Heterogeneous Network Representation Learning Method For Anomaly Detection Of Ethereum Transaction Data

As the second largest blockchain transaction platform in the world,ethereum has improved the scalability of blockchain transaction applications by deploying smart contracts.Anomaly detection research for ethereum transactions can save ethereum accounts from unscrupulous elements and has attracted much attention in recent years.However,with the expansion of transaction scale,various scams that once appeared in traditional cryptocurrency transactions have been transferred to blockchain digital currency transactions such as ethereum,making the anomaly detection research of ethereum transactions confront many challenges.The ethereum transaction network can be mapped to low-dimensional vectors by network representation learning methods,and the resulting vectors can be applied to anomaly detection to improve the efficiency of ethereum transaction anomaly detection.However,there are often multiple transaction information in ethereum,and it is difficult to represent the complex transaction information by homogeneous networks.In addition,anomalous accounts will hide the anomalous transaction information by interacting with normal accounts,which in turn affects the effectiveness of representation learning.To address the above problems,this paper proposes a heterogeneous network representation learning method for anomaly detection of ethereum transaction data.The main research works are as follows:(1)A multiple attribute fusion heterogeneous network representation learning method for anomaly detection of ethereum transaction dataFor the existence of multiple heterogeneous attributes of ethereum transaction data,this paper proposes a heterogeneous network representation learning method based on multiple layers of attribute fusion.The method builds a heterogeneous transaction network by collecting ethereum account transaction data with multiple attributes,and designs a feature extraction strategy for ethereum transaction information to effectively retain the multiple properties of ethereum transactions.In addition,the method fuses heterogeneous transaction attributes in the process of transaction data representation,and learns node representation through a heterogeneous network representation learning method consisting of base embedding,edge embedding and attribute embedding to improve the accuracy of anomaly detection results for ethereum transaction data with multiple transaction features.The experimental results demonstrate that the heterogeneous network representation learning method with multiple attribute fusion outperforms existing algorithms in the application of anomaly detection in the presence of multiple features ethereum transaction datasets.(2)A relational filtering heterogeneous network representation learning method for anomaly detection of ethereum transaction dataFor the behavioral disguise of anomalous accounts in ethereum transaction data,this paper proposes a heterogeneous network representation learning method based on relationship filtering.The method designs a relationship similarity-aware filtering strategy that uses a reinforcement learning component to adaptively select optimal neighbors within each relationship on the basis of relationship filtering to effectively filter anomalous neighbor relationships.In addition,the method proposes a node aggregation method based on heterogeneous graph neural network within and between relations,and adds a residual mapping module to the aggregation process to improve the accuracy of learning results for the representation of ethereum transaction data containing relational artifacts.Experimental results demonstrate that the heterogeneous network representation learning method of relational filtering outperforms existing algorithms for anomaly detection applications on ethereum transaction datasets with artifactual behavior.Finally,the paper summarizes the main research contents of the heterogeneous network representation learning method for anomaly detection of ethereum transaction data,and looks forward to future research directions.