<?xml version="1.0"?>
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:dc="http://purl.org/dc/elements/1.1/"><rdf:Description rdf:about="https://repozitorij.uni-lj.si/IzpisGradiva.php?id=187178"><dc:title>Behavioural Biometrics-Based Context-Adaptive User Authentication</dc:title><dc:creator>Krašovec,	Andraž	(Avtor)
	</dc:creator><dc:creator>Pejović,	Veljko	(Mentor)
	</dc:creator><dc:creator>Baldini,	Gianmarco	(Komentor)
	</dc:creator><dc:subject>user authentication</dc:subject><dc:subject>deep learning</dc:subject><dc:subject>authentication system evaluation</dc:subject><dc:subject>behavioural sensor-based IoT dataset</dc:subject><dc:description>The digital age brought many improvements to daily living; however, it also brought some annoyances, one is the process of user authentication. From passwords to fingerprints, no contemporary method is simultaneously secure, private, and convenient for the user. Authentication based on user behaviour has the potential to cover all three. However, alongside other hurdles, the inconsistency of user behaviour leads to poor security performance. Therefore, in this dissertation, we aim to design a system capable of adapting to behavioural inconsistencies. Simultaneously, we aim to address other concerns in the field. Limited availability of data, heterogeneous evaluation methodology and taxonomy, and inadequate consideration of privacy are all still open topics. First, we collect, process, and publish a novel dataset including about 40 hours of sensor data, set in an internet of things (IoT) environment, and collected from 54 volunteers. Given the multitude of utilised sensing modalities, the dataset's applicability goes beyond cybersecurity as it can equally be applied in the areas of human-computer interaction and cognitive science. Next, we devise a novel privacy-by-design algorithm for training an authentication system and show that it retains the authentication-unrelated information private even when an adversary attempts to infer it. Furthermore, we design, build, and train a behaviour-adaptive authentication system with equal error rates as low as 2%. We achieve this result on data from an unseen user context and it represents at least a 4-fold improvement over the state of the art. Moreover, we propose an evaluation methodology that enables like-to-like comparison of different authentication systems. Lastly, based on the methodology, we build an evaluation toolbox and share it with the research community. The work we accomplish demonstrates that user authentication can indeed be secure, private, and convenient, alleviating everyone of this modern annoyance.</dc:description><dc:date>2026</dc:date><dc:date>2026-09-09 10:35:00</dc:date><dc:type>Doktorsko delo/naloga</dc:type><dc:identifier>187178</dc:identifier><dc:language>sl</dc:language></rdf:Description></rdf:RDF>
