DSMIKervNet: dual stream multi-kernel integration kervolutional neural network for face spoof detection

This study proposes DSMIKervNet, a Dual-Stream Multi-Kernel Integration Kervolutional Neural Network for face anti-spoofing. The model introduces a novel multi-kernel kervolutional layer that integrates polynomial, wave, inverse multiquadric, and Laplacian kernels to learn complex non-linear spoofing patterns. The dual-stream design processes RGB and LBP images in parallel, enhancing both spatial and texture-based feature learning. These are fused using the integrated multi-kernel approach to enhance discriminative power. Extensive experiments on five benchmark datasets, MSU-MFSD, NUAA, LCC-FASD, ROSE-Youtu, and 3DMAD, show that DSMIKervNet achieves up to 15–25% higher accuracy and 40% lower ACER than existing CNN-based methods. These results highlight its superior robustness, adaptability, and potential for real-world biometric authentication systems.