Year: 2026
Type: article
Abstract: The fast integration of digital healthcare platforms, including telemedicine and electronic health records, has created a significant need for secure, contactless authentication methods. Traditional login methods such as passwords and PINs are increasingly insufficient for ensuring the security of sensitive patient data. This paper presents a robust, lightweight Healthcare Authentication System that five biometric modules to address major challenges in facial recognition: spoofing attacks, liveness detection, and facial occlusion due to mask-wearing. The system leverages RetinaFace for face detection, a MobileNetV2 model for mask classification, a lightweight CNN for blink-based liveness verification, and MiniFASNet for spoof detection. For identity matching, the system utilizes InceptionResNetV1 with ArcFace loss, enabling effective recognition of both masked and unmasked faces through a dual-embedding strategy. Each module was evaluated using benchmark datasets: CASIA-FASD for spoofing, RMFD for mask detection, open/closed eye datasets for blink detection, and VGGFace2 (with synthetic masking) for recognition. The system achieved 91% accuracy in blink detection, 85% in spoof detection, 97% in mask classification, and 96% in face recognition. The complete pipeline achieved 88% overall accuracy on the evaluated dataset. The proposed solution offers a practical and scalable approach for secure biometric authentication in healthcare environments, particularly in clinical settings where hygiene, user accessibility, and anti-spoof measures are essential. This work provides a foundation for future research in offline biometric verification, depth-based spoof detection, and multilingual interface design, contributing toward more inclusive and secure healthcare technologies. While we do not fine-tune model backbones, our contribution lies in a clinically viable integration strategy that combines blink-based liveness, anti-spoofing, mask detection, and face verification under a fixed operating point with identity-disjoint evaluation, achieving 88% end-to-end accuracy on public-dataset inputs.
Institutions: Multimedia University, University of Nottingham Ningbo China
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7222460241
FWCI:
Cites:
Cited by:
Related to: