Detecting Mixing Services in Bitcoin Transactions Using Embedding Feature and Machine Learning

Due to the intrinsic anonymity of the Bitcoin network, it is frequently subject to money laundering activities, many of which involve coin mixing services. Coin mixing serves as a technique to obscure the trajectory of funds and is inherently dynamic and complex. Such dynamism results in inconsistent transaction behavior across numerous addresses. Moreover, the attributes and patterns associated with coin mixing fluctuate over diverse temporal spans, complicating its detection. To address this challenge, our research introduces a novel classification technique. Initially, we classify coin mixing addresses by their unique transaction behaviors, capturing a variety of patterns. We then determine the influence scope of each address category within its respective subnetwork, leading to the development of dynamic embedding features. These features adapt in real-time, reflecting the behavioral patterns of the current category. In our final step, we use a multi-classifier ensemble strategy, integrating both embedding features and topological structure insights, thereby bolstering classification precision and resilience. Our experimental findings highlight significant improvements across diverse evaluation benchmarks. Of special note is the efficacy resulting from merging dynamic embedding features with the multi-classifier ensemble methodology, achieving an impressive precision rate of up to 0.9747. This study provides fresh perspectives and approaches to mitigate the challenges of money laundering within the Bitcoin network.