Quadrant-Based Bi-Level Self-Attention Feature Extraction for Face Morphing Attack Detection

Abstract

Face morphing attack is a kind of security breach facing facial biometric systems. Face morphing is the process of blending facial images from two or more different people such that any of the constituent faces can be verified or authenticated as check points with the blended morphed face image. If such blended images are successfully used for verification or authentication under a biometric verification system, then face morphing attack is regarded to have taken place. Existing detection methods often struggle with subtle and localized morphing artifacts. We propose a novel Quadrant-Based Bi-level Self-Attention Feature Extraction for Face Morphing Attack Detection (QBSAF) designed to enhance the detection of face morphing by focusing on low-level pixels localized regions of an image to extract low-level features pertinent to the morphing artefacts. Our method offers several advantages, including improved precision in detecting visual anomalies and coherence across the entire image. Additionally, it highlights important features and patterns that can emerge when considering the full representation. Experimental results demonstrate the effectiveness of QBSAF, with its 98-100% test set accuracy performance across multiple morphing datasets with precision, recall, FI-score and AUC between 0.98 -1.00. The results indicate that QBSAF significantly well in detecting various morphing techniques, including StyleGAN, WebMorph, OpenCV Morph, and FaceMorpher. Moreover, our method maintains high precision and recall even in challenging scenarios, proving its robustness in real-world biometric security applications. These findings suggest that our approach can serve as a reliable solution for enhancing security against face morphing attacks in biometric verification systems.