【Author】
Sun, Xiaowen; Yang, Tan; Hu, Bo
【Source】APPLIED INTELLIGENCE
【Abstract】Coin mixing is a class of techniques used to enhance Bitcoin transaction privacy, and those well-performing coin mixing algorithms can effectively prevent most transaction analysis attacks. Based on this premise, to have a well-functioning transaction analysis algorithm requires coin mixing detection with a high recall to ensure accuracy. Most practical coin mixing algorithms do not change the Bitcoin protocol. Therefore, the transactions they generate are not fundamentally different from regular transactions. Existing coin mixing detection methods are commonly rule-based that can only identify coin mixing classes with well-defined patterns, which leads to an overall low recall rate. Multiple rules could improve the recall in this situation, yet they are ineffective for new classes and classes with ambiguous patterns. This paper considers coin mixing detection as a transaction classification problem and proposes an LSTM Transaction Tree Classifier (LSTM-TC) solution, which includes feature extraction and classification of Bitcoin transactions based on deep learning. We also build a dataset to validate our solution. Experiments show that our approach has better performance and the potential for discovering new classes of coin mixing transactions than rule-based approaches and graph neural network-based Bitcoin transaction classification algorithms.
【Keywords】Bitcoin; Coin mixing detection; CoinJoin; LSTM; Transaction tree; Bitcoin address clustering
【标题】LSTM-TC:高召回率的比特币混合检测方法
评论