Fingerprint liveness detection (FLD) has achieved promising results with the support of a large number of fingerprint data. However, fingerprints involve user privacy, and it is difficult to obtain a large number of fingerprints. Consequently, when fingerprints are scarce or limited, the performance of FLD significantly reduced. Moreover, the performance of existing methods in cross-sensor detection tends to be unsatisfactory when confronted with diverse fingerprint acquisition devices, owing to disparities in their imaging mechanisms. To address these issues, this paper proposes a novel FLD method based on few-shot incremental learning with EMD metric (FIL-FLD). First, to enhance the discrimination between fingerprints, we introduce a class feature-Centric few-shot learning method with Earth Mover’s Distance (EMD) metric, which mitigates the issue of low accuracy due to insufficient fingerprint samples. Next, to improve the detection performance of FLD for unknown sensors, this paper proposes incremantal feature optimization learning. By removing individual features between different types of spoof fingerprints under the same sensor, common features are retained, thereby improving generalization. Finally, this study strategically combines distillation loss and cross-entropy loss to ensure that the model continues to learn from new fingerprint samples without losing old knowledge. Experimental evaluations conducted on the LivDet2015, LivDet2017, and LivDet2019 datasets demonstrate the effectiveness of our proposed method. Notably, our method achieves a 2.54% reduction in ACE compared to the classical BiRi-PAD method. Furthermore, it exhibits robust generalization performance when applied to cross-sensor scenarios.