A Comprehensive Comparison of CNN-Based Deep Learning Architectures for Fingerprint Authentication

This study explores the integration of CNNs in fingerprint biometric systems, tackling the primary challenge of biometric authentication by accurately matching similar pairs while discarding the dissimilar ones. Three different deep learning frameworks are compared. The first model employs custom CNN-based classification model. The second includes transfer learning approach from the VGG-16 network, while the third implements an image similarity learning strategy based on merging CNN sub-networks. To increase confidence in the results, we assess the models using unlabeled data, providing a more accurate representation of the impostor distribution. In contrast to both the first and second models, the third model achieves a competitive EER score of 1.7%, highlighting an effective strategy for authentication with CNN-driven biometric systems.