Summer Research Fellowship Programme of India's Science Academies

Signature replication and fingerprint bio-metric forging using generative adversarial network

Vivek Aswal

Vivekanand Educational Society’s Institute of Technology, Mumbai, Maharashtra, 400074

Professor Siba K. Udgata

School of Computer and Information Sciences, University of Hyderabad, Telangana, 500046


Generative Adversarial Networks (GANs) are one of the most prominent advancements in neural networks and employ a probabilistic approach to generate similar data with some variations compared to that of the target (original) dataset. This is done by achieving equilibrium via an adversarial process between two neural networks, the generator and the discriminator. If GANs are applied in the field of digital identification security systems such as signatures and fingerprint biometrics scanners, then they can be used to test the reliability of modern security identification systems via neural networks. The objective is to generate realistic looking forged images which could deceive the discriminator to treat this fake images as authentic ones. In this research, a DCGAN model has been initially developed based on the popular datasets such as MNIST, Fashion-MNIST and for color images CIFAR10 dataset. Further on to test these GANs on real life situations, these have been modified and implemented on a signature dataset and a fingerprint database of different set of individuals and the results presented are quite encouraging. The quality of the resulting images produced resembles closely to that of the genuine images. To determine the performance of the generated data, three types of datasets have been considered: training set (real images), selected generated fake images (selected manually based on qualitative analysis) and test set (generated images produced by the GAN). The accuracy characteristics of these sets have been compared with that of an established GAN network for MNIST database. However, the accuracy of the discriminator in both the cases is found to be fluctuating. Also the stability of DCGAN to reach the Nash equilibrium in case of identification purposes poses a big challenge for discriminator’s accuracy of the generated fake images.

Keywords: discriminator, generator, process of training, GAN architecture, discriminator accuracy, GAN stability

Written, reviewed, revised, proofed and published with