Research On Phishing Node Detection In The Ethereum Transaction Network

The Ethereum transaction network refers to the virtual network built based on Ethereum accounts and transactions between accounts.In recent years,as the scale of Ethereum transactions continues to expand and the volume of transactions continues to increase,attacks against the Ethereum transaction network have become increasingly rampant,and these attacks seriously threaten the information and property security of Ethereum users.Among them,the most typical attack is phishing fraud attack.The phishing node of the Ethereum transaction network refers to the account with phishing fraud in Ethereum.Traditional phishing fraud detection methods rely on the feature extraction of the phishing medium,ignoring the features of transfer transactions,and therefore cannot be applied to Ethereum phishing fraud detection.This dissertation approaches the detection of phishing fraud from the perspective of transaction networks by utilizing graph comparison learning methods to mine the behavioral patterns of phishing nodes within the Ethereum transaction network.The main work and innovation of this dissertation are as follows:(1)In order to detect phishing fraud in Ethereum,this dissertation proposes GCL-EPD,a phishing node detection model for Ethereum transaction network based on graph contrastive learning.This model identifies hidden phishing nodes by classifying the transaction subgraphs of each transaction network node.Meanwhile,in order to reduce the interference of uneven distribution of transactions on feature learning and reduce the fluctuation of detection results caused by different graph augmentation methods in the process of graph comparison learning,this dissertation proposes a task-based transaction subgraph augmentation method and introduces transaction sequence information in the process of subgraph feature extraction.The experimental results show that the GCL-EPD model outperforms the baseline method in all detection results on the latest datasets constructed in this dissertation,and the stability of the model is also improved compared with other graph contrastive learning methods.(2)To address the balance problem of information redundancy and information loss in transaction subgraph augmentation,this dissertation proposes a fishing node detection model based on automatic augmentation of transaction subgraph.The model,which combines the existing augmentation method of automatic node dropping with the augmentation method of automatic edge removal proposed in this dissertation,can automate the whole process of graph augmentation and effectively improve the effect of graph contrastive learning detection model.For the problem of high-dimensional sparsity of node features in the dataset,this dissertation proposes an automatic node feature dropping method.The method can reduce the dimension of node features in the transaction subgraph,while retaining the key information in the original node features to the maximum extent.The experimental results show that the training efficiency and accuracy of the model are improved after adding the automatic node feature dropping method.