Presentation attack detection: an analysis of spoofing in the wild (SiW) dataset using deep learning models

Abstract

Presentation attacks are executed to attain illegitimate access to the system. They are categorized by their mode of action as a print attack, replay attack, and spoof attack, and by their media of action as iris, biometrics, fingerprint, and face. Though there has been a rise in computational algorithms and models to detect presentation attack, generalization across different datasets remain an essential aspect of performance measure. In this paper, we present presentation attack detection (PAD) and presentation attack types of classification (PATC) models based on convolutional neural networks (CNN). We utilize the different attacks presented on the Spoofing in the wild (SiW) dataset to build these models. The PAD-CNN model is developed with a minimal footprint to optimize training time. High-performing models such as Mobilenet and Inceptionv3 are also used in this research work. In this study, we perform an independent test on images extracted from videos of both seen and unseen subjects. Overall, PAD-CNN performed better or on par with Mobilenet and Inceptionv3, even with less training time. Furthermore, these models were also trained to classify the types of presentation attacks with good results. The benchmarking of these models were done on two different datasets, NUAA photo imposter database and Replay-attack database utilizing transfer learning. Together, these results suggest the robustness and effectiveness of the proposed presentation attack detection models based on CNN on the SiW dataset.