领域自适应场景分类及无监督迁移学习在生物特征活体检测中的应用

There are four types of domain adaptation scenarios based on the different types of target and source domains: unsupervised, supervised, heterogeneous distribution, and multiple source domain problems. In this paper, we designed classification transfer tasks based on practical applications: the source domain data has labels and varies in quantity, while the target domain data is unlabeled and varies in quantity. We simulated various practical applications by combining different source and target domains. We first proposed two models with attention mechanisms for implementing biometric feature liveness detection: the CvT model based on Transformer and the GAT model based on graph attention network. Then, by combining sample adaptive sampling and MDD for aligning the feature space of the target and source domains, we achieved unsupervised model transfer learning in the target domain, greatly improving the transfer effect of the model.