Vol. 7, Issue 1, Part A (2025)

Facial spoofing attacks—such as those involving printed photos, replayed videos, or three-dimensional masks—pose significant risks to biometric security systems by enabling unauthorized access and identity manipulation. Traditional anti-spoofing methods, including manual feature extraction and deep learning models, often face challenges in generalizing across diverse spoofing scenarios due to differences in lighting, facial angles, and domain variability. While Convolutional Neural Networks (CNNs) are commonly used in spoof detection, they frequently fall short in maintaining spatial dependencies among facial features, resulting in higher rates of incorrect acceptances and rejections. To overcome these shortcomings, this study proposes a Capsule Siamese Network (Capsule-SN) as a resilient approach to facial spoof detection. By leveraging dynamic routing between capsules, the Capsule Network maintains the spatial structure of facial features, offering a richer representation of facial texture, depth cues, and intrinsic patterns. The Siamese framework strengthens the model’s capacity to distinguish between authentic and spoofed images by learning similarity metrics between image pairs, thereby reducing sensitivity to specific attack types. Furthermore, the model employs contrastive loss to shape an embedding space that clearly separates genuine from fake inputs.