Publication Date

Spring 5-25-2021

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Fabio Di Troia

Second Advisor

Nada Attar

Third Advisor

Navrati Saxena

Keywords

Biometric authentication, Convolutional Neural Network, Deepfake, Generative Adversarial Networks

Abstract

Biometric systems are referred to those structures that enable recognizing an individual, or specifically a characteristic, using biometric data and mathematical algorithms. These are known to be widely employed in various organizations and companies, mostly as authentication systems. Biometric authentic systems are usually much more secure than a classic one, however they also have some loopholes. Presentation attacks indicate those attacks which spoof the biometric systems or sensors. The presentation attacks covered in this project are: photo attacks and deepfake attacks. In the case of photo attacks, it is observed that interactive action check like Eye Blinking proves efficient in detecting liveness. The Convolutional Neural Network (CNN) model trained on the dataset gave 95% accuracy. In the case of deepfake attacks, it is found out that the deepfake videos and photos are generated by complex Generative Adversarial Networks (GANs) and are difficult for human eye to figure out. However, through experiments, it was observed that comprehensive analysis on the frequency domain divulges a lot of vulnerabilities in the GAN generated images. This makes it easier to separate these fake face images from real live faces. The project documents that with frequency analysis, simple linear models as well as complex models give high accuracy results. The models are trained on StyleGAN generated fake images, Flickr-Faces-HQ Dataset and Reface app generated video dataset. Logistic Regression turns out to be the best classifier with test accuracies of 99.67% and 97.96% on two different datasets. Future research can be conducted on different types of presentation attacks like using video, 3-D rendered face mask or advanced GAN generated deepfakes.

Share

COinS