Potential PhD projects

Supervisory team: Professor Claude Sammut 

Project summary: Our rescue robot has sensors that can create 3D representations of its surroundings. In a rescue, it's helpful for the incident commander to have a graphical visualisation of the data so that they can reconstruct the disaster site. The School of Computer Science and Engineering and the Centre for Health Informatics have a display facility (VISLAB) that permits users to visualise data in three dimensions using stereo projection onto a large 'wedge' screen. 

This project can be approached in two stages. In the first stage, the data from the robot are collected off-line and programs are written to create a 3D reconstruction of the robot's surroundings to be viewed in the visualisation laboratory. In the second stage, we have the robot transmit its sensor data to the VISLAB computers for display in real-time. 

This project requires a good knowledge of computer graphics and will also require the student to learn about sensors such as stereo cameras, laser range finders and other 3D imaging devices. Some knowledge of networking and compression techniques will be useful for the second stage of the project. 

A scholarship/stipend may be available. 

For more information contact: Prof. Claude Sammut

Supervisory team: Wenjie Zhang, Dong Wen, Xiaoyang Wang

Project summary: This project explores the integration of artificial intelligence (AI) techniques with fundamental data processing problems, such as predictive modeling, forecasting, and anomaly detection. The project aims to develop machine learning and deep learning algorithms to gain insights from large volumes of data, which produce novel solutions for various real-world tasks and data types. The research has the potential to revolutionize the way data processing systems are designed, operated, and used in various applications and domains.

A scholarship/stipend may be available.

For more information contact: wenjie.zhang@unsw.edu.au         

Supervisory team: Dr Raymond Louie

Project summary: Accurately predicting disease outcomes can have a significant impact on patient care, leading to early detection, personalized treatment plans, and improved clinical outcomes. Machine learning algorithms provide a powerful tool to achieve this goal by identifying novel biomarkers and drug targets for various diseases. By integrating machine learning algorithms with biological data, you will have the opportunity to push the boundaries of precision medicine and contribute to algorithms that can revolutionize the field.

We are looking for a highly motivated student who is passionate about applying computational skills to solve important health problems. Don’t worry, no specific biological knowledge is necessary, the important thing is you are enthusiastic and willing to learn. Please get in touch if you have any questions. 

A scholarship/stipend may be available. 

For more information contact: Dr. Raymond Louie

Supervisory team: Dr. Aditya Joshi

Project Summary: Discrimination and bias towards protected attributes have legal, social, and commercial implications for individuals and businesses. The project aims to improve the state-of-the-art in the detection of discrimination and bias in text. The project will involve creation of datasets, and development of new approaches using natural language processing models like Transformers. The datasets may include different text forms such as news articles, job advertisements, emails, or social media posts. Similarly, the proposed approaches may use techniques such as chain-of-thought prompting or instruction fine-tuning.

A scholarship/stipend may be available.

For more information, contact aditya.joshi@unsw.edu.au.

Supervisory team: Wenjie Zhang, Dong Wen, Xiaoyang Wang

Project Summary: Large Language Models (LLMs) like GPT are revolutionizing the field of data science. Research in this area is multifaceted, exploring the development, application, and implications of these models. The project aims to utilize the LLMs to solve a wide spectrum of tasks in data science, from data preprocessing to predictive modeling and beyond. The outcome of the project will push the boundaries of data processing techniques, creating more intelligent, efficient, and ethical data science solutions.

A scholarship/stipend may be available.
For more information contact: wenjie.zhang@unsw.edu.au         

Supervisory team: Dr Sasha Vassar

Project Summary: You will be working as part of a team that develops educational large language models, including fine-tuning, design, evaluation and deployment to large audiences.

