DEVELOPMENT AND PERFORMANCE EVALUATION OF A MULTI-LAYER BANK VAULT LASER SECURITY SYSTEM WITH FACE RECOGNITION ACCESS USING RESNET-34

Abstract

The development of a bank vault laser security system with face recognition access using ResNet-34 presents a multi-layered security architecture designed to enhance the protection of high-value assets and confidential information in banking environments. At its base, a laser sensor network controlled by a Raspberry Pi 4 Model B establishes a virtual security perimeter around the vault entrance. Any interruption of the laser beam immediately triggers an audible alarm and updates the LCD status display, forming the first layer of physical intrusion monitoring. For biometric verification, the system employs a face-recognition module built on the ResNet-34. The dataset used for training the model consisted of 7,500 facial images, systematically collected from three subjects, with each subject contributing 2,500 images captured under varying illumination, background conditions, and pose differences. To ensure hardware consistency and avoid inter-device bias, all images were captured using the same in-built camera module. The ResNet-34 model was fine-tuned using the Adam optimizer (learning rate = 0.001, batch size = 32) for 50 training epochs. It achieved a recognition accuracy of 96.7%, precision of 95.8%, recall of 96.1%, and a false acceptance rate (FAR) of 3.3% on the validation set. Facial embeddings produced by the network were matched using the Euclidean distance similarity metric to ensure fast and consistent identity verification. The laser intrusion module demonstrated a response latency of 0.21 seconds and 100% detection reliability under controlled testing.