# PhD Opportunities

## Potential PhD Topic for New Students

- How to Apply (Faculty of Science, Engineering and Built Environment)
- Find a Scholarship
- Apply Online (with approved EOI)

**Principal supervisor: A/Prof Richard Dazeley**

**Contact: https://www.deakin.edu.au/about-deakin/people/richard-dazeley**

**Lab: Machine Intelligence Lab**

**Title: Explainability through Fuzzy Reinforcement Learning**

Reinforcement Learning (RL) observes an environment and determines an action that leads the agent towards its goal. The majority or Deep RL systems use raw state input such as a video image, audio, or other continuous data to represent state information. However, when an RL agent is expected to explain its behaviour it must rely on methods such as saliency maps to identify features of relevance, which have been to been shown to provide a poor communication technique for explanations. Furthermore, providing explanations of the agent’s longer-term intentionality is even more difficult to articulate. One unique alternative approach, called Programmatically Interpretable RL (PIRL) approach raises the possibility of representing state through a programable structure and using this for generating basic explainable functionality.

In this project we will build upon the idea in PIRL by developing an approach where represents the environmental state through an abstracted model using fuzzy rules where the fuzzy sets are generated through environment interactions. These fuzzy rules will provide a structured ontology for interpreting the state, while the rule inference process will provide the agent’s intentionality-based reasoning. The combination of these components will both allow an agent to apply traditional RL learning while also being capable of providing both perception and goal-driven explanations of its behaviour.

This project will be conducted in two stages. The first stage will develop a Fuzzy RL framework capable of learning in traditionally deep learning environments. This agent will then be used to generate explanations of that improves human observers’ mental model of its behaviour. This will involve a quantitative and qualitative study of people predicting the agent’s future behaviour after being trained on past cases.

**Principal supervisor: A/Prof Richard Dazeley**

**Contact: https://www.deakin.edu.au/about-deakin/people/richard-dazeley**

**Lab: Machine Intelligence Lab**

**Title: Explainable Intentionality with Multiobjective Reinforcement Learning**

Researchers have long understood that an AI-based system’s ability to explain its decision is critical to human acceptance, understanding and trust. Recently, with the growth of machine learning based systems there has been a significant increase in work in this domain. Explaining the behaviour of Goal-Driven agents however is currently mostly limited to local decisions rather than explaining the intentionality and temporal nature of the decision. Intentionality, however, is limited in single objective domains. Reward decomposition can provide some degree of justification provision around action preferences but is limited due to the correlation of reward signals.

In this project we will use a multiobjective framework to extract action preferences that allow a comparison of the possible actions against each of the objectives. This will be combined with our approach for transition probability prediction to explain to the user that the selected behaviour increases the opportunity of achieving a particular objective over other actions that lead to alternative and unwanted objectives. For instance, this will allow us to provide the explanation:

*I did X instead of Y because X will still allow (with some probability) me to achieve my primary objective but is unlikely (with some probability) of causing Z (some undesirable outcome).*

The natural extension to this will provide both counterfactual and contrastive explanations and this study will show to what degree human users can develop a mental model learnt from such explanations, allowing them to accurately predict the agent’s behaviour in future environments.

**Principal supervisor: Dr Kerri Morgan**

**Contact: https://www.deakin.edu.au/about-deakin/people/kerri-morgan**

**Lab: Mathematics**

**Title: Graph Polynomials **** **

The Tutte polynomial encapsulates information about a graph including the number of spanning subgraphs, the number of spanning trees, and the number of components and blocks of the graph. Many important graph polynomials can be obtained from partial evaluations of the Tutte polynomial including the flow polynomial which counts network flows and the chromatic polynomial which counts the number of proper colourings of the graph.

An important feature of any polynomial is its roots. The question of which numbers can be roots of a given graph polynomial is of great interest not only to mathematicians and computer scientists but to physicists in the area of statistical mechanics. This interest in statistical mechanics is due to the relationship between the Tutte polynomial and the Potts model partition function. The Potts model is used to study how local interactions in a network affect global network properties. Applications include community detection and identifying structures in social networks, foam behaviour and tumour growth. The limit points of the complex roots of the Tutte polynomial give possible locations of physical phase transitions. This has motivated research into the roots of the Tutte and chromatic polynomials particularly for certain families of graphs.

In the case of the chromatic polynomial, the α + n conjecture states that for any algebraic number α there exists a non-negative integer n such that α + n is a root of some chromatic polynomial. This conjecture is known to be true of roots of quadratic and linear expressions.

In this project, we investigate which algebraic integers can be roots of the Tutte polynomial and/or graph polynomials that are partial evaluations of the Tutte polynomial (for example, the chromatic polynomial). Areas for investigation include:

- Which algebraic integers are roots of a given graph polynomial?
- Can we identify structural properties of a graph that related to algebraic properties of its graph polynomial?
- Does an algorithm exist that, given a polynomial p of degree k, can construct a graph with p (or a shift of p) as a factor? (and implementation of such an algorithm.)
- Can we identify structural operations on a graph that relate to algebraic manipulation of the roots? (For example, performing a join on a graph shifts the roots.)
- Can we find a certificate to show that two graphs have roots that are related in non-trivial ways?
- Can we find a form of certificate to show that a graph is Tutte-unique or chromatically unique?

**Principal supervisor: Dr Kerri Morgan**

**Contact: https://www.deakin.edu.au/about-deakin/people/kerri-morgan**

**Lab: Mathematics**

**Title: Certificates for Network Comparisons **

A certificate gives a sequence of steps that transforms one graph to another graph. Each step in the sequence involves applying an operation on a graph or graphs. Importantly, these operations maintain the properties of interest even while the graph(s) may change. Thus, certificates can be used to show relationships between networks represented by graphs.

