- 學位論文
改良信任模型進行滑鼠動態之連續認證
Continuous Authentication based on Mouse Dynamics with an Improved Trust Model
摘要
現今我們主要使用密碼、指紋、臉部特徵進行認證，一旦登入電腦，身份在會話中永久有效，直到登出，這種方式稱為靜態認證 (static authentication)。然而，若透過靜態認證登入後用戶因故離開且未上鎖電腦，此時入侵者就可以操作其來竊取資料。連續認證 (continuous authentication) 可以有效減緩這個問題。它在會話期間持續監控當前用戶行為，並在偵測到仿冒者時將其鎖定。 基於滑鼠動態的連續認證系統根據當前用戶的每個滑鼠動作判斷其為合法用戶或仿冒者，並透過信任模型 (trust model) 計算出信任度的獎勵或懲罰。在本文中，我們提出一種新的信任模型，當仿冒者動作連續出現時動態地增加對信任度的懲罰，並在合法動作出現時重設懲罰。我們使用隨機森林分類器和所提出的信任模型來建構一個連續認證系統，在公開的滑鼠資料集上以仿冒者動作的平均數量 (average number of impostor action, ANIA) 與合法動作的平均數量 (average number of genuine action, ANGA) 指標來評估其表現。結果顯示，使用我們提出的信任模型可以更快鎖定仿冒者，達到更高的安全性。
並列摘要
Nowadays, we mainly use passwords, fingerprints, and facial features for authentication, once logging in to the computer, the identity is valid during the session until logout, which is called static authentication. However, if a user logs in through static authentication and leaves the computer unlocked for some reason, an intruder can then operate it to steal data. Continuous authentication can mitigate this problem effectively. It monitors user’s behavior during a session and locks out impostors when they are detected. The continuous authentication system based on mouse dynamics determines whether the current user is a genuine user or an impostor for each single mouse action, and calculates reward and penalty for the trust level by trust model. In this thesis, we propose a new trust model that dynamically increases the penalty for trust level when multiple actions are continuously classified as impostor actions and resets the penalty once genuine action occurs. We use a random forest classifier with the proposed trust model to construct a continuous authentication system and evaluate its performance in terms of metrics ”average number of impostor actions (ANIA)” and ”average number of genuine actions (ANGA)” on public mouse dataset. The result shows that using our trust model, continuous authentication system can lock out impostors more quickly and therefore achieve higher security protection.