Incremental Knowledge Distillation for Money Laundering Detection under Blockchain-based Financial Networking

The pseudo-anonymous nature of blockchain transaction systems facilitate illicit money laundering activities, threatening financial network information security. Existing anti-money laundering methods predominantly employ offline learning, rendering them unable to adapt to evolving money laundering patterns. While incremental learning approaches enable continuous model updates, they remain susceptible to catastrophic forgetting. To address these challenges, we propose an online dual-teacher knowledge distillation method, named IKD-BMD, to detect money laundering. Through joint guidance from both benchmark and dynamic teachers to the student model, this approach achieves the dual objectives of preserving existing knowledge and adapting to novel money laundering patterns without requiring access to historical data. Additionally, a periodic reverse distillation mechanism promotes teacher-student collaborative evolution, enabling continuous learning capabilities. Experimental results demonstrate that the proposed method achieves an F1 score of $\mathbf{8 6. 1 3 \%}$, confirming its significant detection efficacy in dynamic blockchain environments.