In this project, we investigate the use of certificates to identify pairs of graphs that can be considered the ‘same’ with respect to a given property (or properties). We will investigate applications of certificates in areas such as identifying equivalence of diagrams arising in software engineering and identifying non-isomorphic networks with the same reliability under random edge failure.

Areas for investigation include:

- Identifying graph operations that do not affect a given property of interest.
- Programs for generating short certificates for a given property.
- Identifying certificates to show that two graphs are essentially the ’same’ with respect to a given property.

**Principal supervisor: Dr Thanh Thi Nguyen**

**Contact: https://www.deakin.edu.au/about-deakin/people/thanh-thi-nguyen**

**Lab: Machine Intelligence Lab**

**Title: Mobile edge caching and cloud-based malware detection using machine learning/deep reinforcement learning**

Mobile edge computing (MEC) is a technique that allows cloud computing functions to take place at the edge nodes of a network. This technology helps to decrease network traffic, reduce overhead and latency when users request to access contents that have been cached in the edges closer to the cellular customer. MEC systems, however, are vulnerable to cyber attacks because they are physically located closer to users and attackers, with less secure protocols compared to cloud servers or database center. On the other hand, one of the most challenging malwares of mobile devices is zero-day attacks, which exploit publicly unknown security vulnerabilities, and until they are contained or mitigated, hackers might have already caused adverse effects on computer programs, data or networks. To avoid such attacks, the traces or log data produced by the applications need to be processed in real time. With limited computational power, battery life and radio bandwidth, mobile devices often offload specific malware detection tasks to security servers at the cloud for processing.

There have been approaches that use machine learning methods, specifically reinforcement learning (RL) methods, to select the defense levels and important parameters such as offloading rate and time, transmission channel and power. As the network state space is large, involving high-dimensional data, the use of traditional RL methods shows disadvantages. This project investigates deep RL methods, which combine deep learning and traditional RL, for cloud-based malware detection and solutions for MEC against jamming and smart attacks and. Jamming attacks can be considered as a special case of denial-of-service attacks, which are defined as any event that diminishes or eradicates a network’s capacity to execute its expected function. The recent development of deep learning can facilitate the use of deep RL for jamming or smart attack handling or mitigation in MEC systems. This project specifically studies security challenges of the MEC systems and proposes a deep RL methods for cloud-based malware detection and solutions against jamming and smart attacks. It is expected that the proposed approach will enhance the security and user privacy of MEC systems and can protect the systems in confronting with different types of smart attacks. Reference paper: https://arxiv.org/abs/1906.05799

**Principal supervisor: Dr Thanh Thi Nguyen**

**Contact: https://www.deakin.edu.au/about-deakin/people/thanh-thi-nguyen**

**Lab: Machine Intelligence Lab**

**Title: Sequential anomaly detection in cyber security using machine learning/deep reinforcement learning**

Anomaly detection is a typical problem in various areas, including data mining, image processing, and also cyber security. There have been numerous techniques proposed for anomaly detection in cyber security, especially for Security Information and Event Management (SIEM) data. These algorithms can be categorized and distinguished into three main types: supervised, semi-supervised and unsupervised anomaly detection. While supervised anomaly detection techniques use data which are completely labelled both normal data and anomalies, training and testing data for semi-supervised methods contain only normal cases. In the other type, i.e. unsupervised, it does not need to use labelled data or even splitting data to training and testing sets. These detection techniques cannot guarantee effectiveness in handling data that have sequential characteristics. Sequential data are an ordered set of continuous events under different types, generated by a wide range of applications. Sequential anomaly detection aims to find out whether a subsequence of an event series is an anomaly.

This project investigates deep reinforcement learning (DRL) algorithms to solve sequential anomaly detection problem in cyber security data sets. The ability to select optimal sequential actions of a DRL method makes it highly capable of solving this problem. The emergence of deep learning and its integration into traditional RL methods have created a class of DRL methods that are able to solve complex, large-scale and high-dimensional problems. The proposed DRL approach is expected to be an advanced anomaly detection method for sequential data, especially in practical cyber security applications, including intrusion detection and response systems. Reference paper: https://arxiv.org/abs/1906.05799

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Next Generation Network Analytics through Quantum Computing**

Quantum computing promises to vastly decrease the run time of many important applications. The catch is that it is not possible to re-use classical algorithms to achieve performance improvement, and new algorithms must be specifically designed for the platform. This research intends to investigate the design of quantum algorithms in the realm of network defence analytics, a problem domain that is struggling to cope with the massive increase in data that modern computer networks produce. Our approach deals with this problem by designing quantum algorithms to improve on the best-known complexity of graph-based network defence algorithms and relies on exploiting the known structure in graph problems to accelerate quantum Monte Carlo algorithms.

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Spectral Methods and Sparisification Graph Analysis on Ethereum**

Research focused on the Bitcoin blockchain illustrated that the underlying graph is a scale-free network. In a similar style of research, but more nuanced, Chen et al. use a graph analysis approach for the analysis of money-flows and smart contract distributions on Ethereum. The authors look at three different dimensions of Ethereum’s logical network distribution, namely the flow of money (Ether) transfers, the patterns for the creation of smart contracts (Contract Creation Graph (CCG)) and the sequences of smart contract invocation. A popular model for detecting community structure in large graphs is the Stochastic Block Model (SBM). The exact parameters to recover the community structure of a SBM has been well studied, and many methods have been proposed to recover a nodes’ community membership. A popular approach is to use spectral methods where the Graph Laplacian L of the given graph is created, and the Fiedler vector of the graph is found. This vector is then used to cluster nodes in the same community. While a robust method, it can be expensive to compute the Fiedler vector exactly. This research examines the types of errors that can be tolerated using spectral methods while still recovering communities. The two sources of error considered are: (i) dropping edges using different sparsification strategies; and (ii) inaccurately computing the eigenvectors. This approach is to be tested on the Ethereum graph so that key communities of interaction on the blockchain can be recovered with dramatic reductions in computation.

