Face the Challenge—Generalization of Presentation Attack Detection

Face recognition is one of the most widely adopted biometric technologies, with applications in mobile devices, banking, and access control. However, its widespread use raises security concerns. One of the most common threats is presentation attacks (PAs), in which adversaries spoof the system using printed photos, videos, or masks. Developing effective Presentation Attack Detection (PAD) methods has become critical, yet generalizing to unseen Presentation Attack Instruments (PAIs) remains a major challenge. This is further complicated by the fact that most public PAD datasets are closed and limited in attack diversity and acquisition conditions. Standard evaluation protocols are typically based on intra- and inter-dataset setups, which may not reflect real-world variability. To address this, we propose analyzing presentation attacks using a novel metric, the Presentation Attack Similarity Index, which quantifies the similarity between different attacks. Based on this, we identify Presentation Attack Similarity Clusters, grouping attacks with high interchangeability. This approach offers deeper insight into PAI relationships, allowing for the strategic selection of representative attacks and the design of more balanced training datasets.