Abstract
Face recognition systems have become prevalent in mobile devices and security applications, increasing the demand for robust face presentation attack detection. Early efforts based on handcrafted features struggled to cope with variations in illumination, pose, and attack modalities, prompting a transition toward deep learning solutions capable of extracting subtle discriminative cues. A novel architecture built upon an EfficientNet-V2 backbone, combined with a Shuffle Attention module and Fourier heads, was developed to capture both spatial and frequency domain characteristics. A dual-path approach processes each input face image through conventional convolutional blocks and a 2D Discrete Fourier Transform path, with dedicated Fourier heads reconstructing frequency maps that reveal minute discrepancies between genuine and spoofed presentations. Experimental evaluation on the Oulu-NPU dataset demonstrates strong performance across four protocols, including robust detection under varying environmental conditions, low error rates with novel attack types, and consistent results across different sensor inputs. Metrics such as APCER, BPCER, and ACER validate the method’s ability to distinguish between live and fake faces reliably. The outcomes suggest that combining spatial and frequency cues addresses limitations observed in earlier approaches, offering valuable insights for deployment in security-sensitive applications and setting a strong foundation for future research in face anti-spoofing.