Detecção de ataques a sistemas de reconhecimento facial utilizando abordagens eficientes de aprendizado de máquina em profundidade

Abstract

Biometrics emerged, in the last decades, as a robust and convenient solution for security systems. However, despite the higher difficulty to circumvent the biometric applications, nowadays, criminals are developing attacks, known as spoofing or presentation attacks, precisely simulating biometric traits of legal users, such as the facial image with high-definition printed photographs. Among the main biometric traits, face is a promising one given its high universality (everyone has a face) and non-intrusive capture. Despite all this, face recognition systems are the ones that most suffer with such frauds given the high availability of facial images of people in the worlwide computer network. In this context, face spoofing detection techniques must be developed and integrated to the traditional face recognition applications in order to preserve their robustness in real scenarios. Deep Learning based methods have presented state-of-the-art performances in many areas, including face spoofing detection. However, the methods proposed in the literature so far present high computational costs, being not feasible in real situations, with significant hardware restrictions. In this context, in this thesis, efficient architectures of deep neural networks for face spoofing detection are proposed. Among the proposed approaches, modifications in the architectures of the Restricted Boltzmann Machines (RBM), generative and efficient models turned into deep discriminative neural networks, as well as modifications in the architecture of the Convolutional Neural Networks (CNN), expanding them in width instead of depth, and a novel training algorithm for CNNs, able to capture local spoofing cues of different parts of the faces, allowed a significant reduction on the amount of parameters and operations required for processing the facial images, as well as a faster convergence of the deep neural networks, allowing them to reach accuracy results, in attack detection, compatible with the state-of-the-art, at lower computational costs.