| 摘要: |
| 针对加密货币兑换阶段的洗钱交易活动，本文提出多特征自适应融合的洗钱交易检测模型FTT-MLD，旨在解决链上交易数据的深层特征提取与融合问题。首先，设计基于FT-Transformer的深层特征提取模块，通过特征标记和自注意力机制有效捕捉交易数据中潜在的复杂模式和高阶依赖关系，显著增强模型对复杂洗钱行为的感知能力；其次，将浅层特征与深层特征横向拼接后输入多层感知机（Multi-Layer Perceptron, MLP）进行加权融合，有效缓解简单拼接带来的特征冗余、冲突和噪声问题，并通过全连接层实现洗钱交易识别任务。实验结果显示，FTT-MLD在比特币和以太坊两种真实交易场景下，分别较已有最佳模型提升了约0.03%–2.4%和0.05%–1.1%，表明所设计的特征提取与融合模块能够有效表征洗钱交易的行为模式和特征，从而实现精准检测。进一步，为增强模型可解释性并验证关键特征的有效性，引入SHAP框架对以太坊交易场景下账户行为进行量化分析，揭示特征对模型预测结果的边际贡献及其交互作用。结论表明，账户主要发送的ERC20代币类型、账户最小发送金额、账户首次与最后一次交易时间间隔、账户接收独立来源地址交易的数量等特征在洗钱交易识别中具有显著判别力，该结论为模型在多维交易数据环境下的有效性提供实证支持，并为金融监管与风险预警机制的优化提供理论基础。 |
| 关键词: 加密货币 洗钱检测 FT-Transformer 多特征融合 关键特征 |
| DOI： |
| 投稿时间：2025-07-18修订日期：2026-04-30 |
| 基金项目:国家自然科学基金 |
|
| FTT-MLD: A Cryptocurrency Money Laundering Detec-tion Method Based on FT-Transformer |
|
lujiarong1, Liaobin2, Liuyi3
|
| (1.Xinjiang University of Finance and Economics;2.Guizhou University of Finance and Economics;3.Xinjiang Medical University) |
| Abstract: |
| To address the money laundering transactions occurring during the cryptocurrency exchange phase, this paper proposes a multi-feature adaptive fusion model for money laundering transaction detection, termed FTT-MLD, which aims to solve the problems of deep feature extraction and fusion from on-chain transaction data. First, a deep feature extraction module based on the FT-Transformer is designed. Through feature tokenization and a self-attention mechanism, it effectively captures potential complex patterns and high-order dependencies in the transaction data, significantly enhancing the model's ability to perceive sophisticated money laundering behaviors. Second, shallow features and deep features are horizontally concatenated and then fed into a Multi-Layer Percep-tron (MLP) for weighted fusion, effectively mitigating the feature redundancy, conflict, and noise caused by simple concatenation. The money laundering transaction identification task is then accomplished through a fully connect-ed layer. Experimental results demonstrate that, in two real-world transaction scenarios—Bitcoin and Ethere-um—FTT-MLD outperforms existing state-of-the-art models by approximately 0.03%–2.4% and 0.05%–1.1%, re-spectively, indicating that the proposed feature extraction and fusion modules can effectively characterize the be-havioral patterns of money laundering transactions, thereby enabling accurate detection. Furthermore, to enhance model interpretability and validate the importance of key features, the SHAP framework is introduced to quantita-tively analyze account behaviors in the Ethereum transaction scenario, revealing the marginal contributions and interactions of features on model predictions. The findings indicate that features such as the primary type of ERC20 tokens sent by the account, the minimum transaction amount sent by the account, the time interval between the ac-count's first and last transactions, and the number of transactions received by the account from distinct source ad-dresses exhibit significant discriminative power in identifying money laundering transactions. These conclusions provide empirical support for the model's effectiveness in a multi-dimensional transaction data environment and offer a theoretical foundation for optimizing financial regulatory and risk early warning mechanisms. |
| Key words: Cryptocurrency Money laundering detection FT-Transformer Multi-Feature fusion key feature |
