In the last decades, the broad development experienced by biometric systems\nhas unveiled several threats which may decrease their trustworthiness. Those\nare attack presentations which can be easily carried out by a non-authorised\nsubject to gain access to the biometric system. In order to mitigate those\nsecurity concerns, most face Presentation Attack Detection techniques have\nreported a good detection performance when they are evaluated on known\nPresentation Attack Instruments (PAI) and acquisition conditions, in contrast\nto more challenging scenarios where unknown attacks are included in the test\nset. For those more realistic scenarios, the existing algorithms face\ndifficulties to detect unknown PAI species in many cases. In this work, we use\na new feature space based on Fisher Vectors, computed from compact Binarised\nStatistical Image Features histograms, which allow discovering semantic feature\nsubsets from known samples in order to enhance the detection of unknown\nattacks. This new representation, evaluated for challenging unknown attacks\ntaken from freely available facial databases, shows promising results: a\nBPCER100 under 17% together with an AUC over 98% can be achieved in the\npresence of unknown attacks. In addition, by training a limited number of\nparameters, our method is able to achieve state-of-the-art deep learning-based\napproaches for cross-dataset scenarios.\n