the wandering sampling process to achieve efficient fusion of node attribute information and connection structure information.Secondly,a blockchain-oriented multi-scale feature extraction method is proposed to combine the features of blockchain for double-sequence multi-scale sampling to capture the feature information of transactions as well as the higher-order relationships implied in the transaction information.Finally,this paper is validated on real blockchain transaction data,and the results show that this method can fuse the attribute information of nodes and connection structure information,and the detection results exceed many classical algorithm models such as Deep Walk and Walklets.(2)To address the problem that existing methods are difficult to effectively use blockchain transaction timing information resulting in low detection accuracy,this paper proposes a dynamic blockchain anomaly transaction detection method based on graph attention network.First,the interrelationship between different transaction attributes is learned through the graph attention mechanism to prevent the neural network from treating different transaction factors equally.Secondly,the high-value information in the time dimension is retained through the gated cyclic unit,and the changing characteristics of blockchain transaction features in the time series are mined.Finally,this paper conducts experimental comparisons with multi-class models on the Bitcoin-OTC and Bitcoin-Alpha datasets to verify that this method can effectively combine the temporal information of nodes and improve the detection accuracy.