Fisher Vectors for Biometric Presentation Attack Detection

Biometric systems have experienced a large development over the last years since they are accurate, secure and in many cases, more user convenient than traditional credential-based access control systems. In spite of their benefits, biometric systems are vulnerable to attack presentations, which can be easily carried out by a non-authorised subject without having a deep computational knowledge. This way, he/she can gain access to several applications where biometric systems are frequently deployed, such as bank accounts and smartphone unlocking. In order to mitigate such threats, we present in this work a study on the feasibility of using the Fisher Vector (FV) representation to spot unknown-attack presentations over different biometric modalities such as fingerprint, face and voice. By learning a common feature space from a set of local features, extracted from known samples, the FVs lead to the construction of reliable discriminative models which can successfully distinguish a bona fide presentation from an attack presentation. The experimental evaluation over publicly available databases (i.e. LivDets, CASIA-FASD, SiW-M and ASVspoof, among others) yields error rates outperforming most state-of-the-art algorithms for challenging scenarios where species, recipies or capture devices remain unknown.