Edge-aware quantum-resilient face recognition using self-supervised DINOv2 and graph Siamese learning
Аннотация
Recent advances in self-supervised vision transformers—particularly DINOv2—have shown impressive capabilities in learning rich visual representations from natural images. However, their potential for building robust and deployable face-recognition systems for Internet of Things (IoT) environments remains largely unexplored. In this work, we introduce a unified architecture that employs pretrained DINOv2 models (small, base, and large) as frozen feature extractors on the Labeled Faces in the Wild (LFW) and CelebA datasets, and couples them with three complementary downstream identity-learning modules: (i) Graph Neural Networks (GNN) operating on similarity graphs of face embeddings, (ii) Siamese networks trained with contrastive and triplet objectives, and (iii) distance-based classifiers using learned metric spaces. We develop an edge–cloud IoT pipeline in which low-power cameras perform lightweight, quantized DINOv2 feature extraction and transmit compact embeddings to a secure edge gateway for real-time matching and decision-making. To ensure long-term resilience, we also propose optimal design elements, similarity search, optimized metric learning, and the integration of quantum-resistant cryptographic mechanisms for template storage and communication. The results show that DINOv2 using GNN-based reasoning and Siamese metric learning can compete with LFW and CelebA approximately 62 ms on Raspberry Pi 4, 28 ms on Jetson Nano, and less than 110 ms end-to-end latency. Altogether, the system achieves approximately 94% bandwidth reduction through compact quantized embedding transmission with better resource budgets comparable to edge IoT systems, leading to quantum-ready and privacy-preserving face recognition systems.