Facial recognition software is becoming a staple of biometric security in public spaces whose urban surveillance systems grow ever more reliant on the technology. However, in unconstrained city environments provided in these datasets, illumination, occlusions, extreme pose and low-resolution imagery are obstacles. In this paper we propose an ArcFace-based facial recognition framework as a means to improve security and reliability for urban surveillance. We extract robust facial features using a ResNet-100 deep backbone and also use additive angular margin loss (ArcFace) to maximize inter class separability and intra class compactness. The performance of the ArcFace enhanced model is examined against state-of-the-art models (VGGFace and FaceNet) and shown to improve accuracy across identification, precision, recall, and F1 score, and lower Equal Error Rate (EER) overall. Experiments reveal that the proposed method, which significantly enhances biometric identification reliability, is able to achieve state-of-the-art recognition accuracy in such difficult surveillance conditions. Going forward, this approach can be extended to handle occluded (masked) faces and will be applied to resource constrained edge deployments. Our study shows, through experiments, the efficacy of applying ArcFace based deep learning towards the enhancement of biometric security in smart city surveillance applications.