Transfer learning for face anti-spoofing detection

Abstract

In recent years, the demand for facial biometric authentication services has increased dramatically. Also, the efforts to cheat this type of system have become more common. In this paper, we propose a single shot CNN-based solution for the face anti-spoofing problem. We trained a deep learning model using transfer learning from a pre-trained VGG16 model. After some pre-processing we rely solely on the network to classify an image. We evaluate several implications of the preprocessing of data, investigate the implications of different amounts of background included in the picture, and the effect of data subsampling. Additionally, we analyze what happens when we sub-sample the training data. We evaluate our results in four publicly available datasets, drawing some insights on the results by using the Grad-CAM algorithm. Our approach is competitive when compared with similar methods. Moreover, we achieved our results while training with a fraction of the original datasets, enforcing that experiments can be run much quicker without sacrificing accuracy.