For more information contact: a.vassar@unsw.edu.au

Eligibility Criteria: 

1. domestic applicants (Citizens or Permanent Residence of Australia and New Zealand)  
2. with first or upper second-class Honours, or an equivalent qualification

Supervisory team: Dr Gelareh Mohammadi, Professor Arcot Sowmya, Dr Gideon Kowadlo

Project summary: The standard model of decision-making in biological systems involves a combination of model-free and model-based reinforcement learning (RL) algorithms. These processes are reflected in the Striatum (model-free) and the Prefrontal Cortex (PFC, model-based). Research shows that the model-free Striatum exerts gating control over the model-based PFC, a relationship captured in the influential PBWM framework (Frank and O'Reilly 2006) within the context of working memory. This intricate functional connectivity underpins decision-making, possibly balancing the strengths of both systems.

In AI, model-free and model-based RL algorithms have achieved significant advancements in applications like game playing and robot control. However, these systems face notable challenges: model-free RL is notoriously data-hungry and struggles with environmental changes, while model-based RL, though more adaptable, is computationally intensive, particularly at decision time. These limitations hinder the efficiency and productivity of AI systems, especially in dynamic and real-time environments.

This project aims to develop a novel RL architecture inspired by the biological interplay between the Striatum and PFC. We propose a "model-free-gated, model-based" recurrent system where the world model provides context/high-level goals to the model-free controller, which in turn exerts gating control over the world model. By integrating the strengths of both approaches, this architecture is designed to enhance the flexibility and efficiency of decision-making processes, reducing the data inefficiency of model-free methods while mitigating the computational burden of model-based planning. Through comparison with human data, we will evaluate this architecture's ability to overcome the limitations of traditional RL systems, ultimately contributing to AI systems that are more productive, adaptable, and capable of making efficient decisions in complex, changing environments.

This project will be conducted in close collaboration with Cerenaut.ai, opens in a new window, an independent research group.

For more information contact: Dr. Gelareh Mohammadi

Eligibility Criteria: 

1. domestic applicants (Citizens or Permanent Residence of Australia and New Zealand)  
2. with first or upper second-class Honours, or an equivalent qualification

Supervisory team: Dr Gelareh Mohammadi, Professor Arcot Sowmya, Dr Gideon Kowadlo

Project summary: The brains of all bilaterally symmetric animals, including humans, are divided into left and right hemispheres. While the anatomy and physiology of these hemispheres overlap significantly, they specialize in different attributes, which contributes to enhanced cognitive and motor functions. Despite this, the principle of hemispheric specialization remains underexplored in artificial intelligence (AI), machine learning (ML), and motor control systems. A preliminary study [Rinaldo24, opens in a new window] demonstrated that it is possible to replicate this type of hemispheric specialization for motor control in AI, where the dominant system excels in trajectory planning, and the non-dominant system specializes in positional control. This study also revealed the potential for exploiting such specialization to improve the performance of simple one-armed motor tasks.

The aim of this project is to extend the research to a two-armed system and more complex tasks, focusing on how hemispheric specialization can enhance productivity and performance in robotic systems. Specifically, we will explore whether the left and right hemispheres can collaborate to improve the performance of a single arm, and how they might enhance task efficiency when each arm performs complementary aspects of a task (e.g., holding an object with the non-dominant hand while the dominant hand performs precise actions). Additionally, we will investigate how smoothly switching between these modes can further optimize robotic performance.

By building a model with left and right neural networks connected via a corpus callosum (interhemispheric communication) to perform motor tasks, and comparing this model to human performance and standard ML approaches, this research will not only contribute to a deeper understanding of why brains are divided into left and right hemispheres but also establish a new principle for motor control in robotics. This approach promises to significantly enhance the efficiency and productivity of robotic systems, leading to more effective and adaptable robots capable of performing complex tasks with greater precision and coordination.

This project will be conducted in close collaboration with Cerenaut.ai, opens in a new window, an independent research group.