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Text Analytics and Transfer Learning**

Transfer learning is a challenging and effective machine learning approach that aims at making use of prior knowledge gained in one domain or genre of text and applying the knowledge learnt in a novel domain. In the case of text classification, transfer learning has been lately approached using recent advances in the natural language processing domain, especially with distributed text embedding structures such as word and sentence embedding. In this project, the candidate will develop new insights in the domain of sentence and paragraph embedding to enhance the effectiveness of transfer learning in the context of text classification. The primary focus of the text classification task will be in the areas of emotion classification and/or medical condition detection from different genres of textual data.

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Rare Events, Outliers, Context Awareness and Bayesian learning**

Rare events and anomalies in the domain of Bayesian Intelligence pose several modelling and reasoning challenges. The inadequacy of prior knowledge and evidence related to such rare data points in combination with the lack of awareness of the data context makes it difficult for traditional Bayesian learning algorithms to effectively learn context-aware conditional probability tables that reflect the significance of outliers. Such outliers are, in some cases, of significant importance and can constitute major portions of human learning and improvement in several application domains, such as elite sports and medical treatment regimes. This research project will therefore focus on better understanding the current Bayesian learning algorithms and developing a context-aware learning technique that can signify and more accurately represent rare data points of interest.

** **

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Real-time Video Analytics in Internet of Things Systems**

The “Internet of Things” (IoT) has facilitated the widespread deployment of sensors to collect data for actionable insights. One such sensor is the smart camera. Today, smart cameras are widely deployed for various purposes, from security monitoring, traffic intersection control, to factory floor monitoring. Recent advances in computer vision, artificial intelligence (AI) and machine learning have opened up opportunities for real-time “video analytics” – a field of study where valuable insights are extracted from the videos collected to benefit science, society and business.

Video analytics are based on classical computer vision techniques and deep convolutional neural networks (CNN). While computer vision models and CNNs have become more accurate and capable, they alone are not enough to extract valuable insights from real-time videos. Unlike other “things” in IoT realm and other text/numeric processing algorithms, video analytics are resource-hungry and computationally-expensive. They require higher bandwidth (up to 5Mbps for HD streams at 60FPS), and lower latency (output of analytics is often used to interact with humans or other systems), higher compute provisioning (faster CPUs to deal with large volumes of data), richer query semantics and stricter security guarantees. When scaled to support multiple distributed cameras, video analytics also result in a large computational footprint.

In this project, we aim to build a scalable, secure, and reliable video analytic system that: (1) characterises the multi-dimensional requirements of video analytics for specific use-cases; (2) captures and process video data efficiently in distributed cameras; (3) dynamically adjusts the parameter configurations of CNN-based video analytics pipelines; (4) uses efficient computer vision algorithms for object detection and video analysis; and (5) efficiently monitors and manages computational/storage resources in a cloud-based infrastructure.

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Subjective Visual Quality-aware Deep Visual Enhancements for Video Conferencing Applications**

Video conferencing has become an essential business critical function. The underlying technologies, however, have not made any major breakthrough in decades. Even with the state-of-the-art equipment and technologies, users nowadays are still experience choppy and intermittent live video streams. The aim of this project is to generate smooth 1080p video in real-time (<100ms) using video captured from ordinary webcam with 720p or lower resolutions. The candidate will develop machine-learning based super-resolution kernels to address the tight latency constraints associated with the video conferencing applications. New video augmentation techniques will also be developed to maintain subjective visual qualities under dynamic network conditions.

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Data Fusion and External (Environmental) Situation Awareness for Autonomous Underwater Vehicles (AUVs)**

Two sources of information can be identified in order to provide external and internal Situation Awareness (SA) for autonomous underwater vehicles (AUVs); extracted knowledge from sensory data and domain knowledge provided by expert. In both cases, it is necessary for information to be diagnosed, classified, and accessed efficiently by the SA and decision-making unit while undertaking a mission. Different types of environmental events and a perimeter of stationary and moving obstacles can be obtained through the various scanning sensors platformed on the AUV. The relationships between the obtained information is used for event extraction and classification, such as recognizing a threatening obstacle, coping unexpected and uncertain disturbances, etc. To provide awareness of the surrounding environment, not only individual obstacles should be identified, but also higher-order relations among different objects must be derived in order to recognize the threat degree and uncertainty associated with an obstacle. Obstacle group classification and recognition, considering their dynamic deployment and behaviour based on their relative distance, angle, and velocity to the AV can explain the threat degree. For threat recognition, some questions can be answered using appropriate data fusion and accurate SA determined, such as:

- what type of object is it? (Multiple sensors can be used for redundancy detection and object identity management).
- what is the purpose and intent of the moving object?
- what is its predicted behaviour going forward in time?
- what is the threat (danger) degree of the object?
- what is the appropriate response when facing this specific object?

Current data fusion models lack the capability to fully support “situational cognition”. To this end, a model is required to imitate human cognition as it relates to internal or external SA to produce an appropriate decision-making process. Designing such a module enhances the overall system capability of threat/disturbance recognition and protection by classifying events and taking appropriate responses.

