On the security and robustness of federated learning with application to smart grid infrastructures

In the past two decades, machine learning has fast emerged as a popular approach to solving many high dimensional problems. It focuses on analysing and interpreting patterns and structures in data to enable learning, reasoning, and decision making. A major benefit of machine learning arises from its ability to produce a model or models that can be applied to solve a wide range of problems. For example, machine learning is commonly used in computer vision systems to detect various objects within images, such as traffic signs at a roadside, without having a developer write code that accounts for every possible way that object can be present within an image. However, machine learning is known to be ``data hungry'' in that it requires vast datasets with significant amount of variation, to produce accurate models and results. Hence, sourcing of this data can have major implications, especially when it comes to private data pertaining to people. This led to the development of federated learning, which is a form of distributed machine learning across many clients who hold their own independent data which is not shared with the central machine learning model. The clients each train a copy of the machine learning model on their own dataset and upload the resulting trained model to a central server. The server aggregates the client models together to produce a new global model, which is sent back to the clients for the next round of training. The federated learning algorithm aims to maintain data privacy by replacing the requirement that distributed learning needs to have the clients share data, instead tasking them with sharing the model. Despite its improvements to privacy, federated learning still has several challenges when it comes to security and robustness. In this thesis, we make several theoretical and analytical contributions to the challenges of robustness, privacy and fairness and their combined effects in the federated learning setting.