For more information contact: Dr. Gelareh Mohammadi

Eligibility Criteria: 

1. domestic applicants (Citizens or Permanent Residence of Australia and New Zealand)  
2. with first or upper second-class Honours, or an equivalent qualification

Supervisory team: Dr Gelareh Mohammadi, Professor Arcot Sowmya, Dr Gideon Kowadlo

Project summary: The brains of all bilaterally symmetric animals, including humans, are divided into left and right hemispheres, each specializing in different cognitive functions. While this principle is well-documented in biology, it remains underutilized in artificial intelligence (AI) and machine learning (ML). According to the Novelty-Routine Hypothesis (NRH), the right hemisphere acts as a 'generalist' that excels in handling novel tasks, while the left hemisphere specializes in routine tasks, with cognitive activity shifting from the right to the left as tasks become more familiar. This natural specialization is particularly relevant to the challenges faced in continual reinforcement learning (RL), where an agent must learn a sequence of tasks while avoiding catastrophic forgetting of previous knowledge.

Current approaches in RL primarily focus on maximizing performance on specific tasks, often neglecting the agent's initial performance on new and unfamiliar tasks. However, in many real-world applications, it is critical that an agent performs competently from the outset, as failures during the learning phase can be costly or dangerous. In a preliminary study [Nicholas24, opens in a new window], we developed a bi-hemispheric RL agent that leverages the generalist capabilities of a right-hemisphere-inspired model to maintain strong initial performance on novel tasks.

The goal of this project is to enhance this model by incorporating interhemispheric communication, mimicking the corpus callosum found in biological brains. This communication channel, shown to be beneficial in bilateral models for motor control [Rinaldo24, opens in a new window], will enable our RL agent to smoothly transition knowledge between hemispheres, further improving its adaptability and performance in continual learning settings. By focusing on graceful task adaptation, this research aims to create AI systems that not only achieve high performance over time but also maintain robust and reliable productivity when faced with new challenges, making them more suitable for deployment in dynamic and safety-critical environments.

This project will be conducted in close collaboration with Cerenaut.ai, opens in a new window, an independent research group.

For more information contact: Dr. Gelareh Mohammadi

Eligibility Criteria: 

1. domestic applicants (Citizens or Permanent Residence of Australia and New Zealand)  
2. with first or upper-second-class Honours, or an equivalent qualification

Project Summary:

This research project aims to enhance biosecurity measures by developing advanced computer vision algorithms for real-time image recognition and automated bait dispensing systems. Leveraging machine learning (ML) and deep learning (DL) methodologies, we will create a robust system capable of identifying invasive species and pests in agricultural settings with high accuracy. The system will enable the real-time recognition of target organisms.

By integrating these cutting-edge technologies, the project seeks to significantly improve biosecurity and productivity in agricultural operations. Automated real-time monitoring and targeted bait dispensing will reduce the reliance on manual labour, lower the use of pesticides, and minimise crop damage, leading to higher yields and more sustainable farming practices. This innovative approach not only addresses critical biosecurity challenges but also contributes to the overall efficiency and resilience of agricultural systems.

This project will be conducted in close collaboration with Intersect Australia, opens in a new window, a not-for-profit organisation established to enhance research productivity, across all disciplines, through the application of advanced information technology and skills. 

For more information contact: Dr Gelareh Mohammadi, opens in a new window

Eligibility Criteria: 

1. domestic applicants (Citizens or Permanent Residence of Australia and New Zealand)  
2. with first or upper-second-class Honours, or an equivalent qualification

This research project aims to implement cutting-edge Data Operations, Machine Learning Operations as well as computer vision and AI methods for land use classification and change detection through remote sensing. Utilising satellite imagery and aerial photographs, we will develop machine learning (ML) and deep learning (DL) models to accurately classify different land use types and monitor changes over time.

