目前傳統的智慧型手機識別機制使用PIN碼、密碼和生物識別方法。問題是已解鎖的智慧型手機會保持解鎖狀態直至再次被鎖定。如此一來，會有一段時間可以使攻擊者竊取未鎖定的手機並竊取手機上的所有資料。連續識別方法需要為智慧型手機提供雙重安全性。非侵入式生物識別機制較好是因為使用者不會對識別有所感覺。然而，使用者行為會隨時間而變化，這會使預測不準確。每隔一段時間重新訓練是一種方法，但會花費許多時間。我們提出三種方法來建立一個能夠適應行為變化而不需重新訓練的模型。第一個方法是條件式領域適應。第二個方法是適應的最近質心分類器。第三個方法是整體學習。 最後，在我們完成所有的實驗後，我們得到的結論是三種方法都有效的處理行為變化，因為他們的EER比較差的基準模型(領域適應)更好，訓練時間比最佳的基準模型(重新訓練)更短。最有效的方法是第二種方法，適應的最近質心分類器，它在提出的三種方法中有最佳的EER及最短的訓練時間，EER平均為2.55%。 ;The current traditional smartphone authentication mechanism uses PIN, password, and biometric-based method. The problem is, the smartphones that were unlocked will stay unlock until it is actively locked again. There always exists a time frame when an attacker can steal the unlocked phone and steal all data on the device. The continuous authentication method needs to give double security to the smartphone. Implicit authentication method behavioral biometric-based is more comfortable because the user will not realize the authentication phase. But, human behavior changes over time that will make the prediction not accurate. Retraining over time can be a solution, but it will take much time. We propose three approaches to build a model that can adapt to behavior changes without retraining. The first approach is conditional domain adaptation. The second approach is the adaptive nearest centroid classifier. And the third approach is an ensemble model. Finally, after we did the experiment, we conclude that all approach in this work effective to handle behavior changes because has better EER than worse baseline (domain adaptation) and shorter training time than best baseline(retraining). The most effective approach was the second approach, adaptive nearest centroid classifier, that have the best EER and shortest training time among the proposed approach in this work with an average EER 2.55%.
