Paper page - AGVBench: A Reliability-Oriented Benchmark of Data Augmentation for Vein Recognition

Source: https://huggingface.co/papers/2607.02271 TL;DR: We open-source AGVBench, the first reliability-oriented benchmark for data augmentation in vein recognition. By systematically evaluating 30 representative augmentation strategies across 5 public palm- and finger-vein datasets using 7 backbone architectures, we demonstrate that accuracy-centric evaluation alone is insufficient for assessing the reliability of biometric recognition systems.

📌 Key Findings:

🔹 Accuracy–Security Decoupling: Multi-image mixing methods (e.g., MixUp and PuzzleMix) generally achieve the strongest recognition performance. However, they are often poorly calibrated and more vulnerable to adversarial perturbations, revealing a clear inconsistency between clean accuracy and adversarial security.

🔹 Degradation from Geometric Transformations: Severe geometric transformations consistently degrade recognition performance, likely due to feature misalignment or spatial cropping, rather than direct disruption of vein topology.

🔹 Modality Variations: The effectiveness of augmentation strategies varies substantially across palm-vein and finger-vein datasets, indicating that no single augmentation strategy universally generalizes across different vein modalities.

🚀 Beyond benchmarking, AGVBench aims to serve as an open research platform for reliability-oriented vein recognition, promoting reproducible evaluation, inspiring the development of next-generation augmentation algorithms, and advancing the deployment of trustworthy biometric systems in real-world applications.

🏠 Project Page:https://advance-veintech-innovators.github.io/AGVBench/

📄 ArXiv Paper:https://arxiv.org/abs/2607.02271

💻 Code & Benchmark:https://github.com/Advance-VeinTech-Innovators/AGVBench