This project not only advances environmental monitoring capabilities but also enhances productivity in land management and planning. By automating the analysis of vast amounts of remote sensing data, we aim to provide timely and actionable insights for policy-making and resource allocation. This will lead to more efficient land use, improved conservation efforts, and better management of natural resources. Furthermore, the project will support climate change adaptation strategies by providing detailed data on habitat shifts and ecosystem alterations, enabling more informed decision-making.

This project will be conducted in close collaboration with Intersect Australia, opens in a new window, a not-for-profit organisation established to enhance research productivity across all disciplines through the application of advanced information technology and skills. 

For more information, contact: Dr Gelareh Mohammadi, opens in a new window

Supervisory team: Dr Raymond Louie, Dr Sara Ballouz

Project Summary: In machine learning, feature selection has become a key step in improving the predictive performance of the algorithm by eliminating redundant variables and selecting for those that are likely critical. In the biomedical field, these features are extremely useful; they can be used for understanding the underlying biology, further validated as biomarkers of disease or clinical diagnostic markers, and as targets for drug therapy. Many feature selection methods exist, but the best approach to use in experiments relating to multi-omics has yet to be assessed. This project will involve the development/assessment of different methods and their application to cancers, autoimmunity, and viral infections.

For more information contact r.louie@unsw.edu.au, s.ballouz@unsw.edu.au

Supervisory team: Dr Yang Song

Project summary: Various types of microscopy images are widely used in biological research to aid our understanding of human biology. Cellular and molecular morphologies give lots of information about the underlying biological processes. The ability to identify and describe the morphological information quantitative, objectively and efficiently is critical. In this PhD project, we'll investigate various computer vision, machine learning (especially deep learning) and statistical analysis methodologies to develop automated morphology analysis methods for microscopy images.

More research topics in computer vision and biomedical imaging can be found here, opens in a new window.

A scholarship/stipend may be available.

For more information contact: Dr Yang Song

Supervisor team: Professor Erik Meijering and Dr John Lock

Project summary: Biologists use multiparametric microscopy to study the effects of drugs on human cells. This generates multichannel image data sets that are too voluminous for humans to analyse by eye and require computer vision methods to automate the data interpretation. The goal of this PhD project is to develop, implement, and test advanced computer vision and deep learning methods for this purpose to help accelerate the challenging process of drug discovery for new cancer therapies. This project is in collaboration with the School of Medical Sciences (SoMS) and will utilise a new and world-leading cell image data set capturing the effects of 114,400 novel drugs on the biological responses (phenotypes) of >25 million single cells.

A scholarship/stipend may be available.

For more information contact: erik.meijering@unsw.edu.au, john.lock@unsw.edu.au

Supervisory team: Dong Wen, Wenjie Zhang

Project summary: Many complex systems and phenomena in the real world can be represented as graphs, such as social networks, biological networks, transportation networks, and communication networks. Under the research theme of Big Data, big graph processing is a key area that draws on concepts from data structure, algorithms, graph theory, distributed systems, parallel computing, machine learning, and database systems to address the unique challenges posed by large-scale graph data. This project aims to develop algorithms, techniques, and systems to efficiently analyze and manipulate big graphs. The research advances knowledge across multiple disciplines and drives innovation in fields ranging from computer science and engineering to biology, sociology, and beyond.

A scholarship/stipend may be available.

For more information contact: dong.wen@unsw.edu.au

Supervisory team: Sri Parameswaran 

Project summary: Reliability is becoming an essential part in embedded processor design due to the fact that they are used in safety critical applications and they need to deal with sensitive information. The first phase in the design of reliable embedded systems involves the identification of faults that could be manipulated into a reliability problem. A technique that is widely used for this identification process is called fault injection and analysis. The aim of this project is to develop a fault injection and detection engine at the hardware level for an embedded processor. 

A scholarship/stipend may be available. 

