The internet of things (IoT) has been attracting growing attention in recent years. As one potential technology used for IoT sensor network, Bluetooth Low Engery (BLE) is becoming one of the most anticipated solutions to establish IoT networks. Whilst BLE based IoT devices are becoming increasingly popular, security becomes an unavoidable concern not only for the devices by themselves but also for the whole network constructed with BLE devices.

In this thesis, we investigate the security issues on devices and we study the potential security issues for BLE-based IoT sensor networks. In order to explore the security on BLE scatternet, we explore how to form a BLE scatternet, how to implement an appropriate routing algorithm specifically for this network and what kind of potential attack this network may encounter often.

The research endeavor makes four contributions to the security field of intrusion detection on devices and IoT sensor networks. First, we propose a hardware-based intrusion detection approach called CONtrol-flow VERification SystEm (CONVERSE), which ensures control-flow integrity by verifying the destination of control-flow branches at runtime. Many techniques exist for an attacker to alter control-flow to trigger malicious behavior, such as stack and heap overflows which overwrite a return address or function pointer. By verifying branch target addresses at runtime, security exploits can be detected as illegal control-flow. Secondly, we propose an approach for scatternet formation and multi-hop routing for BLE scatternet. We define procedures for device discovery, communication between piconets and forming multi-hop scatternet. Thirdly, we improve the routing algorithm for BLE network in step 2, and present BLE Scatternet Battery- Aware Routing (BSBR), a power aware routing mechanism based on Dynamic Source Routing (DSR) which can be applied to BLE-based Mobile Ad-Hoc Networks (MANETs). Furthermore, we study the impact of battery exhaustion attacks on BLE-based networks. In order to defend against this type of attack, we propose an intrusion detection and prevention approach requiring the suspicious nodes to switch roles with its connected nodes after a pre-determined time. If the suspicious node is identified as a malicious one, it will be blacklisted to prevent future attacks.

The rapid advancement of large language models (LLMs) has opened new possibilities across various domains. However, adapting these models for specialized applications remains challenging. This thesis proposes several innovative fine-tuning strategies to address these challenges, particularly focusing on data scarcity, task alignment, and interpretability and control in specialized domains. These strategies include: 1) Customizing loss functions to reflect task-specific priorities. 2) Incorporating specialized embeddings and additional prediction heads combined with Language Modeling (LM) to further align training objectives. 3) Integrating Variational Autoencoders (VAEs) with Transformers to smooth out the embedding space for a better interpretable representation of complex data distributions, enabling the generation of new synthetic data points through sampling from the latent space.

This thesis applies and evaluates these strategies in two critical areas: online health support and cybersecurity applications.In the context of online health support, this thesis develops a novel approach for intent detection in smoking cessation support groups, addressing challenges of noisy, sparse data and overlapping intents. By fine-tuning an LLM with customized loss functions that account for intent priorities and class imbalances, this approach achieves a 95.5% accuracy across 24 intent categories, significantly outperforming existing approaches for online health group intent detection. For online mental health support conversations, this thesis proposes an innovative model for generating empathetic, counseling-style reflections. Through the integration of multiple text-based empathetic factors, including emotion, intent, and three mechanisms for communicating empathy, this model enhances both context and response representations. This approach significantly improves the quality of reflective, empathetic responses, as validated through automated metrics and human evaluations. This advancement has the potential to augment online mental health support services by providing more nuanced and psychologically appropriate responses. In the cybersecurity domain, this thesis presents WebTransVAE, a Transformer-based Variational Autoencoder for generating synthetic web traffic data. This model uniquely combines the strengths of Transformers in handling sequential data with VAEs' ability to create a structured latent space. By implementing techniques such as pooling, denoising, and encoder pretraining, this approach mitigates the issue of posterior collapse. Empirical evaluations demonstrate that the model successfully generates four key variables of HTTP web traffic data through latent manipulation, maintaining a validity rate of 99.6% on average for each generated variable.

These strategies collectively demonstrate significant improvements in adapting LLMs for specialized tasks. In online health support, the enhanced models show a better understanding of user needs and context, generating higher quality counseling responses by better integrating nuances of empathetic reflective listening, potentially leading to more effective interventions. In cybersecurity, the approach enables the generation of more realistic and varied synthetic data, crucial for robust system testing and development.The methodologies presented in this thesis address core challenges of data scarcity, task alignment, and interpretability and control, establishing a framework for adapting language models to diverse specialized domains. This research offers potential solutions for other applications facing similar limitations, paving the way for more effective AI utilization in critical real-world scenarios.

Verification of modern digital systems can consume up to 70% of the design cycle. Verification engineers must create formal properties which reflect correct operation from design specifications or other documents. This process of creating formal correctness properties from textual descriptions is a laborious task which is difficult to automate.

In this work I investigate the creation of formal verification properties from textual descriptions written in natural language. I present two approaches that utilize natural language processing (NLP) and machine learning techniques for the automatic generation of verification properties. In the first approach a set of correctness properties expressed in natural language, called natural language assertions (NLAs), is divided into subsets of structurally similar sentences. A generalized formal property template for each subset is used to provide a mapping from each sentence to a well specified verification property. Experimental results show that this methodology reduces the number of formal properties which must be manually created by up to an order of magnitude.

In the second approach I create a custom attributed formal grammar which captures the English semantics of a temporal tree logic. A translation system is implemented which uses this attribute grammar to perform a semantic parsing of NLAs. Attributes for each grammatical production in the parse tree are then used to build a fully specified formal property for each NLA. Experimental results show that valid formal properties are generated from English NLAs in over 90% of test cases.

In evaluating the translation system it was observed that translation rates for NLAs are strongly dependent on the quality of the formal grammar used. High translation rates require a finely tuned custom grammar. To facilitate the creation of such a grammar I propose a new learning algorithm which automatically generates custom attribute grammars from a small set of NLAs and formal properties. This machine generated grammar is used in the NLA translation system to create formal properties from NLAs taken from two design documents for designs in the same product family. Experimental results show that the learned attribute grammar is of sufficient quality to successfully translate up to 88% of NLAs.