Validation of Federated Unlearning on Collaborative Prostate Segmentation

Description

Machine Unlearning (MU) is an emerging discipline studying methods to remove the effect of a data instance on the parameters of a trained model. Federated Unlearning (FU) extends MU to unlearn the contribution of a dataset provided by a client wishing to drop from a federated learning study. Due to the emerging nature of FU, a practical assessment of the effectiveness of the currently available approaches in complex medical imaging tasks has not been studied so far. In this work, we propose the first in-depth study of FU in medical imaging, with a focus on collaborative prostate segmentation from multi-centric MRI dataset. We first verify the unlearning capabilities of a panel of FU methods from the state-of-the-art, including approaches based on model adaptation, differential privacy, and adaptive retraining. For each method, we quantify their unlearning effectiveness and computational cost as compared to the baseline retraining of a model from scratch after client dropout. Our work highlights a new perspective for the practical implementation of data regulations in collaborative medical imaging applications.

Abstract

International audience

Additional details