For more information contact: sridevan@unsw.edu.au

Supervisory team: Dr Gelareh Mohammadi, Prof. Wenjie Zhang

Project description: Previous studies have shown that cognitive training can effectively improve people's skillsets and emotional capabilities in cognitive deficits. Such training programs are known to enhance the participants' brain health and better prepare them for an independent life. However, the existing conventional technologies for such training are not scalable and lack personalized features to optimize the efficacy. In this project, we will develop a technology platform for automatically acquiring and processing multimodal training data. The project will be conducted in collaboration with Stronger Brains, a not-for-profit organization that provides cognitive training. We aim to develop a fully automated social and cognitive function assessment framework based on multimodal data. Such a framework is essential to establish a  system with less involvement of experts and increase its scalability. The project involves:

• Data collection.
• Developing multimodal predictive models for cognitive functions and affective states in cognitive deficits.
• Developing adaptation techniques to personalize the framework.

Supervisory team: Dr Gelareh Mohammadi, A/Prof. Nadine Marcus

Project description: The fields of Science, Technology, Engineering and Math, otherwise known as STEM, play a key role in the sustained growth and stability of any economy and are a critical component in shaping the future of our society. This project aims to develop new evidence-based guidelines for designing highly effective teaching simulations for a STEM subject that personalizes training to learner proficiency. In particular, we aim to design a novel AI-powered framework for dynamic adaptive learning in STEM educational technology to improve learning outcomes in an accessible and engaging environment. The potential contributions of the project involve:

• Developing a multimodal physio-behavioural AI for rapid assessment of proficiency level.
• Integration of affective state and cognitive load with proficiency level to form a comprehensive cognitive diagnosis and capture the interplay between affective and cognitive processes.
• Establishing dynamic adaptive learning in real-time based on the cognitive diagnosis that responds to the current individual needs of the learner.

Supervisory team: Sanjay Jha, Salil Kanhere 

Project summary: This project aims to develop scalable and efficient one-to-many communication, that is, broadcast and multicast, algorithms in the next generation of WMNs that have multi-rate multi-channel nodes. This is a significant leap compared with the current state of the art of routing in WMNs, which is characterised by unicast in a single-rate single-channel environment. 

A scholarship/stipend may be available. 

For more information contact: sanjay.jha@unsw.edu.au

Supervisory team: Mahbub Hanssan 

Project summary: A major focuses of the Swimnet project will be to look at a QoS framework for multi-radio multi-channel wireless mesh networks. We also plan to develop traffic engineering methodologies for multi-radio multi-channel wireless mesh networks. Guarding against malicious users is of paramount significance in WMN. Some of the major threats include greedy behaviour exploiting the vulnerabilities of the MAC layer, location-based attacks and lack of cooperation between the nodes. The project plans to look at a number of such security concerns and design efficient protection mechanisms (Mesh Security Architecture). 

A scholarship/stipend may be available.

For more information contact: mahbub.hanssan@unsw.edu.au  

Supervisory team: Wen Hu  

Project summary: The mission of the SENSAR (Sensor Applications Research) group is to investigate the systems and networking challenges in realising sensor network applications. Wireless sensor networks are one of the first real-world examples of "pervasive computing", the notion that small, smart and cheap, sensing and computing devices will eventually permeate the environment. Though the technologies still in their early days, the range of potential applications is vast - track bush fires, microclimates and pests in vineyards, monitor the nesting habits of rare sea-birds, and control heating and ventilation systems, let businesses monitor and control their workspaces, etc. 

A scholarship/stipend may be available. 

For more information contact: wen.hu@unsw.edu.au

Supervisory team: Boualem Benatallah, Lina Yao, Fabio Casati

Project summary: This project investigates the significant and challenging issues that underpin the effective integration of software-enabled services with cognitive and conversational interfaces. Our work builds upon advances in natural language processing, conversational AI and services composition.

We aim to advance the fundamental understanding of cognitive services engineering by developing new abstractions and techniques. We’re seeking to enable and semi-automate the augmentation of software and human services with crowdsourcing and generative model training methods, latent knowledge and interaction models. These models are essential for the mapping of potentially ambiguous natural language interactions between users and semi-structured artefacts (for example, emails, PDF files), structured information (for example, indexed data sets), apps and APIs.

