With the rapid advancement of biometric technologies, fingerprint recognition has become one of the most widely used methods due to its low cost and ease of acquisition. However, it remains vulnerable to presentation attacks using forged fingerprints and faces challenges in generalization across different sensors and spoofing materials. To address this, we propose a robust fingerprint liveness detection (FLD) approach. The method applies efficient fingerprint foreground segmentation and local patch extraction as preprocessing, and generates adversarial perturbations to augment the training data, improving resilience against diverse attacks. An improved EfficientNet backbone, enhanced for lightweight structure and real-time performance, is then employed for classification. Experiments on the LivDet2015 dataset demonstrate that our method achieves high accuracy and superior robustness, particularly in cross-sensor scenarios.