Robust palmprint biometric solution for secure mobile authentication

Abstract

Smartphones increasingly rely on biometric authentication for access to financial and personal services, creating a need for palmprint recognition that is accurate, fast, and deployable on device. This paper proposes an end-to-end smartphone palmprint authentication framework that integrates guided mobile image capture, landmark-based region-of-interest (ROI) extraction, and compact embedding inference. A ResNet-18 teacher is first trained with self-supervised contrastive learning to reduce dependence on labeled biometric data, then distilled into a lightweight MobileNetV3 student for efficient mobile deployment. The learned embeddings support both on device verification and large-scale identification using an approximate nearest neighbor index (FAISS). Experiments on a public Kaggle palm dataset achieve 99.2% accuracy with a 0.15% equal error rate (EER). On an iPhone 13, the end-to-end pipeline runs in 87.0 ms with a 12.4 MB student model. For a 1 million-entry gallery, FAISS provides 32 ms query latency while maintaining 99.5% Recall@1. Limitations include evaluation under mostly controlled capture conditions and the absence of an explicit liveness or presentation attack detection (PAD) module; future work will address unconstrained testing and anti-spoofing integration.