The performance of artificial intelligence based learning models is often initially assessed using well established benchmark datasets. However, satisfactory performance on these datasets does not guarantee similar performance in real-world settings where the data may be significantly more imbalanced. For example, in a medical dataset a positive diagnosis may be a relatively rare event and will, therefore, form a minority class within the dataset.
One form of learning across multiple organisations/devices/clients is Federated Learning (FL). It is a mechanism for efficient, collaborative training of a shared model; where participating clients (e.g. edge nodes, sensors, devices) cooperate to train a model by exchanging trained parameters with a central server that coordinates the learning process. The security and privacy of the original data is enhanced as it is not shared with the central server.
Group fairness in machine learning is focused on ensuring fairness across different groups defined by sensitive attributes. However, as federated learning is not carried out on the raw data it has the potential to increase, rather than reduce biases. This gives rise to the need for the development of mechanisms to ensure fairness in FL.
This project will explore the use of federated learning with imbalanced data sets, where the focus of exploring methods to preserving the fairness of FL models while addressing class imbalance within the dataset.