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Internal Situation Awareness (SA) and Fault Tolerance for Autonomous Underwater Vehicles (AUV)**

An AUV should be capable of recognizing any internal failure and accommodating an external situation. A fault-tolerant module helps the system to recognize the minor and major failures of AUV subsystems and take appropriate action accordingly. For example, the vehicle should be able to abort the mission safely in major failures or continue the mission (considering the new circumstances) for any tolerable failure. The internal fault-tolerant system includes three levels of:

- Fault detection (monitors the AUV’s overall hardware and software availabilities),
- Fault recognition (determines the severity of the fault and whether it is tolerable or not),
- Fault accommodation (dealing with the new situation by safely aborting the mission for when faults or risks are intolerable or reconfiguring the AUV control architecture to successfully continue the mission if the fault and risk is tolerable)

In fully autonomous missions, the performance of the onboard sensors and other subsystems are checked continuously during the mission. If a vehicle is damaged, or some of its components fail, mission adaptation will be required to cope with the new restricted capabilities.

The objective of this project is to cover only two initial stages of Fault detection and Fault recognition using appropriate machine learning approaches.

** **

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: Depression estimation from voice source related information**

Depression is a mood state and an aversion to activity that can affect a person’s thoughts, behaviours, feelings, and sense of well-being. Depression and anxiety disorders are evident worldwide, diseases that burden individuals, families and society. Accurate depression classification and estimation are therefore of great importance and have broad application prospects. Recently, some diagnosis of major depression using traditional machine learning has been investigated from data extracted from wearable sensors. Using traditional machine learning, human experts need to extract valuable information and then apply classification methods. However, deep learning methods are recently prominent in different domains. In this project, the candidate will develop a new deep learning model to enhance the estimation of depression using speech source related information.

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: ****Privacy, AI Ethics, and Data Transparency in an AI-driven World**

New privacy regulations, such as the EU General Data Protection Regulation (GDPR), highlight data transparency as a fundamental principle when dealing with personally identifiable information. If not used properly, big data technologies and powerful machine learning algorithms may lead to improper data use. Ethics in AI and data transparency walk hand in hand. Privacy, on the other hand, may have conflicting requirements (e.g., algorithm transparency). Existing data transparency definitions have been debunked on the basis of overlooking data privacy, trustworthiness, fairness, and usage. Examples where existing data transparency measures fail include:

- Provisioning of individual’s data together with data of other individuals to a third party,
- Decisions made about an individual based on a population sample data that does not include the individual’s data, or
- Data used for scientific conclusions. For instance, the origin of data and any changes to improve trustworthiness, ethics, and quality (e.g., COVID-19 tracing apps data).

This project aims to design techniques and tools for data transparency infrastructure, considering the association of usage purposes of data and prevention of misuse while facilitating access, verification, and visibility.

** **

**Principal supervisor: Prof. Peter Eklund**

Contact: https://www.deakin.edu.au/about-deakin/people/peter-eklund

**Lab: Machine Intelligence Lab**

**Title: ****Deep Learning for Insider Threat Detection**** **

A 2019 study reported a 47 percent increase in insider threat incidents over a year with an estimated cost of $11 million to clean-up associated costs. Malicious or accidental, the key indicator of a possible insider attack is the change compared with normal behavior. Complexity, heterogeneity, sparsity, high-dimensionality, and lack of labeled insider threats challenge the traditional machine learning approaches to differentiate between insiders and normal users. Recent literature indicates that deep learning models can improve the detection of insiders. However, lack of labeled data and adaptive attacks are of the several limitations to use deep learning for this. In this project, we aim to investigate this new direction to detect insiders and devise solutions to address the associated challenges.

** **

**Principal supervisor: Dr Sunil Aryal**

Contact: https://www.deakin.edu.au/about-deakin/people/sunil-aryal

**Lab: Smart Networks Lab**

**Title: Unsupervised learning from data with mixed (continuous and discrete) attributes**

Most machine learning algorithms assume data attributes are numeric/continuous. But, real-world problems have both continuous and discrete (ordinal or nominal) attributes resulting in mixed continuous-discrete domains. While there are methods for supervised learning (e.g., decision trees, random forest, Naïve Bayes) that can handle mixed attributes quite well, there are not many methods for unsupervised learning such as clustering and anomaly detection, which can work with mixed attributes data directly. The common practice of handling mixed attributes is converting them into continuous only or discrete only, and using methods designed for continuous or discrete domains only. A continuous attribute can be converted into a discrete attribute through discretisation. A discrete attribute can be converted continuous attributes using one-hot encoding. Converting one type of attributes into another results in information loss, which may result in deteriorating performances of learning algorithms. This project aims to develop unsupervised learning (clustering and anomaly detection) algorithms that can handle mixed attributes directly. The proposed algorithms will be tested against state-of-the-art methods using publicly available real-world datasets with mixed attributes.

**Principal supervisor: Prof. Maia Angelova**

Contact: https://www.deakin.edu.au/about-deakin/people/maia-angelova-turkedjieva

**Lab: Data Analytics Research Lab**

**Title: data driven models of depression**

The aim of this project is to study patterns of depression and association with daily activities in different population groups. We will consider association between social isolation, daily activities, sleep and environmental factors for developing symptoms of depression. The project could focus on the special case of effects of COVID19 and social isolation on wellbeing and the quality of life. Data analytics, machine learning and deep learning techniques, as well as mathematical modelling techniques will be used to reveal patterns from time series data. The project requires a solid mathematical and statistical background, foundation knowledge of data analytics and machine learning, as well as strong programming skills.

