Face Morphing Attack Detection.

Abstract

Failure of facial recognition and authentication system may lead to several unlawful activities. The current facial recognition systems are vulnerable to different biometric attacks. This research focuses on morphing attack detection (MAD). The research proposes a robust detection mechanism that can deal with variation in age, illumination, eye and head gears. A deep learning based feature extractor along with a classifier is adopted. Additionally, image enhancement and feature combination is proposed to augment the detection results. A versatile dataset is also developed that contains Morph-2 and Morph- 3 images, created by sophisticated tools with manual intervention. Morph-3 images can give more realistic appearance and hence difficult to detect. Moreover, Morph-3 images are not considered in the literature before. Professional morphing software depicts a more realistic morph attack scenario as compared to the morphs generated in the previous work from free programs and code scripts. Eight face databases are used for creation of morphs to encompass the variation. These databases are Celebrity2000, Extended Yale, FEI, FGNET, GT-DB, MULTI-PIE, FERET, FRLL. Results are investigated using multiple experimental setups and it is concluded that the proposed methodology gives promising results.