Towards Generalisable Facial Presentation Attack Detection Based on the Analysis of Facial Regions

In the last decade, breakthroughs in the field of deep learning have led to the development of powerful presentation attack detection (PAD) algorithms which reported reliable performance across different realistic scenarios. Typically, most of these techniques analyse the full face to detect attack presentations (APs), ignoring that the attributes or artefacts produced in the fabrication of the attacks vary their location on the face depending on the presentation attack instruments (PAI) species, subject and the environmental conditions. In addition, they still fail to categorise bona fide subjects who inadvertently occlude their face with accessories such as glasses, scarves or masks to prevent respiratory infections. To mitigate these issues, this paper explores the utility of using different facial regions for PAD. In this context, a new metric, Face Region Utility, is proposed, which indicates the usefulness of a particular test region to spot an attack attempt based on another training region. A thorough evaluation in challenging scenarios on well-known databases shows which face regions can successfully substitute the full face to detect APs in scenarios where pristine subjects use some of the mentioned accessories: up to a 67.73% of detection performance improvement is yielded by applying our proposed analysis when pristine subjects wear masks to prevent respiratory infections.