MMFBA_ESTNet: An Enhanced Swin Transformer – Based Multi-Model Finger Biometric Authentication Framework

Abstract

Biometric identification is currently the subject of active research to prevent unauthorized use. Finger biometrics creates spoofing attacks by creating fake biometric finger images. Distinguishing a true biometric image from a fake one with the naked eye is a difficult process. Therefore, the proposed work has introduced a highly effective multi-model-finger biometric authentication (MMFBA) based on finger vein, fingerprint, and knuckle. The proposed model has relied on a Densely connected semantic dilated multi-head attention conventional-based Enhanced Swin Transformer (EST) network i.e. named MMFBA_ESTNet. This provides the best architecture for low-level (edges, corners, lines, and texture), mid-level (combined features), and high-level (complex patterns) feature learning. In addition, multi-head attention helps to establish priorities to improve learning efficiency. The implementation of the proposed profile is performed on the Python platform and evaluated using parameters such as accuracy, precision, F1, area under the curve, false positive rate, false rejection rate, and false acceptance rate. The results of the experiment show that the proposed strategy achieved 0.986 accuracy at a 70% learning rate and 0.99 accuracy at an 80% learning rate.