| 摘要: | 基於比特幣的匿名性與去中心化特點，许多政府和監管機構一直對其持谨慎態度。未来，監管可能會更加嚴格。台灣金管會也計畫於2023年5月訂定台灣虛擬貨幣管理辦法。
對照傳統貨幣，金融機構必須符合國際法規，以確保不向罪犯和恐怖分子提供服務。他們還需要持續監控金融交易以發現可疑行為活動。這些金融機構有許多用來監控和驗證客戶的信息的作業程序，以確認客戶的真實身份。未能檢測到異常交易將導致金融機構造成的嚴重的後果，視情況嚴重程度而定，給予相關機構警告或罰鍰。因此，大部分金融機構使用反洗錢（Anti-money laundering，AML）解決方案進行制裁及觀察名單過濾和篩選，以監控金融網絡內的每筆交易，以確保沒有任何交易可以用於與被禁止的人做生意。近期，金融界和學術界一致認為機器學習可能對交易監控產生重大影響。
因此，本研究採用Kaggle上的比特幣異常交易資料集，進一步探討在比特幣匿名交易的特性下，各種機器學習演算法，包含隨機森林（Random Forest）、邏輯回歸（Logistic Regression）、增強型梯度提升（XGBoost）、梯度提升（Gradient Boosting）與支持向量機（Support Vector Machine）等，對異常交易監控的效率，同時，也因為該資料集的特徵皆做過前置處理，所有特徵名稱皆匿名，故希望透過資料導向方法，以特徵選取方式挑選出對異常交易偵測有較顯著影響的特徵集合。
本研究實驗結果顯示機器學習演算法中，以增強型梯度提升演算法所建立之模型的效率為最佳，隨機森林演算法次之。特徵選取實驗中以交易本身特徵值及交易鄰居節點特徵值等兩個特徵集合對模型效率之影響最為顯著。
;Due to the anonymity and decentralization of Bitcoin, many governments and regulatory agencies have been cautious about it. In the future, regulation may become stricter. The Financial Supervisory Commission of Taiwan also plans to establish a virtual currency regulatory authority in Taiwan in May 2023.
Compared with traditional currencies, financial institutions must comply with international regulations to ensure that they do not provide services to criminals and terrorists. They also need to continuously monitor financial transactions to detect suspicious activities. These financial institutions have many operational procedures to monitor and verify customer information to confirm the real identity of customers. Failure to detect illegal transactions will lead to serious consequences for financial institutions, warnings or fines will be given to relevant institutions depending on the severity of the situation. Therefore, most financial institutions use Anti-Money Laundering（AML）solutions for sanctions and watchlist filtering and screening to monitor every transaction within the financial network to ensure that no transaction can be used to do business with prohibited persons. Recently, the financial community and academia unanimously believe that machine learning may have a significant impact on monitoring.
Therefore, this study uses the Bitcoin abnormal transaction dataset on Kaggle to further explore various machine learning algorithms under the characteristics of Bitcoin anonymous transactions, including Random Forest, Logistic Regression, XGBoost, Gradient Boosting, and Support Vector Machine, etc., for the efficiency of abnormal transaction monitoring. At the same time, because the features of this dataset have been pre-processed, all feature names are anonymous, so it is hoped to select a feature set that has a more significant impact on abnormal transaction detection through data-driven methods.
The experimental results of this research show that the efficiency of the model established by the XGBoost algorithm is the best, followed by the Random Forest algorithm. In the feature selection experiment, the transaction features and aggregation features have the most significant impact on the efficiency of the model. |
