School of Computing and Information Sciences
Mozhgan Azimpourkivi is a Ph.D. candidate in the School of Computing and Information Sciences at Florida International University, co-advised by Bogdan Carbunar and Umut Topkara. She received a Master’s degree in Computer Science from Florida International University in 2015 and a Master’s degree in Information Technology from the Sharif University of Technology, Iran in 2011. Mozhgan’s research interests include usable security, mobile authentication, image data protection, verification in social media, and deep neural networks. Her current research focuses on improving security mechanisms to ensure they are usable and effective in practice.
Mobile and wearable devices are popular platforms for accessing online services. However, the small form factor of such devices, makes a secure and practical experience for user authentication, challenging. Further, online fraud, that includes phishing attacks, has revealed the importance of conversely providing solutions for usable authentication of remote services to online users. In this thesis, we introduce image-based solutions for mutual authentication between a user and a remote service provider. First, we propose and develop Pixie, a two-factor, object-based authentication solution for camera-equipped mobile and wearable devices. We further design ai.lock, a system that reliably extracts from images, authentication credentials similar to biometrics.
Second, we introduce CEAL, a system to generate visual key fingerprint representations of arbitrary binary strings, to be used to visually authenticate online entities and their cryptographic keys. CEAL leverages deep learning to capture the target style and domain of training images, into a generator model from a large collection of sample images rather than hand curated as a collection of rules, hence provides a unique capacity for easy customizability. CEAL integrates a model of the visual discriminative ability of human perception, hence the resulting fingerprint image generator avoids mapping distinct keys to images which are not distinguishable by humans. Further, CEAL deterministically generates visually pleasing fingerprint images from an input vector where the vector components are designated to represent visual properties that are either readily perceptible to human eye, or imperceptible yet are necessary for accurately modeling the target image domain.
We show that image-based authentication using Pixie is usable and fast, while ai.lock extracts authentication credentials that exceed the entropy of biometrics. Further, we show that CEAL outperforms state-of-the-art solution in terms of efficiency, usability, and resilience to powerful adversarial attacks.