For more information contact: b.benatallah@unsw.edu.au or lina.yao@unsw.edu.au

Supervisory team: Helen Paik

Project summary: Micro-transactions stored in blockchain create transparent and traceable data and events, providing burgeoning industry disruptors an instrument for trust-less collaborations. However, the blockchain data and its’ models are highly diverse. To fully utilise its potential, a new technique to efficiently retrieve and analyse the data at scale is necessary.

This project addresses a significant gap in current research, producing a new data-oriented system architecture and data analytics framework optimised for online/offline data analysis across blockchain and associated systems. The outcome will strongly underpin blockchain data analytics at scale, fostering wider and effective adoption of blockchain applications.

A scholarship/stipend may be available.

For more information contact: h.paik@unsw.edu.au

Supervisory team: Fethi Rabhi

Project summary: This project investigates novel architectures & processes to develop AI and machine learning systems for business applications. This includes the use of AutoML and new collaborative “code-free” technologies to simplify AI system design/production within a large enterprise. This project will need a rethink of many traditional software engineering practices in areas of software architecture, development processes and requirements engineering. These issues are all interlinked e.g., adding business objectives may reduce usability and decrease performance, adding more transparency may obscure and decrease trust, and adding more usability may decrease performance. In some cases, ethical and compliance with regulations are other important considerations that need to be taken into account when developing the system.  The main application area is in the financial domain in collaboration with industry partners within the Fintech AI Innovation Consortium, opens in a new window.

For more information contact f.rabhi@unsw.edu.au

Supervisory team: A/Prof. Yulei Sui

Project summary: Modern software repositories are vast, making understanding the source code of a project especially challenging, particularly for legacy code bases. This project aims to design a code language model to automatically generate source code, detect software vulnerabilities, and provide program repair suggestions by understanding the syntax and semantics of code information (e.g., control-flow and data-flows). This project will be based on our group's existing source code analysis and verification tool SVF, opens in a new window. The expected deliverable of this project is an open-source tool that can accept, analyze, and parse user queries to interact with the code language model and SVF, generating high-quality codebases and analyzing large codebases consisting of millions of lines of code. You will work together with our team, including postdocs and PhD students, to conduct exciting research.

For more information contact: y.sui@unsw.edu.au

Supervisory team: Gernot Heiser

Project summary: Project summary: The Trustworthy Systems (TS) group are the creators of seL4, the world's first operating system (OS) kernel with a formal correctness proof. TS continues to conduct research at the intersection of OS, formal methods and programming languages, with the overall aim of producing real-world systems that are provably secure and safe, yet performant.

Specific projects include provable prevention of information leakage through microarchitectural timing channels; OS design and implementation for performance and verification; automatic verification and repeatable verification of OS components; verified compiler for the Pancake systems language; high-assurance worst-case execution-time analysis; provable schedulability of mixed-criticality safety-critical system.

For more information, including availability of scholarships, see https://trustworthy.systems/students/research, opens in a new window, or contact gernot@unsw.edu.au

Supervisory team: Dr Jesse Laeuchli, Dr Arash Shaghaghi, Prof Sanjay Jha

Project summary: Remote and embedded devices are the lynchpin of modern networks. Satellites, Aircraft, Remote Sensors and Drones all require numerous embedded devices to function. A key part of ensuring these devices remain ready to carry out operations is to ensure their memory has not been corrupted by an adversary.

In this project we will explore methods for securing remote devices using early generation quantum computers. These have the ability to work with one or two qubits at a time, and operate with very limited quantum memory, but they still provide access to valuable quantum effects which can be used for security.  

The successful student will have an interest in both cyber-security and quantum computing, with a willingness to explore the mathematics needed to exploit quantum algorithms.

