Decentralized Federated Learning at the Edge

Federated Learning (FL) is a concept in Machine Learning that is based on distributing the workload of training onto the user machines. A local model is provided for each user machine, and the model is trained periodically using the local user data. Federated learning dictates that the contribution of the user to the training would be only the model updates. Therefore, users will send only internal model delta values to a central server. In that way, user data is preserved locally and is inherently secure against potential adversaries in the network. Federated learning involves security and data privacy, as well as distributed learning. Further, the training and inference of FL models often takes place at the edge, as there might be privacy implications of transmitting raw data to the cloud. Further, a more localized deployment of applications can contribute to reducing the wide-area bandwidth usage, as well as the response time, in contrast with a cloud-based deployment.

In contrast to federated learning with a central server, decentralized federated learning is defined in a serverless topology. In DFL, the local updates are sent to an elected client or a middle machine (e.g., one of the edge nodes) to perform the final aggregation and model update and broadcast the new version of the model back to the clients. Unlike FL, DFL is not vulnerable due to the single point of failure. DFL however, is a new concept and needs to be explored in all the aspects that are defined for FL as well, including machine learning, networking, and cybersecurity.  The PhD position will be at the intersection of machine learning, distributed computing, edge computing, and cybersecurity.

As a PhD student, you will be working in the Cloud-to-Edge Lab, in which we conduct cutting-edge research in the broad field of computer systems. You will be expected to engage with all the accessible knowledge including fundamentals, as well as the state-of-the-art, concerning the intersection of the various aspects of the research. Further, you will be expected to develop new systems, models, methods and algorithms; conduct experiments using the lab resources; develop new prototypes and simulators if needed;  and write scientific papers in top-tier venues. As research is often done in teams, it is anticipated that the candidate will be willing to get involved in collaborations with other researchers and/or students.  

To apply, please follow the instructions at: https://www.juliengs.ca/openings-prospective-students/