Certifying Data Removal in Federated Learning

This paper addresses the critical issue of data privacy and the 'right to be forgotten' in vertical federated learning (VFL). It proposes a novel algorithm, FedORA, to efficiently and effectively remove the influence of specific data points or labels from trained models in a distributed setting. The focus on VFL, where data is distributed across different parties, makes this research particularly relevant and challenging. The use of a primal-dual framework, a new unlearning loss function, and adaptive step sizes are key contributions. The theoretical guarantees and experimental validation further strengthen the paper's impact.

• •Proposes FedORA, a novel algorithm for sample and label unlearning in Vertical Federated Learning (VFL).
• •Employs a primal-dual framework to address the challenges of distributed feature architecture in VFL.
• •Introduces a new unlearning loss function that promotes classification uncertainty.
• •Provides theoretical guarantees on unlearning effectiveness and demonstrates comparable performance to training from scratch with reduced overhead.