Honours and Masters project

Large Language Models (LLMs) such as GPT, Llama, Qwen, and Mistral are increasingly used in commercial and academic applications. As more models become available, identifying which model generated a particular response becomes important for copyright auditing, model verification, and AI transparency.

Current fingerprinting methods often rely on manually selected benchmark questions. However, manually designing discriminative questions is time-consuming and may not capture unique behavioral differences between models.

Our research explores novel map representations and projections.

This project seeks to design and trial new map representations for seeing Australian population data sets in new and ideally more effective ways. 

Why is this needed? 

Can a Transformer understand a software patch and predict whether it truly fixes a vulnerability, introduces a new weakness, or leaves the system still exploitable?

This is much more specific than normal vulnerability detection.

Instead of asking:

“Is this code vulnerable?”

we ask:

“Did this security patch actually fix the problem safely?”

In real software projects, security patches are often rushed. A patch may:

This project is not just another IDS project. It sits at the intersection of four powerful areas:

So you will deal with:

Cybersecurity: detecting attacks in IoT and IIoT networks.Federated Learning: training AI without centralizing raw data.Privacy Engineering: measuring and reducing leakage from model updates.Edge AI: making the system lightweight enough for constrained devices.

Evaluation benchmarks are a foundational component of artificial intelligence (AI) research, providing standardized ways to measure and compare the capabilities of AI systems. Benchmarks such as MMLU, GSM8K, HumanEval, and HellaSwag have been instrumental in tracking progress in large language models and related systems. However, benchmark usefulness is not static.

Large Language Models (LLMs) are increasingly used to automatically evaluate other AI systems in tasks such as writing, reasoning, and question answering. This approach called LLM-as-a-Judge is now widely used in research benchmarks and AI development pipelines.

Videos contain rich information about actions, events, interactions, and changes over time. While recent AI models have made strong progress in video understanding, reasoning over complex video content remains challenging, especially when the task requires understanding temporal context or connecting information across different moments.

Foundation models have shown strong reasoning abilities, but their reasoning process is often implicit, difficult to inspect, and not always reliable. This project explores neuro-symbolic reasoning as a way to make AI reasoning more structured, interpretable, and robust.

The project will study how neural models can be combined with explicit logic-based reasoning. A neural model may be used to understand or decompose a problem, while a symbolic reasoning component performs structured inference using logic, rules, programs, or other formal representations.

Ensuring construction quality and safety requires timely detection of defects, yet traditional manual inspections are slow, costly, and inconsistent. This research presents a computer vision-driven solution that automates defect detection using images captured by drones and site cameras. By applying advanced deep learning models, the system are expected to identify cracks, corrosion, and surface irregularities with high accuracy, even under challenging site conditions.

Vision-and-Language Navigation (VLN) has emerged as a promising paradigm for enabling robots to interpret natural language instructions and navigate complex environments. While most prior work focuses on single-agent navigation, this study extends VLN to multi-robot systems, addressing challenges of coordination, communication, and task allocation in dynamic settings.