Research On Bitcoin Transaction Anomaly Detection

As the most successful application of blockchain technology,Bitcoin has attracted much attention in investment and academic fields in recent years.Relying on cryptography,time stamp technology,P2 P network technology,etc.,on the one hand,Bitcoin continues to be hot in the financial investment market and its market value has soared;On the other hand,it is precisely because of the decentralization,anonymity and other characteristics,coupled with the absence of regulatory means,the Bitcoin system has become a tool for criminals to commit crimes,and a series of illegal and criminal cases,such as dark network transactions,extortion and fraud,and money laundering,have broken out frequently,posing severe challenges to public security work and cyberspace security.With the continuous advancement of machine learning technology,there are more and more studies using machine learning technology to detect abnormal Bitcoin transactions,and good results have also been achieved in large-scale transaction data.In this paper,the existing research methods are comprehensively analyzed,and the anomaly detection technology and Bitcoin transaction anomaly detection methods are deeply explored.The anomaly detection of bitcoin transaction is mapped into a machine learning problem of two categories,and the feature extraction and algorithm selection are improved.The specific work is as follows:Firstly,in terms of feature extraction,traditional detection methods regard transactions as isolated nodes in the transaction network,and there are problems such as insufficient mining of interactive information hidden in the network structure and insufficient ability to express abnormal patterns.According to the principle of homogeneity of complex network,a method for detecting abnormal transactions in Bitcoin network based on feature fusion is proposed.With the help of the theoretical knowledge of link prediction,the random walk method of sampling neighborhood node sequences is firstly defined by using node similarity.Then,starting from the original node,a random walk under a certain path selection probability is used to generate a sequence of neighbor nodes,so as to obtain the characteristics of the neighbor nodes as the node implicit information under the interaction relationship.On this basis,in order to play the complementarity of original features and implicit features,the two are mapped to a unified feature expression space.Finally,the final feature set after dimension reduction by stacked automatic encoder is used by RF classifier to verify the effect of the feature extraction method in abnormal transaction detection.Through designing experiments,classical classifiers in anomaly detection fields such as KNN and XGBoost are constructed under different feature extraction methods,and the effectiveness of the detection model and other feature extraction methods constructed in this paper is verified by comparison.Second,in terms of detection algorithms,in view of the insufficient accuracy and generalization performance of traditional single detection algorithms,and the lack of selection and combination of basic learners and the locality of data in existing parallel integration methods,this paper builds a detection model with LSCP algorithm.At the same time,a single algorithm with different research perspectives is used to generate basic learners in the LSCP algorithm to increase the differences between basic learners.After conducting experiments on the data set provided by predecessors and comparing the research methods,the experimental results show that the method has better comprehensive detection ability,and the combination of the feature extraction method in this paper can further improve the overall detection effect.