Secure internet of things using deep learning: applying transformer-based NLP techniques for enhanced IoT security

  • Ahmad Mohammad Al-Zuraiqi

Student thesis: Doctoral Thesis › Doctor of Philosophy

Abstract

The rapid proliferation of Internet of Things (IoT) devices across domains such as wearables, smart home systems, and industrial automation necessitates robust security measures. This study addresses critical IoT security challenges through three interconnected contributions. First, we developed domain-specific datasets that map IoT firmware vulnerabilities to hardware configurations, enriched with metadata and synthetic augmentation to mitigate bias. Static analysis tools, such as the Firmware Analysis and Comparison Tool (FACT), were used to identify recurring vulnerabilities in various categories of IoT devices.

Second, we fine-tuned transformer-based large language models (LLMs) tailored for the IoT domain. By employing advanced alignment techniques, including Contrastive Preference Optimization (CPO) and Kahneman-Tversky Optimization (KTO), these models demonstrated enhanced predictive capabilities and contextual accuracy. Their integration with static analysis workflows resulted in a hybrid framework, improving vulnerability detection and scalability in resource-constrained environments.

Third, a rigorous empirical evaluation was conducted, comparing LLMs with traditional tools across metrics such as computational efficiency, energy consumption, and detection accuracy. This analysis underscores the transformative potential of AI-driven methodologies in improving IoT security.

The findings culminate in a proof-of-concept for secure-by-design IoT systems, aligned with OWASP ISVS standards and offering practical recommendations for real-time adaptability. Future research is directed toward federated learning and cross-modal vulnerability assessments, fostering scalable and ethically sound IoT security frameworks.

Thesis is embargoed until 31 July 2026.
Date of AwardJul 2025
Original languageEnglish
Awarding Institution
  • Queen's University Belfast
SupervisorPaul McMullan (Supervisor) & Desmond Greer (Supervisor)

Keywords

  • IoT security
  • static analysis tools
  • firmware vulnerabilities
  • large language models (LLMs)
  • HALO (Human-Aligned Loss Optimization)
  • contrastive preference Optimization (CPO)
  • direct preference optimization (DPO)
  • Kahneman-Tversky Optimization (KTO)
  • dataset alignment
  • retrieval-augmented Generation (RAG)
  • bias mitigation
  • dual-method framework
  • secure-by-design
  • transformers
  • risk mitigation model

Cite this

Links

Related content

'