Implementation of MobileNetV4 and Efficient Channel Attention in Anti-Spoofing Face Attack Detection

Abstract

Face Anti-Spoofing (FAS) is essential for preventing presentation attacks in biometric systems, yet deploying robust models on mobile devices remains a challenge due to computational constraints. This study proposes a lightweight FAS model integrating the MobileNetV4 architecture with an Efficient Channel Attention (ECA) module. The ECA mechanism is designed to enhance the network’s ability to detect subtle spoofing artifacts, such as texture anomalies, with negligible computational overhead. The model was evaluated using a dataset of 6,400 images, comprising both bona fide and attack presentations. Experimental results demonstrate robust performance, achieving an overall accuracy of 99.69%, 100% precision, and an Average Classification Error Rate (ACER) of 0.25%. Crucially, the model yielded a Bona Fide Presentation Classification Error Rate (BPCER) of 0.00%, ensuring that no genuine users are falsely rejected. While the baseline architecture provided a strong benchmark, the proposed attention-enhanced framework offers a viable trade-off between security and usability, providing a computationally efficient solution suitable for real-time mobile authentication.