近年來，智慧型手機成為彼此生活中最為仰賴的產品，且常儲存個人隱私資訊，因而成為惡意人士的竊取目標之一。雖然目前有提供身分驗證機制為用戶進行身分認證，但多數方法皆屬於一次性驗證，也就是用戶只需要在使用裝置的初始階段通過認證，就可以隨意存取智慧型手機內的任何資料。本研究所採用的連續式身分驗證(Continuous Authentication) 將彌補過往身分驗證機制的不足，其目的為在不打擾用戶的使用情況下，持續收集用戶的行為資料，並定期地在後台進行認證，確保使用者為合法用戶，即行為生物識別(Behavioral Biometrics)。然而，過往研究指出，基於行為的身分驗證系統會遭受到對抗式攻擊(Adversarial Attack) 的影響，也就是假冒攻擊(Impersonation Attack)，惡意人士將偽裝成合法使用者，試圖讓系統誤判，造成系統穩健性(Robustness) 下降。因此，本研究將針對假冒攻擊問題提出兩項能夠提升基於行為身分驗證系統穩健性的方法。即基於行為相似度的分群方法和基於行為變異度的前處理方法，其中分群方法將根據用戶行為相似度來分組，並在移除系統事先指定給群組的容易被攻擊者模仿的群組總體弱特徵後，再進行模型訓練；而前處理方法則基於用戶每項特徵的行為變異度來調整每項特徵資料的權重，使得系統能夠更加容易地分辨合法用戶和攻擊者。透過本研究提出的分群和前處理方法，將可以在不需要為用戶收集攻擊者的模仿資料來執行耗時的篩選弱特徵過程的前提下，達到更好的準確率。;In recent years, smartphones have become essential in our lives, often storing personal information and making them prime targets for malicious people. Current identity verification mechanisms provide authentication, but most are one-time verifications, allowing users to freely access data after initial authentication. This study addresses these shortcomings by using continuous authentication, which continuously collects user behavior data without disturbing the user and performs regular background authentication, known as behavioral biometrics. However, behavior-based authentication systems are vulnerable to adversarial attacks, or impersonation attacks, where malicious individuals pretend to be legitimate users, reducing system robustness. To enhance robustness against these attacks, this study proposes two methods: a similarity-based clustering method and a preprocessing method based on behavioral variability. The clustering method groups users by behavioral similarity and removes weak features that attackers can mimic before model training. The preprocessing method adjusts feature weights based on behavioral variability, making it easier to distinguish between legitimate users and attackers. These methods aim to improve accuracy without the need for collecting attackers’ data for weak feature selection.
