Crossing domains with the inductive transfer encoder: Case study in keystroke biometrics
John Vincent Monaco, Manuel M. Vindiola · 2016
Keystroke biometric samples are often collected under various conditions, such as different device types, increasing levels of practice through repetition, and subject impairment. Cross-domain comparisons, in which query samples are collected under different conditions than the template, generally lead to degraded performance. The difficulty in comparing samples from different domains can be viewed as an inductive transfer learning problem, in which general knowledge of the mapping between a source and target domain can be applied to increase task performance, such as verification accuracy, by transferring source domain samples to the target domain. In this light, we propose the inductive transfer encoder, which utilizes pairwise correspondences from an independent dataset to learn a general transformation between domains. When the transformation is applied to template samples in the source domain, and query samples are in the target domain, increased verification performance is observed. We evaluate four different strategies for establishing the pairwise correspondences between source and target domains, and two cross-domain problems: low vs high practice levels and one vs two typing hands. Empirical results demonstrate that the inductive transfer encoder captures general rules that can be applied to transfer source domain samples to the target domain.