Siamese mViT: Lightweight Biometric Facial Recognition for Edge Devices

Facial recognition has become a widely adopted biometric identification method, revolutionizing authentication process across platforms such as smartphones, and edge IoT devices like smartwatches, drones, autonomous vehicles and smart home security systems. While existing sophisticated models such as Vision Transformers (ViT) deliver impressive accuracy, their reliance on large datasets and heavy computational requirements creates challenges for resource constrained edge devices that struggle to handle their complexity, and large volume of training data. To overcome these limitations, we propose Siamese mViT, a lightweight network that integrates a mobile-optimized Vision Transformer (MobileViT) block into a Siamese architecture for facial authentication. Analysis on the Cross-modal Face-Periocular dataset demonstrate that proposed Siamese mViT achieves 93.12% accuracy with remarkably small dataset of just 190,876 pairs of face and periocular images (1.24GB). Our results demonstrate that Siamese mViT model is both lightweight and accurate, making it well-suited for deployment on edge devices.