University of the Aegean
The Dual Role of AI in Internet of Cloud Privacy and Security: countermeasures, compliance challenges, and LLM-driven inference risks
Abstract
dc:descriptionThe Internet of Cloud (IoC), a convergent architecture synthesizing Internet of Things (IoT) devices with Cloud Computing infrastructure, has emerged as the operational backbone of modern digital societies, enabling unprecedented data exchange across billions of interconnected devices while simultaneously creating complex and escalating privacy and security challenges. This doctoral thesis investigates the multi-layered nature of privacy and security in IoC ecosystems through a programme of six interrelated empirical and theoretical research contributions, each addressing a distinct dimension of the problem space.The thesis begins by establishing the architectural foundations of the IoC and systematically mapping the privacy and security challenges that arise from IoT-cloud convergence, including threats to user data protection, authentication, access control, system integrity, and regulatory compliance. From this foundation, the research transitions to the human dimension of IoC security, examining why existing cybersecurity training frameworks are inadequate for the complex threat landscape generated by IoC ecosystems and proposing an adaptive, role-specific training model that integrates artificial intelligence and gamification to address evolving vulnerabilities.The third contribution presents a comprehensive taxonomy of AI-driven privacy-enhancing techniques for IoT systems, systematically mapping nine AI countermeasures including federated learning, anomaly detection, differential privacy, homomorphic encryption, and secure multi-party computation across four IoT architectural layers. This taxonomy provides both a theoretical framework and practical guidance for deploying AI as a defensive mechanism rather than a risk factor.The fourth contribution reframes this perspective by examining AI, and specifically large language models (LLMs), as potential privacy-risk amplifiers. Through a structured comparative analysis of twenty widely deployed LLMs against ten critical privacy challenges in IoC environments, the research reveals that most commercial LLMs exhibit significant compliance deficiencies across transparency, consent management, data retention, and inference-based leakage domains.The fifth contribution provides empirical evidence that LLMs can infer sensitive demographic attributes (gender, age, profession) and psychological states (depression risk) from user-generated text with high accuracy reaching 93% for gender classification. This demonstrates that privacy leakage can occur even when users do not explicitly disclose protected characteristics, raising fundamental questions about the governance of AI systems in IoC-connected environments.The sixth and final contribution extends this threat model to structured, non-textual data by conducting the first large-scale empirical investigation of LLM-driven inference from structured IoT sensor records. Using a dataset of 65,450 synthetic IoT telemetry records, the research demonstrates that Claude 3 Sonnet and GPT-4o can infer device categories, location contexts, attack vulnerabilities, critical infrastructure types, and privacy risk scores with accuracy levels reaching 87%, without any task-specific training or fine-tuning.Collectively, the six contributions advance a unified research narrative that positions IoC privacy as a multi-layered research challenge requiring simultaneous attention to architecture, human behaviour, AI-based protection, LLM governance, textual profiling, and structured-data de-anonymization. The thesis proposes a layered privacy-by-design framework aligned with GDPR, CCPA, and the EU AI Act, and identifies concrete directions for future research including adaptive privacy frameworks, sentiment-aware IoC adaptation, regulation-aligned LLM deployment standards, and cross-domain inference prevention mechanisms.Keywords: Internet of Cloud (IoC), Internet of Things (IoT), Cloud Computing (CC), privacy, security, cybersecurity training, artificial intelligence (AI), LLMs, de-anonymization, federated learning, GDPR, privacy by design, user profiling.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsouplaki, Asimina
- Τσουπλάκη, Ασημίνα
Subjects
dc:subject × 42- Διαδίκτυο του Νέφους (IoC)
- Ιδιωτικότητα και ασφάλεια δεδομένων
- Τεχνητή νοημοσύνη (ΤΝ)
- Μεγάλα γλωσσικά μοντέλα (LLMs)
- Διαδίκτυο των Πραγμάτων (IoT)
- Προστασία δεδομένων μέσω Τεχνητής Νοημοσύνης
- Επιθέσεις συμπερασμού και αποανωνυμοποίηση
- Κατάρτιση και ευαισθητοποίηση στην κυβερνοασφάλεια
- Σκιαγράφηση προφίλ χρηστών
- Τεχνικές ενίσχυσης της ιδιωτικότητας
- Ιδιωτικότητα εκ σχεδιασμού
- Κανονιστική συμμόρφωση και διακυβέρνηση ΤΝ
- Συμπερασματική ανάλυση δεδομένων αισθητήρων IoT
- Ομοσπονδιακή μάθηση
- Γενικός Κανονισμός Προστασίας Δεδομένων (ΓΚΠΔ)
- Internet of Cloud (IoC)
- Data privacy and security
- Artificial Intelligence (AI)
- Large Language Models (LLMs)
- Internet of things (IoT)
- AI-Driven privacy protection
- Inference attacks and de-anonymization
- Cybersecurity training and awareness
- User profiling
- Privacy-Enhancing Technologies (PETs)
- Privacy by design
- Regulatory compliance and AI governance
- IoT sensor data inference
- Federated learning
- General Data Protection Regulation (GDPR)
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική και Τεχνολογίες, άλλοι τομείς
- Άλλες Κοινωνικές Επιστήμες
- Κοινωνικές Επιστήμες
- Κοινωνικές επιστήμες, διεπιστημονική προσέγγιση
- Other Engineering and Technologies
- Engineering and Technology
- Engineering and Technologies, miscellaneous
- Other Social Sciences
- Social Sciences
- Social sciences, interdisciplinary
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62985