Detecção de ataques em biometria facial utilizando redes neurais convolucionais

• Ano de defesa: 2022
• Autor(a) principal: Sousa Neto, Sandoval Verissimo de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Informática; Programa de Pós-Graduação em Informática; UFPB
• Palavras-chave em Português: Anti-falsificação; Face; Ataques de apresentação; Detecção de ataques; Classi cação de imagens; Redes neurais convolucionais profundas; Face anti-Spoofing; Presentation attacks; Detection; Spoof detection; Image classi cation; Deep convolutional neral network; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/25507

With more services becoming available online by the day, biometric authentication methods such as ngerprints and faces are necessary to provide better security for the user. A person's face is one of it's most critical biometric features, mainly due to the easiness of use, and so it has been increasingly studied in the last years. However, as the use of authentication methods with facial biometrics increases, so does the amount of attack attempts on these systems. The incredible ease of use of facial biometry also comes with the shortcoming that social media makes it may be easier to nd photos and videos of someone and thus use its face to create attacks. Thus it is necessary a system that can detect if a person is real or if it is either a photo or video attack. These applications are known as Face-Anti-Spoo ng systems. This work proposes a spoo ng detection method using Convolutional neural networks. Transfer learning is used for training the model. The impact of di erent types of pre-processing tequiniques was studied. The experiments are made using four datasets widely known in the literature (NUAA, MSU, Replay Attack, OULU). The best results achieve better metrics than some works on literature. With an equal error rating lower than 0; 2% in the best experiment.