Research On Blockchain Malicious Transaction Detection Method Based On Privacy-Preserving Computation

Blockchain technology has penetrated into various fields of market economy and social governance,bringing revolutionary changes to industries such as digital currency,financial transactions,and digital government affairs.Recent studies have shown that the blockchain network is still full of various malicious transactions aimed at financial fraud and money laundering activities.In order to combat these malicious transactions,the key measure is to discover knowledge from existing empirical data,and then detect and even predict malicious transactions in blockchain networks.However,in the process of joint detection of malicious transaction data in the existing technology,problems such as privacy leakage of transaction feature data,contradiction between model usability and privacy,and uncontrolled collaborative modeling behavior are still exposed.First,the privacy of transaction feature data is leaked.The centralized learning mode needs to combine multi-party transaction data to perform blockchain malicious transaction detection tasks,and data collection to centralized servers faces the problem of privacy leakage.Second,the contradiction between model availability and privacy.The graph neural network model shows superiority in malicious transaction detection tasks,but the existing privacy protection technology for the graph neural network model will damage the data accuracy and reduce the availability of the model.Finally,the collaborative modeling behavior is not controlled.The behavior of data holders in the open network is not controlled,and malicious data holders can evade detection by launching poisoning attacks.In response to the above problems,this paper conducts research on blockchain malicious transaction detection technology based on privacy computing.The main contributions are summarized as follows:(1)Privacy-preserving malicious transaction detection algorithm based on differential privacy.Conventional blockchain malicious transaction detection methods adopt a centralized learning mode,and models are carried out on a centralized server by collecting multi-party user data.This method leads to the risk of privacy leakage of user transaction data.To this end,this paper proposes a privacy-preserving malicious transaction detection algorithm based on differential privacy.First,the Gaussian mechanism is used to add differential privacy noise to the transaction feature data to enhance the privacy of the transaction feature data.Second,use the Graph SAGE model to learn important features of transactions in large-scale transaction graph structure data,and capture the capital flow relationship between transactions.Finally,the experimental results on the real dataset Elliptic show that the noise-added transaction feature data can train the available models to jointly detect malicious transactions.(2)Privacy-preserving malicious transaction detection algorithm based on homomorphic encryption.Conventional graph neural network model privacy protection methods need to add differential privacy noise to the data.This kind of noise addition based on confusing ideas will damage the accuracy of the data,which will lead to a decrease in the availability of the model.To this end,this paper proposes a privacy-preserving graph neural network model based on homomorphic encryption.First,design a collaborative training mode that separates linear and nonlinear calculations,and outsource nonlinear calculations to the client to avoid model accuracy degradation caused by polynomial fitting nonlinear calculations.Secondly,a privacy protection algorithm suitable for graph neural networks is designed by adopting Paillier cryptosystem.Finally,experimental results on real datasets demonstrate that the algorithm can train highly available graph neural network models.(3)Collaborative audit mechanism based on blockchain and spectral anomaly detection.Conventional malicious transaction detection methods based on graph federated deep learning does not defend against the malicious behavior of the data holder,and the malicious data holder can launch a poisoning attack to reduce the performance of the model and evade detection.To this end,this paper proposes a collaborative audit mechanism based on blockchain and spectral anomaly detection.First,blockchain technology is used to construct a data sharing framework to ensure the auditability of data,and partitioning technique is used to improve the overall concurrency of the system.Second,combine the spectral anomaly detection model to identify and eliminate anomalous data to train a highly available model.Finally,the experimental results show that the partitioning blockchain can meet the efficiency requirements of data sharing,and the defense method based on the spectral anomaly detection model can resist additional noise attacks.