Abstract <p>The fingerprint-based biometric technology is vulnerable to several kinds of attacks. One of the simplest attacks to carry out on the fingerprint sensor is the presentation attack. Numerous fingerprint presentation attack detection (FPAD) strategies have been put out in recent years. These FPAD techniques have yielded promising results on cross-material datasets. However, when training and testing datasets come from different domains (sensors), the performance of the FPAD approach can degrade by up to 30%. Therefore, to achieve a consistent performance, a robust FPAD approach must learn domain-independent features. We have developed an unsupervised divergence-based domain adaptation (UDDA) method with an adaptive loss function (ALF) to minimize the domain shift in FPAD. The ALF integrates domain divergence loss (DDL) and classification loss. In a cross-sensor scenario, the ALF helps learn domain-invariant features and provides reliable classification of real and fraudulent fingerprints. Experimental results demonstrate that the proposed UDDA approach reduces the cross-sensor average classification error (ACE) by 19.94% on LivDet 2015 and 19.23% on LivDet 2017.</p>