**Principal supervisor: Prof. Maia Angelova**

Contact: https://www.deakin.edu.au/about-deakin/people/maia-angelova-turkedjieva

**Lab: Data Analytics Research Lab**

**Title: Models of Sleep Fragmentation**

This study will investigate patterns of sleep and sleep fragmentation in normal healthy individuals as well as sleep fragmentation which is a result of health or environmental stressors. We will use objective methods to measure the effects based on wearable devices. The project will develop automated methods for sleep classification and sleep stages classification in normal and disturbed sleep. The project requires a solid mathematics and statistics background and sound programming skills.

**Principal supervisor: Prof. Maia Angelova**

Contact: https://www.deakin.edu.au/about-deakin/people/maia-angelova-turkedjieva

**Lab: Data Analytics Research Lab**

**Title: Glucose-Insulin regulation in Type I and II Diabetes. **

This project will develop a dynamic systems model of glucose-insulin regulation based on delay differential equations. The model will be able to distinguish between type I and Type II diabetes. A particular interest will be paid on insulin sensitivity and its possible effects on young and otherwise healthy individuals. The project requires solid mathematical knowledge and sound programming skills. An undergraduate degree in mathematics or physics would be beneficial.

**Principal supervisor: Prof. Maia Angelova**

Contact: https://www.deakin.edu.au/about-deakin/people/maia-angelova-turkedjieva

**Lab: Data Analytics Research Lab**

**Title: Objective measurement of fatigue in athletes and in the workplace**

This project investigates what is fatigue in athletes and how it can be measured objectively using a limited number of wearable devices. It will involve updating and expanding current understanding of fatigue, and developing sound and practical methods for its measurement and automatic detection. This project requires solid programming skills and sound knowledge of statistics.

**Principal supervisor: Prof. Maia Angelova**

Contact: https://www.deakin.edu.au/about-deakin/people/maia-angelova-turkedjieva

**Lab: Data Analytics Research Lab**

**Title: Investigation of patterns of healthy ageing with wearable devices**

This project will investigate changes in physiological and cognitive function as we age and how these changes can be objectively detected using wearable devices. The project will develop measures of physical performance based on measurements from. Wearable devices and automatic methods for classification of healthy from non-healthy (i.e. chronic) conditions. The project requires solid programming skills and sound knowledge of statistics.

**Principal supervisor: Prof. Maia Angelova**

Contact: https://www.deakin.edu.au/about-deakin/people/maia-angelova-turkedjieva

**Lab: Data Analytics Research Lab**

**Title: Knowledge Discovery from Financial Time Series**

The aim of this project is to develop data driven methods for pattern discovery and forecasting of complex time series. The project will involve solid theoretical work on extension and generalisation of existing “distance” and entropy-based measures and metrics for data mining and models for forecasting of time series trends and behaviour. The results will be applied to public financial time series data. The project requires a solid mathematics and statistics background.

**Principal supervisor: A/Prof. Gang Li**

Contact: https://www.deakin.edu.au/about-deakin/people/gang-li

**Lab: Data Analytics Research Lab**

**Title: Data Privacy Compliance Assessment**

Privacy breaches are a regular occurrence in contemporary society with the public becoming increasingly concerned with how their digital data will be handled. Hence, regulatory organizations through the world, are formulating data privacy legislations, such as the EU’s GDPR. Those data related regulations are requiring significant manual effort to determine whether a compliance violation has occurred. This project will develop the theory and an automated system, in which the bounds of privacy loss on the targeted individual can be quantified with heterogeneous privacy settings.

**Principal supervisor: Dr Jingyu Hou**

Contact: https://www.deakin.edu.au/about-deakin/people/jingyu-hou

**Lab: Data Analytics Research Lab**

**Title: Data Privacy Compliance Assessment**

New medical and biological technologies have generated various and massive data that makes it possible and feasible to apply data analytics and computational techniques to discover the knowledge hidden in these massive data resources. In this project, the candidate will explore the available data resources, review state-of-the-art techniques and propose innovative computational approaches and methods to make use of the available data sets for facilitating the diagnosis of one kind of disease such as cancer, Alzheimer’s disease, Parkinson’s disease, or tuberculosis. The available data resources could be next-generation sequencing (NGS) data, microarray data, clinical data, medical images and more. The proposed approaches and methods are expected to be applied to practical applications, as well as be adapted for the diagnosis of other diseases. The success of this project will lay a solid foundation for industrial collaborations and competitive research grant applications.

**Principal supervisor: Dr Daniel Ma**

Contact: https://www.deakin.edu.au/about-deakin/people/daniel-ma

**Lab: Machine Intelligence Lab**

**Title: On the vulnerability of deep learning models to adversarial attacks**

This project explores the varies vulnerabilities of state-of-the-art deep learning models to both white-box and black-box attacks. Most of existing works are done on deep learning-based image classification models, and they were done in a digital setting. This project will extend current works to more boarder areas of deep learning. Powerful attacks can be designed to examine the adversarial vulnerabilities of deep neural networks in applications like face recognition, object segmentation, question answering, natural language processing, speech recognition, video analysis and graph modelling. Furthermore, deep learning based medical systems should also be evaluated against these attacks. Defense against various adversarial attacks under different settings (eg. White-box, black-box and grey-box) will also be developed to protect those systems. Physical-world attacks is also an important part of this project to verify whether those attacks can really happen in real-world scenarios.

