Detection of Blockchain Online Payment Fraud Via CNN-LSTM

Criminal activities such as money laundering, scams, and even terrorist financing are growing common targets for malicious hackers exploiting blockchain-based payment systems, e.g., involving cryptocurrencies such as Bitcoin. Distinguishing between legitimate and suspicious transactions in such activities is difficult without generating many false positives, which is even more complicated with imbalanced datasets and when graph-like structures (e.g., transaction flows) are involved. Existing techniques are not able to effectively and conveniently capture the (sequential and local) patterns of transaction flows.We introduce a hybrid CNN-LSTM model which contains three main building blocks: CNN for mining local features from transaction graphs, LSTM for risk evaluation through a sequence of anomaly forecasting and finally an economic minimization module to realize the minimization of economic losses. Experiments on the Elliptic Bitcoin Dataset with 2% labels for illicit activities among 203,769 transactions demonstrate that our model decreases the expected loss by 45% compared to the rule-based benchmark model while keeping the rate of false positive risks below 0.5% .As such, we are able to seamlessly integrate detection, economic modelling and risk together in the model and show that it is possible to create a tool for real world security analysis of blockchains. CCS CONCEPTS Mathematics of computing ~Probability and statistics ~Probabilistic inference problems