Face Presentation Attack Detection via Spatiotemporal Autoencoder

Wide usage of biometrie authentication systems engages attackers. Face presentation attacks try to deceive the machine learning-based system via malicious inputs to the sensor employing tools such as a face mask or photo. Presentation attacks threaten the deployment of face recognition systems to the real-world since security is a priority for an authentication system. In this paper, we propose a Convolutional Long Short Term Memory (ConvLSTM) based Spatiotemporal Autoencoder model. The proposed model only employs the real samples to train the model by minimizing the reconstruction error. The model learns to reconstruct real samples; thus, reconstruct attack samples with higher reconstruction error. The model classifies whether a sample is genuine or not via the scoring function based on the reconstruction error. The proposed model reduces the half total error rate by 22% compared to the commonly used methods in the literature.