Detection of Face Morphing Attacks Using Deep Learning

The possibility of various illegal acts increases when face recognition and authentication systems fail. Current face recognition systems can be easily compromised by various biometric techniques. This study focuses on attack detection using morphing. An effective detection system is proposed that accounts for differences in age, lighting, eyewear, and headgear. The system uses a deep learning-based feature extractor in conjunction with a classifier. Additionally, we propose combining features and enhancing images to improve detection accuracy. A multipurpose dataset, including Morph-2 and Morph-3 images generated by advanced techniques with human participation, is also under development. Morph-3 images, due to their photorealistic quality, might be difficult to detect. Notably, no prior study has considered Morph-3 images. Previous studies using free programs and code scripts produced morphs that were less realistic than those generated by professional morphing software in a morph attack scenario. Morphs are created using eight unique face databases: Celebrity2000, Extended Yale, FEI, FGNET, GT-DB, MULTI-PIE, FERET, and FRLL to include all possible variations. After assessing the data across several experimental circumstances, we find that the proposed method yields favorable results.