Privacy-preserving Deep-learning Models for Fingerprint Data using Differential Privacy

Abstract

Fingerprint recognition is a widely adopted biometric authentication method that leverages the unique characteristics of fingerprints to identify individuals. Its applications range from access control and authentication to forensic science, making the development of a robust, precise, and secure model for fingerprint recognition and analysis of paramount importance. Recently, deep learning and machine learning models have shown promise in this field, however, the use of these models raises significant privacy concerns as there is a potential for private fingerprint data to be compromised. This research aims to address these concerns by incorporating differential privacy techniques to protect the privacy of fingerprints. It provides evidence that the use of differential privacy technique leads to acceptable trade-off between preserving the privacy of fingerprints and accuracy of fingerprint recognition systems while maintaining robustness against model inversion attacks. With a noise multiplier of 0.01, the verification model attains a good accuracy of 89.32% at privacy budget ε = 1.2, and less than 80% at lower privacy budget for the identification model.