Facial recognition systems have experienced rapid progress with the integration of deep learning architectures, especially convolutional neural networks (CNNs). Among the most prominent approaches, ArcFace and FaceNet are recognized for their high accuracy and robustness in identity verification tasks. However, their comparative performance under real-world, uncontrolled conditions where factors such as illumination, resolution, and occlusion vary remains insufficiently studied. This research presents a systematic comparison between ArcFace and FaceNet using the Analytic Hierarchy Process (AHP) as a decision-support framework. The study employs three publicly available datasets—Labeled Faces in the Wild (LFW), CASIA-WebFace, and Celeb-DF—to evaluate performance in terms of accuracy, precision, recall, F1-score, and equal error rate (EER). Controlled experiments simulate degraded conditions, including low resolution, poor lighting, and partial occlusion. Results reveal that ArcFace consistently outperforms FaceNet in accuracy and F1-score across challenging scenarios, while FaceNet achieves faster inference and a lower false positive rate in some instances. These insights offer practical guidance for selecting optimal models for security, forensic, and identity verification applications.