Avaliação de features de localização para autenticação transparente e contínua: processo e estudo de caso

• Ano de defesa: 2016
• Autor(a) principal: Correia, Márcio André Souto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Autenticação; Biometria; Avaliação de desempenho; Aprendizado de máquina; Computação móvel
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/27526

Given today’s growing number of devices around users, and, at the same time, their faster and frequent interactions with these devices, new security mechanisms have emerged aiming at reducing the time spent by users with authentication as well as raising the security level related to identity verification. In this sense, there are several proposals in the literature with transparent and continuous authentication mechanisms that combine biometric data retrieved from actions that users already do while using mobile devices (e.g. location, screen touch, keystroke, gait, voice, among others). In the literature review performed in this work were found nine proposals that use outdoor location and merge other kinds of biometric features as input to their proposed authentication mechanism. These proposals have in common not only the use of outdoor location but they also fail to evaluate properly each biometric features set individually. Therefore, this work provides a new process for evaluation of biometric features by adapting guidelines of machine learning to perform experiments based on a statistical methodology. This is important to know how the mechanism works, which allows the identification and reuse of features extraction techniques that provide the best performance. Moreover, this process is also used in this work to evaluate and compare the outdoor location features identified in literature. For this evaluation, experiments were conducted with three classification algorithms (C4.5, SVM, and Naive Bayes) available in the WEKA machine learning environment and four datasets, two of which are public (Geolife and MIT Reality). Besides that, twelve measures were collected, being nine efficacy and three efficiency measures. In the analysis of the experimental results, significant variations were found in accuracy, CPU time, and memory regarding all evaluated scenarios. With these results, this work provides evidence of the viability of the proposed process and guides the choice of outdoor location features and learning algorithms that provide better performance for constructing transparent and continuous authentication mechanisms.