Autoencoders for Open-Set Presentation Attack Detection

Face anti-spoofing (FAS) has gained significant attention due to its crucial role in ensuring the security of facial recognition systems. Despite numerous proposed methods for addressing various types of presentation attacks, there remains a gap in the existing literature, as most methods only tackle seen attack types and lack generalizability in real-world scenarios. This paper presents a novel method for face antispoofing, specifically focused on open-set scenarios. The proposed approach consists of an autoencoder exclusively trained to reconstruct bona fide data. The key idea is that features extracted using this model would exhibit poor quality for attack data, resulting in significantly higher reconstruction errors. Consequently, these low-quality features would differ from the well-reconstructed features of bona fide samples. Our method’s key novelty lies in the rich feature learning procedure that focuses on the properties of bona fide samples through an autoencoder formulation. The extracted features are used to train one-class classifiers using only bona fide samples. Experimental results on standard benchmarks such as RECOD-MPAD, WMCA, HQ-WMCA, Replay-Attack, OULU-NPU and MSU-MFSD datasets demonstrate the excellent performance of our proposed approach, placing it among state-of-the-art methods for metrics like Bona Fide Presentation Classification Error Rate (BPCER), Attack Presentation Classification Error Rate (APCER) and Average Classification Error Rate (ACER).