Abstract
Modern security systems include biometric authentication, yet classic features like fingerprints and faces are still vulnerable to fraud. This study introduces PalmGuard, a real-time, forgery-resistant palm-vein identification system using a lightweight, attention-enhanced CNN and unique data approaches. We first publish a cross-device, multilighting vein dataset with GAN-synthesized samples to capture real-world variability. We then employ sensor input. Standardizing adaptive preprocessing pipeline (ROI extraction, contrast enhancement, noise suppression). Using stacked convolutional blocks, channel-wise attention, and contrastive adversarial loss, our basic model decreases Equal-Error-Rate (EER) by 2.5 percent on occluded and low-illumination tests compared to state-of-the-art baselines. PalmGuard, running in under 15 ms on embedded hardware, achieves 2 percent False Accept/Reject Rates, 1.1 percent EER, and 98.7 percent accuracy in five deployment scenarios. Our integration into edge and mobile devices is even more seamless. These findings demonstrate palm-vein identification’s suitability for high-security applications and suggest a path towards scalable, on-device biometric systems.