Due to their convenience and reliable performance, face recognition technologies are now widely implemented in daily applications. However, their vulnerability to presentation attacks, such as printed photos or replayed videos, poses serious security threats, especially when deployed on devices with only RGB cameras. This paper presents a passive, sensor-independent face anti-spoofing solution using a ResNet-50-based convolutional neural network (CNN) trained exclusively on RGB imagery. Leveraging the large-scale Wild Face Anti-Spoofing (WFAS) dataset, which includes over 1.3 million images from 17 attack types, we evaluate the effectiveness of our method using ISO/IEC 30107-3 metrics: Attack Presentation Classification Error Rate (APCER), Bona Fide Presentation Classification Error Rate (BPCER), Average Classification Error Rate (ACER). A comprehensive grid search and ablation study are conducted to optimize training configurations and analyze the impact of data augmentation techniques such as blur and horizontal flipping. Our approach achieves an ACER of 2.99% on the held-out test set, demonstrating high robustness and generalization without relying on depth or infrared data. The results support the feasibility of deploying accurate and secure liveness detection on unimodal devices using only RGB input.