RepGhostFAS: Lightweight and Efficient Face Anti-Spoofing for Internet-of-Things (IoT) and Edge Devices

Face Anti-Spoofing (FAS) or Face Presentation Attack Detection (FacePAD) is essential for securing face recognition systems against presentation attacks. However, existing approaches suffer from poor generalization under domain shift conditions and often incur high computational costs, which limit their discriminative capability and practical deployment, especially in resource-constrained edge and Internet-of-Things (IoT) environments. To address these challenges, we introduce RepGhostFAS, a lightweight and efficient FAS framework that leverages the RepGhost architecture to achieve strong discriminability with low computational cost, marking its first application in the FAS domain. We tackle the domain shift problem using ensemble-based feature fusion strategies that extract complementary information from multiple domains. To further reduce model complexity, we employ knowledge distillation to transfer stronger representations from an ensemble teacher model to a compact student network. Extensive experiments on eight benchmark datasets, namely Replay-Attack, Replay-Mobile, OULU-NPU, ROSE-Youtu, SiW-Mv2, SynthASpoof, SOTERIA, and UniAttackData demonstrate that RepGhostFAS achieves competitive or superior performance while requiring significantly lower computational cost than existing methods. Overall, the proposed framework offers a robust, cost-effective, and deployment-friendly solution for real-world FAS applications on IoT and edge devices.