臺灣博碩士論文加值系統

隨著線上購物盛行，盜刷事件也隨之增加，根據統計，臺灣每年盜刷金額將近二十億。為防止盜刷問題持續發生，近年來有許多研究利用機器學習方法來偵測信用卡詐騙。這些方法遇到的挑戰之一，即是信用卡交易資料為不平衡資料集，詐欺與正常交易的比例相差懸殊。面對不平衡資料集，多數研究使用Undersample，將正常交易筆數減少，使詐騙比例相對提高。此外機器學習需先進行資料前處理，挑選有利於訓練模型的特徵，但由於信用卡交易資料具有許多個人隱私資料，因此取得的資料集欄位可能已經過保密處理，對於機器學習建模有一定的困難度。然而建立深度學習模型前，不須挑選特徵，可使用所有資料欄位建模，對於處理如信用卡交易等具有個人隱私資料建模具有優勢。
本研究中，先將取得的信用卡交易資料進行Undersample，調整詐騙交易與正常交易的比例為50%, 15%, 10%, 5%, 1%來進行比較。由於資料中的欄位名稱除了交易時間與交易金額外，都已經被處理過無法辨識，因此我們隨機挑選資料欄位作為特徵，使用Logistic Regression、Random Forest、Support Vector Machine三種機器學習建模。此外，我們用Keras建立深度學習模型，建立一個輸入層、三個隱藏層的模型，並進行200次訓練週期。在評估預測結果方面，考慮Accuracy，Precision，Recall，F1-Score，以及Matthews correlation coefficient五項具代表性的指標。實驗的結果顯示，深度學習方法在五種不同比例的模型下所得到的預測結果都優於機器學習方法。

With the popularity of online shopping, cases of transaction fraud are also increasing. According to statistics, the amount of losses from credit card fraud is nearly 2 billion annually in Taiwan. In order to prevent fraud events, many studies have used machine learning methods to detect credit card fraud in recent years. One of the challenges of these methods is that the credit card transaction data is highly imbalanced, fraudulent transactions are largely outnumbered by genuine ones. A common strategy for dealing with the problem is to under-sample the legit transaction class in the training set. In addition, machine learning algorithms need to perform data pre-processing first, and select features that are beneficial to the training process. However, due to confidentiality issues, a publicly available credit card data set is usually encrypted or transformed which make it difficult to select proper attributes to train the classifier algorithms. However, there is no need to select features while establishing a deep learning model. The deep learning approach has advantages for processing confidential data such as credit card transactions.
In this study, the credit card transaction data obtained was first under-sampled and the ratio of fraud transactions to legit ones was adjusted to 50%, 15%, 10%, 5%, and 1% for comparison. Since the attribute names in the data set have been transformed except for transaction time and transaction amount. Therefore, we randomly selected the attributes to investigate the performance of Logistic Regression, Random Forest, and Support Vector Machine. In addition, we used Keras to build a deep learning model which includes an input layer and three hidden layers, and conducted 200 times of training. Five metrics were used to report the performance of fraud detection classifiers including Accuracy, Precision, Recall, F1-Score, and Matthews correlation coefficient. The experimental results show that the deep learning method, in most circumstance, is outperform the machine learning methods.

目錄
致謝i
摘要ii
Abstractiii
目錄 v
圖目錄 vi
表目錄 vii
第一章 緒論 1
11 研究動機 1
12 研究分析 1
13 論文架構 3
第二章 文獻探討 4
21 處理不平衡資料集之方法 4
22 偵測信用卡詐騙相關研究 5
23 深度學習簡介 6
第三章 研究方法與實驗 7
31 Undersample 7
32 三種機器學習方法 8
33 深度學習 – Keras 9
第四章 實驗結果 10 41 評估指標 10
42 預測結果 11
第五章 結論 18
參考文獻 19

圖目錄
圖 1 Keras 多層感知器模型 9
圖 2 機器學習的 Accuracy 12
圖 3 機器學習的 Precision 12
圖 4 機器學習的 F1-Score 13
圖 5 機器學習的 Recall 13
圖 6 機器學習的 MCC 14
圖 7 Keras 與 RF 的 Accuracy 14
圖 8 Keras 與 RF 的 Precision 15
圖 9 Keras 與 RF 的 Recall 15
圖 10 Keras 與 RF 的 MCC 16
圖 11 Keras 與 RF 的 F1-Score 16
表目錄
表 1 訓練資料集與測試資料集 7
表 2 Confusion Matrix 10

1.Sebastian Raschka. (2017)．Python機器學習(初版九刷) (劉立民、吳建華)．臺灣:博碩文化股份有限公司．(原著出版年:2015)。
2.林大貴。(2017)．TensorFlow+Keras深度學習人工智慧實務應用(初版)．臺灣:博碩文化股份有限公司。
3.尼爾森：不到四成台灣消費者每天消費偏好塑膠貨幣付款。Retrieved from http://www.nielsen.com/tw/zh/press-room/2014/newsTaiwanPayment0225.html
4.金融監督管理委員會 - 金融監督管理委員會-防範信用卡網路交易盜刷機制。Retrieved from https://www.fsc.gov.tw/ch/home.jsp?id=96&parentpath=0,2&mcustomize=news_view.jsp&dataserno=201802130007&aplistdn=ou&dtable=News
5.財團法人聯合信用卡處理中心公開資料 - 各發卡機構通報之詐欺金額。Retrieved from https://www.nccc.com.tw/wps/wcm/connect/zh/home/openinformation/Business?WCM_PI=1&WCM_Page.bcdb178c-e429-4c73-b2bb-43c4fbe133e5=1
6.Fighting Financial Fraud with Targeted Friction – Airbnb Engineering & Data Science – Medium. Retrieved from https://medium.com/airbnb-engineering/fighting-financial-fraud-with-targeted-friction-82d950d8900e
7.Awoyemi, J. O., Adetunmbi, A. O., & Oluwadare, S. A. (2017). Credit card fraud detection using machine learning techniques: A comparative analysis. 2017 International Conference on Computing Networking and Informatics (ICCNI).
8.Bahnsen, A. C., Stojanovic, A., Aouada, D., & Ottersten, B. (2013). Cost Sensitive Credit Card Fraud Detection Using Bayes Minimum Risk. 2013 12th International Conference on Machine Learning and Applications.
9.Bhattacharyya, S., Jha, S., Tharakunnel, K., & Westland, J. C. (2011). Data mining for credit card fraud: A comparative study. Decision Support Systems,50(3), 602-613.
10.Kho, J. R., & Vea, L. A. (2017). Credit card fraud detection based on transaction behavior. TENCON 2017 - 2017 IEEE Region 10 Conference.
11.Pozzolo, A. D., Caelen, O., Johnson, R. A., & Bontempi, G. (2015). Calibrating Probability with Undersampling for Unbalanced Classification. 2015 IEEE Symposium Series on Computational Intelligence.
12.Randhawa, K., Loo, C. K., Seera, M., Lim, C. P., & Nandi, A. K. (2018). Credit Card Fraud Detection Using AdaBoost and Majority Voting. IEEE Access,6, 14277-14284.
13.M. Fahmi, A. Hamdy and K. Nagati, "Data Mining Techniques for Credit Card Fraud Detection: Empirical Study," Sustainable Vital Technologies in Engineering & Informatics, pp. 1-9, 2016.