A Comprehensive Framework for Palm Vein Anti-Spoofing With Preprocessing Pipeline, Dataset, and Benchmark

Non-contact palm-vein recognition has been widely adopted in security-critical applications owing to its contact-free acquisition paradigm and exceptional discriminative power. However, these systems remain susceptible to a spectrum of presentation attacks (PAs), creating significant security risks that require urgent mitigation. Progress on palm-vein anti-spoofing is currently impeded by three fundamental gaps: (i) the absence of an open-source, end-to-end pipeline for preprocessing liveness-detection data; (ii) a lack of publicly available datasets specifically tailored to anti-spoofing evaluation; and (iii) the unavailability of benchmark studies employing standardized protocols and reference implementations. In light of these key issues, we make the following four contributions. Firstly, we propose a new open-source preprocessing pipeline that can significantly improve model performance, reducing errors by up to 53.3%. Secondly, we introduce PVASD, a new largest known dataset comprised of 1,187,519 images belonging to 5,515 subjects, which consists of 880,241 live palm vein images along with 307,278 spoof attack images including 16 different attack types – both 2d (printed stuff) and 3d (gloves and prosthesis models) attacks captured under a variety of environments using off-the-shelf commercial-grade sensors at five resolutions. Lastly, comprehensive benchmarks are also created by evaluating three types of representative methods namely classical image classification models, face anti-spoofing methods adapted from face domain and anomaly-detection-based approaches, while our experimental results reveal unique characteristics intrinsic only to palm vein spoof attacks, which will hopefully provide valuable guidance to researchers for further investigation. Furthermore, we also expanded PVASD by adding 20,000 spoof samples generated by artificial intelligence, and evaluated the vulnerability of the existing models to deepfake attacks. We will release our preprocessing pipeline, dataset, and benchmark codes at https://github.com/valhongli/PVASD to advance future reproducible studies and accelerate palm-vein anti-spoofing algorithmic research.