A carregar...
Publicação
SmartyFlow - Biometria Facial Robusta para Identificação Virtual
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 33.94 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
O roubo de identidade é um problema crescente na nossa sociedade em geral. Deste
modo, é necessário garantir que os métodos de autenticação existentes sejam seguros
contra ataques de apresentação. Nesta tese pretende-se estudar métodos de autenticação
com base em biometria facial, mais especificamente, verificação facial. Trata-se de um
método que, apesar de moderno, é igualmente vulnerável a ataques de segurança, em
particular ataques de falsificaçãodo rosto. Ultimamente, têm surgido abordagens que
utilizam a verificação da vivacidade para detetar tais ameaças.
Assim, no contexto desta tese, a vivacidade será detetada através de um vídeo da face
de um indivíduo, utilizando o seu ritmo cardíaco estimado através de Eulerian Video
Magnification (EVM). Ritmo cardíaco este que é posteriormente classificado recorrendo
a dois tipos de redes neurais profundas diferentes: Convolution Neural Network (CNN)
e Temporal Convolutional Network (TCN). Utilizando esta técnica de deteção, é possível garantir maior resiliência a ataques de apresentação, pois o ritmo cardíaco é uma característica fisiológica dificilmente falsificável.
Para além de classificar o sinal do ritmo cardíaco estimado, procurou-se desenvol ver uma forma eficiente de melhorar ainda mais a robustez dos modelos implementa dos ao detetar os ataques de apresentação. Para isso, com base no Treino Adversarial desenvolveu-se a Deep Convolutional Generative Adversarial Network (DCGAN) que per mite a criação de sinais cardíacos artificiais.
Como resultado concluiu-se que a rede TCN é mais apropriada para esta tarefa (obtendo 90,17 de eficácia sem sinais artificiais) e que a introdução de sinais artificiais produzidos pela DCGAN permitem de facto melhorar a robustez do modelo (obtendo 93,55 de eficácia).
Identity theft is an ever-increasing problem in our society. Thus, it is necessary to ensure that the existing authentication methods are secure against presentation attacks. The proposed thesis aims to study authentication methods based on facial biometrics, more specifically, facial verification. Nonetheless, despite being a rather modern method, it is also vulnerable to security attacks, in particular, to face spoofing. Several approaches have recently emerged that use liveness checks to detect such threats. So, in the context of this thesis, liveness will be detected through a video of an individual’s face, using its estimated heart rate estimated through EVM. The heart rate is then classified using two different types of deep neural networks: CNN e TCN. By using this detection technique, it is possible to ensure a higher level of resilience to presentation attacks, considering that heart rate is a physiological characteristic that is difficult to forge. Besides classifying the estimated heart rate signal, an efficient way to increase the robustness of the implemented models in detecting presentation attacks was developed. To achieve this, on the basis of Adversarial Training, the DCGAN was developed, which allows the creation of artificial heart signals. As a result it was concluded that the TCN is more appropriate for this task (achieving 90,17 efficacy without artificial signals) and that the introduction of artificial signals produced by DCGAN can in fact improve the robustness of the model (achieving 93,55 efficacy).
Identity theft is an ever-increasing problem in our society. Thus, it is necessary to ensure that the existing authentication methods are secure against presentation attacks. The proposed thesis aims to study authentication methods based on facial biometrics, more specifically, facial verification. Nonetheless, despite being a rather modern method, it is also vulnerable to security attacks, in particular, to face spoofing. Several approaches have recently emerged that use liveness checks to detect such threats. So, in the context of this thesis, liveness will be detected through a video of an individual’s face, using its estimated heart rate estimated through EVM. The heart rate is then classified using two different types of deep neural networks: CNN e TCN. By using this detection technique, it is possible to ensure a higher level of resilience to presentation attacks, considering that heart rate is a physiological characteristic that is difficult to forge. Besides classifying the estimated heart rate signal, an efficient way to increase the robustness of the implemented models in detecting presentation attacks was developed. To achieve this, on the basis of Adversarial Training, the DCGAN was developed, which allows the creation of artificial heart signals. As a result it was concluded that the TCN is more appropriate for this task (achieving 90,17 efficacy without artificial signals) and that the introduction of artificial signals produced by DCGAN can in fact improve the robustness of the model (achieving 93,55 efficacy).
Descrição
Palavras-chave
Batimento cardíaco Biometria facial CNN Deteção de vivacidade Falsificação do rosto GAN