Domain free deep learning based security models for cyberphysical systems

Abstract

With the developments in digital age and growing interest in IoT, a variety of institutions and organizations have started to digitalize their systems. As a consequence of these digitalizations, security of collecting, accessing and transferring great amounts of private data via internet connection have became an important issue. In particular, protection of data collected, transmitted and stored on cyberphysical systems (CPS) such as security systems have gained great importance. Recently, many studies have been conducted using state-of-the-art Deep Learning (DL) algorithms for security systems. However, despite their groundbreaking results, most of these studies either are biased to some particular datasets or too complex and computationally-expensive to be used in real time. Moreover, DL algorithms require a lot of input data to extract the most informative feature representations and become disadvantageous in real situations, where imbalances among classes and unlabelled samples in input data are quite common. Therefore, first goal of this dissertation is to conduct a comprehensive research and to study AI-based new approaches for two different domains of security-themed systems: biometric systems and cybersecurity. In particular, new Capsule-based feature representations for these domains are investigated in detail and these representations are compared with their equivalent state-of-the-art algorithm-based models for the first time. Second goal is to conduct an experiment on Transfer Learning (TL) for cybersecurity, where features are in time-domain and benchmark datasets do not share sufficient common feature space with each other like image-domain counterparts such as biometric systems to use pre-trained network in 1D. In addition, possible scenarios are examined to adapt security systems into different domains and generalize by using available benchmark datasets with different traffic collection as well as feature spaces.