Enhancing deep discriminative feature maps via perturbation for face presentation attack detection

摘要: Abstract Face presentation attack detection (PAD) in unconstrained conditions is one of the key issues face biometric-based authentication and security applications. In this paper, we propose a perturbation layer — learnable pre-processing for low-level deep features to enhance discriminative ability PAD. The takes candidate Convolutional Neural Network (CNN), corresponding hand-crafted an input image, produces adaptive convolutional weights layer. These determine importance pixels proposed adds very little overhead total trainable parameters model. We evaluated with Local Binary Patterns (LBP), without color information, on three publicly available PAD databases, i.e., CASIA, Idiap Replay-Attack, OULU-NPU databases. Our experimental results show that introduction CNN improved performance, both intra-database cross-database scenarios. also highlight attention created by its effectiveness general.