Research On Bitcoin Abnormal Address Detection Method Based On Transaction Structure Graph Analysis

With the development of blockchain technology,cryptocurrencies represented by Bitcoin are used by more and more people around the world.Bitcoin is based on cryptography theory and peer-to-peer network(Peer-to-Peer)technology to ensure the security,openness and transparency of currency circulation.Bitcoin transaction records are designed to be anonymous,and users use addresses instead of real identities for transactions.This anonymous recording method provides Bitcoin with anonymity and privacy protection,but it also makes Bitcoin transactions more difficult to trace,making it an ideal choice for criminals,resulting in a large number of Bitcoins used for extortion,money laundering,illegal drugs,weapons,etc.transactions and other illegal transactions.Bitcoin addresses involved in these illegal transactions are usually called abnormal addresses.Therefore,research on Bitcoin abnormal address identification technology can help regulators better understand the risks and loopholes in the Bitcoin trading market,and provide strong support for cracking down on related criminal activities.At present,the methods for Bitcoin abnormal address identification are mainly heuristic-based address clustering and machine learning-based address clustering methods.These methods have two main problems.On the one hand,traditional heuristic address clustering algorithms rely on manual Participation,lack of flexibility,more sensitive to noise in features,which leads to low recognition rate.Although the address clustering method based on machine learning can solve the problems of traditional methods,due to the lack of full use of Bitcoin transaction information,the input features cannot fully and accurately reflect the relationship between addresses,so the model has high false positives rate and omission rate.On the other hand,the data set used for Bitcoin abnormal address identification has a data imbalance problem,which leads to a low recognition ability of the classification model for abnormal address samples.In the current solution,the traditional oversampling method may lead to overfitting,while the undersampling method may lose information.In addition,cost-sensitive learning requires manual setting of cost weights,which is not automatic enough.Bitcoin abnormal address identification is realized by analyzing and modeling the Bitcoin transaction network,so the data set has graph structure characteristics.Since the graph structure data can effectively reflect the transaction association of Bitcoin in the market,it plays an important role in identifying abnormal addresses.Based on the above considerations,this paper will conduct research from the following two aspects:1.Aiming at the problem that the existing bitcoin abnormal address identification method does not make full use of all the characteristics of abnormal bitcoin addresses generated during the transaction process,resulting in a low recognition rate,this paper proposes a bitcoin abnormal address identification based on transaction network feature enhancement Method,a Bitcoin anomaly address detection model TNF-AARM is constructed.First,convert the bitcoin transaction data into a complex network,analyze the characteristics of the bitcoin transaction network from the perspective of the complex network,and propose an improved Page Rank-based node importance feature construction method based on these characteristics,and introduce the bitcoin transaction amount and frequency correlation to get the new Page Rank value.Then,the graph data is learned by graph neural network,and graph embedding vectors are extracted to add to the feature set.Finally,the extracted feature set is applied to a variety of machine learning algorithms,including the construction and evaluation of classifiers such as support vector machines,logistic regression,random forests,and extreme gradient boosting trees(XGBoost).Experimental results show that the feature extraction method proposed in this paper can significantly improve the performance of classifiers and perform consistently across different classifiers.Especially with the XGBoost classifier,the F1 score increased from 0.83 to 0.94,and the AUC value increased from 0.88 to 0.95,achieving the best results.Compared with the feature extraction methods commonly used in the existing Bitcoin illegal transaction detection,the method in this paper performs better in overall detection performance.2.Aiming at the data imbalance problem of bitcoin abnormal transactions,a data imbalance processing method Embedding-SMOTE based on graph embedding is proposed.This method combines the graph embedding technology Graph Sage based on the graph neural network.First,the nodes of the graph structure data are converted into graph embedding feature vectors,and then the minority class nodes are synthesized by the SMOTE method to balance the data distribution,and then the weighted inner product decoder is used to calculate The correlation between nodes guides the generation of edges;then,the nodes and edges are converted into new graph embedding feature vectors through the Graph Sage module;finally,the feature set is sent to the tree model XGBoost for node classification.In order to verify the effectiveness and scalability of this method,this paper applies it to the Bitcoin dataset and other imbalanced graph datasets,and conducts experiments.The experimental results show that the average improvement in AUC value and F1 score is 0.105 and0.095,respectively.