Face Recognition-Based Login Security System Using Deep Learning

Abstract

Password-based authentication remains the dominant access-control mechanism for digital systems, yet it suffers from well-documented weaknesses including weak or reused credentials, phishing, credential leakage, and the cognitive burden of password management. This study proposes and evaluates a face recognition-based login security system built on deep learning to address these limitations through a biometric factor that is intrinsic to the user. The system follows a verification pipeline consisting of webcam image acquisition, preprocessing, face detection, alignment, deep feature extraction, and similarity-based decision making. A convolutional neural network (CNN) of the FaceNet family maps each detected face to a compact 128-dimensional embedding, and authentication is performed by comparing the live embedding against an enrolled template using a Euclidean-distance threshold. A lightweight browser-based prototype was implemented with HTML, JavaScript, and the face-api.js library to demonstrate real-time, client-side operation without specialized hardware. Using a simulated single-identity evaluation across 300 login attempts spanning frontal, off-angle, low-light, eyewear, occlusion, and impostor scenarios, the system achieved an accuracy of 97.0%, precision of 97.96%, recall of 96.0%, and an F1-score of 96.97%, with a False Acceptance Rate of 2.0%, a False Rejection Rate of 4.0%, and an average login time of 1.28 seconds. The results indicate that deep-learning face recognition is a practical and accurate alternative to passwords, while highlighting residual challenges related to illumination, pose, and presentation (spoofing) attacks that motivate the integration of liveness detection and multi-factor authentication.