Attention-Guided GANs for Robust Face Anti-Spoofing Detection in Unconstrained Environment

Face anti-spoofing serves as an essential security element to protect biometric authentication systems when operating in unconstrained environments with various lighting effects and obstructed or changing facial positions. This work presents an Attention-Guided GAN architecture that incorporates CBAM functionality to boost the separation of authentic face images from spoofs. The model deploys spatial and channel-wise attention to enhance the visibility of important facial areas including eyes and nose and mouth because this is where spoofing artifacts tend to occur. The model features conditional elements from GAN and PatchGAN Discriminator together with EfficientNet-Lite classifier for extracting detailed features while achieving accurate spoof prediction. The training and testing process occurs on MSU-MFSD before moving to CASIA-FASD for cross-domain testing. Experimental results show that the model achieves 98.12% accuracy with 1.56% APCER and 1.88% ACER on MSU-MFSD whereas maintaining 92.45% accuracy in cross-dataset evaluation. The experimental findings confirm that the proposed structure proves to be reliable for practical use conditions. This model shows high efficiency in processing power needs and quick inference operations which enables its use on edge devices. The study introduces an attention-enabled GAN approach for face anti-spoofing which serves to combine high accuracy together with practical real-world operation and superior efficiency.