Conventional and Neural Architectures for Biometric Presentation Attack Detection

Facial biometrics, which enable an efficient and reliable method of person recognition, have been growing continuously as an active sub-area of computer vision. Automatic face recognition offers a natural and non-intrusive method for recognising users from their facial characteristics. However, facial recognition systems are vulnerable to presentation attacks (or spoofing attacks) when an attacker attempts to hide their true identity and masquerades as a valid user by misleading the biometric system. Thus, Facial Presentation Attack Detection (Facial PAD) (or facial antispoofing) techniques that aim to protect face recognition systems from such attacks, have been attracting more research attention in recent years. Various systems and algorithms have been proposed and evaluated. This thesis explores and compares some novel directions for detecting facial presentation attacks, including traditional features as well as approaches based on deep learning. In particular, different features encapsulating temporal information are developed and explored for describing the dynamic characteristics in presentation attacks. Hand-crafted features, deep neural architectures and their possible extensions are explored for their application in PAD. The proposed novel traditional features address the problem of modelling distinct representations of presentation attacks in the temporal domain and consider two possible branches: behaviour-level and texture-level temporal information. The behaviour-level feature is developed from a symbolic system that was widely used in psychological studies and automated emotion analysis. Other proposed traditional features aim to capture the distinct differences in image quality, shadings and skin reflections by using dynamic texture descriptors. This thesis then explores deep learning approaches using different pre-trained neural architectures with the aim of improving detection performance. In doing so, this thesis also explores visualisations of the internal representation of the networks to inform the further development of such approaches for improving performance and suggest possible new directions for future research. These directions include interpretable capability of deep learning approaches for PAD and a fully automatic system design capability in which the network architecture and parameters are determined by the available data. The interpretable capability can produce justifications for PAD decisions through both natural language and saliency map formats. Such systems can lead to further performance improvement through the use of an attention sub-network by learning from the justifications. Designing optimum deep neural architectures for PAD is still a complex problem that requires substantial effort from human experts. For this reason, the necessity of producing a system that can automatically design the neural architecture for a particular task is clear. A gradient-based neural architecture search algorithm is explored and extended through the development of different optimisation functions for designing the neural architectures for PAD automatically. These possible extensions of the deep learning approaches for PAD were evaluated using challenging benchmark datasets and the potential of the proposed approaches were demonstrated by comparing with the state-of-the-art techniques and published results. The proposed methods were evaluated and analysed using publicly available datasets. Results from the experiments demonstrate the usefulness of temporal information and the potential benefits of applying deep learning techniques for presentation attack detection. In particular, the use of explanations for improving usability and performance of deep learning PAD techniques and automatic techniques for the design of PAD neural architectures show considerable promise for future development.