** **

**Principal supervisor: Dr Daniel Ma**

Contact: https://www.deakin.edu.au/about-deakin/people/daniel-ma

**Lab: Machine Intelligence Lab**

**Title: Weakly supervised learning with deep neural networks**

This project aims to explore the robustness of machine learning models to noisy annotations/labels. Most of existing works are done for simple image classification tasks. The extension of existing works to object segmentation, and natural language processing will be a part of this project. Weakly supervised learning plays an interesting role in machine learning due to its connections to other learning schemes, such as federated learning and adversarial training. In other words, weakly supervised learning is between supervised learning and unsupervised learning. With the emerging of unsupervised learning techniques like BYOL, it requires deeper understandings as for why unsupervised learning can work as good as supervised learning. Weakly supervised learning can provide an interesting angle into those questions. On the other hand, applying weakly supervised learning techniques in other learning schemes often lead to surprising results.

**Principal supervisor: Dr Sergiy Shelyag**

Contact: https://www.deakin.edu.au/about-deakin/people/sergiy-shelyag

**Lab: Data Analytics Research Lab**

**Title: Theory and applications of delay-differential equations**** **

Delay differential (and, generally, functional differential) equations (DDEs) and their systems appear in natural and artificial phenomena, when the behaviour of a system explicitly depends both on its current state and its history in some functional form. Among such systems are communication networks, systems of biological and physiological regulation, population growth, infection spread, epidemics and pandemics, devices with actuators and delayed feedback, business cycle models in economics, decision making. Unlike ordinary differential equations and their systems, which are finite-dimensional in phase space, DDEs are infinitely-dimensional. Inclusion of a delay in a dynamical system can lead to rather complicated dynamics, (sometimes unwanted) oscillations and even chaos. Analysis of DDEs is generally more involved, in part due to the structure of the corresponding characteristic equations, and often not allowing for an analytical treatment. Numerical solution of such equations is also not trivial due to propagating discontinuities and strict requirements to interpolation techniques. Nevertheless, in the recent years advances in understanding of DDEs, and analytical and computational approaches to their solution have been achieved.

In this project, we will advance theory and applications of the delay-differential equations and their systems with variable time-dependent coefficients in models of glucose-insulin interaction and of neurophysiological regulation of sleep. We will study the mathematical properties of the model systems with the aim to provide detailed explanations for the physiological processes behind sleep disturbances and disruptions of glucose-insulin regulation system with the aim to support design of mitigation strategies for a range of sleep and metabolic disorders.

**Principal supervisor: Dr Sergiy Shelyag**

Contact: https://www.deakin.edu.au/about-deakin/people/sergiy-shelyag

**Lab: Data Analytics Research Lab**

**Title: Anomaly detection in very low frequency radio wave time series**

Very low frequency radio waves from standard radio transmitters can travel vast distances around the Earth. On their way, they collect the information about the medium, through which they propagate. This information includes noise, regular variations, such as day-night, or seasonal changes, as well as the transient information on large-scale geophysical anomalies and events, such as earthquakes, volcano eruptions, tsunamis, bushfires.

A network of receivers has been created around the globe to receive those radio waves. Currently, the big question is how to understand and analyse the very low frequency radio wave propagation data in a hope to better understand and forecast geophysical hazards. In this project, we will be looking at ways to answer this big question. We will be using a variety of methods for time series analysis on already existing time series in order to find the best ways to post-predict known large-scale geophysical events. Our aim will be to find and analyse the signatures of those events in the data before they occurred at the Earth surface.

**Principal supervisor: Dr Sergiy Shelyag**

Contact: https://www.deakin.edu.au/about-deakin/people/sergiy-shelyag

**Lab: Data Analytics Research Lab**

**Title: Machine learning for space weather forecasting**

Large-scale space weather events, such as solar flares and coronal mass ejections, have potential to cause geomagnetic storms in near-Earth environment. These storms can be enormously damaging both to space infrastructure, such as communication satellites, and to the ground-based technology, such as local communications, electrical grids, gas and oil pipelines. It is estimated that an event, similar to the geomagnetic storm of 1859 (“Carrington event”), would nowadays cause massive, multi-trillion economic damage worldwide due to catastrophic long-term, widespread electric outages. Therefore, an urgent need to reliably forecast large-scale space weather events exists in order to mitigate their potential negative economic impact.

In this project, we will use a number of machine learning techniques, including deep learning, for solar far-side imaging and predictions of solar magnetic field configurations, prone to producing the solar flares. Multiple data sources, which include time series and imaging data from multiple space- and ground-based space weather and solar observational instruments, and modern data fusion techniques will be used to develop more robust and reliable methods, algorithms and tools for space weather forecasting.

**Principal supervisor: Dr Nayyar Zaidi**

Contact: https://www.deakin.edu.au/about-deakin/people/nayyar-zaidi

**Lab: Machine Intelligence Lab**

**Title: Large Scale Effective and Explainable Machine Learning**

Machine Learning algorithms have matured deeply in the last decade. Their applications are prevalent in every aspect of our daily life as they constitute a key part in any Artificial Intelligence System. We believe that a hallmark of a good learning algorithm lies in two things a) how quickly it learns from the data and b) how easily it can explain its predictions. Keeping these two criteria in mind, this project investigates learning effective yet efficient learning algorithms. This project will investigate limited pass learning algorithm such as restricted Bayesian Networks to learn an effective yet a low-biased model. Bayesian Networks have this property that they are easily interpretable and hence explainable, this project will investigate this property under the umbrella of deep learning. The project will also investigate knowledge fusion in model building stage for better predictions. The outcome of the project will be a model that is comparable or better in performance to state of the art models such as XGBoost, but more efficient in terms of training and testing, as well as having state of the art capability in model explanation.

