Research On Blockchain Anomaly Transaction Detection Based On Graph Neural Network

In recent years,with the widespread popularity and development of cross-border ecommerce and Internet finance,the blockchain field has become a hot spot,while more and more money laundering and economic crimes are shifting from the traditional financial field to the blockchain field,making the current money laundering crimes in the blockchain industry increasingly serious,so it is especially important to study the abnormal transaction detection model for blockchain data.However,due to the complexity and dynamic nature of blockchain data,there is no solution that can be fully applied to realistic scenarios,and the existing solutions suffer from high false alarm and leakage rates in practical applications.In contrast,graph neural networks capture the local structure and global features in graph data by defining the neighbor relationship between nodes,which makes it highly accurate in processing complex graph data and can maintain strong robustness in dealing with noisy and missing data.Therefore,this thesis proposes a graph neural network-based approach for exploring potential correlations between blockchain transaction features,and after the model is trained to learn to aggregate input features more effectively in the local neighborhood of nodes to determine whether a transaction is abnormal or not.The main research components of the thesis are as follows:(1)The design and analysis of weighted sampling GraphSAGE blockchain abnormal transaction detection modelAt present,relatively little research has been conducted on blockchain abnormal transaction detection,and most of the existing models are trained by traditional machine learning algorithms.There are fewer studies related to using graph neural networks for abnormal transaction detection of blockchain data.To solve this problem,this thesis applies the traditional model Graph SAGE in graph neural networks to blockchain abnormal transaction detection,and improves the existing methods for the shortcomings and problems in sampling,aggregates the node features within the target node neighborhood with weighted sampling,uses weighted cross entropy as the loss function,and designs the weighted sampling Graph SAGE blockchain abnormal transaction detection model.Experimental results show that our model achieves the best F1 score in terms of prediction accuracy with lower false alarm rate and leakage rate compared with other models,and outperforms other models in terms of robustness.(2)The design and analysis of dynamic EvolveGraphSAGE blockchain anomaly transaction detection modelThe existing models in graph neural networks for blockchain transaction data for node classification tasks are mainly designed based on convolutional graph neural networks,random forests,etc.These traditional models do not take into account the dynamic situations such as frequent appearance and disappearance of nodes and possible appearance of new nodes after training,which leads to the problem of poor training effect of static models.To address this problem,this thesis proposes the Evolve Graph SAGE model.The model proposes to update the parameters of the graph neural network model by using recurrent neural networks to capture the dynamic characteristics of the changing graph data structure.The performance of the model is verified on a public blockchain transaction dataset,and the experimental results improve the F1 score by about 5% compared to other models when concept drift occurs,indicating that our proposed dynamic model outperforms the comparative static model in terms of detection performance.