Meta Face Anti-spoofing with Regularization and Convex Optimization

摘要

Face anti-spoofing plays an important role in preventing face presentation attacks. Numerous face anti-spoofing methods have been proposed, but they mostly cannot generalize well to cross-dataset. In this paper, we regard face anti-spoofing as a domain generalization (DG) problem and use a newly meta- learning method named meta-learning with regularization and convex optimization to solve this problem. To make our meta- learning work focus on a more generalized direction and get more general distinction clues, we use domain knowledge to regularize the feature space instead of with only binary class labels. Besides, to avoid the bad performance of using a nearest-neighbor classifier at a small data scale, we choose a linear classifier to classify, and to overcome the challenge of computation, we use the method of convex optimization. Experiments on three public datasets show that our proposed method is effective and performs the state-of- the-art results.