Dissimilarity vector representation for retinal template protection

Biometric systems extract features from biometric samples. The extracted features are called “biometric templates”. Traditional biometric systems store biometric templates of the users inside their databases. Biometrics, although convenient for the purpose of human recognition, contain private information about individuals. From leaked biometric data, private information can be derived. Therefore, it is essential to protect biometric templates. Because biometric data is fuzzy, it cannot be protected using conventional encryption methods. In the past 20 years, researchers have proposed many biometric template protection schemes that can deal with fuzziness in the captured biometric data. Many of these schemes fail in providing sufficient information to show that these schemes protect privacy of their users. ISO/IEC 24745 introduces and updates the standard requirements for evaluating the privacy protection during development of biometric template protection schemes. This thesis reviews retina as a biometric characteristic. It discusses the privacy concerns related to the use of biometric data and outlines the motivations behind developing a retinal template protection scheme. A comprehensive literature review on biometric template protection schemes, and the challenges and concerns related to them is provided in our work. We introduce cohort-based dissimilarity vectors for protection of retinal images. The proposed scheme extracts sparse spatial graph templates from the vascular pattern in retinal images. We review the recognition performance of the proposed retinal template protection scheme and compare it with the performance of the conventional retinal recognition system that compares the unprotected graph templates. Our results show that the drop in recognition accuracy in the protected domain is at most 1−2% which is very encouraging. This is a desired property for a biometric template protection scheme to have comparable performance in recognition for protected and unprotected templates. A known inverse attack exists which can reconstruct face images using their respective comparison scores with over 70% success rate. We evaluate its success rate when applied to reconstruct biometric samples/templates from our protected templates, i.e., our cohort-based dissimilarity vectors. The results of this evaluation demonstrate that templates/images reconstructed by this attack do not look similar enough to the original attacked templates/images to be able to penetrate the biometric system. The success rate of this attack is bellow 0.3% in all of our experiments. This suggests that our protected templates are irreversible which is compliant with the first ISO/IEC 24745 requirement. We show that it is possible to generate a large retinal image dataset applying generative adversarial networks (GANs) using a limited training set with only approximately 800 training images. Generating this dataset was essential since for evaluating unlinkability of the protected templates, access to a large number of retinal samples was required, and no publicly available high quality databases currently exist. Finally, we use “the general framework to evaluate unlinkability of biometric template protection schemes” to evaluate the local and global linkability levels of our proposed template protection scheme. Applying this general framework allowed comparison of thelinkability level of our proposed scheme with that of stat-of-the-art schemes (developed for other biometric characteristics). The results of our experiments show comparable low linkability scores to the stat-of-the-art schemes which is compliant with the second ISO/IEC 24745 requirement. Overall, the proposed system in this thesis meets all the desired and required ISO/IEC 24745 properties for developing a biometric template protection scheme. This scheme is, to the best of our knowledge, the first retinal template protection scheme in the literature.