Deakin-Coventry Cotutelle - Bottom-up swarm computing for workload placement in edge-cloud continuum

This is a doctoral Cotutelle project between Deakin University (Australia) and Coventry University (United Kingdom). The successful PhD Student will be awarded a scholarship from Deakin University with the supervision team being drawn from Deakin University and Coventry University. The PhD Student will graduate with two testamurs, one from Deakin University and one from Coventry University, each of which recognises that the program was carried out as part of a jointly supervised doctoral program. The PhD Student is anticipated to spend a minimum 6 months and a maximum of 12 months (with approval) of the total period of the program at Deakin University, with the remainder of the program based at Coventry University.

The program is for a duration of 3.5 years and scheduled to commence in May 2025.

This project will conduct research on resource allocation in federated learning by jointly considering communication, computation, and data heterogeneity.
Edge intelligence has been recognised as a key technology for next-generation wireless networks such as 6G. Driven by the recent success of mobile edge computing, edge intelligence pushes computation-intensive artificial intelligence tasks from the centralised cloud to distributed base stations at the wireless network edge, to efficiently utilise the massive data generated by numerous edge devices. However, the large volume of data at edge devices must be properly processed (potentially jointly processed with sensed data from other sensors such as camera and Lidar) via artificial intelligence models in a swift manner, in order to support applications with ultralow-latency sensing, communication, computation, and control requirements.

Federated learning has emerged as a promising solution, where edge devices can iteratively exchange their locally trained artificial intelligence models to update the desired global model in a distributed manner, while preserving data privacy at each edge device. However, given the scarcity of spectrum resources, the wireless communication for exchanging model parameters is recognised as the performance bottleneck for federated learning. To tackle this issue, the over-the-air federated learning technique has recently been proposed, which allows distributed edge devices to concurrently transmit their local gradient or model updates for 'one-shot' aggregation at each communication round. In this way, communication and computation are seamlessly integrated to enhance the communication efficiency in federated learning, thereby supporting various low-latency and communication-efficient applications. However, realising the potential benefits of over-the-air federated learning poses several technical challenges due to the resource and data heterogeneities at distributed edge devices, as follows.

1. Resource limitation and heterogeneity: While distributed learning tasks are generally resource-consuming, the participating edge devices only have limited computation and communication resources, and their available resources may vary significantly. On one hand, the local iteration latencies of edge devices may differ significantly due to distinct computation capabilities. On the other hand, the over-the-air aggregation errors highly depend on the channel conditions and the transmit power at distributed edge devices.
2. Data heterogeneity: Data heterogeneity is another key factor affecting the performance of over-the-air federated learning. Specifically, in practical wireless networks, training datasets are distributed at different edge devices in a non-i.i.d. manner, and the number of available data samples may be highly unbalanced among different edge devices.

The aforementioned resource limitation and heterogeneity issues essentially lead to distinct uploading time between different devices and edge servers. As a result, if the edge server uses synchronised aggregation of the model updates of the edge devices, the edge device with the longest delay dominates the communication time of a single round. In addition, the aforementioned data heterogeneity issue causes the data collected at different devices to have different distributions, depending on the application scenarios, locations, and user behaviours. Therefore, it is important to conduct research on resource allocation in over-the-air federated learning.

Applications close 5pm, Wednesday 15 January 2025

Please be aware that screening for this advert will commence immediately and the scholarship may be awarded prior to the closing date.

This scholarship is supported by Coventry University, is available over 3.5 years and includes:

• Stipend of £18,622 per annum (2024 rate)
• A Tuition Fee Waiver
• Travel Support Package including one return economy airfare to Deakin University to support residency period in Australia
• Student visa and health insurance costs for period of residency at Deakin University in Australia

Deakin University will also provide a full tuition fee waiver for a duration of up to 4 years.

To be eligible you must:.

• be either a domestic or international candidate. Domestic includes candidates with Australian Citizenship, Australian Permanent Residency or New Zealand Citizenship.
• meet the PhD entry requirements of both Deakin University and Coventry university, including English language proficiency requirements
• be enrolling full time
• be able to physically locate to both Coventry University (UK) and Deakin University (Australia)

Please refer to the research degree entry pathways page and Coventry’s research entry criteria page for further information.