Face Morphing Attack Detection and Localization Based on Feature-Wise Supervision
Le Qin, Fei Peng, Min Long
Abstract
To strengthen the security of face recognition systems to morphing attacks (MAs), many countermeasures were proposed. However, in the existing face morphing attack detection (MAD), the deep networks trained by classical score-level losses are weak in characterizing the intrinsic morphing patterns of different MAs, and they also cannot be directly applied to differential MAD scenarios. To this end, this paper presents a method for detecting and locating face MAs by the use of feature-wise supervision. It constructs the fine-grained classification loss on the basis of different morphing patterns, and designs the similarity-based and distance-based differential losses according to the properties of differential MAD scenarios. The experimental results and analysis show that the fine-grained classification loss can locate the local morphed areas after detecting MAs, while the differential losses are able to improve the generalization ability of MAD methods to unseen MAs, and can enhance the robustness of MAD methods to low-resolution and non-frontal probe face images.