Research On Ethereum Scam Accounts Detection Method Based On Machine Learning

Blockchain is an emerging technology with the advantages of decentralization,tamper-proof and traceability,which has been absorbing the attention of many researchers since it came into being.Ethereum,as a representative of blockchain 2.0,not only brings convenience to users,but also has hidden dangers of fraudsters using the anonymity of blockchain to carry out fraudulent activities,such as scam externally owned accounts and scam contracts,which may bring immeasurable economic losses to investors.At present,there are few fraud detections researches on the Ethereum blockchain in the world,and the existing detection technologies to identify fraudulent behaviors are less than satisfactory.Therefore,this thesis uses machine learning technology to propose two methods detecting typical fraud accounts and contracts based on Ethereum dataset,so as to protect users' property and promote the healthy development of blockchain ecosystem.The main research works are as follows:(1)A fraudulent account detection method for external fraud account is proposed based on Light GBM.Considering current fraudulent account detection methods have low detection rate and poor performance,we collected and labeled 2223 fraudulent accounts and 5805 non-fraudulent accounts after retrieving transaction and block data from Etherscan.Then we further combined 14 handcrafted features extracted based on transaction history information with 86 statistical features automatically extracted by featuretools.Those features describe the user's transaction behavior in a more comprehensive way.With the help of Light GBM model,our experiments showed that the F1 score reached 94.92%,was higher than the detection result of using only a single feature and can detect scam accounts with better efficiency and effectiveness.(2)A scam contract detection method for smart contract is also proposed based on attention capsule network.In most cases,deploymenting a smart contract needn't source code's supporting,so it is difficult to detect its source code features.In order to describe a smart constract behavior,this thesis collected and downloaded the bytecode and application binary interface(ABI)required for smart contract deployment,then extracted the sequence,the frequency and call relationship to describe the behavior of them,and finally converted those into RGB images.To reduce the impact of data imbalance on the detection work,Fancy PCA was used for data augmentation,and the processed RGB images were inputted to the attention capsule network(SE-Caps Net)model.Compared with the existing research work,the F1 score of our method reached 98.38% is higher than existed methods and our method could effectively detect fraudulent contracts during the contract deployment stage.(3)Feature analysis of ethereum scam accounts is shown.According to the two detection methods proposed above,the importance of account and contract features were evaluated through experiments.It is found that the transaction time interval,total transaction value,and number of transactions are the important factors for distinguishing fraudulent accounts,and the frequency of opecode has a great impact on scam contract detection.In addition,the case analysis on Kaggle fraud datasets and honeypot contract datasets were conducted to verify the effectiveness of the proposed methods,and the experimental results showed that these detection methods were not only suitable for a specific type of fraud,but also can achieve good results in other fraud detection tasks.