Fingerprint Template Invertibility: Minutiae vs. Deep Templates

Much of the success fingerprint recognition is attributed to minutiae-based representation. It was believed that minutiae templates could not be inverted obtain a high fidelity image, but this assumption has been shown false. The deep learning resulted in alternative representations (embeddings), hope they might offer better accuracy as well non-invertibility network-based templates. We evaluate whether suffer from same reconstruction attacks show while template can produce image matched its source are more resistant than In particular, reconstructed images yield TAR about 100.0% (98.3%) @ FAR 0.01% for type-I (type-II) using state-of-the-art commercial matcher, when tested on NIST SD4. corresponding attack performance matcher yields less 1% both and type-II attacks; however, network, achieve 85.95% (68.10%) attacks. Furthermore, what missing previous inversion studies an evaluation black-box performance, which we perform 3 different matchers. conclude generated by inverting highly susceptible white-box evaluations, evaluations comparatively evaluations.