預測使用者行為以輔助身份辨識的融合統計方法

Abstract

已有許多研究提出利用生物特徵辨識技術來輔助傳統帳號密碼方式的身份認證。然而，不管是哪種生物特徵，往往都會隨著時間而有所改變，特別是行為特徵。因此，在本論文中，除了使用者敲鍵行為本身的特徵外，亦將敲鍵行為的改變趨勢視為使用者的另一項特徵，藉此預測使用者目前最可能的行為。本篇論文提出結合高斯模型、自動回歸預測模型，及統計學習理論中的隱藏馬可夫模型，以建立敲鍵行為特徵預測的統計機率模型，利用此機率模型分析使用者登入帳號密碼的時間資訊是否符合該身份擁有的特徵，藉此降低身份冒用的風險。實驗結果顯示，本論文所提出的方法可使錯誤辨識率降至2.19%，相較於過去其他相關研究（通常高於3%），甚至我們之前的研究成果（2.54%）都來得更準確。特別是當使用者敲鍵行為資料具有特定變動趨勢時，藉由行為預測可使辨識準確率有效提升。

Biometric verification mechanism has been used to complement the traditional password-based authentication system. However, the biometrics may change over time, especially for behavioral biometrics. In this thesis, the keystroke typing characteristics are not the only features used, but also the tendency of change in keystroke behavior. The most likelihood behavior of user will be predicted to verify the user. In this paper, we propose a fusion model for predictive keystroke analysis inspired by Gaussian Model, Autoregressive Model, and Hidden Markov Model. This model predicts the keying behavior of a user based on his past statistical information. Results of the experiment showed that, the EER could down to 2.19%, which is better than other works in literature to our knowledge (generally higher than 3%), and even better than our previous work (2.54%). Especially as users type with some trend or regularly, their identified accuracy could be enhanced by predicting their keying behavior.