**Principal supervisor: Dr Sutharshan Rajasegarar**

Contact: https://www.deakin.edu.au/about-deakin/people/sutharshan-rajasegarar

**Lab: Data Analytics Research Lab**

**Title: Computer vision for emotion recognition**

Human emotion are often expressed using facial emotions and bodily gestures. Detecting emotional behaviors is important in the field of Artificial Intelligence (AI), such as for improving human machine interactions, and behavioral detection. Accurately detecting the various emotional behaviors from videos is challenging due to various factors, such as background noise and the scene complexity with multiple objects and people. The aim of the project is to propose novel deep learning based algorithms for accurately profiling emotions from videos and other modalities. Students who have prior exposure to machine learning and achieved high scores, such as H1 (honors) in their prior studies, are encouraged to apply. Further, having peer reviewed publications in good venues (journals, conferences) in the related area will increase the chances of securing a scholarship.

**Principal supervisor: Dr Sutharshan Rajasegarar**

Contact: https://www.deakin.edu.au/about-deakin/people/sutharshan-rajasegarar

**Lab: Data Analytics Research Lab**

**Title: Learning Activities of Daily living from Wearable Devices**

Wearable devices, such as wrist-worn accelerometer sensors are becoming popular nowadays, and mostly used for detecting physical activities. Activities of daily living are often complex and involve combinations of several bodily movements. Detecting such activities from sensor data is challenging because of the complex nature of the movements involved in performing certain activities; for example, drinking while walking. This require analyzing signals and learning patterns of movements for accurately detecting them. The aim of this project is to propose novel suits of classical machine learning and deep learning based algorithms to learn the patterns of movements accurately and detecting the complex activities using the sensor data. Students who have prior exposure to machine learning are encouraged to apply. Further, having peer reviewed publications in good venues (journals, conferences) in the related area will increase the chances of securing a scholarship.

**Principal supervisor: Dr Sutharshan Rajasegarar**

Contact: https://www.deakin.edu.au/about-deakin/people/sutharshan-rajasegarar

**Lab: Data Analytics Research Lab**

**Title: Crowd Behavior Analysis and Anomaly Detection**

Crowd behavior analysis from videos is important for leaning anomalous crowd movement patterns, surveillance and emotion behavioral studies. Devising automated models for crowd behavior analysis is challenging, and only a limited research has been done in the past in this context. It has challenges in terms of detecting collective and individual behaviors in a crowded scene from videos and handling occlusions and partial images from the scenes. The aims of this project include surveying the existing state-of-the-art methodologies in detecting crowd behaviors, and proposing novel suits of deep learning based intelligent crowd behavior analysis frameworks. Students who have prior exposure to machine learning are encouraged to apply. Further, having peer reviewed publications in good venues (journals, conferences) in the related area will increase the chances of securing a scholarship.

**Principal supervisor: Dr Sutharshan Rajasegarar**

Contact: https://www.deakin.edu.au/about-deakin/people/sutharshan-rajasegarar

**Lab: Data Analytics Research Lab**

**Title: Learning the Focus of Attention to Detect Distributed Coordinated Attacks**

Cyber security has become a strategic national priority given the dependence of modern society on Internet-based services. While cyber defences have improved, attackers are also becoming more sophisticated in the design of their attacks in order to evade defences. Distributed coordinated attacks have grown, in which attackers use a large number of infected hosts that are widely distributed across the Internet to generate malicious traffic in a coordinated manner. Cyber security analysts need to detect and respond to the coordinated attacks as soon as possible, to minimize the damage attackers can inflict. However, the growth in highly distributed attacks that span multiple networks has meant that massive volumes of data need to be analysed. While machine learning techniques can help filter the data, we need techniques that can automatically provide a focus of attention for analysts on the most relevant observations. The aim of the project is to propose novel suite of deep learning-based algorithms that can focus the search of security analytics techniques, which substantially improve the accuracy and efficiency of distributed coordinated attack detection in high volume distributed security data streams. Students who have prior exposure to cybersecurity and machine learning are encouraged to apply. Having peer reviewed publications in good venues (journals, conferences) in the related area will increase the chances of securing a scholarship.

**Principal supervisor: Dr Guillermo Pineda Villavicencio**

Contact: https://www.deakin.edu.au/about-deakin/people/guillermo-pineda-villavicencio

**Lab: Mathematics**

**Title: Combinatorics of polytopes**

A *polytope* is the intersection of all convex sets containing a finite set of points in the Euclidean space, its *vertices*. The dimension of a polytope is the maximum number of affinely independent points in the polytope minus one; the maximum number of affinely independent points in R^{d} is *d+1*.

A polytope is structured around other polytopes, its faces. A *face* of a polytope P is P itself, or the intersection of P with a hyperplane that contains P in one of its closed halfspaces. In a d-dimensional polytope, or d-polytope for short, the vertices are its 0-dimensional faces, and the *edges* are its 1-dimensional faces. The vertices and edges of a polytope form *graph*, where two vertices are adjacent if they belong to the same edge.

The project aims to gain insights into graphs of polytopes, and to pinpoint to what extent they explain properties of polytopes. Two questions make these aims concrete.

- Given the graph of a polytope, what properties can we infer about the polytope?
- Given a class of polytopes, what properties characterise their graphs?

The outcome of the project will be a framework to answer these two questions. We investigate graph-theoretical topics such as connectivity, colourings, linkedness of graphs of polytopes. These are well-established topics within graph theory, with dedicated books, and chapters in most textbooks (Diestel, 2017), but not yet in polytope theory.

