On Self-Supervised Learning of Vision Transformers for Fingerprint Presentation Attack Detection
Nadeem, Maryam
Nadeem, Maryam
Author
Supervisor
Department
Machine Learning
Embargo End Date
Type
Thesis
Date
2023
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
With the rapid growth of fingerprint-based biometric systems, it is imperative to ensure the security and reliability of these systems, given the immense security vulnerabilities that they are exposed to. Fingerprint presentation attack detection (FPAD) is an example of a threat to these systems. Various studies done on fingerprint attack detection have shown their significance in this area. Despite all the work done over the past years, a shortcoming of these detection systems is that since fingerprint images captured from different fingerprint sensors exhibit unique characteristics because of the various fingerprint sensing technologies, sensor noises, and resolutions these methods fail to generalize well on unseen data and thus are still prone to attacks. Therefore, it is important to enhance the generalizability of these FPAD algorithms in cross-sensor and cross-material settings. In this work, we introduce two self-supervised learning approaches using Masked Image Modeling (MIM) and Image-text alignment based techniques utilizing vision transformer backbones to create robust models for improving cross-sensor performance without any knowledge of the target sensor. Although both the approaches demonstrate excellent generalizability, we show that MIM based framework outperforms the text-alignment based framework for the task of FPAD.
Citation
M. Nadeem, "On Self-Supervised Learning of Vision Transformers for Fingerprint Presentation Attack Detection", M.S. Thesis, Machine Learning, MBZUAI, Abu Dhabi, UAE, 2023.