Centralized Access Management for Vertical Housing Using Edge Computing and Deep Learning

Abstract

The implementation of security systems in vertical housing often has a choice between high infrastructure costs from decentralized hardware and privacy risks from cloud solutions. This study presents a prototype for a centralized access management system utilizing edge computing (Intel NUC) as a local server to authenticate residents at various access points. The system uses Frigate NVR for lightweight real-time object detection and the ArcFace Deep Learning model for facial recognition. It processes all biometric data locally to protect privacy. We used a dataset of three registered subjects to test the experiment. The tests looked at how well the system worked at different distances (1 to 5 meters), in different lighting conditions (daylight and infrared), with different types of facial occlusions (medical masks), and with 2D spoofing attacks (print and digital media). Using a confusion matrix over 50 random test samples that included both authorized users and unknown intruders, the system got a global accuracy of 80.0%. The system also had a Genuine Acceptance Rate (GAR) of 86.6%. The system was very stable when it was 1 to 2 meters away, but it didn't work as well in extreme conditions. With an average CPU usage of 46.87% and physical control latency via the MQTT protocol of less than 0.2 seconds, resource efficiency was kept up. These results show that the proposed edge architecture can work as a responsive and computationally efficient prototype for smart apartment security. They also show that liveness detection needs to be improved in the future to reduce the risk of digital spoofing.