隨著線上拍賣的盛行，相關詐騙事件也開始層出不窮。學者們對此紛紛提出各種不同的詐騙偵測方法，期能協助消費者安心使用便利的網路拍賣平台。他們提出的方法雖然有效，但似乎只針對現行犯，對於潛伏期的詐騙者則無法有效處理。但就詐騙防治而言，事發於未然最為重要，等受害者真正出現才發出警訊，恐怕事倍功半。有鑑於此，本研究針對潛伏期詐騙者，發展了一套有效的偵測方法。首先，為了增加偵測結果的準確性，我們提出了一套新的分類屬性集。接著，為了分辨潛伏期的詐騙者，本研究以終點回溯的方式來切割交易歷史紀錄，擷取詐騙者在潛伏時期的特徵。為進一步提升偵測效能，本研究更發展了一套二階段篩選方法，結合不同的回溯模型進行連續過濾。我們由Yahoo! Taiwan蒐集實際的交易資料進行實驗。由實驗結果可知，配合回溯切割方法，本研究提出之屬性集所建立之偵測模型的召回率與精度均超過85%，優於前人研究。當使用二階段偵測流程時，精度與召回率更可提升至89%以上。上述結果顯示，本研究提出之方法對於潛伏詐騙者確有良好的偵測能力，若能實際應用在網路拍賣上，必能協助消費者避開詐騙，安心進行交易。 The online auction fraud is becoming a serious problem in recent years. In Yahoo!Taiwan, about 12,000 fraud cases have been reported in the past two years. Researchers have proposed a lot of useful detection methods to help the trader in avoiding online auction frauds. However, these methods proposed in the related work do not consider an important issue in fraud detection, that is, they are not designed for detecting latent fraudsters. In view of fraud prevention, we need recognize the fraudsters before they become fraudster actually. To detect latent fraudsters effectively, this study first proposed a set of new attributes to model the fraudsters. Then, an endpoint-backtrack method is developed to build detection models for latent fraudsters. In addition, a two-phased detection flow is designed to improve the overall accuracy. To demonstrate the effectiveness of the proposed method, we collect real transaction data from Yahoo! Taiwan auction sites and conduct a series of experiments. The results show that, in comparison with other work, our method can provide better precision and recall rate for latent fraudster detection.
