Design and Implementation of Hybrid CNN-Transformer Architecture for Face Anti-Spoofing

Abstract

Biometric face recognition systems are progressively being used in security-sensitive areas and are vulnerable to presentation attacks (PAs) such as printed photograph recaptures, video replay attacks and 3D silicone mask attacks. Face Anti-Spoofing (FAS) also known as liveness detection is the first line of defence against such attacks. In this paper, we introduce HybridLivenessNet, a novel deep neural architecture that combines MobileNetV3-Small as a computationally efficient convolutional feature extractor with a multi-head Transformer encoder, thus capturing diverse and complementary representations at the local texture and global spatial levels. The proposed approach is trained and tested on the LCC FASD dataset with 11,247 annotated facial images in real and spoofed classes. During a 100-epoch training process with AdamW optimization, ReduceLROnPlateau adaptive scheduling strategy, strong dropout regularization and multi-pattern data augmentation, HybridLivenessNet attains its peak validation accuracy (98.47%) at epoch 9 and provides stable performance ranging from 98.03% to 98.27% from epoch 20 till 100, which suggests a stable convergence without over-fitting. The method offers a good balance of computational cost and detection performance, making it ideal for implementation in real-life systems with limited resources. In addition, we analyse and discuss training dynamics, learning rate decay, and epoch-wise accuracy curves.