ConFUSE: Confusion-based Federated Unlearning with Salience Exploration

The increasing scale and complexity of deep neural networks, coupled with heightened privacy concerns, has under-scored the importance of developing techniques that align with privacy regulations such as the GDPR and CCPA. These laws mandate the 'right to be forgotten', which presents a significant challenge in the context of Federated Learning (FL). FL models trained collaboratively without sharing private data, necessitate efficient unlearning methods that allow for the deletion of specific data without retraining from scratch, which is both computationally and communicatively demanding. This paper introduces a novel framework named ConFUSE, designed to address the multi-faceted challenges of machine unlearning within FL by incorporating neuroscientific principles into a confusion-based technique for memory degradation. This approach enables targeted data erasure at various levels-instance, feature, and client-without the need for knowledge distillation, thus preserving the model's integrity and reducing the computational burden on clients. We evaluate the effectiveness of our method using three benchmark datasets, demonstrating its efficiency and adaptability in FL environments, thereby ensuring compliance with privacy laws and enhancing the model's fairness and reliability.