Learning and Unlearning Your Data in Federated Settings

Federated learning is a distributed machine learning approach, allowing multiple data owners to collaboratively train a machine learning model while never revealing their data sets to other data owners. It is seen as a promising approach towards achieving data privacy, and it has already proven useful in several ubiquitous applications, such as predictive spelling.

Machine unlearning is another emerging machine learning sub-field, focusing on the need to minimize, and/or fully remove a data input from a training data set, as well as from the trained machine learning model. The need for machine unlearning comes as a direct response to recent regulations, such as the GDPR's right to erasure, and the right to be forgotten. In this proposal, we focus on the question of machine unlearning in the context of federated learning. Due to the distributed and collaborative nature of federated learning, simply removing a data input from a training set, and re-training the model typically is not possible. More commonly, different approaches, referred to federated unlearning need to be used. We introduce and analyze several such federated unlearning approaches, in terms of their ability to unlearn and performance. We also provide guidance for a practical federated unlearning method.