Eligibility: Domestic Candidates only, PhD only

For more information contact Dr Jesse Laeuchli or Dr Arash Shaghaghi.

Supervisory team: Ron van der Meyden 

Project summary: The technology of cryptocurrency and its concepts can be broadly applicable to range of applications including financial services, legal automation, health informatics and international trade. These underlying ideas and the emerging infrastructure for these applications is known as ‘Distributed Ledger Technology’. 

A scholarship/stipend may be available. 

For more information contact: meyden@cse.unsw.edu.au  

Supervisors:

• Dr Gelareh Mohammadi (g.mohammadi@unsw.edu.au)
• Prof. Wenjie Zhang (wenjie.zhang@unsw.edu.au)

Project description:

Previous studies have shown that cognitive training can effectively improve people's skillsets and emotional capabilities in cognitive deficits. Such training programs are known to enhance the participants' brain health and better prepare them for an independent life. However, the existing conventional technologies for such training are not scalable and lack personalized features to optimize the efficacy. In this project, we will develop a technology platform for automatically acquiring and processing multimodal training data. The project will be conducted in collaboration with Stronger Brains, a not-for-profit organization that provides cognitive training. We aim to develop a fully automated social and cognitive function assessment framework based on multimodal data. Such a framework is essential to establish a  system with less involvement of experts and increase its scalability. The project involves:

1. Data collection.
2. Developing multimodal predictive models for cognitive functions and affective states in cognitive deficits.
3. Developing adaptation techniques to personalize the framework.

Supervisors:

• Dr Gelareh Mohammadi (g.mohammadi@unsw.edu.au)
• A/Prof. Nadine Marcus (nadinem@unsw.edu.au)

Project description:

The fields of Science, Technology, Engineering and Math, otherwise known as STEM, play a key role in the sustained growth and stability of any economy and are a critical component in shaping the future of our society. This project aims to develop new evidence-based guidelines for designing highly effective teaching simulations for a STEM subject that personalizes training to learner proficiency. In particular, we aim to design a novel AI-powered framework for dynamic adaptive learning in STEM educational technology to improve learning outcomes in an accessible and engaging environment. The potential contributions of the project involve:

1. Developing a multimodal physio-behavioural AI for rapid assessment of proficiency level.
2. Integration of affective state and cognitive load with proficiency level to form a comprehensive cognitive diagnosis and capture the interplay between affective and cognitive processes.
3. Establishing dynamic adaptive learning in real-time based on the cognitive diagnosis that responds to the current individual needs of the learner.

Supervisor: Dr Rahat Masood (rahat.masood@unsw.edu.au)

Supervisory team: Prof Salil Kanhere (CSE - UNSW), Suranga Seneviratne (USyd), Prof Aruna Seneviratne (EE&T – UNSW)

Project description:

Children start using the Internet from a very early age for entertainment and educational purposes and continue to do so into their teen years and beyond. In addition to providing the required functionality, the online services also collect information about their users, track them, and provide content that may be inappropriate such as sexually explicit content; content that promotes hate and violence, and other content compromising users’ safety. Another major issue is that there is no established mechanism to detect the age of users on online platforms hence, leading children to sign up for services that are inappropriate for them. Through this research work, we aim to develop an age detection framework that can help detect children’s activities on online platforms using various behavioural biometrics such as swipes, keystrokes, and handwriting. The core of this project revolves around the ground-breaking idea that “User Touch Gestures” contain sufficient information to uniquely identify them, and the “Touch Behaviour” of a child is very different from that of an adult, hence leading to child detection on online platforms. The success of this project will enable online service providers to detect the presence of children on their platforms and offer age-appropriate content accordingly.

