| dc.description.abstract | The traditional authentication mechanisms such as Personal Identification Number (PIN) and password had proven to be insecure as they can be cracked or duplicated easily. To address this issue, the advancement of research on digital security mechanism has seen an increasing interest on the implementation of biometric security system. The objective of biometric-based authentication mechanism is to utilize either physical or behavioral feature to authenticate the genuine users. Even though most biometric authentication systems use physical data as the feature, the physical feature is still prone to the spoofing attack. In this work, a smart-phone authentication system that uses biometric data, specifically behavioral data, to authenticate user is proposed. Using behavioral data as a feature gives the benefit that the authentication key will adapt to the user characteristic so the users do not need to remember the authentication key. The behavioral feature is also difficult to be spoofed and imitated. Moreover, the proposed authentication system is designed based on the non-intrusive approach. The proposed non-intrusive authentication mechanism offers a better solution as the system monitors user′s behavior in the background.
Generally, most researches in the biometric security mechanism use batch learning approach to train authentication system. In this work, online learning approach is adopted to offer a novel method to build the user model over a period of time. Not only mimics the real world situations, online learning approach offers a faster training time when it is compared to batch learning. In the experiment results, the online learning approach has an around 75\% faster training time compared with the previous works in this field.
Another benefit of online learning approach is the capability of the model to be adaptive to user behavior change. In the experiment results, the online learning approach shows lower Equal Error Rate (EER) than the model that is never updated. The online learning approach also shows the model improvement over period of time. Comparing the result of the updated models in online learning approach over a period of time, the current user model has a lower EER than the previously updated model. In this work, the performance of proposed online learning approach is also compared to batch learning approach using box plot statistical analysis. From the experiment results, it showed that the first quartile of proposed work is better than the batch learning and the proposed work has median value near to the result of batch learning. | en_US |