Using deep learning techniques and genetic-based feature extraction for presentation attack mitigation

Abstract

Abstract— Biometric authentication systems are becoming more prevalent for commercial use with computers and smart devices. Biometric systems also have several vulnerable points that can be exploited by a hacker to gain unauthorized access to a system. Replay attacks focus on capturing feature extractors (FEs) during transmission, decrypting, and replaying for illegal access. The Genetic and Evolutionary Feature Extraction (GEFE) technique, developed at North Carolina A&T State University, recently showed promising results in mitigating replay attacks in combination with a feature selection algorithm. Biometric-based presentation attacks, the focus of this work, is another biometric system vulnerability primarily focused on presenting a biometric sample of quality to illegally gain access to secured data. Recently, deep learning techniques to mitigate presentation attacks have shown promising results. However, the accuracy of deep learning-based biometric presentation attack detection (PAD) methods are limited by the quality of the samples provided. In absence of large sets of original biometric sample data, data augmentation has been shown to be successful in generating synthetic biometric image data and improving the performance of deep learning techniques applied. The novelty of this paper lies in the following two aspects: First, a data augmentation technique with Generative Adversarial Networks (GANs) is used to generate comparative synthetic (spoofing) dataset. With the proliferation of deep fakes in media, this technique should provide insight on the GAN technique often used. Once properly trained, the synthetic images are used to create spoofing datasets. Second, the GEFE technique is used in combination with the GANs to generate improved anti-spoofing feature extractors optimized to mitigate presentation attacks. The combination of GEFE and GANs is used to identify those discriminative biometric features used to mitigate synthetic presentation attacks. The GEFE ​+ ​GAN technique outperforms the LBP and GEFE techniques alone in overall identification and verification results on spoofing datasets.