Supervisors:

• Dr Rahat Masood (rahat.masood@unsw.edu.au)
• Prof. Salil Kanhere (CSE - UNSW)

Project description:

Users unintentionally leave digital traces of their personal information, interests and intents while using online services, revealing sensitive information about them to online service providers. Though, some online services offer configurable privacy controls that limit access to user data. However, not all users are aware of these settings and those who know might misconfigure these controls due to the complexity or lack of clear instructions. The lack of privacy awareness combined with privacy breaches on the web leads to distrust among the users in online services. Through this research study, we intend to improve the trust of users on the web and mobile services by designing and developing user-centric privacy-preserving solutions that involve aspects of user privacy settings, user reactions and feedbacks on privacy alerts, user behavioural actions and user psychology. The aforementioned factors will be first used in quantifying privacy risks and later used in designing privacy-preserving solutions. In essence, we aim to improve privacy in mobile and web platforms by investigating various human factors in: i) privacy risk quantification and assessment, and ii) privacy-preserving solutions.

Supervisors:

• Dr Yang Song (yang.song1@unsw.edu.au) - primary
• Prof. Maurice Pagnucco - joint

Project description:

Deep learning techniques have shown great success in many applications, such as computer vision and natural language processing. However, in many cases, purely data-driven approaches would provide suboptimal results, especially when limited data are available for training the models. This dependency on large-scale training data is well understood as the main limitation of deep learning models. One way to mitigate this problem is to incorporate knowledge priors into the model, similarly to how humans reason with data; and there are various types of knowledge priors, such as data-specific relational information, knowledge graphs, logic rules and statistical modelling. In this PhD project, we will investigate novel methods that effectively integrate knowledge priors and commonsense reasoning with deep learning models. Such models can be developed for a wide range of application domains, such as computer vision, social networks, biological discovery and human-robot interaction.

Supervisors:

• Dr Yang Song (yang.song1@unsw.edu.au) - primary
• Prof. Maurice Pagnucco - joint

Project description:

Deep learning models are typically considered a black-box, and the lack of explainability has become a major obstacle to deploy deep learning models to critical applications such as medicine and finance. Explainable AI has thus become an important topic in research and industry, especially in the deep learning era. Various methods for explaining deep learning models have been developed, and we are especially interested in explainability in graph neural networks, which is a new topic that has emerged very recently. Graph neural networks are becoming increasingly popular due to their inherent capability of representing graph structured data, yet their explainability is more challenging to explore with the irregular and dynamic nature of graphs. In this PhD project, we will investigate novel ways of modelling explainability in graph neural networks, and apply this to various applications, such as computer vision, biological studies, recommender systems and social network analysis.

Supervision team

•  Dr Jiaojiao Jiao (jiaojiao.jiang@unsw.edu.au) – primary
•  Prof. Sanjay Jha (sanjay.jha@unsw.edu.au) – joint
•  Stephen Doherty (s.doherty@unsw.edu.au) - secondary

Background

Most cyber threat intelligence platforms provide scores and metrics that are mainly derived from open-source and external sources. Organisations must then figure out if and how the output is relevant to them.

Research problems

• Dynamic threat risk/exposure score

Continuous monitoring and calculation of an organisation’s ‘Threat Risk’ posture score using a range of internal and external intelligence.

• Customised/targeted newsfeed

A curated cyber and threat newsfeed that is relevant to an organisation. The source of the newsfeed will leverage the internal and external analysis from the first question. The output will include information that helps users understand and digest their organisation’s threat posture in a non-technical manner.

Proposed approaches

We propose to develop dynamic GNN models for discovering dynamic cyber threat intelligence from blended sources. GNN has achieved state-of-the-art performance in many high-impact applications, such as fraud detection, information retrieval, and recommender systems, due to their powerful representation learning capabilities. We propose to develop new GNN models which can take blended intelligence sources into account in the threat intelligence prediction. Moreover, many GNN models are static that deal with fixed structures and parameters. Therefore, we propose to develop dynamic GNN models which can learn the evolution pattern or persistent pattern of dynamic graphs.