智慧型手機已成為使用者日常執行各種活動的智慧行動裝置，而智慧型手機也逐漸需要受到保護以防個人資料遭到非法存取。過去提出使用兩階段認證：傳統認證與行為認證來加強安全性，然而在行為認證方面，使用者於不同的使用情境時常有不同的操作行為，隨著時間推移，資料逐漸變得不可靠，且因模型無法應對使用行為改變而使合法使用者進行認證時失敗。因此，使用者行為模型必須隨時重新訓練以解決這個問題。 重新訓練模型之方法會使用所有過去模型的資料來訓練新的模型，但這會造成程序耗竭且佔用大量資源。因此，我們提出一種輕量且快速的分類器：具有池化策略的簡化線性判別分析，在快速的訓練時間且不犧牲等錯誤率的狀況下來解決使用者行為改變的問題。在這份研究中，我們比較四種方法來評估我們所提出的模型，分別是M0-B SVM當作最佳基準模型、無重新訓練的M0-W1 LDA與M0-W2 SVM當作最差基準模型、以及具有池化策略的M1-Simplified LDA為我們提出的模型。實驗結果顯示我們提出的模型M1的等錯誤率為0.085，比經6次迭代的M0-W1的0.146及M0-W2的0.104還要好，而M0-B的等錯誤率為0.043。 不過M1僅有0.044秒的訓練時間及，0.029秒的測試時間，快於其他三種發法。M0-W1的訓練時間為0.344秒，M0-B的訓練時間為0.614秒，而M0-W2的測試時間為0.074秒。 ;Smartphones are one of the smart devices that become a daily driver for users to perform various activities that has to be secured on their smartphone to protect these personal data from unauthorized access. Two-step authentication using traditional and behavioral authentication was proposed to strengthen the security. However, in behavioral authentication, users tend to use their smartphone on different posture on different activities, it will make user data unstable and change over time and will cause the authentication process failed to predict as a legitimate user because the previous authentication model cannot handle user behavior change. Thus, the user model has to be retrained over time to handle the problem. Retraining approach keeps all previous data to retrain the new model, it will cause the exhausting process and will take a large number of resources. Hence, we proposed a lightweight and fast classifier using Simplified LDA with pooling strategy to handle user behavior change with faster training time without sacrificing the Equal Error Rate (EER). In this research, we compared four approaches to evaluate the proposed model, M0-B SVM as the best baseline model, M0-W1 LDA and M0-W2 SVM non-Retraining as the worst baseline model, and M1-Simplified LDA with pooling strategy as our proposed model. The experiment result shows that the proposed model M1 has EER 0.085, better than M0-W1 which only got EER 0.146 and M0-W2 got 0.104 in the 6th iteration and approached M0-B EER 0.043. M1 also only need 0.044 seconds on training time and 0.029 seconds on testing time, faster than M0-W1 0.344 seconds and M0-B 0.614 seconds training time and M0-W2 0.074 seconds on testing time.
