Face recognition is widely used for biometric authentication in applications such as mobile devices, financial services, intelligent surveillance, access control, and border security. However, face recognition systems remain vulnerable to presentation attacks, including printed photographs, replay attacks, and three-dimensional masks. Although recent deep learning-based face anti-spoofing (FAS) methods have achieved substantial improvements, many existing approaches still involve considerable computational cost, provide limited emphasis on fine-grained spoof artifacts, and face challenges in effectively integrating heterogeneous representations. To address these limitations, this paper proposes a Lightweight Face Anti-Spoofing Framework with Multi-Modal Representation, Spoof Artifact Enhancement, and Adaptive Feature Fusion. The framework starts from a single RGB facial image and constructs complementary Depth and Near-Infrared (NIR) representations using a Depth and Near-Infrared Construction Module (DNCM), rather than requiring dedicated Depth or NIR sensors. A shared EfficientNetV2 backbone is then employed to extract features from the three representations with reduced computational redundancy. The proposed Spoof Artifact Enhancement Module (SAEM) emphasizes subtle spoof-specific visual cues, while Cross-Modal Consistency Learning (CMCL) reduces representation discrepancies across the constructed modalities. Subsequently, the Adaptive Feature Fusion Module (AFFM) dynamically weights the refined representations according to their discriminative contribution. Extensive experiments on CelebA-Spoof, CASIA-SURF, HQ-WMCA, and SiW-M demonstrate the effectiveness of the proposed framework. The framework achieves accuracies of 98.16%, 97.54%, 99.08%, and 97.18%, with corresponding ACER values of 2.87%, 4.36%, 2.03%, and 4.18%, respectively. Across five independent runs, statistical analysis further indicates consistent performance with significant improvements over the selected baseline. In addition, the framework requires only 9.8 million parameters and 2.1 GFLOPs and achieves an inference speed of 82 FPS, demonstrating a favourable balance between detection effectiveness and computational efficiency. The results indicate that multi-modal representation combined with explicit spoof artifact enhancement and adaptive feature fusion can provide an efficient solution for robust and real-time face anti-spoofing.

Face recognition systems are increasingly deployed in security-critical applications, but remain vulnerable to presentation attacks such as printed photographs, replay videos, and three-dimensional masks. Although recent face anti-spoofing methods exploit complementary information from RGB, depth, and near-infrared (NIR) representations, existing approaches primarily focus on feature fusion and do not explicitly model the intrinsic consistency relationships among 
these representations. This limitation can reduce robustness against sophisticated spoofing attacks and cross-dataset variations.
This paper proposes a novel Cross-Modal Consistency Learning (CMCL) framework for robust face anti-spoofing using the original RGB input together with RGB-derived depth and NIR representations. The depth and NIR representations are constructed from RGB inputs during the modality preprocessing stage and subsequently processed together with RGB using independent Swin Transformer encoders. A cross-modal attention fusion module adaptively integrates complementary appearance, geometric, and spectral information, while a consistency learning module explicitly encourages feature coherence for genuine samples and emphasizes cross-modal discrepancies associated with spoof attacks. The consistency module is used only during training and removed during inference, avoiding additional deployment overhead.
Extensive experiments are conducted on CASIA-SURF, WMCA, CelebA-Spoof, and MSU-MFSD using intra-dataset, cross-dataset, ablation, feature-space, modality-consistency, and statistical analyses. The proposed CMCL achieves an ACER of 0.70% and accuracy of 99.3% on CASIA-SURF, while achieving an ACER of 2.55% and accuracy of 97.9% on WMCA. The results demonstrate improved spoof detection performance, representation discrimination, and cross-dataset generalization compared with the evaluated state-of-the-art methods. The proposed framework provides a consistency-driven and computationally practical approach for robust face anti-spoofing.