Federated Edge Learning with Hybrid CNN-Transformers for Secure, High-Accuracy Face Recognition in Resource-Constrained IoT Systems

Traditional face recognition systems struggle to balance accuracy, privacy, and computational efficiency in IoT environments. This work addresses these challenges by proposing a novel framework integrating hybrid CNN-Transformer models and federated edge learning. The architecture synergizes convolutional layers for localized feature extraction and transformer blocks for global facial context modeling, achieving state-of-the-art accuracy with minimal parameters. Privacy is preserved through federated learning, where models are trained decentralizedly across IoT nodes without raw data transmission. Further, dynamic network pruning, quantization-aware training, and hardware-aligned kernel optimizations ensure real-time inference on edge devices like Raspberry Pi 4 and NVIDIA Jetson Nano. Rigorous evaluation on the LFW, CASIA-WebFace, and custom IoT-Face-MINI dataset demonstrates 98.2% accuracy (mean of 5-fold cross-validation), surpassing lightweight CNNs (MobileFaceNet: 97.8%) and cloud-based ResNet-50 (98.0%) while reducing latency by 62% and energy use by 35%. Ablation studies confirm the hybrid model’s superiority in handling occlusion and illumination variations. Statistical analysis (p < 0.01) validates significance against benchmarks. The system’s reproducibility is ensured via open-sourced code, Dockerized training pipelines, and detailed hyperparameter documentation.By unifying accuracy, efficiency, and ethical design, this framework empowers applications like secure smart homes and patient monitoring. The work critically addresses trade-offs between biometric precision and edge deployment constraints, setting a precedent for trustworthy IoT-AI systems.