Wavelet Based Convolutional Neural Networks for Face Anti-spoofing

With increase in the number of systems using facial recognition as the preferred mode of bio-metric authentication, attacks against such systems have also increased, thereby needing good and accurate anti-spoofing mechanisms to detect the same. In this paper we propose a powerful combination of Convolutional Neural Networks (CNNs) along with Discrete Wavelet Transform (DWT) and transfer learning. DWT is used as a pooling mechanism for down sampling of feature maps. Wavelet transformation decomposes the feature maps into low and high frequency sub-bands. The high frequency sub-bands are discarded and the low frequency sub-band is carried forward in the network. We train the VGG16 and VGG19 networks by cutting off the network at some suitable point, extracting the features using pretrained ImageNet weights and then applying DWT using various wavelets. The models are trained on two datasets- LCC-FASD database and NUAA database. Our experimental results show that the combination of CNN and DWT produces promising results on both the datasets, outperforming the current state of the art results, with lesser number of parameters.