Unsupervised Learning for Face Anti-Spoofing Models

Abstract

Distinguishing images of bonafide (genuine) faces from Presentation Attacks (PA) or spoofs attracts increasing attention in industry due to the wide application of reliable automatic authentication. The anomaly detection problem in high dimensional data such as images can be addressed as a supervised or unsupervised learning problem, but when spoof data is sparse and not all spoof domains can be represented in training the unsupervised model can generalize knowledge to unseen domains to a higher extent. The unsupervised anomaly detection paradigm is essentially a one-class learning problem to distinguish the distribution of bonafide images from everything else. In this thesis we explore the mathematical framework and implementation of encoderdecoder- based deep learning models to learn the distribution of real images and classify spoofs. This thesis is preliminary work on the end goal of designing a lightweight unsupervised anti-spoofing model to run in Smart Eye automotive software. The researched models reproduce input images by utilizing a latent space embedding fitted to the distribution of bonafide images. The latent space along with the reconstructed image is used to classify spoofs. The results of the experiments show promise but are not yet at the level feasible for implementation in production devices. The models have been evaluated in terms of the spoof precision and recall as well as the embedding of the spoofs in the latent space. Further, some models use Gaussian Mixture Models (GMM) of the latent space to determine the spoof affiliation, though these results are inconclusive.