This work examines the practical deployment of an embedding-based face recognition andauthentication system built using pretrained deep learning models. The system combinesRetinaFace for face detection and FaceNet512 for facial embedding generation through theDeepFace framework. Identity verification is carried out by comparing embeddings usingcosine similarity with a fixed threshold value of 0.7. The full pipeline was implemented as a RESTful API using FastAPI so that registrationand authentication can be performed through simple API requests. Evaluation was conducted using a small controlled identity set consisting of 10 registered individuals. For eachidentity, repeated authentication trials were performed under indoor lighting conditions. Additional tests were also carried out using printed and digital photographs to observe whethernon-live inputs would be incorrectly accepted. The current implementation produced an average authentication time of approximately11 seconds per request when executed on CPU-only hardware. No false acceptance wasobserved during photograph-based attempts. The results indicate that while the systemfunctions reliably at small scale, feature extraction and face detection remain the dominantcomputational bottlenecks.
Srinivasan et al. (Wed,) studied this question.