3D Face Anti-Spoofing With Dense Squeeze and Excitation Network and Neighborhood-Aware Kernel Adaptation Scheme

Abstract

Face anti-spoofing is a critical challenge in biometric security systems, where sophisticated spoofing techniques pose significant threats. To enhance the effectiveness and efficiency of the methods employed for face anti-spoofing, this paper presents a new lightweight 3D face anti-spoofing framework characterized by several advanced mechanisms. To this purpose, the proposed architecture introduces DenseNet, a Squeeze and Excitation mechanism, and a new computational component called Neighborhood-Aware Kernel Adaptation (NAKA) that adaptively modifies 3D convolution kernels according to spatial proximity. Initially, an adaptive thresholding-based wavelet decomposition is employed for image denoising, followed by cross-channel attention to improve feature learning. Finally, Multiple Instance Learning (MIL) is used to address face anti-spoofing for the first time by modeling the spatial and temporal variations across facial areas. We validate our framework on two publicly available datasets: CelebA-Spoof and CASIA-SURF. We compared the performance of our proposed framework with several state-of-the-art methods using Classification accuracy, Precision, Recall, F1-score, ACER, APCER, and BPCER. Our model realizes 99.62% on the CelebA-Spoof and 99.86% on CASIA-SURF datasets. The proposed approach realized superior results in terms of high classification accuracy (99.62% and 99.86%), precision (99.85% and 99.83%), recall (99.39% and 99.84%), F-score (99.62% and 99.84%), ACER (0.0038/ 0.0014), FPR (0.0015/ 0.0013), APCER (0.0015/0.0016), and BPCER (0.0061/0.0013). These results are compared with 10 state-of-the-art methods to show the effectiveness of our approach in outperforming existing methods. The global comparative results reveal that the proposed approach is relatively effective in determining masked and true face scans.