Research On Virtual Currency Criminal Account Recognition Based On Graph Representation Learning

Because virtual currency transactions do not require real-name registration and have no national boundaries,more and more criminal gangs use virtual currency for funds transferring to avoid bank supervision and transnational money controls.The most used virtual currency in current criminal activities is USDT-TRC20 issued in TRON,so mining the suspect’s account from the transaction data in TRON is of great significance in combating the virtual currency crime.Existing research mainly focuses on Bitcoin ransom address identification and Ethereum phishing fraud account detection.The transaction network constructed by existing research methods is a homogeneous graph with only a single account type and a single transfer type.The graph neural network model designed for homogeneous graphs ignores characteristics such as account type,transaction type,and transaction pattern.The prediction results are mostly anonymous accounts without associated registration information,and cannot be used as valid clues.In addition,the prediction results of existing research methods are lack of interpretability,which limits their application in actual case handling scenarios.In response to the above issues,this paper constructs a crime dataset in TRON,and proposes a virtual currency criminal account identification scheme to help the police seek clues to crime.The specific work is as follows:(1)This thesis propose an account classification model based on transaction relation convolution network,which models transaction relation features such as counterparty,transaction frequency,transaction time distribution,and transfer type.Virtual currency accounts are divided into four types: anonymous accounts,contract accounts,exchange deposit accounts,and personal deposit accounts.Experiments have proven the effectiveness of the model for modeling transaction relation features,and compared to the existing best classification methods,the average F1-score has increased by 0.0473.(2)This thesis proposes a fund transfer account identification model based on transaction hash subgraph representation learning,which identifies fund transfer accounts from anonymous accounts.By constructing the transaction hash subgraph for the account and learning its representation,the transaction pattern features hidden in the transaction network are mined,and the probability of the account belonging to a criminal fund transfer account is predicted based on the transaction pattern features of the account.Experiments have proven the effectiveness of constructing the transaction hash subgraph to model transaction pattern features.In the task of identifying funds transfer accounts,this model improves the F1-score by 0.1137 compared to the best classification method based on transaction relation features modeling.(3)This thesis proposes an interpretable suspect account inference model based on Knearest neighbor,which infers suspect accounts from personal deposit accounts.By using transaction relation features and transaction pattern features,calculate the transaction network subgraph structure and transaction mode similarity between accounts,to find K accounts that are most similar to the target account,and calculate the probability that the target account belonging to suspect account based on the proportion of suspect accounts.Provide an explanation for the prediction results by showing the structural comparison of the transaction subgraph between the target account and the suspect account,as well as the probability that the counterparty belongs to a fund transfer account.The prediction and interpretation ability of the model was verified through experiments.