Face Spoof Attack Detection Using Hybrid Machine Learning Approach

Contemporary face recognition approaches are being threatened by spoofing attacks with static face images, videos and realistic masks. It is challenging for many existing methods to capture the subtle physical differences of legit and impostor face inputs in video data. This paper presents a videobased anti-face-spoofing system that integrates visual analysis and physiological signal monitoring. It starts off by resizing and normalizing every video frame, feature extraction through a pre-trained Vision Transformer (ViT) model where removal of its last classification layer dwells on higher-level visual representations. At the same time, it also analyzes the green channel out of frames for remote photoplethysmography (rPPG) signals - imperceptible fluctuations associated with heartbeats - that provide an additional layer of biometrics. The features extracted from both sources are then combined and reduced by PCA. Then, classification is carried out by a stacked ensemble of hypercolumns with KNN, SVM, and XGBoost (the classifier of the LightGBM). The system achieved 98.25% testing accuracy, also demonstrating its robustness to identify fake faces. But beyond accuracy, it includes protections such as a hashing mechanism to confirm code integrity. It offers fast verification even with just one frame per video, a fit for banking, phones, and security.