Multiattention-Net: A Novel Approach to Face Anti-Spoofing with Modified Squeezed Residual Blocks

Abstract

Introducing a novel Attack-agnostic Face Anti-spoofing framework, this paper addresses the challenge of determining the authenticity of a captured face in face recognition systems. Current methods, trained on existing fake faces, often lack generalization and perform poorly against unseen attacks. The proposed framework presents a fresh approach to face anti-spoofing, leveraging modified squeezed residual blocks and attention mechanisms. Convolutional layers within the Multiattention-Net architecture capture spatially hierarchical features, enhancing feature representation and improving the network’s sensitivity to critical features. These spatial features are refined through a dual attention block to emphasize important features. The squeeze-and-excitation (SE) mechanism further enhances the representation by recalibrating channel-wise responses to emphasize informative features, incorporating global average pooling and channel-wise excitation. The Multiattention-Net achieves a balanced trade-off between feature richness and computational efficiency, demonstrating superior performance in face anti-spoofing tasks. Experimental results on benchmark datasets validate the effectiveness of this approach, highlighting its potential for real-world applications in security and biometric authentication.