Due to the environmental differences, many face anti-spoofing methods fail to generalize to unseen scenarios. In light of this, we propose a unified unsupervised and semi-supervised domain adaptation network (USDAN) for cross-scenario face anti-spoofing, aiming at minimizing the distribution discrepancy between the source and the target domains. Specifically, two modules, i.e., marginal distribution alignment module (MDA) and conditional distribution alignment module (CDA), are designed to seek a domain-invariant feature space via adversarial learning and make the features of the same class compact, respectively. By adding/removing the CDA module, the network can be easily switched for semi-supervised/unsupervised setting, in which sense our method is named with “unified”. Moreover, the adaptive cross-entropy loss and normalization techniques are further incorporated to improve the generalization. Extensive experimental results show that the proposed USDAN outperforms state-of-the-art methods on several public datasets.

由于环境差异，许多面部反欺骗方法无法推广到看不见的场景。有鉴于此，我们提出了一种用于跨场景人脸反欺骗的统一的无监督和半监督域自适应网络（USDAN），旨在最大程度地减少源域和目标域之间的分布差异。具体来说，有两个模块，即边际分布对齐模块（MDA）和条件分布对齐模块（CDA）旨在通过对抗学习寻求领域不变的特征空间，并使同一个类的特征分别紧凑。通过添加/删除CDA模块，可以轻松地将网络切换为半监督/无监督设置，在这种情况下，我们的方法称为“统一”。而且，自适应交叉熵损失和归一化技术被进一步结合以改善泛化。大量的实验结果表明，在一些公共数据集上，拟议的USDAN优于最新方法。