From: Blockchain abnormal behavior awareness methods: a survey
Service behavior | Awareness difficulty | Awareness methods | Technical details | Applicable scene | Awareness accuracy | Awareness speed | Advantage | Disadvantage | Precision | Recall | F1 | Data using |
---|---|---|---|---|---|---|---|---|---|---|---|---|
Ponzi schemes | Easy | Based on supervised learning | Using XGBoost to classify the Ethereum transaction flow graphs (Chen et al. 2018) | Blockchain with sufficient transaction scale | High | Fast | High precision | Unsatisfactory recall rate | 94.00% | 81.00% | 86.00% | Ethereum transaction data |
Based on mixed information model construction | Fusion of transaction records, contract source code and on chain information to build a Ponzi scheme behavior model (Bartolettiet al. 2020) | Blockchain with sufficient transaction volume and support for smart contracts | – | – | Comprehensive analysis dimensions | Multi-dimensional data is weakly correlated | Find extra 184 ponzi schemes on Ethereum | |||||
Based on unsupervised learning | Smart contract operation code and account feature extraction based on deep neural network (Zhang and Lou 2021) | All blockchains that support smart contracts | High | Fast | High precision and recall rates | High data set dependence | 99.60% | 96.30% | 97.92% | Ethereum transaction data combined with contract code data | ||
Money laundering | Easy | Based on unsupervised learning | Compare the transaction flow graph between the normal transactions and money laundering transactions (Hu et al. 2019) | Blockchain with sufficient transaction scale | High | Medium | High precision | Unable to identify blockchain application with insufficient transaction scale | 92.74% | 97.37% | 95.00% | Bitcoin partial transaction datasets |
Based on supervised learning | Use graph neural network to identify and visualize money laundering transaction sequence diagrams (Weber et al. 2019) | Popular crypto-currency platform | High | Medium | High robustness and suitable for all kinds of popular crypto-currency platform | Unsatisfactory accuracy | 97.10% | 67.50% | 79.00% | |||
Based on active learning | Use less data to achieve recognition (Lorenz et al. 2020) | Blockchain with limited data and manpower | Medium | Medium | Independent of data set size | Robustness needs to be verified | With the number of labeled samples, the best F1: 83.00% |