Research On Abnormal Transaction Behavior Recognition Algorithm For Ethereum

With the widespread application of blockchain technology,the cyberspace security issue of phishing has also appeared in the emerging blockchain cryptocurrency ecosystem.Because phishing fraud in cryptocurrency transactions has its own unique characteristics compared to traditional phishing,many existing phishing detection algorithms are not usable.Therefore,considering that the graph nodes in the Ethereum transaction graph itself do not have any characteristics(all information exists in transactions)and how to identify phishing users,this paper discusses the graph node classification and graph classification model in Ethereum phishing fraud detection.This not only makes the research results more complete and reliable,but also provides an effective new idea for the detection of phishing fraud in blockchain cryptocurrency networks.In the graph node classification model,XGBoost is used to classify users on the basis of manually extracted features,and it is confirmed that the time factor plays an important role in the identification of Ethereum phishing users.Then,the nonlinear features of blockchain cryptocurrency transaction graph(graph structure features)are added on the basis of the manual features.The results of Graph Embedding experiment show that the effect is better after adding graph structure features.Since the majority of nodes in the Ethernet transaction graph are normal users and the proportion of phishing nodes is very low,there is a serious imbalance problem in the graph node classification.In order to solve this problem,this paper integrates the Graph SMOTE model based on considering the temporal characteristics in Ethereum transactions.The model not only solves the imbalance problem more scientifically,but also improves the classification effect.And its F1,AUC and ACC scores all reachIn the graph classification model,based on graph convolutional networks,This paper have researched and built a high-performance model for detecting blockchain cryptocurrency phishing fraud.This model divides the constructed blockchain cryptocurrency transaction graph into “Sender” and “Receiver” graphs,according to the sending and receiving directions.Then,the edge features in the graph are transferred.Finally,after using a double-layer graph convolution network for feature learning,the feature are sent to the classifier for fraud detection.After completing training on the actual dataset collected from Ethereum,the model achieved an accuracy of 88.02% and an F1 score of 88.14% on the test data,which had a better performance than that of the other graph classification models.