Özet:
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.
