On feasibility of GAN-based fingerprint morphing

Morphing of two fingerprints is shown to be feasible when using a model-based minutia-oriented approach, in which original fingerprint images are cut to two almost equal parts. The morphed fingerprint is a result of assembling two parts of different fingerprints along a cut line. It is important that each part of an original fingerprint in the morphed fingerprint contains enough minutiae to enable the successful matching between the morphed and both original fingerprints. The major drawback of this approach is that the resulting fingerprint often does not appear realistic. Another way to morph fingerprints is exploiting neural generative models. The projections of fingerprints onto the latent space of the generator network are blended and the resulting latent vector is fed to the generator network. In contrast to the model-based approach, a morphed fingerprint almost always appears realistic, but there is no guarantee that it matches successfully both original fingerprints, unless the identity prior is included into the generation process. This paper discusses the advantages and pitfalls of fingerprint morphing using generative adversarial networks (GAN). We experimentally show that GAN-based fingerprint morphing is feasible for creating double-identity fingerprints but fails to anonymize fingerprints i.e. create new virtual identities.