The large development of social networks has unveiled security concerns related to potential attacks on biometric systems. In particular, numerous examples have shown how a non-authorised subject can easily download a photo or video of a target individual, and use it for impersonation. In order to address those security threats, Fisher Vector-based Presentation Attack Detection (PAD) solutions have shown a reliable detection performance.
In order to improve the generalisation capability of the Fisher Vector (FV) approach to unknown attacks, the thesis investigated the combination of FV with filters learnt by different end-to-end deep learning-based approaches. Those convolutional filters have shown to be more powerful than traditional handcrafted descriptors in terms of sample description. In particular, the work established a benchmark among three different filters-selection methods. The experimental evaluation was conducted over freely available databases namely the CASIA Face Antispoofing, Replay-Mobile, MSU-FASD, and OULU-NPU. All experimental results are reported in compliance with the international ISO/IEC 30107-3 for Biometric PAD. Those results were, in turn, achieved over several challenging scenarios in which either the attack types (i.e., Presentation Attack Instrument species) or the capture devices used in the recapture of the samples are unknown.