**Principal supervisor: Prof. Gleb Beliakov**

Contact: https://www.deakin.edu.au/about-deakin/people/gleb-beliakov

**Lab: Mathematics**

**Title: Optimisation and decision making without additivity**

This project aims at developing mathematical and computational methods to optimise cost functions, or to make decisions based on multiple criteria, where there is significant interaction between the criteria or the objectives. The theory of capacities, or cooperative games, handles the interactions between the criteria by considering the worth of all possible combinations of criteria, which are aggregated by the Choquet integral. The model is non-additive, in the sense that the combination of some inputs can be larger or smaller than their sum. The power of this theory creates a challenge of computational complexity, an exponentially large number of parameters. This project will explore strategies for reducing model parameters, efficient computational optimisation techniques and also applications of capacities and fuzzy integrals. This project will involve mathematics, operations research, computational modelling and data analysis.

**Principal supervisor: Dr Sergey Polyakovskiy**

Contact: https://www.deakin.edu.au/about-deakin/people/sergey-polyakovskiy

**Lab: Mathematics**

**Title: Advanced Decomposition Techniques for Multi-Component Optimisation Problems**

Real-life optimisation problems often consist of several sub‐problems of different nature. Not only do they combine several optimisation aspects into a single problem, but they also emanate from the compounded complexity of conflicting issues in numerous areas like logistics, planning and manufacturing. Solving them requires a thorough understanding of both their compounded and their individual natures. As traditional optimisation methods may demonstrate only limited efficiency for such problems, designing advanced decomposition approaches hybridizing several algorithmic techniques to handle their specificity and non‐linear behaviour intrinsic to them appears promising.

The focus of this research is on perspective decomposition methods, search strategies, and learning as a way to tune the search process at the runtime. On the application side, it aims to develop state‐of‐the‐art solution techniques for a number of multi-component optimisation problems. For a technique to be high-performing, it is of vital importance to be capable to adaptively select sub-problems to solve in a way that ensures fast convergence towards an optimal solution. Tuning the search process appears beneficial, but a posteriori decision-making based on a subset of test instances tends to produce ambiguous settings unsuitable for the whole set. It can ignore promising decisions as the search progresses and thus seriously affect the efficiency of the entire search. More sophisticated decision methods using learning mechanisms during the search should guarantee advanced performance, but require designing new methods to measure the search performance itself. Finding such techniques is a part of this research.

**Principal supervisor: A/Prof. Vicky Mak**

Contact: https://www.deakin.edu.au/about-deakin/people/vicky-mak

**Lab: Mathematics**

**Title: Multi-Orbit Satellite Constellation Resource Optimization**

Multi-orbit satellite constellations, in low, medium, and geosynchronous earth, i.e., LEO, MEO and GEO provide exciting and unparalleled opportunities in the satellite communications domain. The multi-orbit constellations will provide unparalleled advantages, including among others the ubiquitous coverage to polar regions, maritime and aerospace with ultra-low latency, edge-processing for 5G and IoT networks, quantum key distribution, and effective disaster management. An important prerequisite for realizing these promises of multi-orbit constellations is to optimally manage the space and ground resources and the communications schedules among orbits and ground. This research project will devise techniques for dynamic resource management in multi-orbit satellite constellations.

The research student will investigate the issues and constraints in the multi-orbit constellations. The students will be responsible for (1) deriving mathematical models and algorithms for the dynamic management of ground stations and inter-orbit communications; (2) using explainable AI to analyze and establish relations between (i) the data traffic patterns based on weather conditions, geographical locations and demographics; and (ii) space and ground resources; (3) help devise decision-making tools for satellite operators based on steps 1 and 2; (4) develop visualization tools for dynamic resource management; (5) run a comprehensive set of experiments.

**Principal supervisor: A/Prof. Vicky Mak**

Contact: https://www.deakin.edu.au/about-deakin/people/vicky-mak

**Lab: Mathematics**

**Title: EV fast-charging network optimisation**

Building a national electric vehicle fast-charging network is one of the national high priority infrastructure initiatives. By 2040, electric vehicles (EVs) are projected to account for 70% to 100% of new vehicle sales and at least 30% of the vehicle fleet in Australia. According to the Electric Vehicle Council, lack of access to charging stations has been identified by around two-thirds of motorists as a key barrier to the adoption of EVs. Australia currently has less than 2,000 public charging stations, of which approximately 250 are fast‑charging. This research is to derive techniques for decision makers to optimise the national electric vehicle fast-charging network.

The research the student will be completing include: 1) analysing the current network use pattern of existing EV users, 2) predicting the network use pattern of EV users by 2040 and beyond, 3) deriving mathematical models and algorithms for finding the optimal locations of EV fast-charging stations based on a given set of parameters, 4) using simulation to generate parameter sets for the optimisation problems in 3) and run a comprehensive set of numerical experiments, and 5) using explainable AI to analyse the relations between the parameters and solutions and derive a decision tool for recommending a small number of decision plans for users.

**Principal supervisor: A/Prof. Vicky Mak**

Contact: https://www.deakin.edu.au/about-deakin/people/vicky-mak

**Lab: Mathematics**

**Title: Optimising self-organising mobile agents **

Given a task or a set of tasks, finding the optimal number of mobile agents required to complete the task(s) and the optimal sequence of procedures to carry out for each of the agents, is a highly complex optimisation problem. We consider the case for a set of self-organising mobile agents where the complete nature of the task(s) is not known in advance, and the agents are required to process the information acquired about the task(s) “on the fly” and derive an optimised plan to carry out the task(s). The outcome of the research can be applied to the construction, mining, and many other sectors where using mobile agents can significantly improve the safety of human workers.

The research that the student will be conducting is 1) to derive a generic scenario of application and working with industry (construction, mining, etc) to identify three use cases, 2) a framework for solving the two optimisation problems: the optimal number of mobile agents required to complete the task(s) and the optimal sequence of procedures to carry out for each of the agents, and 3) implement the new techniques for solving problem instances from the three industry use cases.