Biometric authentication has become increasingly prevalent in modern security systems, but there are still limitations in terms of accuracy and reliability. One major challenge is the quality of the biometric data itself, which can be affected by various factors such as sensor lens issues or algorithm training. As part of my undergrad thesis at IIT Mandi, I worked on a solution using Generative Adversarial Networks (GANs) to improve the quality of biometric data.

Specifically, we focus on fingerprint recognition and propose a method for mapping low-quality fingerprint images to high-quality ones of the same subject using CycleGAN. We also incorporate a Siamese loss function in addition to the cyclic loss to further improve the results. We demonstrate the effectiveness of our approach through qualitative and quantitative results using different GAN models and datasets. Additionally, we show how our approach can improve fingerprint matching using a trained Siamese Network and Minutiae Cylinder code.

Overall, our project highlighted the potential for GANs to improve biometric authentication and addresses a significant challenge in the field of computer security.

Link to the report